FINANCIAL Inside the issue:

Announcement of

ANALYSTS the 2022 Graham

and Dodd Award
Winners

JOURNAL ®
Volume 79 Number 1
First Quarter 2023

REDEFINING THE OPTIMAL

5
RETIREMENT INCOME STRATEGY
David Blanchett, CFA
ALLOCATING TO THEMATIC
18
INVESTMENTS
Koye Somefun, Romain Perchet, Chenyang Yin,
and Raul Leote de Carvalho

37
TARGETING MACROECONOMIC
EXPOSURES IN EQUITY
PORTFOLIOS: A FIRM-LEVEL
MEASUREMENT APPROACH FOR
OUT-OF-SAMPLE ROBUSTNESS
Mikheil Esakia and Felix Goltz SUPPLY CHAIN CLIMATE
58
EXPOSURE
Greg Hall, Kate Liu, Lukasz Pomorski, and
Laura Serban

77
TRADE INFORMATIVENESS IN
MODERN MARKETS
Samarpan Nawn, CFA, and Gaurav Raizada
99
OPTION PRICING VIA BREAKEVEN
VOLATILITY
Blair Hull, Anlong Li, and Xiao Qiao
 Financial Analysts Journal

58 Supply Chain Climate Exposure

CONTENTS Greg Hall, Kate Liu, Lukasz Pomorski, and

Laura Serban

All companies, even green ones, can

Volume 79 Number 1 | First Quarter 2023
PERSPECTIVES
5 Redefining the Optimal Retirement
Income Strategy
have material climate exposure. The
authors propose a better measure of
supply-chain climate risks that captures
price movements, shows performance
patterns, and can be applied to portfolio
decisions.

David Blanchett, CFA 77 Trade Informativeness in Modern

This paper introduces a cohesive series
of models designed to improve retirement
income projections. The framework can
produce guidance that differs from the
advice generated by models that use
more basic assumptions.
Markets
Samarpan Nawn, CFA, and Gaurav Raizada
The authors demonstrate that trade
informativeness is highest for institutions
and lowest for the retail group of traders.
Algorithmic trading generally provides
advantages versus manual trading,
especially in certain scenarios.

18 Allocating to Thematic Investments

Koye Somefun, Romain Perchet,
Chenyang Yin, and Raul Leote de Carvalho 99 Option Pricing via Breakeven
Volatility
Allocating to thematic investments can be Blair Hull, Anlong Li, and Xiao Qiao
difficult because they have exposures to
multiple risk factors. The authors propose Calculating breakeven volatility based on
using a framework based on robust S&P 500 options data, this study builds
portfolio optimization to make themes fit a predictive model that provides a link
in traditional multi-asset portfolios. between option characteristics and
breakeven volatility. A simulated trading
strategy demonstrates the model’s
economic value.
RESEARCH
37 Targeting Macroeconomic
Exposures in Equity Portfolios: A
Firm-Level Measurement Approach
for Out-of-Sample Robustness
Mikheil Esakia and Felix Goltz

The authors propose a method that uses

firm-level measures of macroeconomic
exposures to produce better estimations
for equity strategies versus standard
methods. This approach has applications
for equity portfolio construction.

UFAJ_I_79_1_GN_pages.indd 1 16-01-2023 14:42:26

 FINANCIAL EXECUTIVE EDITOR William Goetzmann
Yale School of Management, USA
MANAGING EDITOR Luis Garcia-Feijóo, CFA, CIPM
ANALYSTS® Florida Atlantic University and EIA Economic Index
Associates, USA

JOURNAL
A publication of CFA Institute
CO-EDITORS Nicole Boyson
Northeastern University, USA
Daniel Giamouridis
Bank of America Merrill Lynch, UK

ADVISORY COUNCIL EDITORIAL BOARD

Stephen J. Brown Steven Thorley, CFA Richard B. Evans Russ Wermers

Monash Business School and BYU Marriott School of Business, Darden Graduate School of Smith School of Business,
New York University Stern School, USA Business, University of Virginia, University of Maryland, USA
USA USA
Maria Vassalou Fanesca Young, CFA
Joanne M. Hill Vassalou Capital Management, William W. Jennings, CFA GIC, USA
CBOE Vest, USA USA US Air Force Academy, USA
Maureen O’Hara K.C. John Wei Sebastien Page, CFA
Cornell University, USA Hong Kong Polytechnic University, T. Rowe Price, USA
Hong Kong SAR, China

CFA INSTITUTE OFFICERS

Margaret Franklin, CFA

President and CEO
Paul Andrews
Managing Director, Research,
Advocacy, & Standards

Issues of the Financial Analysts Journal are available online at www.tandfonline.com/ufaj.

UFAJ_I_79_1_GN_pages.indd 2 16-01-2023 14:42:30

 CFA INSTITUTE BOARD OF GOVERNORS 2022–2023

Chair Daniel J. Fasciano, CFA Punita Kumar-Sinha, CFA

Mark Lazberger, CFA GW&K Pacific Paradigm Advisors LLC
Omnia Capital Partners United States India/United States
Australia
Oyebanji Fehintola, CFA Yimei Li, CFA
Vice Chair Africa Finance Corporation ChinaAMC
Tricia Rothschild, CFA Nigeria China
United States
Mei Gao, CFA Geoffrey Ng, CFA
CFA Institute President and CEO IDG Capital Fortress Capital Asset Management
Margaret Franklin, CFA China/United States Malaysia
CFA Institute
Canada Joanne Hill Maria Wilton, CFA
Cboe Vest Victorian Funds Management
Marshall Bailey, CFA United States Corporation
MUFG Securities EMEA plc Australia
United Kingdom Heinz Hockmann
Lovell Minnick Partners
Alexander Birkin Germany
Ernst & Young
United Kingdom

Abstract/indexed in: Business Periodicals Index.

Financial Analysts Journal (Print ISSN 0015-198X, Online ISSN 1938-3312) is published quarterly in January, April, July, and October for a total of 4 issues per year by
Taylor & Francis Group, LLC, 530 Walnut Street, Suite 850, Philadelphia, PA 19106, +1 215 625 8900. Periodicals postage paid (Permit no. 571-200) at the Post Office
in Charlottesville, VA, and additional mailing offices.
US Postmaster: Please send address changes to Financial Analysts Journal, c/o The Sheridan Press, PO Box 465, Hanover, PA 17331.
Subscription records are maintained at Taylor & Francis Group, 4 Park Square, Milton Park, Abingdon, OX14 4RN, UK.
Subscription information: For more information and subscription rates, please see tandfonline.com/pricing/journal/ufaj. Taylor & Francis journals are available in
a range of different packages, designed to suit every library’s needs and budget. This journal is available for institutional subscriptions with online only or print &
online options. This journal may also be available as part of our libraries, subject collections, or archives. For more information on our sales packages, please visit:
librarianresources.taylorandfrancis.com.
For support with any institutional subscription, please visit help.tandfonline.com or email our dedicated team at subscriptions@tandf.co.uk.
Subscriptions purchased at the personal rate are strictly for personal, non-commercial use only. The reselling of personal subscriptions is prohibited. Personal
subscriptions must be purchased with a personal check, credit card, or BAC/wire transfer. Proof of personal status may be requested.
Back issues: Taylor & Francis Group retains a current and one-year back issue stock of journals. Older volumes are held by our official stockists to whom all orders
and enquiries should be addressed: Periodicals Service Company, 351 Fairview Ave., Suite 300, Hudson, New York 12534, USA. Tel: +1 518 537 4700; email: psc@
periodicals.com.
Ordering information: To subscribe to the Journal, please contact: T&F Customer Services, Informa UK Ltd, Sheepen Place, Colchester, Essex, CO3 3LP, United
Kingdom. Tel: +44 (0) 20 7017 5544; email: subscriptions@tandf.co.uk.
Taylor & Francis journals are priced in USD, GBP and EUR (as well as AUD and CAD for a limited number of journals). All subscriptions are charged depending on
where the end customer is based. If you are unsure which rate applies to you, please contact Customer Services. All subscriptions are payable in advance and
all rates include postage. We are required to charge applicable VAT/GST on all print and online combination subscriptions, in addition to our online only journals.
Subscriptions are entered on an annual basis, i.e., January to December. Payment may be made by sterling check, dollar check, euro check, international money
order, National Giro or credit cards (Amex, Visa and Mastercard).
Copyright © 2023 CFA Institute. All rights reserved. CFA®, Chartered Financial Analyst®, GIPS®, and Financial Analysts Journal® are registered trademarks
owned by CFA Institute. CFA Institute is a Virginia not-for-profit corporation devoted to the advancement of investment management and security analysis. To
view a complete list of CFA Institute trademarks and a guide for the use of our marks, please visit our website at www.cfainstitute.org. No part of this publication
may be reproduced, stored, transmitted, or disseminated, in any form, or by any means, without prior written permission from Taylor & Francis Group, to whom all
requests to reproduce copyright material should be directed, in writing.
Disclaimer: CFA Institute and our publisher Taylor & Francis Group, LLC (“T&F”), make every effort to ensure the accuracy of all the information (the “Content”)
contained in our publications. However, CFA Institute and our publisher T&F, our agents, and our licensors make no representations or warranties whatsoever as
to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views
of the authors, and are not the views of or endorsed by CFA Institute or our publisher T&F. The accuracy of the Content should not be relied upon and should
be independently verified with primary sources of information and any reliance on the Content is at your own risk. CFA Institute and our publisher T&F make no
representations, warranties or guarantees, whether express or implied, that the Content is accurate, complete or up to date. CFA Institute and our publisher T&F,
shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused aris-
ing directly or indirectly in connection with, in relation to or arising out of the use of the Content. Full Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
Taylor & Francis Group, LLC, grants authorization for individuals to photocopy copyright material for private research use, on the sole basis that requests for such
use are referred directly to the requestor’s local Reproduction Rights Organization (RRO). The copyright fee is exclusive of any charge or fee levied. In order to
contact your local RRO, please contact International Federation of Reproduction Rights Organizations (IFRRO), rue du Prince Royal, 87, B-1050 Brussels, Belgium;
email ifrro@skynet.be; Copyright Clearance Center Inc., 222 Rosewood Drive, Danvers, MA 01923, USA; email info@copyright.com; or Copyright Licensing Agency,
90 Tottenham Court Road, London, W1P 0LP, UK; email cla@cla.co.uk. This authorization does not extend to any other kind of copying, by any means, in any form,
for any purpose other than private research use.
Permissions: See help.tandfonline.com/Librarian/s/article/Permissions
Submission Information: See tandfonline.com/ufaj
All Taylor & Francis Group journals are printed on paper from renewable sources by accredited partners.
January 2023

UFAJ_I_79_1_GN_pages.indd 3 16-01-2023 14:42:30

Financial Analysts Journal | A Publication of CFA Institute
https://doi.org/10.1080/0015198X.2022.2129947
Redefining the Optimal
Retirement Income
Strategy
David Blanchett, CFA
David Blanchett, CFA, is the head of Retirement Research, PGIM DC Solutions, Newark, NJ. Send correspondence to David Blanchett at
david.blanchett@pgim.com.
This paper introduces a cohesive
series of models designed to
improve retirement income projec-
tions. First, the retirement income
goal (i.e., liability) is decomposed
based on assumed spending elasti-
city (e.g., “needs” and “wants”).
Second, spending is assumed to
evolve throughout retirement using
a dynamic withdrawal strategy lev-
eraging the funded ratio concept.
Third, optimal strategies are deter-
mined using an expected utility
model based on prospect theory,
which also yields a client-friendly
outcomes metric. Overall, this
framework can result in advice and
guidance that is notably different
than models using more basic (and
common) assumptions, especially
approaches relying on probability of
success-related metrics.
Keywords: financial planning;
retirement; wealth management
Disclosure: PGIM DC Solutions is currently
developing a series of solutions based on
this research and methodologies.
PL Credits: 0.75
Volume 79, Number 1
Perspectives
etirement is seldom as simple as assumed in research and finan-
R cial planning tools. Too often, the retirement spending goal is
assumed to be some constant (static) amount, in today’s dollars
(i.e., in real terms), where the efficacy of a given strategy is determined
using metrics such as probability of success, which is the frequency
with which the goal is completely accomplished in a given simulation.
These flawed assumptions can result in estimates for required savings
or retirement spending in research and financial planning tools that are
significantly different than if a more realistic model is used.
In this paper, a cohesive series of models are introduced that are
designed to improve retirement income projections. The models in
this research are far more evolutionary, rather than revolutionary,
given the decades of existing research in the retirement income space
on these topics. This research is primarily focused on functional imple-
mentation, where the respective models introduced are designed to
specifically address some of the more obvious shortfalls in existing
models in a way that can be realistically (and practically) implemented.
First, we decompose the retirement spending goal (liability) into two separ-
ate goals: needs and wants, which reflects the fact that retirees typically
have varied levels of elasticity (or required certainty) associated with differ-
ent types of expenditures. For example, spending on travel is generally
more flexible than spending on healthcare. Second, we introduce a model
where spending levels (i.e., portfolio withdrawals) evolve throughout retire-
ment based on how the retiree’s funded ratio (i.e., financial situation)
changes over time. This approach can explicitly incorporate nonconstant
cash flows, which is a key weakness of most existing approaches. Third, an
expected utility model based on prospect theory is introduced to determine
optimal strategies that better capture the expected satisfaction associated
This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted use, distribution, and reproduction in any medium, provided the original
work is properly cited.
© 2022 The Author(s). Published with license by Taylor & Francis Group, LLC. 5
Financial Analysts Journal | A Publication of CFA Institute
with accomplishing different levels of spending in retire-
ment. The certainty-equivalent spending estimate from
our utility model (the “goal completion score”) is a client-
friendly outcomes metric that can be used to relay the
efficacy of a given strategy to a household.
Overall, this analysis suggests that approaches that
more accurately capture the retirement liability can
result in advice and guidance that is notably different
than models using more basic (and common) assump-
tions, especially approaches relying on probability of
success-related metrics. Collectively, this research
provides a framework that can help households make
better decisions about retirement-related issues such
as optimal spending levels, optimal portfolio risk lev-
els, and optimal allocations to guaranteed income.
Redefining the Retirement
Income Goal
Early research on retirement spending predominately
focused on estimating “safe” initial withdrawal rates.
Bengen’s (1994) research on the “4% Rule” is perhaps
the most widely known example.1 With initial withdrawal
rates, the noted percentage is multiplied by the respect-
ive balance and that initial withdrawal amount (not the
initial percentage) is assumed to be withdrawn from the
portfolio for the duration of retirement (i.e., 30 years),
typically adjusted annually for inflation. For example,
with a 4% initial withdrawal rate, if the balance at retire-
ment is $1 million, the withdrawal in the first year of
retirement would be $40,000 and that initial amount
($40,000) would be increased by inflation until death.
Past research determining optimal portfolio withdrawal
rates has generally used some type of stochastic (i.e.,
random) component, typically with respect to the
return series, based on actual historical time-series data
or some type of Monte Carlo simulation. Early research
predominately relied on historical U.S. returns, although
more recent research has considered international
time-series data (e.g., Pfau 2010). There has also been
an increased focus on using expected (or forecasted)
returns given the relatively low interest rate environ-
ment today (e.g., Fink, Pfau, and Blanchett 2013).
Optimal outcomes in financial planning tools are typic-
ally determined based on achieving some acceptable
probability of success (e.g., 90%), which is defined by
the user. The probability of success is the percentage
of runs (or trials) where the retiree accomplishes the
respective retirement goal. Probability-of-success met-
rics completely ignore the magnitude of failure, which
is a notable weakness of the metric.
6
Portfolio withdrawals, which are assumed to reflect
retiree spending, are generally static and change only
based on inflation and are not adjusted based on
portfolio performance. This is obviously overly sim-
plistic and consistent with life-cycle consumption
smoothing only under a very limited set of implaus-
ible preference parameters (Milevsky and Huang
2011). In reality, retirees headed toward certain ruin
(i.e., portfolio depletion) would likely adjust with-
drawals to prolong the life of the portfolio.
A number of concepts that are commonly used when
modeling retiree liabilities and determining optimal
retirement strategies have been borrowed from best
practices in the defined benefit (DB) space (e.g., the
concept of liability driven investing, or LDI). Unlike in
DB plans, though, which have “hard” liabilities that are
legally mandated, retirees typically have some control
over expenses (i.e., they could be perceived as “soft,”
to some extent, as discussed by Idzorek and Blanchett
[2019]). In other words, retirees typically have the
ability to cut back on spending as situations warrant.
The “pain” associated with cutting back on spending
depends to some extent on the respective expend-
iture. While some expenditures, such as healthcare,
may be relatively painful for retirees to scale back on
(and would therefore be considered relatively inelas-
tic), other expenditures, such as travel, could be
reduced more easily as situations warrant (and would
be considered more elastic). Therefore, understanding
the composition of retiree spending is important
when assessing the efficacy of the common modeling
assumption that all retiree expenditures are inelastic.
To better understand the composition of retiree
expenditures, data from the 2020 Interview file of the
Consumer Expenditure Survey (CES)2 was reviewed.
The analysis only includes respondents between the
ages of 65 and 80 (inclusive) where the household is
coded as being retired. Expenditures are categorized
as being either elastic or inelastic. Figure 1 includes
the average percentage of total household expendi-
tures that are estimated to be inelastic for various
total expenditure groups.
While approximately 65% of total retiree expendi-
tures are estimated to be inelastic (i.e., “needs”), on
average, households with higher total expenditures
clearly have higher allocations to more elastic expen-
ditures, on average (i.e., the more retirees spend, the
more flexibility they appear to have with respect to
what they are spending money on). There is also an
increasing level variability in the breakdown of inelas-
tic expenditures as total expenditures increase.

Figure 1. Inelastic Expenditures as a Percentage of Total Expenditures
100
83
Inelastic Expenditure % of
80
Total Expenditures
60
40
20
0
<$20k
While these findings are not especially novel, since it
is intuitive that some degree of retiree spending
should be flexible, relatively few financial planning
tools explicitly incorporate the notion of spending
flexibility into the retirement forecast. In other
words, financial planning tools overwhelmingly
assume that retirees are unable and unwilling to
adjust spending throughout retirement, despite rela-
tively straightforward evidence to the contrary, espe-
cially among households with higher expenditures
levels (who are more likely to use financial advisors).
While there are decades of research exploring various
approaches to adjusting portfolio withdrawals over
time, commonly referred to as dynamic spending (or
withdrawal) rules, these approaches are generally diffi-
cult to implement and can significantly complicate
financial planning tools. For example, when implement-
ing some type of dynamic spending approach, not only
do a series of rules need to be developed to determine
how spending should (or will) evolve over time, but the
outcomes metric also needs to be able to incorporate a
range of potential scenarios, which “break” more com-
mon binary metrics, like the probability of success.
The unique contribution of this research to the litera-
ture is that it attempts to address these various
retirement modeling issues simultaneously in a rela-
tively straightforward series of cohesive models.
Dynamic Spending
The literature on dynamic spending rules has evolved
over the last few decades and can generally be split
into two groups. For the first group, the emphasis is on
the rules themselves: Bengen (2001) suggests a floor/
ceiling approach; Pye (2001) suggests withdrawing the
lower of the previous amount (in real terms) or the
Volume 79, Number 1
Redefining the Optimal Retirement Income Strategy
78
73 74
70
54 51
45
41
$20k- $40k- $60k- $80k- $100k- $125k- $150k- >=$200k
$40k $60k $80k $100k $125k $150k $200k
Total Household Expenditures
amortized current portfolio value using the length of
the remaining distribution period and expected port-
folio rate of return; Guyton (2004) suggests guardrails;
Stout and Mitchell (2006) develop a model based on
both portfolio performance and remaining life expect-
ancy; and Blanchett and Frank (2009) link adjustments
to the ongoing likelihood of failure.
Recently, Waring and Siegel (2015, 2018) utilized an
intertemporal budget constraint, accompanied by a regu-
larly recalculated spending amount, to develop a fairly
simple spending approach with little or no chance of ruin.
The budget constraint, coupled with revisiting the annuiti-
zation decision annually, allows this method to incorpor-
ate volatile portfolios that are invested in risky assets.
The other line of early research on spending is quite dif-
ferent in its origin, coming out of stochastic dynamic-pro-
gramming and utility-maximizing (life cycle) models, often
in the context of the intertemporal asset pricing litera-
ture. One of the earliest life cycle models was developed
by Merton (1969). Another example is Rubinstein’s
model (1976), which explicitly derives a spending rule in
the multi-period asset pricing context. While many of the
dynamic spending models appear in the context of asset
pricing, there are some that were developed explicitly for
the spending problem, notably Milevsky and Huang
(2011) and Habib, Huang, and Milevsky (2017).
Relatively few, if any, of the dynamic withdrawal rules
that have been introduced in past research have actu-
ally been implemented in financial planning tools used
today. This can largely be attributed to the fact that
the approaches are either too computationally complex
or they cannot handle the myriad of household situa-
tions that exist in real life (or both). The existence of
nonconstant cash flows significantly increases the com-
plexity of a simulation. For example, a scenario with
7
Financial Analysts Journal | A Publication of CFA Institute
spouses retiring at different ages, claiming Social
Security benefits at different ages, and receiving future
sources of guaranteed income (e.g., a longevity annuity
starting at age 85) is something that most dynamic
spending rules are simply not designed to consider.
In summary, while many dynamic spending rules have
been developed for research purposes, they are not
robust enough to capture the actual varied situations
for retiree households that exist in the real world.
Our dynamic spending rule is specifically designed to
be robust enough to capture the varied situations
(and cash flows) of retirees while being something
that can be implemented in practice. The model is
based on the concept of the “funded ratio,” which is
conceptually similar to an approach introduced by
Kaplan and Blanchett (2021).
The funded ratio is a metric commonly used to describe
the health of pension plans, but it can more generally
be used to estimate the overall financial situation for
any goal (i.e., retiree consumption). The funded ratio is
the total value of the assets, which includes both cur-
rent balances and future expected income, divided by
the liability, which would be all current and future
expected spending. A funded ratio of 1.0 would imply
that an individual has just enough assets to fully fund
that goal. A funded ratio greater than 1.0 implies that
the individual has a surplus, while a funded ratio of less
than 1.0 implies that an individual has a shortfall.
For our analysis, we estimate funded ratios for two
goals: “needs” (inelastic) spending and “wants”
(elastic) spending. The model could easily be
expanded to consider additional spending goals
(e.g., “wishes” and/or “dreams”) if further differenti-
ation is desired.
More technically, the formula used to estimate the
needs funded ratio (N FRt ) is included in Equation (1),
where T is the total number of simulation years; At is
total portfolio value at year t; cpst is the cumulative
probability of surviving to year t; Pt includes all
future inflows, which are primarily assumed to be
pension benefits such as Social Security retirement
benefits; n gt is the assumed needs income goal for
that specific year (t); and r is the assumed discount
rate (which can vary by cash flow).
P cps P
At þ Tt¼0 ð1þrt tÞt
N FRt ¼ PT cps n gt
p
(1)
t utility model introduced in the following section.
t¼0 ð1þrn Þt
The wants funded ratio (W FRt ), which is included in
Equation (2), is conceptually identical to the needs
funded ratio calculation (in Equation (1)), but the
8
assets available to fund the wants spending goal
(w g) are reduced by the needs income goal liability.
0 PT cpst Pt PT cpst n gt 1
At þ 
B t¼0 ð1þrp Þ t t¼0 ð1þrn Þt C
W FRt ¼ max@ PT cpst w gt , 0A
t¼0 ð1þrw Þt
(2)
The initial spending level for each of the two goals
(needs and wants) would be based on their respective
portion of the total retirement goal. For example, if
the initial retirement income goal was $100,000, of
which 70% is assumed to be needs, the initial “needs”
withdrawal amount would be $70,000 and the initial
“wants” withdrawal amount would be $30,000. Those
spending amounts could evolve in future years based
on the adjustments applied to the respective target
values or other key assumptions (e.g., inflation).
This research uses relatively simple spending adjust-
ment schedules to adjust needs and wants spending
throughout the simulation. The needs goal is assumed
to be completely inelastic and increases only with
inflation annually (i.e., does not vary depending on
portfolio performance). This replicates how the entire
retirement spending goal is commonly modeled today
(i.e., effectively static in real terms) and is consistent
with more general pension liabilities (which are “hard”
in nature). In reality, though, even the needs spending
would likely be adjusted if the retiree becomes signifi-
cantly underfunded and therefore models used in
practice may want to consider this.
For wants spending, the assumed spending level is
adjusted based on the funded ratio for the end of
the previous year for that trial. The specific wants
spending adjustment schedule used for this analysis
is included in Table 1.
Looking at the schedule in Table 1, if the wants
spending amount was $50,000 and the wants funded
ratio was 1.40, the wants spending amount would
increase by 2% in the following year (to $51,000).
The rate of adjustment in Table 1 clearly varies
depending whether the retiree is overfunded versus
overfunded, where the changes are larger (in abso-
lute terms) if the retiree is underfunded. This effect-
ively assumes diminishing marginal utility of
consumption, which is consistent with the expected
This dynamic spending adjustment used for this
research is intentionally simplistic. Alternatively, an
equation, or series of equations, could be used to
determine adjustments. For example, a time component

Table 1. Real Spending Adjustment
Thresholds by Funding Ratio Level
for Wants Spending
Funded Ratio Change %
0.00 –20%
0.25 –15%
0.50 –10%
0.75 –5%
1.00 0%
1.25 2%
1.50 4%
1.75 8%
2.00 10%
could also be considered (e.g., remaining duration of
goal), in addition to funded status, type of goal (e.g.,
needs versus wants), portfolio risk level, and so on. The
goal of this research is not to opine on optimal model-
ing parameters, since these are going to be at least
somewhat endogenous to the other modeling assump-
tions (e.g., asset returns, mortality assumptions, dis-
count rates), but rather to provide a general framework
that can easily be iterated on (and improved by others).
The dynamic spending model is conceptually similar to
some existing approaches, yet is more functional
because it is more comprehensive in nature. For
example, Waring and Siegel (2015) determine spending
dynamically based on an annually recalculated virtual
annuity. While the Waring and Siegel approach should
work reasonably well in a projection with constant cash
flows, it focuses on the portfolio assets and does not
jointly consider liabilities as well as future potential
cash flows (e.g., income from a deferred income annu-
ity) and therefore is unlikely to yield optimal guidance
across a more robust set of retiree scenarios.
Asset/Liability Mapping
One potential benefit associated with splitting up the
retirement income goal is that it allows more targeted
portfolios to be created, where assets can be
“mapped” to the respective spending liability. This has
attractive behavioral implications and is consistent
with other approaches such as bucketing, which is a
time segmentation approach. For example, if the
assumed target funded ratio for needs spending
(N FRt ) is one, the amount of portfolio assets (At )
required to fund that respective spending goal could
be determined and monies could be invested specific-
ally to fund that goal. If available funds exceed the
Volume 79, Number 1
Redefining the Optimal Retirement Income Strategy
needs liability they could be allocated to wants spend-
ing, and to the extent a surplus exists (i.e., the wants
goal is overfunded), remaining monies could even be
invested in an “excess” portfolio (or just be invested in
a wants portfolio). This concept is depicted visually in
Figure 2 for three different hypothetical households.
In Figure 2 we see that the potential portfolio alloca-
tions vary materially depending the asset and liability
composition. The underlying portfolios for the
respective liabilities could vary both in terms of the
allocation to risky assets (e.g., equities) as well as
underlying asset classes (e.g., weights to nominal ver-
sus inflation-linked bonds).
This idea of segmenting assets based on the objective
is very much related to the growing field of goals-
based investing. Thaler (1985) introduced the idea of
mental accounting (“buckets”), and Shefrin and
Statman (2000) more formally introduced the idea of
behavioral portfolio theory. Creating distinct portfolios
for each goal may result in strategies that are more
behaviorally efficient, something that has been dis-
cussed by Brunel (2003), Nevins (2004), and Chhabra
(2005), among others, and results in portfolios that
are not necessarily any less efficient than strategies
employing more traditional methods depending on the
evaluation metrics used (Das et al. 2010).
The “matching” aspect of the funded ratio methodology
to respective spending goals may be attractive to asset
managers who want more targeted individualized port-
folios and may be explored further in future research.
Redefining the Optimal Outcome
Retirees clearly have the ability to adjust spending
(i.e., portfolio withdrawals) during retirement (i.e., the
retirement goal is not entirely inelastic), and we have
introduced a model that can dynamically adjust with-
drawals and that accounts for varying levels of
spending elasticity. Next, we introduce a utility model
to more accurately capture the expected satisfaction
of accomplishing a retirement goal.
While utility metrics are relatively common in academic
research on optimal retirement strategies (e.g., used by
Yaari [1965] in his pioneering research on the potential
value of annuities) and are far more adaptive than suc-
cess-related metrics, they are relatively rare in financial
planning tools. Utility functions can be somewhat
unstable and often do not yield outcomes metrics that
would be generally accessible to financial advisors and
especially clients, two potential reasons for their gen-
eral lack of popularity among non-academics.
9
Financial Analysts Journal | A Publication of CFA Institute

Figure 2. Portfolio Allocations Based on Asset and Liability Structure

The most common utility functions in retirement Total utility is based on the percentage of the overall
research assume Constant Relative Risk Aversion goal that is defined as a need (n), or is nondiscretion-
(CRRA), which is based on the idea of diminishing ary, as well as the assumed income risk aversion level
marginal utility. In this framework, each additional unit (c) for that level income. The risk aversion parameter is
of some good (e.g., consumption) is assumed to result based on whether spending is satisfying a needs goal
in more satisfaction (utility), but at a decreasing rate. or wants goal, based on Equations (3), (4), and (5).
This means that retirees want to spend as much as ! !
i 1
possible, but they must balance the desire to spend If  n then ut ¼ it
n g (3)
g cn
more with a potential shortfall, where shortfalls are ðgðnÞÞ
typically more painful than additional spending. i i
If > n and  1 then
g g
More complex utility functions distinguish risk aversion 0 00 1 11
1 (4)
in a single period from intertemporal risk aversion by ut ¼ @ðgðnÞÞ þ @@ ð1  nÞAAg
using a recursive utility function described by Epstein ðit ðgðnÞÞÞ A
cw
ðgðgðnÞÞÞ
and Zin (1989), where volatility in consumption within 0 1
a given run or assumed lifetime (defined as the elasti- 1
Else ut ¼ @ ce Ag (5)
city of intertemporal substitution) is considered separ- it
g
ately from the volatility of consumption across runs
(generally defined as general risk tolerance).
Although the model only assumes two primary types
For this research, we use a utility function based on of goals (needs and wants), the model could easily be
prospect theory, a field introduced by Kahneman and extended to consider additional types of spending
Tversky (1979). Prospect theory suggests that individu- goals (e.g., wishes).
als evaluate outcomes based on a specific reference
point and that deviations from this value (positive and If we assume an income risk aversion levels of 4, 1,
negative) are valued differently (i.e., losses receive more and 0.25 for cn , cw , and ce , respectively, we can gen-
weight than gains). More specifically, prospect theory erate expected utility values for various levels of
suggests that individual utility functions are concave in spending. These are included in Figure 3, where the
the domain of gains and convex in the domain of losses. assumed “need” (n) is either 50% or 90% of the total
retirement income goal, in Panels A and B, respectively.
Our research uses a utility model with three specific
ranges, where the utility derived from a given level of Figure 3 demonstrates that the expected level of utility
spending (i.e., consumption) depends on how that from a given spending level using this model can vary
spending relates to the retiree’s total income goal for significantly depending on the composition of a retire-
that respective year. More specifically, the expected ment goal (i.e., needs versus wants), even if the same
utility (u) of a given amount of income (i) for a given income level is achieved. For example, if we assume the
year (t) is based on the respective total income goal (g). total income goal is $100,000 and the retiree realizes

10
 Redefining the Optimal Retirement Income Strategy

Figure 3. Utility of Income for Various Goal Funding and Needs Levels

A. Need = 50% of Goal B. Need = 90% of Goal

1.20 1.20
1.00 1.00
0.80 0.80
Utility

Utility
0.60 0.60
0.40 0.40
0.20 0.20
0.00 0.00
0% 25% 50% 75% 100% 0% 25% 50% 75% 100%
% of Total Retirement % of Total Retirement
Goal Funded Goal Funded

Need Want Need Want

$50,000 (i.e., 50% of the target), the result would be
0.5 units of utility if the need represents 50% of the
goal (Panel A) versus only approximately 0.1 units of
utility if the need represents 90% of the goal (Panel B).
These differences demonstrate how two retirees with
similar total spending goals (and income levels) could
have very different retirement funding strategies
depending on the composition of their assets and
overall funded status. For example, allocating to guar-
anteed income would generally be more attractive for
retirees whose “needs” represent a higher portion of
the spending goal to reduce the uncertainty (and asso-
ciated shortfall), especially for those who do not
already have guaranteed income sources (e.g., Social
Security retirement benefits) to cover needs spending.
Given the expected utility for a level of spending in a
given year (ut ), estimated using Equations (3), (4), and/
or (5), we can estimate the constant amount of income
with the same utility as the actual income path (II). This
would be the utility-equivalent constant income level
and provides insight into the elasticity of intertemporal
substitution. This is given by Equation (6), where ut is
the utility-adjusted income in year t, qt is the probabil-
ity of surviving to at least year t, T is the final year of
retirement (i.e., the length of the assumed simulation),
and q is the investor’s subjective real discount rate.
0 P 1 This metric provides context around the percentage of
T t
q ð 1 þ qÞ u the goal completed, adjusted for utility. The goal com-
B t¼0 t t
C
II ¼ @ PT t A
(6)
q
t¼0 t
ð1 þ q Þ
To assume a fixed retirement period (e.g., 30 years), a
user would set qt to one and T to the target length
of retirement. If a user wanted each year to be
considered equally in the calculation, q would be set
to zero, where the resulting statistic would be a sim-
ple average of annual income utility levels.
The overall expected utility of a given strategy is
measured using the CRRA utility function with a risk
tolerance parameter (k) as noted in Equation (7), where
R is the number of paths (or runs), pi is the probability
of path i occurring (which is set to 1/R, i.e., the num-
ber of runs), and Ut would be the certainty-equivalent
of the stochastic utility-adjusted income II:
" #k1
XR
k
Ut ¼ pi ðIIi Þ (7)
i¼1
When considering multiple strategies (e.g., spending lev-
els, portfolio risk levels, annuity allocations) the optimal
strategy would be the one that maximizes Ut :
We generate an outcomes metric that should be rela-
tively accessible to investors by dividing the cer-
tainty-equivalent utility-adjusted income (Ut ) by the
income goal (g), yielding a metric we refer to as the
“goal completion score” as noted in Equation (8).
Ut
Goal Completion Score ¼ (8)
g
pletion score can be used to convey the overall effi-
cacy of a given strategy to a retiree, that is in the spirit
of more common metrics used in financial plans, such
as the probability of success (higher is better, with a
target of 100), but is more holistic in that it considers
preferences around spending elasticity.

Volume 79, Number 1 11

Financial Analysts Journal | A Publication of CFA Institute
The utility model developed in this paper is focused on
accomplishing a single type of goal. In reality, a house-
hold may have multiple financial goals that could differ
significantly in terms of timing, size, and scope (e.g.,
retirement income, bequest, college funding). With
respect to this model, each goal should likely be con-
sidered separately; however, assets available for fund-
ing each goal could be considered jointly, which would
require some ordering logic given potential competing
priorities. Alternatively, specific accounts could be allo-
cated to fund each goal without overlap. While in the-
ory, all goals could be considered jointly, as a single
goal, this utility model is designed for goals that are
relatively similar over time. Nevertheless, there are
ways to adjust the respective weights if multiple goals
are considered, for example, goals based on how the
income utility values are weighted and combined.
One additional potential concern with the utility
function is that it would generally suggest that
underfunded households (i.e., those headed toward
ruin) would prefer a riskier portfolio, which may not
be practical. This issue is not necessarily unique to
this model though. Therefore, while this model is
designed to more accurately capture retiree preferen-
ces than other more common metrics (e.g., the prob-
ability of success), it is not without its weakness.
Analysis
This section provides additional context on the model
in a stochastic setting. For our base scenario, we
assume the retiree household to be a joint couple,
Figure 4. Distribution of Simulation Outcomes
$120
Real Annual Total Income ($0,000s)
Wants
$100
$80
$60
Needs
$40
$20
$0
0 5
Percentiles:
12
male and female, both aged 65, retiring immediately.
The retiree couple has an $80,000 total retirement
income goal, that is assumed to increase annually by
inflation, of which 70% is classified as being a “need”
($56,000 of the total $80,000 goal).
The retiree couple will receive $40,000 in annual Social
Security benefits (which increase with inflation) and will
therefore have to generate the remaining $40,000 from
$1 million in assumed total retirement savings (i.e.,
which is consistent with the often-noted “4% Rule”).
Key model parameters are included in Appendix A and
the assumed equity glide path in retirement is
included in Appendix B. Figure 4 provides insight into
the distribution of income at various percentiles, as
well as context around where the respective income
level falls across the needs and wants income goals.
We can see that for the vast majority of runs, the
“needs” goal is fully achieved for the entire duration
of retirement. Even for the worst one in 20 runs (the
fifth percentile) the goal is still accomplished up until
the last year of retirement, where the total income
generated is $42,263, or 75% of inflation-adjusted
needs goal for that year. Note, had the income
amount not been adjusted downward throughout the
simulation, the portfolio would have been depleted
sooner, potentially much sooner based on the trial.
While total spending is expected to decline, in real
terms, for approximately 30% of scenarios moving
through retirement (with the spending level below the
initial $80,000, in today’s dollars), the reductions largely
only affect the “wants” spending goal, and in
10 15 20 25 30
Year of Retirement
5th 25th 50th 75th 95th
 Redefining the Optimal Retirement Income Strategy

approximately 70% of scenarios total real income is

actually expected to increase during retirement. Viewing Table 2. Optimal Retirement Strategies
the continuum of potential retirement outcomes through Using Goal Completion Score as
this lens (i.e., as the distribution of potential income) pro- the Outcomes Metric
vides a very different perspective than one using a trad-
itional analysis with a static spending assumption where Need % of Goal
the outcome is defined based on the probability of suc- 30% 50% 70% 90%
cess. For example, the probability of success for this
simulation is 70.2% (assuming a static real income goal), A. Optimal initial withdrawal rate (%)
while the goal completion score is 98.2%. Pension Income $10k 4.1 4.0 4.0 3.8
$30k 4.5 4.4 4.2 3.9
While a probability of success of 70.2% is not neces- $50k 5.1 4.8 4.6 4.0
sarily a bad outcome, it does not convey the same $70k 5.4 5.3 5.0 4.0
level of (high) certainty as a 98.2% goal completion $90k 5.3 5.4 5.4 4.0
B. Optimal portfolio static equity allocation (%)
score. In other words, while the probability of success
Pension Income $10k 50 40 40 25
implies a decent amount of uncertainty regarding the
$30k 55 55 45 35
retirement income goal (29.8% probability of failure, $50k 65 60 45 40
or adjustment, depending on one’s perspective), the $70k 70 65 50 40
actual likelihood of the retirees in this scenario not $90k 75 75 55 45
largely accomplishing their retirement income goal, C. Optimal annuity allocation (%)
especially the amount they need, is incredibly low. Pension Income $10k 50 50 50 50
$30k 25 30 35 40
This model can lead to notable differences in recom- $50k 10 15 30 30
mendations around factors like safe initial withdrawal $70k 0 0 15 25
rates, optimal portfolio equity allocations, and recom- $90k 0 0 5 10
mendation to annuities across various household scen-
arios. We demonstrate this in various panels in Table 2,
which are based on an additional set of analyses.
For example, in Table 2, optimal initial withdrawal
In Panel A, we solve for the initial withdrawal rate that rates (Panel A) and equity allocations (Panel B)
yields a 95% goal completion score3 for various scen- are highest for retirees who have higher levels of
arios. In Panel B, we solve for the optimal portfolio guaranteed income and more elasticity with
equity allocation, assuming a 5% initial withdrawal respect to retirement spending, while annuity alloca-
rate. In Panel C, we solve for the optimal allocation to tions (Panel C) are highest for retirees who have
an immediate annuity, assuming a 5% initial withdrawal lower levels of guaranteed income and less elasticity
rate and a nominal annuity payout rate of 4.75% (with with respect to retirement spending using this model.
a maximum potential allocation of 50%).
The differences in Table 2 can have notable implica-
As a reminder, if the outcomes metric is the probability tions for things like required retirement savings. For
of success, then the percentage of the goal that is example, focusing on withdrawal rates (Panel A), an
defined as needs and the relative amount of guaranteed initial withdrawal rate of 3.8% (low pension income
income (holding the portfolio value constant) has no and high needs goal) effectively requires 36.8% more
impact on optimal initial withdrawal rates, equity alloca- savings than an initial withdrawal rate of 5.2% (high
tion, or annuity allocations, especially if taxes are pension income and low needs goal). This is a stag-
ignored. In other words, the values in Panels A, B, and C gering difference. Alternatively, the higher with-
should not vary if the probability of success is the out- drawal rate implies the household could spend
comes metric. This is because the probability of success significantly more in retirement. Additionally, the
focuses entirely on the marginal role of the portfolio potential benefit of allocating to an annuity (Panel C)
when it comes to funding the liability and does not con- varies, whereby retirees who already have high levels
sider magnitude of failure (or allow for dynamic adjust- of guaranteed income and a relatively flexible spend-
ments). The results in Table 2, though, clearly ing goal clearly benefit less from annuitization.
demonstrate that a more nuanced perspective of house-
hold assets and preferences can have a significant Overall, the results of this section suggest that using
impact on optimal retirement strategies if appropri- the retirement income models introduced in this
ately considered. piece, which incorporate dynamic withdrawals and

Volume 79, Number 1 13

Financial Analysts Journal | A Publication of CFA Institute
spending elasticity and can account for expected util-
ity across income levels, can lead to notably different
retirement recommendations for retirees, especially
when compared to outputs generated from probabil-
ity of success-related metrics.
Conclusions
Despite significant advances in computing power and
a relatively extensive body of research on the nature
of retirement, assumptions in retirement research
and income planning tools have evolved only mod-
estly over the last 30 years. This lack of evolution has
created an environment where many financial advi-
sors and online calculators provide overly simplistic
advice on one of the most important (and expensive)
decisions made by households: how much they need
to save for retirement (and can spend during it).
This research builds off of decades of existing litera-
ture in the retirement income space and introduces a
series of models that can improve retirement income
projections by better understanding the elasticity of
spending, modeling changes in spending based on
changes in finances, and determining the optimal
strategy using a utility function that better quantifies
Appendices
Appendix A. Modeling Assumptions
Assumption Regarding
Income goal
% of goal that is need
Risk aversion factor for needs spending
Risk aversion factor for wants spending
Risk aversion factor for excess spending
Length of projection/end of retirement
Number of simulation runs/trials
Probability of survival
Discount rate for EOIS calculation
Risk aversion factor for certainty-equivalent
Nominal discount rate for needs spending
Nominal discount rate for wants spending
Nominal discount rate for inflows (guaranteed income)
Inflation
Equity, annual arithmetic return
Equity annual standard deviation
Bonds, annual arithmetic return
Bonds, annual standard deviation
Stock/bond correlation
Taxes are ignored for the analysis
EOIS ¼ Elasticity of intertemporal substitution.
14
the expected satisfaction across outcomes. While
there are a number of other existing approaches simi-
lar to those introduced in this research, most of these
approaches are relatively narrowly focused and cannot
easily be used to generate a more comprehensive per-
spective on retirement modeling. This research is
unique in introducing models specifically designed to
work together that be easily configured and allow for
a variety of client scenarios and user assumptions.
There are a number of potential extensions associ-
ated with this research, such as extending the model
to include additional goals (e.g., wishes), determining
optimal portfolio allocations for the different types of
“mapped” goals (e.g., needs and wants), determining
the optimal parameters to use with the funded ratio
dynamic spending rule, as well as exploring the more
general parameters in the model.
In closing, given the significant time and energy asso-
ciated with funding retirement, it is important that
the estimate be as accurate as possible. This research
demonstrates that existing retirement planning
frameworks (and tools) have notable deficiencies that
should be addressed to ensure retirees (and savers)
receive the best guidance and advice possible.
Variable Value
g $80,000 (varies)
n 70% (varies)
cn 4
cw 1
ce 0.25
T 30
R 1,000
qt 1 (fixed retirement period)
q 0
k 2
rn 5%
rw 4%
rp 3%
2%
7.5%
20%
2.5%
5%
0.0
 Redefining the Optimal Retirement Income Strategy

Appendix B. Equity Glide Path

100%

80%
Equity Allocation

60%

40%

20%

0%
50 60 70 80 90 100
Age

Editor's Note
Submitted 11 February 2022
Accepted 26 September 2022 by William N. Goetzmann

Notes
1. Other notable research includes Milevsky, Ho, and 3. This is a subjective target and can obviously vary by user
Robinson (1997) and Cooley, Hubbard, and Waltz (1998). or client.

2. https://www.bls.gov/cex/pumd.htm

References
Bengen, William. 1994. “Determining Withdrawal Rates Using
Historical Data.” Journal of Financial Planning 7 (4): 171–80.
Bengen, William. 2001. “Conserving Client Portfolios during
Retirement, Part IV.” Journal of Financial Planning 14 (5): 110–9.
Blanchett, David, and Larry Frank. 2009. “A Dynamic and
Adaptive Approach to Distribution Planning and Monitoring.”
Journal of Financial Planning 22 (4): 52–66.
Brunel, Jean. 2003. “Revisiting the Asset Allocation Challenge
through a Behavioral Finance Lens.” The Journal of Wealth
Management 6 (2): 10–20. doi:10.3905/jwm.2003.320479.
Journal of Financial and Quantitative Analysis 45 (2): 311–34.
doi:10.1017/S0022109010000141.
Epstein, Larry G., and Stanley E. Zin. 1989. “Substitution Risk
Aversion and the Temporal Behavior of Consumption and
Asset Returns: A Theoretical Framework.” Econometrica 57 (4):
937–69. doi:10.2307/1913778.
Finke, Michael, Wade D. Pfau, and David M. Blanchett. 2013. “The 4
Percent Rule Is Not Safe in a Low-Yield World.” Journal of Financial
Planning 26 (6): 46–55. https://www.financialplanningassociation.
org/article/4-percent-rule-not-safe-low-yield-world

Chhabra, Ashvin. 2005. “Beyond Markowitz: A Comprehensive Guyton, Jonathan. 2004. “Decision Rules and Portfolio
Wealth Allocation Framework for Individuals.” The Journal of Management for Retirees: Is the ‘Safe’ Initial Withdrawal Rate
Wealth Management 7 (4): 8–34. doi:10.3905/jwm.2005.470606. Too Safe?” Journal of Financial Planning 17 (10): 54–62.

Cooley, Phillip, Carl Hubbard, and Daniel Waltz. 1998.
“Retirement Spending: Choosing a Sustainable Withdrawal
Rate.” Journal of the American Association of Individual Investors
20 (2): 16–21.
Das, Sanjiv, Harry Markowitz, Jonathan Scheid, and Meir
Statman. 2010. “Portfolio Optimization with Mental Accounts.”
Habib, Faisal, Huaxiong Huang, and Moshe Milevsky. 2017.
“Approximate Solutions to Retirement Spending Problems and
the Optimality of Ruin.” Working paper, version March 31 2017.
Idzorek, Thomas, and David Blanchett. 2019. “LDI for Individual
Portfolios.” The Journal of Investing 28 (1): 31–54. doi:10.3905/
joi.2019.28.1.031.

Volume 79, Number 1 15

Financial Analysts Journal | A Publication of CFA Institute
Kahneman, Daniel, and Amos Tversky. 1979. “Prospect Theory:
An Analysis of Decision under Risk.” Econometrica 47 (2):
263–91. doi:10.2307/1914185.
Kaplan, Paul, and David Blanchett. 2021. “A Dynamic Spending
Rule for All Retirees.” Working Paper.
Merton, Robert. 1969. “Lifetime Portfolio Selection under
Uncertainty: The Continuous-Time Case.” The Review of
Economics and Statistics 51 (3): 247–57. doi:10.2307/1926560.
Milevsky, Moshe, Kwok Ho, and Chris Robinson. 1997. “Asset
Allocation via the Conditional First Exit Time or How to Avoid
Outliving Your Money.” Review of Quantitative Finance and
Accounting 9 (1): 53–70. doi:10.1023/A:1008278910581.
Milevsky, Moshe, and Huaxiong Huang. 2011. “Spending
Retirement on Planet Vulcan: The Impact of Longevity Risk
Aversion on Optimal Withdrawal Rates.” Financial Analysts
Journal 67 (2): 45–58. doi:10.2469/faj.v67.n2.2.
Nevins, D. 2004. “Goals-Based Investing: Integrating Traditional
and Behavioral Finance” Journal of Wealth Management 6 (4):
8–23. doi:10.3905/jwm.2004.391053.
Pfau, Wade. 2010. “An International Perspective on Safe
Withdrawal Rates from Retirement Savings: The Demise of the
4 Percent Rule?” Journal of Financial Planning 23 (12): 52–61.
Pye, Gordon. 2001. “Adjusting Withdrawal Rates for Taxes and
Expenses.” Journal of Financial Planning 14 (4): 126–36.
16
Rubinstein, Mark. 1976. “The Strong Case for the Generalized
Logarithmic Utility Model as the Premier Model of Financial
Markets.” The Journal of Finance 31 (2): 551–71. doi:10.2307/
2326626.
Shefrin, Hersh, and Meir Statman. 2000. “Behavioral Portfolio
Theory.” The Journal of Financial and Quantitative Analysis 35
(2): 127–51. doi:10.2307/2676187.
Stout, Gene, and John Mitchell. 2006. “Dynamic Retirement
Withdrawal Planning.” Financial Services Review 15:117–31.
Thaler, Richard. 1985. “Mental Accounting and Consumer
Choice.” Marketing Science 4 (3): 199–214. doi:10.1287/mksc.4.
3.199.
Waring, Barton, and Laurence Siegel. 2015. “The Only
Spending Rule You Will Ever Need.” Financial Analysts Journal
71 (1): 91–107. doi:10.2469/faj.v71.n1.2.
Waring, Barton, and Laurence Siegel. 2018. “What Investment
Risk Means to You, Illustrated: Strategic Asset Allocation, the
Budget Constraint, and the Volatility of Spending during
Retirement.” The Journal of Retirement 6 (2): 7–26. doi:10.
3905/jor.2018.1.041.
Yaari, Menahem. 1965. “Uncertain Lifetime, Life Insurance, and
the Theory of the Consumer.” The Review of Economic Studies
32 (2): 137–50. doi:10.2307/2296058.
Perspectives Financial Analysts Journal | A Publication of CFA Institute
https://doi.org/10.1080/0015198X.2022.2112895

Allocating to Thematic
Investments
Koye Somefun, Romain Perchet , Chenyang Yin, and
Raul Leote de Carvalho
Koye Somefun is the head of Multi-Asset and Solutions in the Quant Research Group at BNP Paribas Asset Management,
Herengracht, Amsterdam, The Netherlands. Romain Perchet is the head of Multi-Assets in the Quant Research Group at BNP Paribas
Asset Management, Paris, France. Chenyang Yin is an analyst in Multi-Assets in the Quant Research Group at BNP Paribas Asset
Management, Paris, France. Raul Leote de Carvalho is the deputy head of the Quant Research Group at BNP Paribas Asset
Management, Paris, France. Send correspondence to Romain Perchet at romain.perchet@bnpparibas.com.

We introduce the notion of Introduction

themes as an additional invest-
ment dimension beyond asset
classes, regions, sectors and
styles, and propose a framework
to allocate to thematic invest-
T asset classes: sovereign bonds, corporate bonds, equities and
ments at a strategic
asset allocation level. Allocating
to themes requires discipline
because thematic investments are
not only exposed to the theme
but also to the traditional risk fac-
tors. Our approach uses a frame-
work based on robust portfolio
optimisation, which accounts for
the expected excess return from
the exposure to the theme and
from exposures to traditional risk
factors. We provide an example
to illustrate how thematic invest-
ments fit in traditional multi-
asset portfolios.
Keywords: asset allocation; core
satellite; portfolio construction;
portfolio optimisation; thematic
investing; themes
Disclosure: The authors report no
conflicts of interest. The authors alone
are responsible for the content and
writing of the paper.
PL Credits: 0.75
raditionally, investors categorise their investments into different
others. In each of these, they tend to distinguish between
regions and sectors. Portfolio construction takes this classification into
account, with investors seeking diversification across asset classes
because doing so can meaningfully reduce the risk of significant losses
as each asset class can respond differently to market conditions and
changes in the economic environment. Additionally, investors seek to
diversify across sectors and regions in each asset class because this
helps to mitigate systematic risks caused by factors affecting asset
returns from a given industry or region.
Style factors may also be used, in particular for equities (Fama and
French 1992; Haugen and Baker 1996) but also for corporate bonds
(Houweling and van Zundert 2017; Israel, Palhares, and Richardson
2018; Heckel et al. 2019) and sometimes for other asset classes
(Asness, Moskowitz, and Pedersen 2013). For example, value factors
identify assets that are thought to be trading at a discount relative to
their respective fundamental value. Momentum, another style, relies
on factors that identify assets by trends in either their prices or in
their fundamental values. Styles are used because they also tend to
explain some of the correlation of returns of the assets exposed to
the same underlying factors.
Thematic investing brings a new dimension that transcends the
classifications based on regions, sectors and styles. Themes are
structural trends expected to significantly impact economies and
We are grateful to the editor and to the reviewers for their comments and sugges-
tions that helped us to significantly improve the manuscript. We are also grateful for
valuable comments from Alexander Bernhardt, Clement Dupire, Daniel Morris,
Edward Lees, François Soupe, Frederic Surry, Guy Davies, Jane Ambachtsheer,
Marco Barbaro, Michael Victoros, Pieter Oyens, Tarek Issaoui, Thomas Heckel, Ulrik
Fugmann, and Yannick Leite Velho.

18 © 2022 CFA Institute. All rights reserved. Volume 79, Number 1


redefine business models. Thus, it is natural to
expect themes to play a role in explaining the
returns and risks of investments with greater
exposure to those structural trends. However, not
all themes matter equally, and thematic investing
offers more than just a single compelling opportun-
ity. The goal of thematic investing is to identify the
assets that have their returns impacted by the
structural changes underlying the theme. These
changes come about through megatrends that
shape societies: demographic shifts, social or attitu-
dinal changes, environmental impact, resource scar-
city, economic imbalances, transfer of power,
technological advances and regulatory or political
changes. An investment theme aims to transform
one such megatrend into a relevant investment
opportunity.
According to Morningstar (2021), over the three
years through March 2021, collective assets under
management in thematic funds more than tripled to
USD 595bn from USD 174bn worldwide. While the-
matic portfolios have mainly focused on equities,
there is also a growing trend to define themes
in fixed income or even in other asset classes. In a
survey of institutional investors, Walmsley
(2021) reports that approximately 60% of respond-
ents are already using sustainable thematic invest-
ment strategies in fixed income, as are 50%
in alternatives. Despite the growing interest in the-
matic investing, literature is still scarce, in particular
when it comes to proposing a robust framework to
allocate to thematic investments at strategic alloca-
tion level.
In this paper, we contribute to the literature on the-
matic investments in three ways. First, we give an
introduction to the topic, highlighting what should be
considered as thematic investing and what it should
not be confused with. In particular, we defend why
we believe that themes can bring an additional
dimension to an asset’s risk decomposition. Second,
we propose a framework to allocating to thematic
investments which are based on robust portfolio
optimisation and that takes into account the expo-
sures of thematic investments to the traditional risk
factors used in asset allocation. This framework
offers a robust solution to the problem of allocating
both traditional assets and to thematic investments
in a given portfolio. Third, we illustrate the imple-
mentation of the framework with an example that
goes beyond allocating only to equity thematic
investments by also including thematic fixed income
investments.
Volume 79, Number 1
Allocating to Thematic Investments
What Is Thematic Investing?
We see themes as structural trends expected to sig-
nificantly impact economies and business models.
Thematic investing is about investing in the assets
that are more likely to benefit from value creation
derived from those themes. Taking into account the
growing interest in this type of investment it is nor-
mal to expect some confusion about what should
and should not be considered as thematic invest-
ments. In the remainder of this section, we define
what we see as thematic investing and how it differs
from just investing in growth stocks, or in selected
sectors or industries, and also how it differs from
more broad concepts such as ESG investing (taking
environmental, social and governance factors expli-
citly into account) or long-term investing.
Is Thematic Investing Just about
Investing in Growth Companies? This is a
question asked often. The short answer is no. First,
themes do not necessarily relate only to equities or
corporate bonds. For example, some themes may be
played via other assets such as public infrastructure
investments. Second, when it comes to equities or
corporate bonds, this is also not necessarily the case.
Themes can be relevant to companies in all stages of
their growth cycle. Indeed, some themes may relate
more to disruption through innovation, for example,
disruptive tech, often involving smaller capitalization
growth firms, whereas other themes such as environ-
mental sustainability may see a much larger spectrum
of companies taking the lead, including larger capital-
ization incumbents. What is important when looking
for companies exposed to a theme is that the
selected firms have a significant exposure or commit-
ment of their business to activities related to the
theme and either generate or plan to generate a sig-
nificant part of their revenues from selling products
or services related to the theme.
Is Thematic Investing Just a Wager on a
Particular Industry or Sector? Again, the
short answer is no. For example, a theme may favour
a sub-industry or a group of companies that can dis-
rupt an industry because of some cost or size advan-
tage; or because of new products or services
(particularly if these require patents or licences); or
due to regulatory protection; or because such groups
of companies are motivated by the megatrends that
shape societies. Anything that can give a group of
companies a significant advantage that disrupts entire
industries can be useful to identify themes. For this
reason, some narrower themes may come with some
19
Financial Analysts Journal | A Publication of CFA Institute
exposure to some sectors but thematic investments
should not be just a bet on a given sector
or industry.
In recent years, globalisation, regulation and de-regu-
lation trends, digitalisation and innovation, and
energy transition are examples of themes that have
played an important role in changing the balance of
forces that can make certain industries more or less
susceptible to disruption.
Thematic, ESG, and Long-Term Investing
Are Easily Confused. It is also important to
make the difference between thematic investing, ESG
investing and long-term investing.
While thematic investing entails a focus on trends
that often materialise in the medium to longer term,
this is not always the case. Some themes may run
their course faster and find themselves discounted
by the market quicker than others. It is important to
evaluate to what extent a theme still remains a good
investment opportunity. Some themes of the past
may have reached maturity and it is important to
judge to what extent they remain an attractive
investment opportunity. Moreover, the granularity of
themes may be important. This is because some
themes within a broader theme can run their course
while the overarching drivers remain relevant and
investable, simply with a different set of companies.
In such cases, a bottom-up investment strategy may
be a compelling approach to allocate to sub-themes
within a given theme. Thematic investing can be
more subtle than just finding undervalued opportuni-
ties to hold over long investment horizons.
A similar argument can be made for thematic versus
ESG investing. Not all ESG investing is necessarily
thematic investing. There is a difference between
companies simply pursuing sustainable practices and
those deriving or committed to derive a significant
component of their revenues from services and prod-
ucts related to sustainable themes such as energy
transition or environmental sustainability. Investors
turn to ESG in search of investments that integrate
material environmental, social or governance factors
in the pursuit of greater financial risk-adjusted
returns, which is something that transcends thematic
investing: ESG integration does not necessarily focus
on investing in a specific sustainable theme
or themes.
A desire to drive change in society can also be a
motive for thematic investing. A good example is
energy transition, a necessary step to address climate
20
change, which is perhaps the biggest challenge
humanity has ever faced and is already having a pro-
found impact on weather systems and economic out-
comes. Investing around this theme means finding
companies well positioned to generate a measurable
impact while profiting from this structural change.
Indeed, most impact investing is thematic investing.
It is worth noting that we would not consider inves-
ting in the market capitalisation index re-weighted
after excluding just a few of the biggest carbon emit-
ters as an example of thematic or impact investing.
Such an approach is typically just a low tracking error
climate risk mitigation strategy designed to limit
downside risk from the transition to low carbon,
rather than a higher tracking error investment in
energy transition leaders.
What Is the Appeal of Thematic
Investing? Thematic investing provides investors
with the means to focus on investments well posi-
tioned to profit from transformative changes, identi-
fying the current and future champions for each
trend. Those champions should benefit from an out-
sized impact from value creation. It is also a means
to stay out of assets more negatively exposed to
such trends. The flipside of the coin is that investors
may need to accept choppy returns in the shorter
term from less diversified portfolios exposed to
themes, as the outsized returns tend to be less fre-
quent and can arrive suddenly. Moreover, too much
exposure to high concentration thematic investments
may also result in investors missing out by not inves-
ting in companies not exposed to their particular
chosen themes. Portfolio construction is thus of
paramount importance when it comes to integrating
thematic investing in a larger portfolio.
Allocating to Thematic Investing. Thematic
investing may be appealing, but how can we add it
to a portfolio? There is no single answer as it
depends on a number of parameters such as the size
and the level of diversification of the portfolio. One
simple approach is to rely on a core-satellite frame-
work adding thematic investments to the satellite.
Scherer (2015) proposes the use of this framework
for deciding about the allocation to passive and
active managers. Here we extend the application to
an allocation that includes thematic investments. This
is an adequate solution for many investors, in par-
ticular for large portfolios investing in low-capacity
themes, or for smaller investors who prefer to limit
their active allocations to thematic investments.

In any case, to do this effectively it is important to
investigate the impact of adding thematic invest-
ments to the overall risk and return characteristics of
the aggregate portfolio. Therefore, it is preferable
and more straightforward to integrate thematic
investments directly in the overall allocation. To do
so, it is necessary to understand each thematic
investment in terms of expected risk, return and
interaction with other investments. While this may
not be easy, quantitative approaches can help, as we
shall discuss below.
Understanding What Is behind Each
Theme. Allocating to thematic investments requires
a deep understanding of how they interact with
other investments. If thematic investments are con-
structed from investing in stocks and bonds then
their expected returns, risk and correlations can be
derived (i) from their exposure to those traditional
asset classes, (ii) from their exposure to the trad-
itional risk factors of those asset classes such as
regions, styles or sectors and (iii) from an additional
specific component related to the theme. In other
words, we need to investigate the extent to which
thematic investments bring something new on top of
their exposures to traditional assets. This is all that is
required to integrate thematic investments into our
strategic asset allocation process.
How does this work in practice? The more common
themes can usually be represented by existing the-
matic benchmark indices, e.g., those provided by the
standard index vendors. Moreover, those indices are
often used as benchmarks for the thematic mutual
funds where active management aims at outperform-
ing such benchmarks. Finally, sometimes those indi-
ces are also replicated by exchange-traded funds
(ETFs). Thus, thematic benchmark indices are a good
starting point.
However, while using thematic benchmark indices
brings the advantage of standardisation, return dis-
persion of investments related to a theme can be
large because there is less consensus on which assets
should be part of a portfolio representing the theme
than there is for regions, sectors or perhaps even
style factor benchmarks. One alternative is to use
benchmark indices for the same theme from more
than one index provider. Another alternative is to
rely on the return time series of the funds selected
to invest in a given theme, combining them into a
representative return series. When more than one
benchmark index or fund is used, machine learning
algorithms such as Lasso regressions1 against trad-
itional risk factors could be used to efficiently select
Volume 79, Number 1
Allocating to Thematic Investments
and combine the returns of such funds into a return
time series representative of the theme.
Once the times series of returns representing a
theme is created, we can assess how it relates to
other assets selected for the portfolio. In the robust
portfolio optimization approach employed in the next
section, the optimizer allocates by taking into
account the systematic exposures of the non-the-
matic and thematic investments to a set of risk fac-
tors whilst potentially adding the expected alpha
from the thematic exposure. Hence, we estimate the
exposures of the time series of returns representing
the theme to the set of risk factors used in the risk
model. The expected returns of thematic investments
are constructed by taking into account the returns
derived from systematic risk exposures to the risk
factors and from an additional thematic alpha sized
as a function of how much residual risk is found. The
thematic alpha should be forward-looking.
For themes that relate to specific sectors of the
economy, it is not unusual to see those thematic
funds only benchmarked against a relevant sector
index. In such cases, it is important to assess the
expected added value from biasing the selection
towards the companies more exposed to the theme
when compared to the sector and to take into
account that an exposure to the sector will itself cre-
ate tracking error and excess returns relative to the
broad equity market. Making sure that the thematic
investment generates sufficient residual risk relative
to sectors is important because otherwise the the-
matic investment will be nothing more than just a
sector bet. Diversifying across themes that are biased
towards different sectors is a way of minimising the
problem of sector exposures.
For equity thematic investments, valuations also
need to be checked, in particular that the stocks
behind a given theme are not overpriced. If valua-
tions are high even after correcting for any sector
exposures, that may suggest that the theme has run
its course. Negative exposure to quality factors not
explained by sector tilts is another example of a style
exposure that needs to be taken into account when
forecasting thematic returns. In this respect, Blitz
(2021) found that thematic indices proposed by S&P
and MSCI, two index providers, invested mainly in
high volatility, low quality and high valuation stocks.
Such style biases, if not just caused by sector biases,
are likely to create a drag on returns that need to be
counter-balanced with sufficiently high thematic
alpha. If the risk not explained by traditional risk fac-
tor exposures is small then, even if the thematic
21
Financial Analysts Journal | A Publication of CFA Institute

alpha is positive, there is a high likelihood that the
alpha is small and that such thematic indices
under-perform.
Here we address some of the concerns raised by
Blitz (2021). Namely, we used thematic indices from
other providers than just S&P and MSCI, we diversi-
fied by combining thematic indices from different
providers for the same theme, we chose themes with
different levels of beta, we chose themes with low
exposures to the factors highlighted by Blitz (2021)
and with significant risk derived from more specific-
ally from the theme. Moreover, we also investigated
sector exposures, which we believe can be more
important for equity thematic investments than
style exposures.
Allocating to Thematic
Investments
We shall now go through an example of how to allo-
cate to thematic investments.
The first step towards an allocation to thematic
investments starts with the selection of themes and
vehicles to invest in the selected themes, typically
mutual funds or ETFs replicating thematic indices.
Here we consider five themes: (i) disruptive tech, (ii)
environmental sustainability, (iii) sustainable water,
(iv) sustainable food and (v) energy transition. We
selected these themes just as examples to illustrate
the implementation of the proposed framework for
the allocation of thematic investments. Their choice
was motivated by the fact that we could find the-
matic indices from different providers to represent
them and, as we shall see later, because a sufficiently
large portion of the risk of those indices is not
explained by traditional risk factors. That means that
these thematic indices do bring some different sour-
ces of risk to the portfolio.
Each theme is represented by two equally weighted
thematic equity indices. While more indices could
have been used for each theme, we chose to keep
the example simple. The environmental sustainability
theme is also represented separately by a fixed-
income thematic index, which allows us to allocate to
this theme not only using equities but also using
fixed income. In Table 1, we show the indices
selected to represent the themes.

Table 1. Thematic Indices Used to Represent the Selected Themes

Theme Benchmark Ticker Bloomberg Start_date

Equities Disruptive tech Morningstar Exponential Tech MSEXPONU Index 4-Mar-2015

Net Total Return USD Index
Nasdaq CTA AI NTR Index NQINTELN Index 24-Mar-2008
Sustainable water MSCI Global Sustainable NGESW Index 1-Sep-2010
Water Net Return Index
FTSE Environmental EOWR Index 13-Nov-2008
Opportunities Water
Technology Index
Sustainable food MSCI ACWI IMI Food MXACIFDP Index 30-Nov-2016
Revolution Index
Foxberry Tematica Research FXBYFOOD Index 20-Mar-2015
Sustainable Future of Food
NTR Index
Energy transition ECPI Global Renewable Energy GALPLRWN Index 3-Jan-2011
Liquid NTR
FTSE Environmental EORE Index 13-Nov-2008
Opportunities Renewable
and Alternative Energy Index
Environmental MSCI ACWI IMI Efficient MXACIEEG Index 31-May-2013
sustainability Energy Index Net USD
MSCI Global Environment Net M1CXUNB Index 15-Jul-2016
Return USD Index
Fixed income Environmental Bloomberg MSCI Global Green GBGLTRUU Index 14-Oct-2014
sustainability Bond Index Unhedged USD

Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ.

22

In the next two sections, we use time series analysis
to investigate the extent to which the themes repre-
sented by these index combinations show persistent
exposures to traditional risk factors. In Section
“Exposure of Thematic Investments to Regions,
Sectors and Styles,” we use Lasso regressions to find
the exposures of these themes to regions, styles and
sectors represented by a large number of indices. In
Section “Exposure of Thematic Investments to
Traditional Macro Risk Factors,” we use regressions
to find the exposures of these themes to the macro
risk factors from the asset allocation risk model used
in the portfolio optimization.
Exposure of Thematic Investments to
Regions, Sectors and Styles. In this section,
we investigate the exposure of these themes to
regions, styles and sectors. In our empirical analysis
we used 128 equity indices and 46 fixed income indi-
ces representing regions, styles and sectors, which
can be found in Table A2 in Appendix A. In our ana-
lysis, we first estimated the beta of the equity the-
matic indices and its residual returns from a CAPM
type regression using the MSCI AC index to repre-
sent the global equity market. We did the same for
the bond thematic index using the Bloomberg
Barclays Global Aggregate Index to represent the glo-
bal bond market. Finally, we did the same for all
regional, style and sectors indices. In this step, we
removed the exposure to the market from all those
indices used as explanatory variables in regressions.
We then used Lasso regressions to find the regions,
styles and sectors with significant contribution to
each theme. These regressed the residuals of the
thematic indices on the residuals of the region, style
and sector indices. We used data from June 2017
through May 2022. The details of the methodology
are given in Appendix A.
All market data was sourced from Bloomberg. We
used weekly net total returns in USD calculated from
Wednesday to Wednesday. All index returns were
converted into USD when required. Most global the-
matic indices are reported in USD2 without hedging
the currency risk. The results of our analysis can be
found in Table 2. In this table, we retained only indi-
ces for which significant exposures were found. For
this reason, the vast majority of the indices used in
the Lasso regression are not included in the table
and can be found only in Appendix A.
The results from Table 2 show that only disruptive
tech and environmental sustainability tend to have
high beta. All other themes have neutral or lower
beta and lower volatility. Sustainable water and
Volume 79, Number 1
Allocating to Thematic Investments
sustainable food have the lowest volatility, signifi-
cantly lower than that of the high beta equity
themes. In Table 2, we also include the volatility of
the residual returns after removing the beta compo-
nent from returns. This ranges from 6.7% to 13% for
equity themes. The fixed income theme is only 2%,
relatively small compared to the volatility of
the index.
Table 2 also sheds light on the importance of each of
the three orthogonal components of thematic indices
returns. Sustainable water equities, energy transition
equities and environmental sustainability equities do
not have any significant exposure to style factors.
Sustainable food equities have a small exposure to
USA Small Cap Growth while disruptive technology
equities are slightly more exposed. We also found a
number of sector exposures for some of these
themes. Nevertheless, it is important to highlight that
most of these systematic exposures contribute rela-
tively little to risk. The contributions of regions,
styles and sectors to the risk of the beta adjusted
excess returns of these thematic indices are small
except for disruptive tech equities and environmental
sustainability fixed income.
It is important to note that the methodology pre-
sented in this section is quite general and can be
used to analyze other thematic indices or funds. It
can be helpful for the selection of themes or bench-
mark indices to represent themes.
Exposure of Thematic Investments to
Traditional Macro Risk Factors. Here we
look at the exposures of the themes to the factors of
a traditional risk factor model used in asset allocation
problems. Following the methodology proposed by
Bass, Gladstone, and Ang (2017) and Issaoui et al.
(2021), we constructed a statistical factor model
using Principal Component Analysis (PCA) to select
the set of most significant statistical risk factors from
the correlation matrix of monthly local currency
returns of 17 global major assets. The risk model was
estimated with returns from 2003 through 2021. The
list of 17 assets is given in Table A1 in Appendix A.
We identified six statistical (but interpretable) risk
factors that correspond to the first six eigenvectors
of the correlation matrix and their corresponding
eigenvalues. These explain about 87% of the vari-
ance, which is consistent with the finding of Bass
et al. (2017) and Issaoui et al. (2021). Table A1 in
Appendix A shows the six eigenvectors represented
as long-short portfolios.
23
24
Table 2. Exposure of Themes to the Market Factor, to Sectors, and to Styles
Equities Fixed Income
Disruptive Sustainable Sustainable Energy Environmental Environmental
Themes Tech Water Food Transition Sustainability Sustainability

Beta Against global market index 1.28 0.93 0.96 1.01 1.26 1.34
Volatility of In excess of cash 22.7% 18.0% 17.2% 19.5% 24.5% 8.3%
thematic returns In excess of beta exposure 8.5% 9.5% 6.7% 9.9% 13.0% 2.0%
In excess of beta, region, style and sector exposures 7.0% 5.8% 2.7% 2.3% 0.0% 1.8%
Explained by regions, style and sectors exposures 4.8% 7.6% 6.2% 9.7% 13.0% 1.0%
Contribution to From regions, styles and sectors exposures 68.3% 36.8% 16.2% 5.5% 0.0% 75.7%
volatility of In excess of regions, styles and sectors exposures 31.7% 63.2% 83.8% 94.5% 100.0% 23.0%
thematic returns
in excess of beta
R2 of regression Thematic returns in excess of beta on 68% 37% 16% 5% 0% 76%
On indices Region, style and sector returns in excess of beta
Financial Analysts Journal | A Publication of CFA Institute

selected
by Lasso
Beta on styles MSCI USA Small Cap Growth 0.39 0.25
Beta on sectors MSCI Japan Info Tech 0.15
MSCI Emerging Markets Info Tech 0.15
STOXX Europe 600 Oil Gas 0.78
MSCI US Industrials 0.48
MSCI US Info Tech 0.36
MSCI US Utilities 0.26
MSCI Canada Energy 0.33
Bloomberg Barclays Pan Euro Corporate Health Care 0.33
Bloomberg Barclays Pan-European Agg. Local Authorities 0.16

Note: The calculations are based on weekly returns and follow the framework detailed in the appendix. The volatility is annualized with the square root of the time frequency.
Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ. Authors’ calculations.
 Allocating to Thematic Investments

Table 3. Exposures of Themes to Traditional Risk Factors from Asset Allocation Risk Model
Equities Bonds
Environmental Environmental
Disruptive Sustainable Sustainable Energy Sustainability Sustainability
Themes Tech (%) Water (%) Food (%) Transition (%) (%) (%)

Risk factors
Market risk 15.7 8.2 8.3 10.7 15.1 1.7
Duration
3.8
3.8
3.9
4.2
4.3 3.4
EM/Commodities 2.0
0.6
0.6
2.5 2.6
0.6
Corporate spreads
1.9
0.4
0.4
1.1
2.3 0.2
US 3.3 3.8 3.9 2.0 4.1
0.4
Asia/Japan
0.7
3.6
3.6
1.3
1.5
0.3
Systematic volatility 16.7 10.5 10.6 12.1 16.6 3.9
Specific volatility 5.0 4.0 4.0 2.9 5.1 0.8
Thematic volatility 7.9 10.1 8.1 9.7 12.1 0.4
Total volatility 19.2 15.1 13.9 15.8 21.2 4.0

Note: Exposures are measured in units of volatility of returns. The risk model is estimated using returns from Jan-2003 through
Dec-2021. The projections of thematic indices on the factors of the risk model are calculated from a regression against 17 core
assets returns using 5 years of weekly returns from Jun-2017 to May-2022.
Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ. Authors’ calculations.

The risk factors were named according to the eco-
nomic interpretation of the underlying long-short
portfolios in Table A1. The market risk factor is a
portfolio invested mainly in risky assets such as equi-
ties and corporate bonds while excluding government
bonds. The duration factor is a portfolio invested
mainly in interest rate-sensitive assets, including gov-
ernment bonds, corporate credit and emerging debt.
The corporate spreads factor is a portfolio invested
in corporate bonds and short equities and treasuries.
The EM/commodity factor is a portfolio mainly
invested in emerging market assets and commodities.
The US and the Asia/Japan factors are portfolios
mainly invested in US and Asia/Japan assets, respect-
ively. These risk factors are orthogonal to each other
in the sense that the correlation of returns resulting
from the long-short linear combinations of the under-
lying equity, fixed income and commodity indices is
exactly zero in the sample used to construct the
risk model.
In Table 3, we show the exposures of the themes to
these six macro risk factors. Themes represented by
equity indices have a large component of their total
risk explained by the market risk factor. They also
have a small negative exposure to the duration risk
factor, a small positive exposure to the corporate
spreads risk factors and a small positive exposure to
the US risk factor. Disruptive tech equity, energy
transition equity and environmental sustainability
equity also has a small negative exposure to the
Asia/Japan risk factor. The disruptive tech equity and
environmental sustainability equity themes have add-
itional small positive exposures to the EM/commod-
ities factor while the other themes have a small but
negative exposure to this factor. The environmental
sustainability bond theme has a larger exposure to
the duration risk factor and a small exposure to the
market risk factor. Exposures to all other factors are
negative but small.
It should not come as a surprise that some of the
risk of the themes is explained by exposures to the
traditional risk factors in the risk model, as shown in
Table 3. The systematic volatility refers to the risk
arising from exposures to the six statistical factors in
the model. The specific volatility in Table 3 is meas-
ured by taking into account that the risk model relies
only on the first six eigenvectors of the variance
covariance matrix and that 13% of total variance is
left unexplained. Finally, the thematic volatility is the
risk component that is specific to the themes and is
neither accounted for by the systematic risk or the
specific risk of traditional asset classes. The thematic
volatility is significant for all equity themes as shown
in Table 3.
It is interesting to highlight that in both Tables 2 and
3, disruptive tech equities exhibit a smaller propor-
tion of thematic risk than other equity themes.
These exposures play an important role in determin-
ing the final allocation to themes. The fact that the

Volume 79, Number 1 25

Financial Analysts Journal | A Publication of CFA Institute
themes come such exposures to traditional risk fac-
tors needs to be taken into account when deciding
about the optimal allocation to each theme.
Strategic Asset Allocation to Thematic
Investments. We considered five different risk
profiles in our example. This starts with the conser-
vative profile for the most risk-averse investors and
goes all the way to the aggressive profile, for those
prepared to take the most risk. The three intermedi-
ate risk profiles are moderately conservative, moder-
ate and moderately aggressive, in order of falling
risk aversion.
We separate assets into four high-capacity core
assets and twelve lower-capacity satellite assets.
The core assets are Equity Europe, Equity North
America for equities and Bond EUR Investment
Grade Aggregate and Bond USD Investment Grade
Aggregate, which include treasuries and corporate
bonds, for fixed income. In turn, the twelve satellite
assets are divided into two groups. The first is made
of six diversification assets: Equity Eurozone Small
Cap, Equity North America Small Cap, Equity Pacific
Japan, Equity Global Emerging Markets, Equity
Listed Real Estate Pan Europe, Bond Global
Emerging Market. The second is made of one secu-
lar theme, Disruptive Tech Equities, and four sus-
tainable themes, Sustainable Water Equities,
Sustainable Food Equities, Energy Transition
Equities and Environmental Sustainability Equities
and Bonds. This last theme will be invested using
equities and bonds.
We shall skip the information about the core and
diversification asset exposures to the traditional fac-
tors in the asset allocation risk model. However, as
should be expected, these were taken into account
in the portfolio construction along with the expo-
sures in Table 3 for the thematic investments.
We used long-term estimates of asset returns going
as far back as 1954 for the core and the diversifica-
tion assets in the portfolio construction step. Long-
term Sharpe ratio of traditional asset classes varies
between 0.3 and 0.4. These were used to construct
long-term expected returns for the six traditional risk
factors in our risk model since they are just linear
combinations of traditional assets. For the diversify-
ing assets and for the thematic investments, the
expected returns were constructed by taking into
account their exposures to the traditional risk factors
over the last five years. For the thematic invest-
ments, we added an additional expected positive
alpha by assuming that the thematic risk will be
26
remunerated with an information ratio of 0.3. A dif-
ferent information ratio could have been used
depending on the level of confidence that a given
theme will be able to add alpha in the future. While
we believe most investors choose to allocate to the-
matic investments because they expect positive
alpha, in fact, even a zero information ratio could
have been used. In that case, thematic investments
would only add diversification, much like diversifica-
tion assets, which would simply reduce the size of
their allocation in portfolios.
The estimated returns for each core asset, for each
diversifying asset and for each thematic investment,
along with the asset allocation risk model, were used
as inputs for the portfolio construction. We used a
robust optimisation approach to portfolio construc-
tion. Unlike mean-variance optimisation, robust opti-
misation takes into account the uncertainty in the
asset return estimates and significantly reduces the
sensitivity of the final portfolio to small changes in
expected returns. The description of the framework
based on robust optimisation can be found in Yin,
Perchet, and Soupe  (2021). Details about how the
allocation of robust portfolios is constructed from
the robust optimisation algorithm was investigated
by Perchet et al. (2016).
The results from applying this framework to the con-
struction of three portfolios for each of the five risk
profiles are shown in Tables 4 and 5. For each risk
profile, we constructed three different portfolios. The
first, A, is allowed to invest only in core assets. The
other two, B and C, are based on a core-satellite
approach where the portfolio is also allowed to
invest in diversification assets for B, and in diversifi-
cation assets and thematic investments for C.
Investments in diversification assets and in themes
represent the satellite of such portfolios. An aversion
to tracking error relative to the portfolio of the same
risk profile invested only in core assets, A, regulates
the size of the satellite. The portfolios are con-
strained to be fully invested, i.e., neither leverage nor
short-selling is allowed.
 Portfolio A—Core assets only
As should be expected, for portfolios invested in
core assets only, A, the allocation to riskier assets
increases as the risk aversion is reduced. While the
conservative portfolio invests almost 60% in cash,
the riskier aggressive portfolio invests only in equi-
ties. The falling Sharpe ratio is a consequence of the
fully invested constraint. Nevertheless, despite the
lower Sharpe ratios, the riskier portfolio is still
 Table 4. Asset Weights for Five Strategic Asset Allocation Portfolios with Different Risk Profiles
Portfolio Weights (%)

Satellite

Volume 79, Number 1

Core Diversification Thematic Investments

Equities Bonds Cash Equities Bonds Equities Bonds Equities

EUR USD Euro North Global Listed Global

North Investment Investment Zone America Pacific Emerging Real Estate Emerging Disruptive Sustainable Sustainable Energy Global Env.
Risk Profile Europe America Grade Agg. Grade Agg. Cash Small Cap. Small Cap. Japan Markets Pan Europe Markets Tech Water Food Transition Environment—Equities Sustainability—Bonds

A Conservative 6.1 7.3 10.2 17.6 58.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Moderately 17.3 20.8 22.7 39.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
conservative
Moderate 27.4 32.8 12.9 26.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
Moderately 36.9 43.9 3.3 15.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
aggressive
Aggressive 49.2 50.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
B Conservative 4.4 3.8 13.6 14.5 55.4 1.3 1.8 0.7 0.8 1.5 2.2 0.0 0.0 0.0 0.0 0.0 0.0
Moderately 14.0 13.8 26.0 33.2 0.0 2.2 3.8 0.8 1.0 2.3 2.8 0.0 0.0 0.0 0.0 0.0 0.0
conservative
Moderate 23.6 24.2 15.9 20.9 0.0 2.8 4.8 0.9 1.1 2.5 3.2 0.0 0.0 0.0 0.0 0.0 0.0
Moderately 32.9 34.5 6.1 10.0 0.0 3.1 5.3 0.9 1.1 2.6 3.4 0.0 0.0 0.0 0.0 0.0 0.0
aggressive
Aggressive 44.0 40.9 0.0 0.0 0.0 4.0 5.2 1.9 1.6 2.3 0.0 0.0 0.0 0.0 0.0 0.0 0.0
C Conservative 2.9 0.0 6.4 8.2 53.6 1.0 0.0 0.2 0.0 1.3 1.6 3.0 1.6 1.6 2.3 1.1 15.1
Moderately 11.8 8.9 16.9 23.3 0.0 2.0 0.0 0.4 0.0 2.4 1.9 3.6 2.5 2.8 3.0 1.8 18.7
conservative
Moderate 21.2 19.2 7.5 11.5 0.0 2.6 0.0 0.4 0.0 2.6 1.9 3.8 2.8 3.3 3.3 2.1 17.7
Moderately 30.4 29.6 0.0 0.0 0.0 2.8 0.0 0.5 0.0 2.7 2.2 3.9 3.0 3.6 3.4 2.3 15.7
aggressive
Aggressive 42.1 35.6 0.0 0.0 0.0 3.9 0.8 0.9 0.0 2.2 0.0 5.2 1.9 1.9 2.8 2.6 0.0

Note: Risk model estimation from Jan-2003 through Dec-2021 using monthly USD returns. The projections of thematic indices on the factors of the risk model are based on regres-
sions using five years of weekly returns from Jun-2017 to May-2022. Estimates of long-term returns rely on data from Jan-1954. For illustration purposes only. Past performance is
not indicative of future performance.
Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ. Authors’ calculations.

27
Allocating to Thematic Investments
28
Table 5. Portfolio Statistics and Aggregated Asset Weights for the Five Strategic Asset Allocation Portfolios with Different Risk
Levels in Table 4
Portfolio Statistics Portfolio Weights (%)
Satellite
Annualized Tracking
Excess Return Annualized Sharpe Error Relative Thematic
Risk Profile over Cash (%) Volatility (%) Ratio to Core (%) Equities Bonds Cash Core Satellite Diversification Investments
Financial Analysts Journal | A Publication of CFA Institute

A Conservative 0.8 1.8 0.46 0.0 13.4 27.8 58.8 100.0 0.0 0.0 0.0
Moderately conservative 2.3 5.0 0.45 0.0 38.1 61.9 0.0 100.0 0.0 0.0 0.0
Moderate 3.3 8.0 0.41 0.0 60.3 39.7 0.0 100.0 0.0 0.0 0.0
Moderately aggressive 4.3 11.0 0.39 0.0 80.8 19.2 0.0 100.0 0.0 0.0 0.0
Aggressive 5.2 14.0 0.37 0.0 100.0 0.0 0.0 100.0 0.0 0.0 0.0
B Conservative 1.1 2.2 0.51 0.7 14.3 30.3 55.4 91.7 8.3 8.3 0.0
Moderately conservative 2.6 5.4 0.48 0.8 38.0 62.0 0.0 86.9 13.1 13.1 0.0
Moderate 3.6 8.3 0.44 0.9 60.0 40.0 0.0 84.6 15.4 15.4 0.0
Moderately aggressive 4.6 11.3 0.41 0.9 80.5 19.5 0.0 83.5 16.5 16.5 0.0
Aggressive 5.3 14.0 0.38 0.7 100.0 0.0 0.0 85.0 15.0 15.0 0.0
C Conservative 1.4 2.3 0.62 0.8 15.2 31.2 53.6 71.1 28.9 4.2 24.7
Moderately conservative 3.0 5.4 0.55 1.1 39.1 60.9 0.0 60.9 39.1 6.6 32.5
Moderate 4.1 8.3 0.49 1.1 61.5 38.5 0.0 59.4 40.6 7.5 33.0
Moderately aggressive 5.0 11.3 0.45 1.2 82.1 17.9 0.0 60.0 40.0 8.2 31.8
Aggressive 5.7 13.9 0.41 1.0 100.0 0.0 0.0 77.7 22.3 7.9 14.4

Note: Based on the same calculations used for Table 4. For illustration purposes only. Past performance is not indicative of future performance.
Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ. Authors’ calculations.
 Allocating to Thematic Investments

Table 6. Risk Exposures of Selected Asset Allocation Portfolios from Table 4

Conservative Moderate Aggressive
Portfolios A (%) B (%) C (%) A (%) B (%) C (%) A (%) B (%) C (%)

Risk factors
Market risk 1.6 2.1 2.2 7.4 7.8 7.9 12.5 12.6 12.6
Duration 0.3 0.5 0.5
1.3
1.1
1.1
4.2
4.2
4.1
EM/Commodities
0.4
0.4
0.3
1.6
1.5
1.4
2.5
2.5
2.3
Corporate spreads
0.4
0.3
0.3
1.0
0.9
0.9
1.2
1.3
1.3
US 0.3 0.2 0.3 1.2 1.1 1.1 1.6 1.6 1.6
Asia/Japan
0.2
0.2
0.2
0.6
0.6
0.7
0.7
0.7
0.6
Systematic volatility 1.8% 2.2 2.3 7.8 8.2 8.2 13.5 13.6 13.6
Specific volatility 0.2 0.2 0.0 1.7 1.4 1.3 3.4 3.1 3.0
Thematic volatility 0.0 0.0 0.4 0.0 0.0 0.6 0.0 0.0 0.6
Total volatility 1.8 2.2 2.3 8.0 8.3 8.3 14.0 14.0 13.9

Note: Based on the same calculations used for Table 4.

Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ. Authors’ calculations.

expected to deliver higher returns than the less risky in A ranging from 0.8% for the conservative portfolio
portfolios invested only in core assets. to 1.2% for the moderately aggressive portfolio.

It is interesting to see that some sustainable thematic

 Portfolio B—Core assets and diversifica-
investments in C simply replace some of the diversifi-
tion satellite
cation assets in B, as is the case for Global Equity
If we allow for a satellite invested in diversifying
Emerging Markets which no longer shows up in the
assets, then we find higher expected returns for the
portfolios. The allocation to core assets is also
core-satellite portfolios than for their counterparts
affected, with the allocation to core equities and
invested only in core assets. The size of the satellite
core fixed income significantly reduced. Finally,
in portfolio B increases with falling risk aversion from
Environmental Sustainability bonds, less risky than
about 8% for the conservative portfolio to more than
equity thematic investments, take a larger weight in
15% for the riskier portfolios. The allocation to riskier
the satellite. Of course, the choice of themes and
assets in the satellite increases with falling risk aver-
their exposures to traditional risk factors, as well as
sion, for example if the aggressive portfolio no longer
the expected returns used as an input, play a key
invests in emerging bonds. As expected, all Sharpe
role in explaining these results. The fact that the-
ratios and expected returns are higher for this more
matic investments come with an additional positive
diversified portfolio in B than for their respective
alpha plays an important role in the final allocations.
counterparts in A. The tracking error resulting from
the satellite ranges from 0.7% for the conservative
Strategic Risk Allocation to Thematic
and aggressive portfolios to 0.9% for the moderate
Investments. In Table 6, we show the factor risk
and the moderately aggressive portfolios. Long only
exposures of the portfolios in Table 4 for the
constraints make it more attractive to reduce slightly
Conservative, Moderate and Aggressive risk profiles.
the tracking error for the aggressive portfolio.
The results are calculated by linearly combining the
exposures of core, diversification and thematic assets
 Portfolio C—Core assets, diversification and
to the model risk factors. For the risk not explained
thematics satellite
by exposures to the model risk factors we distinguish
When thematic investments are also allowed, C, then between that arising from either specific risk from
we find that the Sharpe ratios and expected returns core or diversification assets, which we call specific
increase further when compared to their respective volatility in Table 6, and that from thematic assets,
counterparts in B or A. The size of the satellite port- which we call thematic volatility.
folio increases further, to about 29% for the conser-
vative portfolio and almost 41% for the moderately We find that as we allow portfolios to diversify, from
aggressive portfolio, resulting in a tracking error rela- A to B, the exposures to risk factors are about the
tive to their counterparts invested only in core assets same. Portfolios B are just more diversified, with

Volume 79, Number 1 29

Financial Analysts Journal | A Publication of CFA Institute

Table 7. Tracking Error of Portfolios C Relative to Portfolios B of Similar Risk Profile in

Table 4
Moderately Moderately
Portfolio C—Portfolio B Conservative (%) Conservative (%) Moderate (%) Aggressive (%) Aggressive (%)

Risk factors
Market risk 0.1 0.0 0.0 0.0 0.0
Duration 0.0 0.0 0.0 0.0 0.1
EM/Commodities 0.1 0.1 0.1 0.1 0.2
Corporate spreads 0.0 0.0 0.0 0.0 0.0
US 0.1 0.0 0.0 0.0 0.1
Asia/Japan
0.1 0.0 0.0 0.0 0.0
Systematic tracking error 0.1 0.1 0.1 0.1 0.2
Specific tracking error 0.1 0.1 0.2 0.2 0.0
Thematic tracking error 0.4 0.6 0.6 0.7 0.6
Tracking error 0.4 0.6 0.7 0.7 0.7

Note: Based on the same calculations used for Table 4.

Source: Bloomberg, ECPI, Foxberry, FTSE, Morningstar, MSCI and NASDAQ. Authors’ calculations.

lower specific volatility, in particular for the
Moderate and Aggressive profiles. However, port-
folios C add specific risk (thematic volatility) derived
from allocations thematic investments. This is not
surprising since we attribute a positive alpha to this
risk—the portfolio will seek to mimic the exposures
to other risk factors while creating exposure to the-
matic volatility. However, even if we had not done
so, thematic investments would still appear in the
portfolios for as long as they would also bring add-
itional diversification.
Finally, it is also instructive to investigate the track-
ing error of portfolios C relative to portfolio B for
the different risk profiles, respectively. As shown
in Table 7, most of the tracking error created
with the introduction of thematic investments is
made of thematic volatility. As expected, the port-
folio optimiser seeks to allocate to thematic assets
so as to create an exposure to the specific risk of
thematic assets to which we attributed a posi-
tive alpha.
Conclusions
Themes are structural trends expected to significantly
impact economies and redefine business models. For
this reason, they should have an impact on the
returns and risks of investments with greater expos-
ure to those structural trends. Thematic investing
allows portfolios to earn excess returns generated
from assets that have their returns impacted by the
structural changes underlying relevant themes.
Thematic investing can be seen as an additional
dimension in portfolios that transcends the classifica-
tions regions, styles and sectors while not being fully
independent from them. Constructing portfolios with
thematic investments requires adequate risk manage-
ment in order to assess by how much a given theme
is exposed to traditional risk factors and to assess
how much alpha a given theme is likely to add to the
portfolio in excess of such exposures. We provide
guidance on how to investigate the exposures of
themes to traditional risk factors. We also propose a
framework based on robust optimisation as an allo-
cating tool capable of handling the different risk
exposures of assets and themes while taking into
account the expected returns from the different
assets and the expected alphas for themes. With this,
we give a detailed example with an allocation to five
different themes played in equities and bonds, and
show how the allocation changes according to the
level of risk aversion of the investor. The results
shown in the paper are for illustrative purposes only,
not meant for investment advice, and illustrate how
the framework can be used to allocate to thematic
investments.
Extensions of this work could include the use of
more detailed risk models. For example, when
designing an allocation for equity only portfolios we
may envisage using a risk model that explicitly
includes styles and sectors as factors. The framework
here proposed can be easily adapted for the purpose.
That would require more inputs, but would allow for
better allocating to themes with stronger exposures
to styles or sectors than the examples used here.

30
 Allocating to Thematic Investments

Appendix A: Asset Allocation Model Risk Factors

Table A1. Statistical Risk Factors from Risk Model

Statistical Factors
Market Duration Emerging Corporate Asia/Japan
Tickers Risk (%) (%) /Commodities (%) Spreads (%) US (%) (%)

Equity EMU NDDLEURO Index 28.2
14.7
25.6
16.5 3.1
10.3
Equity EMU Small Cap NCLDEMU Index 29.1
12.0
21.2
2.9 0.6
12.5
Equity UK NDDLUK Index 27.5
9.3
10.1
23.2
4.1
22.1
Equity USA NDDUUS Index 28.8
12.8
2.4
19.6 8.1
9.4
Equity USA Small Cap RU20INTR Index 27.0
16.7
3.3
16.8 12.4
15.4
Equity Pacific Japan NDDLJN Index 21.8
20.9
17.4
18.2
14.9 79.7
Equity NDUEEGF Index 28.2
5.5 27.8
17.6 6.6 14.0
Emerging Global
Bond EUR Sovereign LEATTREU Index 3.0 47.8
32.5
22.9
38.4 10.0
Bond EUR Corporate LECPTREU Index 21.8 33.5
19.0 25.6
30.6 9.4
Investment Grade
Bond EUR High Yield LF88TREU Index 28.3 2.4
2.4 51.2 1.9 8.8
Bond USD Sovereign LUATTRUU Index
6.6 49.6 7.8
39.3 18.1
10.8
Bond USD Corporate LUACTRUU Index 19.5 42.5 6.0 14.3 12.5 2.8
Investment Grade
Bond USD High Yield LF89TRUU Index 28.8 3.8 9.6 40.9 21.1
2.0
Bond EMD HC JPGCCOMP Index 26.0 26.4 19.4 4.3 19.2 5.8
Sov Global
Bond EMD LC JGENVUUG Index 24.4 13.1 38.4
25.1 22.8 15.9
Sov Global
Diversification Real TRNHUE Index 25.4 1.6
32.6
0.9
7.3
36.1
Estate Pan Europe
Diversification BCOMXAL Index 17.9
8.4 56.3
1.7
72.1
19.9
Commodity Global

Note: Risk model estimation from Jan-2003 through Dec-2021 using monthly USD returns based on a PCA model.
Source: Bloomberg (Bond and Commodities), FTSE EPRA (Listed Real Estate), MSCI (Equities except US Small Caps) and FTSE
Russell (US Small Caps). Authors’ calculations.

Empirical Framework for Region, Style,
and Sector Exposures. The returns of thematic
indices in excess of cash, XRthematic , can be decom-
posed into three components:
XRthematic ¼ XRthematic
market þ XRregion style sector þ XRpure theme
thematic thematic
(1)
where XRthematic
market is a term representing what is
explained by the exposure to the equity market as a
whole, XRthematic
region style sector is a term representing what is
explained by regions, styles and sectors and
XRthematic
pure theme is what cannot be explained from
correlations with the equity market as a whole
or with regions, styles or sectors. In order to
estimate each component while making sure they are
orthogonal to each other, the following methodology
was used:
 Step 1: The thematic index i returns in excess of cash
were regressed against the global market index (MSCI
ACWI Index is used for equity indices and Bloomberg
Barclays Global Aggregate Index for bond indices)
returns in excess of cash:
XRthematic; i ¼ a þ bmarket
thematic; i XRmarket þ ethematic; i
market
(2)
Where XRthematic; i the vector of excess returns of the-
matic i over cash, ethematic
market; i the vector of residual from the
regression and XRmarket the vector of excess returns of
global market index over cash.

Volume 79, Number 1 31

Financial Analysts Journal | A Publication of CFA Institute

 Step 4: The pure thematic returns dpure theme

thematic, i is calculated
 Step 2: The returns of each regional, style and sector
index j in excess of cash are regressed against the returns as the residual of the linear regression of ethematic
market on the
of global market index in excess of cash: selected j region, style or sector residual returns:

 pure theme
emarket
thematic; i ¼ bregion style sector; j eregion style sector; j þ dthematic; i
market market
(5)
XRregion style ector; j ¼ a þ bmarket
region style sector; j XRmarket

þ emarket
region style sector; j (3)
Lasso regressions used in Step 3 reduce overfitting
Where XRregion style sector, j the vector of excess returns of since 128 equity and 46 fixed income explanatory
regional, style and sector index j and emarket
region style sector, j rep- variables were considered while there are only 260
resents the vector of residual from each regression, weekly returns in each regressed thematic time ser-
respectively. ies. When applying Lasso regression, we chose to
estimate the k shrinkage parameter using the meth-
odology proposed by Friedman, Hastie, and
 Step 3: A Lasso regression, as originally proposed by
Tibshirani (2010) retaining the k that yields the
Tibshirani (1996), is used to find the region, style and sector
residuals relative to the global market, as in Step 2, that mean-squared error from cross-the validation pro-
have the highest and most significant contributions towards cess that is within one standard error of the min-
explaining the thematic residuals relative to the global mar- imum. The complete list of region, style and sector
ket, as in Step 1: indices used in this framework can be found in
X
T X
J X
J   Table A2.
ðemarket 2 bj 
thematic; i; t
bj emarket
region style sector; j; t Þ þk (4)
t¼1 j j¼1

Table A2. List of Region, Style, and Sector Equity Indices and Bond
Indices Used in Lasso Regressions
Index for Regions Ticker

MSCI EMU MXEM Index

MSCI UK NDDLUK Index
MSCI Europe Small Cap NCEDE15 Index
MSCI USA NDDUUS Index
MSCI USA Small Cap NCUDUS Index
MSCI Japan NDDLJN Index
MSCI Pacific Ex Japan NDDUPFXJ Index
MSCI Emerging Markets NDUEEGF Index
Index for Styles Ticker

MSCI EMU Growth M7EM000G Index

MSCI EMU Value M7EM000V Index
MSCI EMU Large Value MLVLEMUN Index
MSCI EMU Small Value MSVLEMUN Index
MSCI EMU Large Growth MLGLEMUN Index
MSCI EMU Small Growth MGSLEMUN Index
MSCI UK Small Cap NCLDUK Index
MSCI UK Value NDLVUK Index
MSCI UK Growth NLDGUK Index
MSCI USA Value M1US000V Index
MSCI USA Growth M1US000G Index
MSCI USA Quality M1USQU Index
MSCI USA Momentum M1USMMT Index
MSCI USA Mid Cap M1USMC Index
MSCI USA Large Cap M1USLC Index
MSCI USA Small Cap Value MXUS00SV Index
continued

32
 Allocating to Thematic Investments

Table A2. List of Region, Style, and Sector Equity Indices and Bond
Indices Used in Lasso Regressions (continued)
Index for Styles Ticker

MSCI USA Small Cap Growth MXUS00SG Index

MSCI USA Small Cap Multi-Factor M1USDM Index
MSCI Japan Small Cap NCLAJN Index
MSCI Japan Value NAVLJN Index
MSCI Japan Growth NAGLJN Index
MSCI Japan Quality M4JPQU Index
MSCI Japan Small Growth M7JP00SG Index
MSCI Japan Large Growth M7JP0002 Index
MSCI Japan Large Value M7JP0001 Index
MSCI Japan Small Value M7JP00SV Index
MSCI Pacific Ex Japan IMI Value MIVLPXJN Index
MSCI Pacific Ex Japan IMI Growth MGILPXJN Index
MSCI Pacific Excluding Japan Small Cap GCALPXJ Index
MSCI Pacific Ex Japan Small Growth MGSLPXJN Index
MSCI Pacific Ex Japan Small Value MSVLPXJN Index
MSCI Pacific Ex Japan Large Value MLVLPXJN Index
MSCI Pacific Ex Japan Large Growth MLGLPXJN Index
MSCI Pacific Ex Japan Large Cap MLCLPXJN Index
MSCI Emerging Markets Large Cap M1EFLC Index
MSCI Emerging Markets Mid Cap M1EFMC Index
MSCI Emerging Markets Small Cap M1EFSC Index
MSCI Emerging Markets Value M1EF000V Index
MSCI Emerging Markets Growth M1EF000G Index
MSCI Emerging Markets SMID Value M1EF0009 Index
MSCI Emerging Markets Large Growth M1EF0002 Index
MSCI Emerging Markets Large Value M1EF0001 Index
MSCI Emerging Markets Mid Value M1EF0003 Index
MSCI Emerging Markets Mid Growth M1EF0004 Index
MSCI Emerging Markets Small Value M1EF0005 Index
MSCI Emerging Markets Small Growth M1EF0006 Index

Index for Sectors Ticker

MSCI Japan Consumer Discretionary MXJP0CD Index

MSCI Japan Consumer Staples MXJP0CS Index
MSCI Japan Information Technology MXJP0IT Index
MSCI Japan Energy MXJP0EN Index
MSCI Japan Financials MXJP0FN Index
MSCI Japan Industrials MXJP0IN Index
MSCI Japan Real Estate MXJP0RE Index
MSCI Japan Telecommunications and Services MXJP0TC Index
MSCI Japan Utilities MXJP0UT Index
MSCI Japan Health Care MXJP0HC Index
MSCI Japan Materials MXJP0MT Index
MSCI Emerging Markets Consumer Discretionary M1EF0CD Index
MSCI Emerging Markets Consumer Staples M1EF0CS Index
MSCI Emerging Markets Energy M1EF0EN Index
MSCI Emerging Markets Financials M1EF0FN Index
MSCI Emerging Markets Health Care M1EF0HC Index
MSCI Emerging Markets Industrials M1EF0IN Index
MSCI Emerging Markets Information Technology M1EF0IT Index
MSCI Emerging Markets Materials M1EF0MT Index
continued

Volume 79, Number 1 33

Financial Analysts Journal | A Publication of CFA Institute

Table A2. List of Region, Style, and Sector Equity Indices and Bond
Indices Used in Lasso Regressions (continued)
Index for Sectors Ticker

MSCI Emerging Markets Telecommunications and Services M1EF0TC Index

MSCI Emerging Markets Utilities M1EF0UT Index
MSCI EMU Materials NDLUMAT Index
MSCI EMU Consumer Discretionary NDLUCDIS Index
MSCI EMU Consumer Staples NDLUCSTA Index
MSCI EMU Telecommunications and Services NDLUTEL Index
MSCI EMU Energy NDLUENR Index
MSCI EMU Health Care NDLUHC Index
MSCI EMU Utilities NDLUUTI Index
MSCI EMU Industrials NDLUIND Index
MSCI EMU Information Technology NDLUIT Index
MSCI EMU Financials NDLUFNCL Index
MSCI EMU Real Estate NDLURE Index
STOXX Europe 600 Automobiles SXAR Index
STOXX Europe 600 Banks SX7R Index
STOXX Europe 600 Construction SXOR Index
STOXX Europe 600 Health Care SXDR Index
STOXX Europe 600 Insurance SXIR Index
STOXX Europe 600 Media SXMR Index
STOXX Europe 600 Oil and Gas SXER Index
STOXX Europe 600 Technology SX8R Index
STOXX Europe 600 Telecommunications SXKR Index
STOXX Europe 600 Energy S600ENR Index
STOXX Europe 600 Industrials SXIDUR Index
STOXX Europe 600 Utilities SX6R Index
STOXX Europe 600 Basic Materials SXBSCR Index
MSCI UK Consumer Discretionary MXGB0CD Index
MSCI UK Consumer Staples MXGB0CS Index
MSCI UK Energy MXGB0EN Index
MSCI UK Financials MXGB0FN Index
MSCI UK Utilities MXGB0UT Index
MSCI UK Industrials MXGB0IN Index
MSCI UK Materials MXGB0MT Index
MSCI UK Information Technology MXGB0IT Index
MSCI UK Health Care MXGB0HC Index
MSCI UK Telecommunications and Services MXGB0TC Index
MSCI US Energy M1US0ENE Index
MSCI US Consumer Discretionary M1US0CDE Index
MSCI US Consumer Staple M1US0CSE Index
MSCI US Healthcare M1US0HCE Index
MSCI US Materials M2US0MTI Index
MSCI US Industrials M1US0INI Index
MSCI US Financials M1US0FNI Index
MSCI US Information Technology M1US0ITI Index
MSCI US Telecommunications and Services M1US0TCI Index
MSCI US Utilities M1US0UTI Index
MSCI Australia Materials GICS Sector MXAU0MT Index
MSCI Australia Energy GICS Sector MXAU0EN Index
MSCI Australia Telecommunications and Services MXAU0TC Index
MSCI Canada Materials MXCA0MT Index
MSCI Canada Energy MXCA0EN Index
MSCI Canada Telecommunications and Services MXCA0TC Index
continued

34
 Allocating to Thematic Investments

Table A2. List of Region, Style, and Sector Equity Indices and Bond
Indices Used in Lasso Regressions (continued)
Index for Sectors Ticker

EPRA Europe RPRA Index

EPRA USA RUUS Index
EPRA Japan RYJP Index

Index for Bonds Ticker

Bloomberg Barclays Euro Agg Government LEEGTREU Index

Bloomberg Euro Govt Inflation-Linked Bond All Maturities BEIG1T Index
Bloomberg Euro Aggregate Corporate LECPTREU Index
Bloomberg Euro High Yield Bond LF88TREU Index
Bloomberg US Government LUAGTRUU Index
Bloomberg US Treasury Inflation Notes LBUTTRUU Index
Bloomberg US Corporate LUACTRUU Index
Bloomberg US Corporate High Yield LF98TRUU Index
ICE BoAML Japan Government G0Y0 Index
ICE BoAML Japan Corporate JC00 Index
ICE BoAML Asian Dollar High Yield Corporate ACHY Index
J.P. Morgan Emerging Markets Bond Global JPEICORE Index
J.P. Morgan Emerging Local Markets JPPUELM Index
Bloomberg Barclays Pan Euro Corporate Health Care H31360EU Index
Bloomberg Barclays Pan-European Aggregate Utility I02552EU Index
Bloomberg Barclays Euro Corporate 500 Industrials Technology H04374EU Index
Bloomberg Barclays US Agg Industrial LUAITRUU Index
Bloomberg Barclays US Agg Utility LUAUTRUU Index
Bloomberg Barclays US Agg Finance LUAFTRUU Index
Bloomberg IG Health Care I00383US Index
Bloomberg IG Building Materials I00360US Index
S and P USD Global IG Corporate Bond Information Technology SPUGIITT Index
Bloomberg IG Consumer Noncyclical Total Return I00379US Index
Bloomberg IG Consumer Cyclical Total Return I00367US Index
Bloomberg IG Energy Total Return I00388US Index
Bloomberg IG Communications Total Return I00711US Index
Bloomberg Barclays Pan-European Aggregate Industrial I02551EU Index
Bloomberg Barclays Pan-European Aggregate Finance I02550EU Index
Bloomberg Barclays Pan-European Aggregate Local Authorities I09540EU Index
Bloomberg Barclays Pan-Euro Corporate Basic Industrial H31335EU Index
Bloomberg Barclays Pan-Euro Corporate Chemicals H31336EU Index
Bloomberg Barclays Pan-Euro Corporate Metals and Mining H31337EU Index
Bloomberg Barclays Pan-Euro Corporate Paper H31338EU Index
Bloomberg Barclays Pan-Euro Corporate Capital Goods H31339EU Index
Bloomberg Barclays Pan-Euro Corporate Aerospace Defense H31340EU Index
Bloomberg Barclays Pan-Euro Corporate Building Materials H31341EU Index
Bloomberg Barclays Pan-Euro Corporate Diversified Manufacturing H31342EU Index
Bloomberg Barclays Pan-Euro Corporate Construction Machinery H31343EU Index
Bloomberg Barclays Pan-Euro Corporate Packaging H31344EU Index
Bloomberg Barclays Pan-Euro Corporate Consumer Non-Cyclical H31356EU Index
Bloomberg Barclays Pan-Euro Corporate Beverage H31357EU Index
Bloomberg Barclays Pan-Euro Corporate Food H31359EU Index
Bloomberg Barclays Pan-Euro Corporate Environmental H31345EU Index
Bloomberg Barclays Pan-Euro Corporate Pharmaceuticals H31361EU Index
Bloomberg Barclays Pan-Euro Corporate Energy H31365EU Index
Bloomberg Barclays Pan-Euro Corporate Technology H31371EU Index

Volume 79, Number 1 35

Financial Analysts Journal | A Publication of CFA Institute
Editor's Note
Submitted 16 February 2022
Accepted 10 August 2022 by William N. Goetzmann
Notes
1. Lasso (least absolute shrinkage and selection operator) is
a regression analysis method that performs both
variable selection and regularisation in order to
enhance the prediction accuracy and interpretability of
the resulting statistical model. Regularisation is the
process of adding information in order to prevent
References
Asness, C. S., T. J. Moskowitz, and L. H. Pedersen. 2013.
“Value and Momentum Everywhere.” The Journal of Finance 68
(3): 929–85. doi:10.1111/jofi.12021.
Bass, R., S. Gladstone, and A. Ang. 2017. “Total Portfolio
Factor, Not Just Asset, Allocation” The Journal of Portfolio
Management 43 (5): 38–53. doi:10.3905/jpm.2017.43.5.038.
Blitz, D. 2021. “Betting against Quant: Examining the Factor
Exposures of Thematic Indexes.” The Journal of Index Investing
12 (3): 5–14. doi:10.3905/jii.2021.1.111.
Fama, E. F., and K. R. French. 1992. “The Cross-Section of
Expected Stock Returns.” The Journal of Finance 47 (2):
427–65. doi:10.1111/j.1540-6261.1992.tb04398.x.
Friedman, J., T. Hastie, and R. Tibshirani. 2010. “Regularization
Paths for Generalized Linear Models via Coordinate Descent.”
Journal of Statistical Software 33 (1): 1–22. doi:10.18637/jss.
v033.i01.
Hastie, T., R. Tibshirani, and M. Wainwright. 2015. “Statistical
Learning with Sparsity: The Lasso and Generalizations.” In
Monographs on Statistics & Applied Probability. Boca Raton:
Chapman & Hall/CRC.
Haugen, R. A., and N. Baker. 1996. “Commonality in the
Determinants of Expected Stock Returns.” Journal of Financial
Economics 41 (3): 401–39. doi:10.1016/0304-
405X(95)00868-F.
nie, and R. Leote de
Heckel, T., Z. Amghar, I. Haik, O. Laple
Carvalho. 2019. “Factor Investing in Corporate Bond Markets:
Enhancing Efficacy through Diversification and Purification!”
The Journal of Fixed Income 29 (3): 6–21. doi:10.3905/jfi.2019.
1.074.
36
overfitting or solve an ill-posed problem (for more on
Lasso regressions see Hastie, Tibshirani, and
Wainwright 2015).
2. The major thematic index providers including S&P Kensho,
MSCI and FTSE report their thematic indices in USD while
ECPI in EUR.
Houweling, P., and J. van Zundert. 2017. “Factor Investing in
the Corporate Bond Market.” Financial Analysts Journal 73 (2):
100–15. doi:10.2469/faj.v73.n2.1.
Israel, R., D. Palhares, and S. Richardson. 2018. “Common
Factors in Corporate Bond Returns.” Journal of Investment
Management 16:17–46.
re, F. Soupe
Issaoui, T., R. Perchet, O. Retie , C. Yin, and R.
Leote de Carvalho. 2022. “Mass Customization of Asset
Allocation.” Journal of Investing 31 (3): 73–97. doi:10.3905/joi.
2021.1.217.
Morningstar. 2021. “Explore the Global Thematic Funds
Landscape.” https://www.morningstar.com/lp/global-thematic-
fund-landscape.
Perchet, R., X. Lu, R. Leote de Carvalho, and T. Heckel. 2016.
“Insights into Robust Robust Portfolio Optimisation:
Decomposition into Mean-Variance and Risk-Based Portfolio.”
Journal of Investment Strategies 6 (1): 1–24.
Scherer, B. 2015. “Core–Satellite Investing: Budgeting Active
Manager Risk.” In Portfolio Construction and Risk Budgeting,
edited by B. Scherer, Chapter 14, 5th ed. London: Risk Books.
Tibshirani, R. 1996. “Regression Shrinkage and Selection via the
Lasso.” Journal of the Royal Statistical Society: Series B
(Methodological) 58 (1): 267–88. doi:10.1111/j.2517-6161.
1996.tb02080.x.
Walmsley, D. 2021. Targeting Impact: Integrated ESG and the
Role of Thematic Strategies in Asset-Owner Portfolios. Stamford,
CT: Coalition Greenwich, CRISIL.
. 2021. “A Practical Guide to
Yin, C., R. Perchet, and F. Soupe
Robust Portfolio Optimisation.” Quantitative Finance 21 (6):
911–28. doi:10.1080/14697688.2020.1849780.
Financial Analysts Journal | A Publication of CFA Institute
https://doi.org/10.1080/0015198X.2022.2150500
Targeting Macroeconomic
Exposures in Equity
Portfolios: A Firm-Level
Measurement Approach for
Out-of-Sample Robustness
Mikheil Esakia and Felix Goltz
Mikheil Esakia is quantitative research analyst at Scientific Beta, Nice, France. Felix Goltz is research director at Scientific Beta and associate
researcher at EDHEC Business School, Nice, France. Send correspondence to Felix Goltz at felix.goltz@edhec.edu
We propose firm-level measures of
exposures to macroeconomic risks
that substantially improve out-of-sam-
ple robustness compared to standard
estimation approaches. Systematic
equity strategies constructed from
T economic risks in equity investing. Managing macroeconomic
such measures offer more consistent
macro exposures out of sample than
strategies that allocate across sectors
or equity-style factors. We do not
find significant cost to the perform-
ance of such systematic strategies in
exchange for targeting exposures to
macroeconomic risks, such as interest
rates, term spread, credit spread, or
inflation. Our methodology can be
used to construct equity portfolios
for investors who have hedging
demands or active views regarding
macroeconomic conditions.
Keywords: macroeconomic risks;
out-of-sample exposure; systematic
equity strategies
Disclosure: No potential conflict of interest
was reported by the author(s).
PL Credits: 2.0
Volume 79, Number 1
Research
Introduction
his paper proposes a method of targeting exposures to macro-
risks is an important task for equity investors because their port-
folios may come with substantial exposures to not only stock market
risk but also broader macroeconomic risks, such as interest rate risk,
inflation risk, or recession risk.
Beyond the exposure that the broad market index may have to such
macroeconomic factors, investors may pick up cross-sectional differ-
ences in exposure. There is ample evidence that such differences in
stocks’ macro exposures can be substantial. For example, Chen, Roll,
and Ross (1986) found that portfolios sorted on market-capitalisation
had different exposures to various macroeconomic variables, including
the term spread and the credit spread. Similar findings were docu-
mented by Petkova (2006) across double-sorted portfolios on book-
to-price and market capitalisation and by Boons (2016) across
stock-level returns. Such implicit exposures will not necessarily align
with what an investor would target when considering macroeconomic
risk exposures explicitly. If an investor is already exposed to a given
macroeconomic risk outside their equity allocation, the total portfolio
will not be well diversified and suffer losses if a given macroeconomic
risk materialises. For example, an increase in credit spreads often sig-
nals a slowdown in economic activity and a negative impact on labour
income. Investors who take above-average exposure to credit risk in
We would like to thank No€el Amenc, two anonymous referees, and the editors of
the journal for helpful comments and suggestions.
© 2022 CFA Institute. All rights reserved. 37
Financial Analysts Journal | A Publication of CFA Institute
their equity portfolios might load up on exposure
they already have through their labour income.
Similarly, the liabilities of an investor increase with
declining interest rates. Investors with liabilities thus
should consider how their equity portfolio reacts to
changes in interest rates.
In practice, managing macro risks is often done via
allocation across asset classes. There are various
studies that focus on constructing mimicking port-
folios of different economic variables through
dynamic asset allocation (see, e.g., Jurczenko and
Teiletche 2020 and Swade et al. 2021). In contrast,
this paper focusses on a single asset class. Focussing
on equity portfolios poses challenges because expos-
ure to macroeconomic risks may be less diverse and
more difficult to measure. However, investors may
have preference to manage macroeconomic risks in
equities for different reasons. For example, equities
have delivered the largest premium historically
amongst traditional asset classes. Investors could be
better off if they can achieve their macro risk objec-
tives whilst collecting the long-term equity premium,
especially during the times when compensation for
other asset classes, such as long-term bonds or infla-
tion-protected securities, are at historical lows.
In a related paper, Chousakos and Giamouridis (2020)
assess the premia in equity markets associated with
macro risks, such as growth, fragility, and volatility.
Whilst their paper focusses on pricing and perform-
ance enhancements from new types of macro fac-
tors, our paper focusses on measurement of
exposures and robustness out of sample to long-
established macro-state variables.
A better understanding of how to measure exposures
reliably is valuable for investors. Reliable measure-
ment of how different stocks are exposed to macro-
economic risks would allow investors to build equity
portfolios that hedge undesired macro risks. In add-
ition, targeting macroeconomic exposures is useful
for investors who have views on the future eco-
nomic conditions.
However, this is easier said than done. One of the
major challenges to efficiently managing macro risks
in equity portfolios is to reliably estimate the expo-
sures. Most academic studies are concerned with ex-
post analysis over the long term. Common estimation
techniques available in the literature do not consider
out-of-sample reliability. For example, one of the ear-
liest influential studies on this subject by Chen, Roll,
and Ross (1986) applied the ordinary least-squares
(OLS) approach to monthly returns to obtain betas to
38
different macroeconomic variables, such as inflation,
growth, term spread, and credit spread. Bali, Brown,
and Caglayan (2014) used monthly estimates of
macroeconomic uncertainty to measure firm-level
exposures using OLS regression. Levi and Welch
(2017) find that roughly two-thirds of the papers
published in the top academic finance journals
between 2013 and 2015 estimate betas using a
monthly frequency over samples of 1 to 5 years.1
Whilst the use of simple estimation procedures is
suitable for ex-post analysis of investment strategies,
the reliability of out-of-sample exposures is the key
concern when investors’ objective is to target these
macro exposures in their portfolios.
The empirical evidence shows that estimating the
relation between equities and macroeconomic varia-
re, and Signori (2012) authored
bles is hard. Ang, Brie
one of the few studies specifically focussing on the
out-of-sample reliability of macro betas. They find
that a portfolio of stocks with high inflation betas
over the recent past has inflation betas that are
indistinguishable from zero out of sample. Fama and
French (1997) emphasise that exposures of industry
portfolios to equity risk factors are measured with a
high margin of error. Boons (2016) acknowledges
that estimating exposures to macroeconomic varia-
bles is difficult and proposes various statistical adjust-
ment to achieve out-of-sample reliability.2 In general,
naïve estimation techniques that are common in
academic studies, such as standard regression betas
estimated on a low frequency, are not suitable for
out-of-sample measurement of risk exposures.
In light of the difficult measurement of macro expo-
sures, investment decisions in practice are often dis-
cretionary. For example, some actively managed
funds explicitly aim at strong outperformance in high
inflationary periods. However, discretionary choices
to target macro exposures are hard to evaluate given
their lack of transparency and replicability.
To target macro exposures of equity portfolios, the
approach of choice in investment practice is to allo-
cate across predefined asset classes, such as sectors
or factors. However, whilst such building blocks are
readily available, they have never been designed to
reliably capture macro risks. Using such blunt tools
may in fact compound the measurement challenges
when trying to target macro exposures.
In contrast to popular practice, we propose a system-
atic approach that is transparent and replicable. We
also go beyond analysing sector differences and

instead exploit the firm-level heterogeneity of risk
exposures. Our approach relies on three ingredients.
First, we choose forward-looking macroeconomic
variables that are derived from financial markets. This
allows us to identify links between stock-level
returns and macroeconomic risks. Commonly used
macroeconomic variables, such as realised inflation or
growth, are backward looking and do not reflect
investor expectations. They also come at a lower fre-
quency that aggravates the challenge of exposure
measurement. Our macroeconomic variables are
derived from asset prices, such as fixed-income
securities. They are available at higher frequency and
naturally incorporate investor expectations, thus
improving the reliability of exposure measurement.
Second, we apply robust measurement tools to esti-
mate stock-level macro exposures out of sample. Our
estimation technique is superior to the naïve
approach that uses OLS on data that come at a rela-
tively lower frequency.
Third, we construct dedicated equity portfolios from
stock-level exposures in a fully systematic way. We
show that our stylised strategies can provide stronger
out-of-sample macro exposures than other popular
approaches in the industry that allocate across preex-
isting equity products, such as sector or factor indi-
ces. This finding is consistent across all the macro
variables that we test.
In addition to capturing macroeconomic exposures in
a reliable fashion, our stylised strategies come with
performance in line with market returns, irrespective
of the macroeconomic risks they target. Therefore,
such strategies provide access to the long-term
equity premium along with targeting the macroeco-
nomic risks that an investor might be con-
cerned with.3
Selecting Forward-Looking Variables. In
general, realised quantities of fundamental economic
measures are not suited when analysing movements
in asset prices. In liquid markets, information is
quickly reflected in prices. Prices of financial assets
that are claims to future cash flows depend on
investors’ expectations. Therefore, looking at realised
measures of economic variables that lack information
about future economic conditions does not allow to
capture the relation between asset prices and eco-
nomic conditions. Instead, we rely on macro-state
variables that are forward looking and quickly incorp-
orate available information. For example, if investors’
expectations about future interest rates change, this
Volume 79, Number 1
Targeting Macroeconomic Exposures in Equity Portfolios
information is reflected in asset prices, in particular
treasury bond prices. Similarly, the aggregate credit
spread reflects investors’ risk tolerance and whether
future economic conditions are expected to alter
firms’ ability to meet their obligations.
We primarily rely on U.S. data. However, our
approach can be easily extended to other geograph-
ical areas where the respective macro variables are
available. Below, we list the macro variables that we
use in our analysis:4
1. Short-term interest rates, defined as a yield on
3-month U.S. Treasury bills.
2. Term spread, defined as a difference between
yields on 10-year and 1-year U.S. Treasury bonds.
3. Credit spread, defined as a difference between
yields on Moody’s Corporate Baa and Aaa bonds.
4. Expected inflation, defined as a difference
between the yields of nominal 10-year Treasury
bonds and 10-year Treasury Inflation Protected
Securities (TIPS).
Note that all four macro variables are derived from
asset prices and thus reflect investors’ perception on
economic conditions. Amenc et al. (2019) provide a
protocol for selecting relevant state variables and
show that such variables5 are useful in predicting
aggregate economic conditions and provide informa-
tion on the aggregate risk tolerance of investors.6
Short-term interest rates are highly influenced by
monetary policy actions, and they are also considered
as a proxy for the risk-free rate. Surprises in short
rates not only reflect central bank actions but also
reflect the demand on safe assets in a “flight to qual-
ity” (Longstaff 2004). The term spread is one of the
most often used macro variables to forecast eco-
nomic growth and recessions (Campbell 1987;
Estrella and Trubin 2006). It is also considered a gen-
eral measure of compensation for exposure to shocks
on long-term discount rates.7 The default spread sig-
nals risk aversion during adverse economic condi-
tions, when equity and bond investors require higher
compensation for bearing risk. Fama and French
(1989) and Keim and Stambaugh (1986) provide
empirical evidence for such a link. Finally, expected
inflation determines the purchasing power of future
cash flows. It therefore has a direct impact on asset
prices, especially those that produce cash flows
adjusted for inflation (such as TIPS). The four state
variables provide information about aggregate
expectations about different dimensions of eco-
nomic conditions.
39
Financial Analysts Journal | A Publication of CFA Institute
Also note that our macro-state variables are available
at high frequency, which is not the case for back-
ward-looking economic fundamental measures, such
as realised inflation or growth. High-frequency data
are beneficial in statistical estimation procedures, as
we will discuss in detail later in this section.
Surprises Matter, Not Levels. Another
important aspect of our methodology is to use sur-
prises, or innovations, in macro variables instead of
levels. The current value of assets already reflects
information that is known to investors today. As new
information arrives, asset prices, including stock pri-
ces, adjust accordingly. Therefore, we are interested
in surprises. In fact, the observed level of a macro-
economic variable that was fully anticipated by
investors will not lead to different price reactions
across stocks.
Consider the example of surprises in expected inflation.
In Figure 1, we plot the levels as well as innovations in
breakeven inflation over a selected period around the
2008/2009 global financial crisis. More specifically, we
consider a period from 2006 to 2012. Looking at the
level of breakeven inflation, we see that in December
2008, breakeven inflation is very low (close to zero).
From December 2008 to July 2009, inflation rises
strongly to about 1.7% but remains at levels that are
far below earlier levels of about 2.5%. In the first half
of 2009, inflation levels were low but innovations in
breakeven inflation were actually high, as the U.S.
economy rebounded from the financial crisis.
Now consider a stock which benefits from high levels
of inflation and suffers when they are low. During the
first half of 2009, all else being equal, you would
expect such a stock to achieve positive performance
because markets would be likely to reprice it whilst
moving from a very low inflation environment to one
of moderate inflation (though the price of such a stock
might still be below its prior value, as inflation was
much higher at 2.5% before 2008). Still, the inflation
sensitivity of such a stock will show up in price move-
ments over the short term. To pick up the sensitivity
of stock returns with inflation, we need to look at the
contemporaneous changes in breakeven inflation that
were not anticipated by investors, rather than looking
at the level of inflation. Of course, this argument also
applies to other macro variables, such as the term
spread, credit spread, and interest rates.
In general, the current value of assets already reflects
information that is known to investors, such as the
current values of macro variables. As new informa-
tion arrives, stock prices will adjust accordingly.
40
When it comes to estimating anticipated changes in a
macroeconomic variable, there are various methods
available. The standard approach followed in litera-
ture is to use a vector auto-regressive model (VAR).
In our case, we can simply use the change of macro
variables as it delivers similar results to a more
involved model, such as a VAR model.8 As shown in
Figure 1, the VAR innovations are not distinguishable
from simple first-order differences in breakeven infla-
tion. We find a similar picture when looking at other
macro variables.
Statistical Estimates of Macro Exposures.
We now describe the necessary ingredients for
robust estimates of macro exposures. Our objective
is to exploit differences in stock-level macro expo-
sures without altering access to the market premium.
Therefore, we control for the market exposure when
we estimate stock-level macro exposures. We start
from a simple bivariate linear model:
Ri, t
RRF, t ¼ ai þ bMKT
i ðRMKT, t
RRF, t Þ


þ bmacro
i Macro t
EðMacro tjt
1 Þ
þei, t (1)
The first independent variable corresponds to the
equity market premium and the second to innova-
tions in the relevant macro variable. In addition, we
take several important steps to ensure that expo-
sures estimated on past data stay reliable out
of sample.
First, we use observations at weekly frequency. The
data that we rely on are also available at daily fre-
quency, but we use weekly observations to avoid
problems due to differences in closing times between
bond and equity markets. Levi and Welch (2017)
show that using such higher-frequency data provides
substantial improvements in accuracy of estimated
betas over using monthly returns data. Having higher
frequency leads to more observations, which helps to
reduce estimation error. This is, however, very
uncommon in the literature, especially when working
with macro variables (see Levi and Welch 2017).
Second, we manage to account for recent dynamics
in macro betas whilst maintaining deep historical
samples for estimation. Estimation problems face a
basic trade-off between sample size and reactivity to
changes in exposures. Using historical data that are a
decade old provides a large sample but does not
seem ideal as the given firm might have drastically
changed over this period. On the other hand, relying
on a short window of estimation, such as only the
 Targeting Macroeconomic Exposures in Equity Portfolios

Figure 1. Innovations in Macro Variables Are Different from Levels: The Case of
Breakeven Inflation

most recent year, will lead to imprecise estimates of beta estimates. The idea behind this adjustment is
due to small sample size. We overcome this trade-off to shrink estimated coefficients to a prior. We
by a using long-term history of stock-level returns choose the cross-sectional average exposure esti-
(20 years if available) and attributing decreasing mates as a prior, which is close to 0 most of the time
importance to observations as we go further back in as we control for the market beta in the model. The
history.9 Our methodology differs from commonly intensity of shrinkage towards the prior will depend
used rolling-window approaches in investment prac- on the magnitude of estimation error for a respective
tice because we fit the model using the weighted- beta estimate.
least squares method.
nX o !
T lnð2Þ
min e 2
 e
jT
tj 260
(2) r2i, TS r2i, TS
t¼1 t
bfinal
i ¼ bprior
i þ bestimate
i 1

r2i, TS þ r2XS r2i, TS þ r2XS
lnð2Þ
The term 260 specifies that, as T approaches infinity, (3)
half of the total weights are attributed to the first
260 weeks (i.e., 5 years). Our estimation approach where by TS we refer to sampling variance and by XS
benefits from a large sample whilst also capturing the we refer to the cross-sectional variance across esti-
recent dynamics of macro exposures. mated beta coefficients.
Third, we explicitly account for estimation risk at the The final estimate10 is the weighted average of prior
firm level. Treating macro exposures of identical mag- beta and estimated beta. The higher the estimation
nitude for two stocks as equal would ignore estima- error, the more importance is given to the prior, and
tion risk. Even if macro betas for two stocks are vice versa.
estimated to be identical in magnitude, they may dif-
fer in terms of the uncertainty around the point esti- We can summarise the estimation process in
mate. Therefore, we also account for the differences three steps:
in uncertainty across stocks and adjust macro expo-
sures that are estimated imprecisely. Vasicek (1973)  The first key ingredient of our methodology is to
proposes Bayesian shrinkage to improve the accuracy use surprises in forward-looking variables.

Volume 79, Number 1 41

Financial Analysts Journal | A Publication of CFA Institute
 Second, we employ various statistical adjustments
to generate robust estimates of stock-level
macro exposures.
 Third, we use the estimated stock-level macro
exposures to create dedicated building blocks
instead of allocating across existing portfolios
(such as sector or factor portfolios).
In the following section, we analyse how our estima-
tion approach performs in terms of out-of-sample
reliability and how it compares with alterna-
tive approaches.
Assessing the Reliability of Macro
Exposures Estimates
We evaluate the reliability of macro exposure esti-
mates by constructing mimicking portfolios. For each
macro variable, we create a long-short portfolio that
buys stocks with the highest macro exposure esti-
mates (top 30%) and sells stocks with the lowest
macro exposures estimates (bottom 30%). The equal-
weighted long-short portfolio is formed each quarter
and held until the next quarterly rebalancing. The
estimation is only based on data available at the
rebalancing date, and the realised macro exposures
are computed over the holding period. Because the
macro mimicking portfolio buys stocks with high
exposure and sells stocks with low exposure, we
expect that the mimicking portfolio will demonstrate
positive realised macro exposure when we go out of
sample. If that is the case, we can conclude that ex-
ante macro exposure estimates are reliable (and
stocks in long leg have higher out-of-sample expo-
sures than stocks in the short leg).
Our analysis starts in June 1970 and spans over
52 years, ending in June 2022. The equity universe
consists of the largest 500 stocks in CRSP universe11
and is reviewed each quarter on the third Friday of
March, June, September, and December. We focus
on the United States, which represents a significant
share of global equity markets in terms of market
capitalisation. The data for U.S. macro variables cover
a long history, whilst data for other developed coun-
tries cover substantially shorter periods.12
The data for breakeven inflation start only in January
1997. Therefore, when evaluating portfolios that
mimic changes in breakeven inflation, we start the
analysis in March 2002 because we require at least 5
years of data for estimation of macro exposure. It is
42
possible to create an inflation proxy that also covers
the earlier sample period and thus allows for long-
term analysis. In particular, we created an economic
tracking portfolio that invests in broad asset classes
to mimic surprises in breakeven inflation. We find
that the conclusions of our analysis remain intact
when using such an inflation proxy over a long-
term period.13
The macro exposures are estimated in a bivariate
model that includes the market factor and the corre-
sponding macro factor (see equation 1). We estimate
realised exposures using standard OLS as our aim is
not to produce out-of-sample reliable estimates, but
to observe what happens ex-post, over the long-
term period.
Before evaluating whether our approach allows cre-
ating mimicking portfolios with the desired exposure
out of sample, we briefly test mimicking portfolios to
backward-looking variables that are commonly used
when analysing the sensitivity of equity returns to
economic conditions. More specifically, we look at
economic growth proxied by industrial production
index, as it is available at a monthly frequency (unlike
other measures, such as GDP, which is available at a
quarterly frequency). For example, Chen, Roll, and
Ross (1986) use industrial production growth as one
of the factors in their asset pricing model. Ung and
Luk (2016) look at the realised growth in GDP to
measure sensitivity of different equity-style factors
to economic conditions. Devarajan et al. (2016) also
propose using GDP growth to measure the cyclical
variation of equity factors.
We also look at inflation, proxied by realised changes
in the consumer price index. We apply a naïve esti-
mation approach that uses 5 years of recent data at
a monthly frequency to estimate stock-level betas to
a given macro variable.
Indeed, whilst economic growth and realised inflation
are key forces describing macroeconomic conditions,
asset prices would typically lead and reflect informa-
tion about growth and inflation before it is observ-
able through the respective measures (see Fama,
1981). This is one of the reasons that the relation
between asset returns and economic variables is
often found to be weak.
The results in Table 1 show that exposure estimates
to backward-looking variables are unreliable, confirm-
re,
ing earlier results on inflation betas of Ang, Brie
and Signori (2012). The realised exposures of mimick-
ing portfolios to these variables are not distinguish-
able from zero.
 Targeting Macroeconomic Exposures in Equity Portfolios

This analysis illustrates that drawing on commonly
used variables and estimation techniques is not suit-
able for measuring macro exposures. Using back-
ward-looking variables does not support robust
estimation of firms’ sensitivity to economic condi-
tions. Even though economic growth and inflation
could be important drivers of asset prices, backward-
looking information on these dimensions is unlikely
to be relevant for stock returns. Choosing the right
proxy is essential to establish the link between asset
prices and economic fundamentals.
We now move to analysing mimicking portfolios that
rely on different exposure measures to forward-look-
ing variables. Table 2 shows realised macro expo-
sures as well as the corresponding t-statistics
adjusted for heteroscedasticity and autocorrelation.
The exposures are reported for different mimicking
portfolios that rely on different estima-
tion techniques.
In the first row, we show realised macro exposures
for mimicking portfolios that use naïve estimation
techniques. More specifically, macro betas are esti-
mated on a monthly frequency over the recent 5
years, using the OLS method. As noted by Levi and
Welch (2017), this naïve estimation approach is com-
monly used in academic studies. The results suggest
that some portfolios created using naïve estimation
do not come with statistically significant exposures
out of sample. In particular, the mimicking portfolio
for the credit spread comes with out-of-sample
exposure that is far from being statistically signifi-
cant, with a t-statistic of only 1.44. However, the
realised exposures to all variables are positive and
more reliable than the exposures we obtained when
using the naïve estimation approach to estimate
exposure to backward-looking variables. This con-
firms that our macroeconomic variables contain more
relevant information than backward-looking variables,
such as the growth and realised inflation.
The second row of Table 2 shows the results when
using weekly returns for estimation, but without any
further adjustments, such as dynamic weights or
Bayesian shrinkage. Therefore, the only difference
with the naïve approach is the frequency at which
the macro betas are estimated. The realised expo-
sures of macro mimicking portfolios that use weekly
returns are much stronger. We see positive and
highly significant realised exposures for all macro var-
iables. On average, we observe that realised macro
betas increase from 1.71 when using a monthly esti-
mation frequency to 2.40 when using a weekly fre-
quency. Moreover, the t-statistics increase from 2.53
to 4.17 on average. This increase is substantial and
clearly indicates that estimating exposures using a
weekly frequency is superior to using a
monthly frequency.
The third row shows the results obtained when
applying the weighted least-squares (WLS) approach
using a weekly frequency (but still without any
shrinkage). Compared to OLS, using WLS improves
the realised exposures to all four variables, as shown
by higher t-statistics. On average, the magnitudes of
out-of-sample exposures as well as corresponding
t-statistics are higher when using WLS compared
to OLS.
The final row shows the outcome when using weekly
frequency and all statistical adjustments, namely the
dynamic weighting of past observations and Bayesian
shrinkage. Compared to the third row, where we do
not apply the shrinkage, we find that results are
mostly similar across all the macro variables.
However, we still observe mild improvements when
looking at the average realised betas and

Table 1. Out-of-Sample Exposures of Mimicking Portfolios to Backward-

Looking Variables
United States
30 June 1970–30 June 2022 Realised Macro Beta t-Statistic

Mimicking portfolios to backward-looking variables using naïve estimation techniques

Industrial Production Growth
0.01
0.28
Consumer Price Index Growth (CPI) 0.23 0.97

Notes: The realised macro exposures are estimated in a bivariate regression that includes the market factor and the
corresponding macro variable. Standard errors are adjusted for heteroskedasticity and autocorrelation following
Newey and West (1987). We follow Andrews (1991) when determining data-dependent bandwidth/lag truncation
parameters, which are estimated in a first-order autoregressive model (AR(1)) with a maximum likelihood method.
Statistical significance is indicated by a single asterisk () for results that are significant on a 5% basis and double
asterisks () for results that are significant on a 1% basis.

Volume 79, Number 1 43

Financial Analysts Journal | A Publication of CFA Institute

Table 2. Reliability of Macro Beta Estimates Using Different Regression Specifications: Out-
of-Sample Betas with Targeted Macro Variables
Estimation Estimation Bayesian Breakeven
Frequency Method Shrinkage Short Rates Term Spread Credit Spread Inflation Average

Monthly OLS No 1.07 1.20 1.84 2.73 1.71

(4.08) (2.56) (1.44) (2.04) (2.53)
Weekly OLS No 0.83 1.25 2.16 5.37 2.40
(5.23) (4.33) (3.90) (3.23) (4.17)
Weekly WLS No 0.83 1.34 2.21 5.32 2.43
(5.62) (4.51) (3.98) (3.36) (4.37)
Weekly WLS Yes 0.81 1.34 2.28 5.36 2.45
(5.49) (4.49) (4.12) (3.45) (4.39)

OLS: ordinary least-squares; WLS: weighted least-squares.

Notes: The analysis has been done from 19 June 1970 to 30 June 2022. The only exception is breakeven inflation, for which the
analysis starts in 31 March 2002. The out-of-sample macro exposures are estimated in a bivariate regression that includes the
market factor and innovations in corresponding macro variable. The estimation frequency and frequency when calculating realised
exposures are the same. Standard errors are adjusted for heteroskedasticity and autocorrelation following Newey and West
(1987). We follow Andrews (1991) when determining data-dependent bandwidth/lag truncation parameters, which are estimated
in a first-order autoregressive model (AR(1)) with a maximum likelihood method. Statistical significance is indicated by a single
asterisk () for results that are significant on a 5% basis and double asterisks () for results that are significant on a 1% basis.

corresponding t-statistics. The limited impact of
shrinkage is not surprising in our setting because we
are forming portfolios, which already reduces the
estimation error compared to stock-level estimates.
To summarise, we have seen that our statistical
approach to estimate macro betas leads to reliable
classification of high- and low-exposures stocks. The
mimicking portfolios for each macro variable have
positive and highly significant exposures out of sam-
ple. This is not the case when using a naïve estima-
tion technique that is commonly used in academic
studies. The average t-statistic of realised macro
exposures is 2.53 for the naïve approach, whilst
using our statistical approach leads to an average t-
statistic of 4.39 and a minimum t-statistic of 3.45.
We provide additional plots to show the relation
between (market-beta adjusted) returns of mimicking
portfolios and innovations in the corresponding
macro variables. We rely on our preferred estimation
approach, which uses weekly returns and applies
Bayesian shrinkage and dynamic weighting (WLS).
Figure 2 shows returns of mimicking portfolios on a
scatterplot where the horizontal axis corresponds to
weekly innovations in macro variables, and the verti-
cal axis corresponds to weekly returns. The plots also
indicate correlation coefficients. Across different
mimicking portfolios, the correlations of returns with
shocks in the corresponding macro variable are
above 10%. The inflation mimicking portfolio yields
the highest correlation of around 25%. Visual inspec-
tion of these scatterplots also suggests that our
findings about out of sample macro exposures are
not driven by outliers.
Designing Equity Portfolios from
Stock-Level Measures of
Macro Exposures
We have already seen that our macro exposure esti-
mates can reliably distinguish between stocks with
high and low exposure to macro risks. This allows us
to construct long-only portfolios that come with
exposure to the broad stock market index as well as
with targeted macro exposures. We test stylised
strategies that simply select 30% of the stocks with
the highest or lowest macro exposure estimates and
weight them equally across all selected stocks. Note
that these strategies simply correspond to the long
and the short legs of the macro mimicking portfolios
introduced above. We refer to such strategies as
macro dedicated portfolios. This gives us four port-
folios that correspond to high exposure or Macro
Exposure(þ) to a given variable and four low expos-
ure or Macro Exposure(
) portfolios for the same
four state variables.
An alternative to building portfolios from stock-level
exposures would be to use a set of existing portfolios
and allocate across them based on the estimated
macro exposure. Investment practice often uses sec-
tor rotation or factor allocation to gain desired macro
exposures, as opposed to building dedicated port-
folios from stock-level information. We provide a

44

Figure 2. Relation between Returns on Macro Mimicking Portfolios and Macro Variables
comparison with these alternative approaches to
illustrate the importance of creating macro-dedicated
portfolios. Note that, when we test allocation across
factors or sectors, we use the same robust estimation
method for macro exposures that we use to build
dedicated portfolios.
The first approach that we compare allocates equal
weights across two out of six equity-factor portfolios
with the most desirable macro exposures. We con-
sider six factor portfolios: mid-cap, value, high
momentum, low volatility, high profitability, and low
investment. The single-factor portfolios select 30% of
the stocks with the best factor scores and equally
weights them. We refer to such strategies as factor
allocations. The second approach allocates across
sectors following the Thomson Reuters Business clas-
sification.14 We select the three sectors with the
most suitable macro exposure from a total of 10 sec-
tors. Sector indices weight stocks by market cap, and
Volume 79, Number 1
Targeting Macroeconomic Exposures in Equity Portfolios
we attribute equal weights to each sector. We refer
to such strategies as sector allocations.
Table 3 shows the realised macro exposures for
these different approaches. Panel A shows results for
macro dedicated equity portfolios. The results sug-
gest that dedicated macro portfolios constructed
from stock-level data achieve out-of-sample expo-
sures that are in line with the target. For example,
the short-rates (þ) strategy has a realised beta of
0.20, which is statistically significant with a t-statistic
of 2.49. Moreover, all macro dedicated portfolios
have statistically significant exposures out of sample,
whether targeting positive or negative exposure.
If we look at panel B of Table 3, we see quite different
results. Even though factor allocation strategies mostly
display exposures that have the desired sign, they are
often statistically insignificant. For example, the alloca-
tion strategy is unsuccessful at capturing positive
45
Financial Analysts Journal | A Publication of CFA Institute

Table 3. Realised Macro Exposures of Different Equity Macro Strategies

United States
19 June 1970–30 June 2022 Short Rates Term Spread Credit Spread Breakeven Inflation

Panel A: macro dedicated portfolios

Macro Exposure þ 0.20 1.07 1.31 4.56
(2.49) (6.39) (4.44) (4.68)
Macro Exposure

0.60
0.26
0.97
0.81
(
6.49) (
2.08) (
3.13) (
2.64)
Differences in Macro Exposures 0.81 1.33 2.28 5.37
(5.25) (4.96) (3.94) (4.22)
Panel B: factor allocation
Macro Exposure þ 0.01 0.48 0.35 1.71
(0.12) (3.71) (0.99) (4.15)
Macro Exposure

0.28 0.07
0.48 0.20
(
3.12) (0.54) (
1.67) (0.93)
Differences in Macro Exposures 0.29 0.41 0.84 1.51
(2.32) (2.27) (1.82) (3.23)
Panel C: sector allocation
Macro Exposure þ 0.43 0.84 0.51 6.41
(3.69) (3.63) (0.97) (5.33)
Macro Exposure

0.51
0.36
0.88 0.55
(
3.71) (
1.45) (
1.49) (0.87)
Differences in Macro Exposures 0.94 1.19 1.39 5.86
(5.21) (3.55) (1.75) (4.31)

Notes: The analysis has been done from 19 June 1970 to 30 June 2022. The only exception is breakeven inflation, for which the
analysis starts in 31 March 2002. The out-of-sample macro exposures are estimated in a bivariate regression that includes the
market factor and innovations in corresponding macro variable. Standard errors are adjusted for heteroskedasticity and autocorrel-
ation following Newey and West (1987). We follow Andrews (1991) when determining data-dependent bandwidth/lag truncation
parameters, which are estimated in a first-order autoregressive model (AR(1)) with a maximum likelihood method. The null hypoth-
esis that the differences in macro exposures are zero is tested by Welch’s test. The Welch’s t-statistics may not equal to the t-sta-
tistics of the null hypothesis that the macro mimicking portfolio has zero exposure, reported in Table 2. Statistical significance is
indicated by a single asterisk () for results that are significant on a 5% basis and by double asterisks () for results that are sig-
nificant on a 1% basis.

exposures to short rates and credit spread. In addition,
factor allocations were unable to achieve negative
exposures to breakeven inflation and the term spread.
Sector allocations produce relatively better perform-
ance than factor allocation, as shown in panel C of
Table 3. The exposures are in line with the objective
for three out of four variables, but sometimes they
are not statistically significant. Some exposures are
the opposite of what was targeted: The sector alloca-
tion targeting negative exposure to breakeven infla-
tion comes with positive out-of-sample exposure.
Our results clearly indicate that selecting stocks
based on their macro exposures leads to more con-
sistent and pronounced exposures out of sample
than allocating across sectors or factors. This may
not be surprising, as sectors or factors were not
designed to proxy certain macroeconomic risks,
whilst this is the case for macro dedicated portfolios.
Macro exposures of sectors or factors are not
heterogenous enough to create portfolios with reli-
able macro sensitivity out of sample. Industry practi-
ces that try to achieve these macro exposures
through sector rotation or factor allocation strategies
might be problematic. We find that the cross-section
of stocks provides more diverse exposures that allow
creating portfolios with stronger out-of-sample macro
exposures, consistently across all macro variables.
As an alternative to selecting a fixed number of
stocks according to estimated exposures, one may
also consider the statistical significance of exposure
estimates when selecting stocks. We also test mim-
icking portfolios where the long and short legs are
formed by selecting only those stocks that have stat-
istically significant exposure estimates at the
5% level.15
Before forming portfolios, we first consider whether
we have enough stocks in our universe at each point
in time with either positive or negative exposures to

46

the given macro variable. Having a sufficient number
of stocks is necessary to ensure that our test port-
folios are not overly concentrated. Figure 3 below
reports the percentage of stocks in our universe that
have statistically significant exposures. We observe
that, at times, stocks with significantly negative or
positive exposure to a given macro variable only
make up less than 5% of the universe. This means
that the mimicking portfolio would be highly concen-
trated. To avoid extreme levels of concentration, we
impose that at least 5% of the stocks are selected.
Table 4 repeats the analysis from and reports out-of-
sample exposures of macro dedicated portfolios. We
Figure 3. Percentage of Stocks with Statistically Significant Macro Exposures across Time
Volume 79, Number 1
Targeting Macroeconomic Exposures in Equity Portfolios
report the out-of-sample exposures of portfolios that
select a fixed number of stocks (30%) in panel A,
whilst panel B focusses on portfolios that only select
stocks with statistically significant macro exposures
(but at least 5% of stocks).
As expected, we find that more aggressive stock
selection improves the magnitude of out-of-sample
betas on average. However, there is no improve-
ment when looking at average t-statistics of realised
exposures. Using a fixed threshold of selecting 30%
of stocks allows for better diversified portfolios,
which leads to more out-of-sample robustness than
forming concentrated portfolios of stocks that have
47
Financial Analysts Journal | A Publication of CFA Institute

the most significant exposures on a stand- exposures become small and statistically insignificant
alone basis. in most cases.

Why Do Equity-Style Factors Not Suffice
to Manage Macro Risks? The factor allocation
strategies in Table 3 do not allow managing macro
risks as efficiently as macro dedicated strategies.
However, it has been documented that different fac-
tors do react differently to economic conditions (see
Amenc et al. 2019). Below, we show why—despite
this heterogeneity—factor allocation is not a suffi-
cient tool to manage macro exposures. We consider
six long-only factors constructed in a highly liquid
universe, namely the mid-cap, value, momentum, low
volatility, high profitability, and low investment factor
portfolios. In addition to the long-term realised macro
exposures, we report ex-ante estimated exposures at
each quarter. Table 5 reports realised long-term
exposures as well as descriptive statistics for ex-ante
exposures estimated each quarter. For reference, we
also report the exposure of macro dedicated port-
folios for each macro variable.
As expected, realised macro exposures are different
across equity factors, as shown in Table 5. For
example, the low-volatility portfolio comes with a
credit spread beta of 0.58, whilst the high-profitabil-
ity portfolio has beta of
0.61 to innovations in term
spread. However, once we look at the average expo-
sures across six equity factor portfolios, the realised
Second, dedicated macro portfolios are always able
to achieve stronger realised exposures than any of
the six equity factors, whether positive or negative.
This suggests that if investors’ objective is to gain
pronounced exposure to a given macro variable, they
would be better off using macro dedicated strategies.
In addition, selecting amongst equity factors does not
allow gaining both positive and negative exposures
to three out of four macro variables. For example,
we find that four out of six factor portfolios come
with significantly positive exposure to the term
spread, but none of the factors provide significantly
negative exposure. This is true not just for the term
spread but also for short-rates and breakeven infla-
tion. Hence, using standard equity factors not only
delivers relatively weak macro exposures, but it also
renders some exposure targets infeasible.
Last, but not least, we find that the macro exposures
of equity-style factors are not necessarily consistent
across time. We find that the estimated ex-ante
macro exposures of most equity factors take both
positive and negative values in different periods of
our sample. For instance, the 25th percentile of ex-
ante betas to the credit spread are negative for all
six factors, whilst the 75th percentiles are positive
for all factors. This implies that, for 25% of the quar-
ters in our sample, a factor has a positive credit

Table 4. Reliability of Macro Beta Estimates Using an Alternative Method to Form

Mimicking Portfolios
United States
19 June 1970–30 June 2022 Short-Rates Term Spread Credit Spread Breakeven Inflation Average

Panel A: macro dedicated portfolios—30% EW selection

Macro exposure þ 0.20 1.07 1.31 4.56 1.78
(2.49) (6.39) (4.44) (4.68) (4.50)
Macro exposure

0.60
0.26
0.97
0.81
0.66
(
6.49) (
2.08) (
3.13) (
2.64) (
3.59)
Panel B: macro dedicated portfolios—EW significant exposures (min 5%)
Macro exposure þ 0.38 1.22 1.73 5.65 2.25
(2.60) (4.78) (3.12) (5.23) (3.93)
Macro exposure

0.72
0.60
0.81
0.86
0.75
(
5.31) (
2.37) (
1.61) (
0.77) (
2.51)

Notes: The analysis has been done from 19 June 1970 to 30 June 2022. The only exception is breakeven inflation, for which the
analysis starts in 31 March 2002. The out-of-sample macro exposures are estimated in a bivariate regression that includes the
market factor and innovations in corresponding macro variable. The estimation frequency and frequency when calculating realised
exposures are the same. Standard errors are adjusted for heteroskedasticity and autocorrelation following Newey and West
(1987). We follow Andrews (1991) when determining data-dependent bandwidth/lag truncation parameters, which are estimated
in a first-order autoregressive model (AR(1)) with a maximum likelihood method. Statistical significance is indicated by a single
asterisk () for results that are significant on a 5% basis and double asterisks () for results that are significant on a 1% basis.
EW ¼ equal weighted.

48
 Table 5. Ex-Ante and Realised Exposures of Equity-Style Factors
Average Macro Macro
19 June 1970–30 Low High Low across Exposure Exposure
June 2022 Mid-Cap Value Momentum Volatility Profitability Investment 6 Factors (þ) (
)

Volume 79, Number 1

Exposures to short rates
Realised exposure 20.16* 20.28** 20.06 20.32** 0.07 20.24** 20.17 0.20* 20.60**
Mean ex-ante exposure
0.12
0.17
0.12
0.22 0.09
0.18
0.12 1.45
1.75
Min
0.75
0.64
0.60
0.76
0.31
0.68
0.50 0.35
3.84
25th percentile
0.37
0.39
0.35
0.40
0.07
0.40
0.30 0.61
2.41
Median
0.13
0.19
0.18
0.26 0.04
0.22
0.15 0.98
1.45
75th percentile 0.00
0.03 0.01
0.02 0.28
0.08 0.04 2.26
1.05
Max 1.11 1.35 1.12 0.44 0.71 1.34 0.39 4.84
0.81
Exposures to term spread
Realised exposure 0.42** 0.67** 0.06 0.26* 0.12 0.44** 0.33* 1.07** 20.26*
Mean ex-ante exposure 0.21 0.28 0.18 0.10 0.17 0.22 0.25 2.04
1.59
Min
0.80
0.85
0.73
0.75
0.77
0.77
0.74 0.97
4.75
25th percentile
0.07 0.05
0.13
0.24 0.03
0.02 0.05 1.26
2.43
Median 0.19 0.27 0.17 0.08 0.14 0.15 0.23 2.05
1.12
75th percentile 0.52 0.52 0.47 0.46 0.31 0.50 0.50 2.42
0.67
Max 0.96 0.98 1.04 0.93 0.95 0.92 0.88 4.80
0.27
Exposures to credit spread
Realised exposure 20.22 0.04 0.14 0.58* 20.61** 0.29 0.04 1.32** 20.97**
Mean ex-ante exposure 0.11 0.30 0.22 0.46
0.16 0.32 0.35 3.86
3.43
Min
1.77
1.85
1.78
1.18
1.93
1.41
1.55 2.21
7.12
25th percentile
0.64
0.41
0.41
0.17
0.53
0.24
0.22 2.92
4.96
Median 0.29 0.51 0.29 0.54
0.09 0.52 0.49 3.76
3.40
75th percentile 0.65 0.80 0.77 0.96 0.24 0.81 0.74 4.73
1.77
Max 2.30 2.24 2.28 2.47 1.27 2.00 1.87 6.52
1.21
Exposures to breakeven inflation (31 March 2002–30 June 2022)
Realised exposure 0.23** 0.20* 0.36* 0.34 0.20** 0.42** 0.29** 4.54** 20.82**
Mean ex-ante exposure 1.30 1.39 0.86 0.27 1.01 1.00 0.65 5.37
2.58
Min 0.39 0.17
0.51
0.49 0.11 0.07 0.24 3.50
4.19
25th percentile 0.76 0.62 0.30
0.08 0.79 0.45 0.36 4.56
2.93
Median 1.40 1.57 0.75 0.33 1.00 1.13 0.45 5.23
2.53
75th percentile 1.75 1.91 1.40 0.63 1.31 1.28 0.91 5.82
2.04
Max 2.29 3.02 2.46 1.04 1.83 2.59 1.63 8.67
1.78

Notes: The equity-style factor portfolios select 30% of the stocks with the best factor scores and equally weights them. The equity universe consists of 500 largest stocks in the
U.S. and is reviewed quarterly, on the third Friday of March, June, September, and December. The percentile values are taken from the time series of ex-ante beta estimates for a
given equity factor and a given macro variable. Realised exposures were computed over the full sample. Realised macro exposures are estimated in a bivariate regression using

49
Targeting Macroeconomic Exposures in Equity Portfolios

weekly observations, where explanatory variables are the market excess return and innovations in a given macro variable. Ex-ante exposure measures are estimated following our
proposed methodology. Statistical significance is indicated by a single asterisk () for results that are significant on a 5% basis and double asterisks () for results that are significant
on a 1% basis. We test statistical significance only for full-sample realised exposures. Values set in bold are realised (i.e., ex-post) exposures. Realised exposures of dedicated macro
portfolios are highlighted in yellow when they are unique in the sense that none of the equity style factors provide significant exposures with the same sign.
Financial Analysts Journal | A Publication of CFA Institute

spread beta, but for 25% of the quarters the same
factor has a negative beta. We observe similar
instability across time in ex-ante exposure to other
macro variables, at least for some of the
equity factors.
Because their macro exposures are relatively weak
and unstable over time, standard equity factors may
not be suitable tools to build portfolios that target
macro exposures. In contrast, dedicated macro port-
folios allow for more effective targeting of
macro exposures.
Conditional Performance of Dedicated
Macro Exposure Portfolios. We now show
how desired macro exposures translate into returns
of the macro dedicated portfolios. We split the sam-
ple into positive and negative surprises (top/bottom
25%) for each given macro variable. A portfolio that
targets positive exposure to a given macro variable
should outperform the market when the surprise in
this variable is positive, and vice versa. Figure 4
reports average annualised returns adjusted for mar-
ket exposure for all dedicated macro portfolios in
respective conditions of macro surprises. For
example, the term spread (þ) portfolio outperforms
its market benchmark by around 12 percentage
points per year when surprises in term spread are
positive. Similarly, we see the opposite for term
spread (
) portfolio, which underperforms the market
in positive surprises but outperforms when it is
meant to (i.e., when the surprises in term spread are
negative). We see that all portfolios outperform in
macro conditions that they target. The only excep-
tion is the short-rates (þ) portfolio that comes with
statistically insignificant, albeit positive, outperform-
ance during the times of positive interest rate sur-
prises. The inflation (þ) portfolio demonstrated a
very strong performance when expected inflation
rises: The annualised returns are more than 20 per-
centage points higher than that of the market port-
folio during positive inflation shocks.
The dedicated macro strategies deliver economically
large and statistically significant outperformance
when innovations in the targeted variable are in the
desired direction, with the exception of short rates
(þ). Whilst it is beyond the scope of this paper to
analyse particular applications of such portfolios, it is
clear that the reliable conditionality of macro

Figure 4. Conditional Relative Performance of Macro Dedicated Portfolios

The plot shows average annualised returns of macro dedicated portfolios after adjusting for their market exposure (CAPM beta).
The analysis was done from 19 June 1970 to 30 June 2022 for all portfolios, except for the inflation portfolios, for which the ana-
lysis starts on 31 March 2002. Positive (negative) surprise periods are defined as calendar weeks (Friday to Friday) when the change
in a given macro variable was amongst the top (bottom) 25% of historical observations. Returns that are statistically significant at a
5% level are presented in bars with solid fill, whilst returns that are not statistically significant at 5% are presented in bars filled
with diagonal stripes.

50

dedicated portfolios can be of great use for investors.
For example, an investor with a large allocation to
credit could offset the losses by investing their
equity portfolio in a credit spread (þ) strategy, thus
outperforming the market during the periods of posi-
tive surprises in credit spread. As another example,
for an active investor who thinks that inflation
expectations will rise unexpectedly in the future,
using an inflation (þ) portfolio can provide strong
performance for their equity portfolio in times when
surprises in breakeven inflation are positive.
Overall, the results suggest that macro dedicated
portfolios can provide strong and economically sig-
nificant outperformance over the market portfolio in
desired macro conditions. Investors who are con-
cerned with substantial movements in certain macro-
economic variables can benefit by investing in the
corresponding macro dedicated portfolio.
Performance and Risk of Dedicated
Macro Exposure Portfolios. The macro dedi-
cated portfolios deliver positive outperformance over
the market portfolio during times of surprises, either
positive or negative, in a targeted macro variable. We
now analyse the unconditional performance and risk
of these strategies. As a reference, we use the cap-
weighted market portfolio.
Table 6 reports unconditional performance of various
macro dedicated portfolios. The results suggest that
the performance of macro-dedicated portfolios is not
markedly different from the reference index. The
statistical significance test of the differences in
Sharpe ratios indicates that only the credit spread (þ)
portfolio has a Sharpe ratio greater than the market
portfolio. Risk-adjusted performance of all other
macro dedicated portfolios is statistically indistin-
guishable from that of the reference index. The devi-
ations in relative returns are small, ranging between
0.53% and 1.76%. The p values of relative returns
are all above 5%, meaning that none of these return
differences are statistically significant.
Note that we have omitted the inflation portfolios
from this analysis because they are only available
staring in March 2002. Table 7 reports the same ana-
lysis for inflation portfolios, but over the period start-
ing on 31 March 2002 and ending on 30 June 2022.
Our findings are very similar to those above. The dif-
ferences in Sharpe ratio as well as relative returns
are statistically insignificant. Overall, we do not find
that any macro dedicated strategies bear significant
cost relative to the market index in our sam-
ple period.
Volume 79, Number 1
Targeting Macroeconomic Exposures in Equity Portfolios
Table 8 shows the results of seven-factor regressions
that include the market as well as six long-short
equity factors widely used in the literature. We find
that market betas of dedicated macro portfolios range
between 0.87 and 1.06. Exposures to other equity
factors are small in magnitude for most strategies and
rarely exceed 0.30 in magnitude. Another pattern that
we observe consistently across all strategies is a posi-
tive exposure to the size factor, which is expected
because we weight the selected stocks equally.
Table 8 also provides interesting insights regarding
the multifactor alphas of macro dedicated portfolios.
In particular, we find that seven out of eight port-
folios have an alpha that is statistically indistinguish-
able from zero. The credit spread (þ) portfolio has a
positive and significant alpha in the multifactor
model. None of the eight portfolios display a signifi-
cantly negative multifactor alpha over the sample
period. This suggests that investors, who had
focussed on targeting their desired macro exposure,
would not have born a cost of underperformance
(negative alpha), regardless of which of the eight pos-
sible exposures they had targeted.
Our results imply that equity investors can target
desired macro exposures without paying a premium.
To further evaluate the potential performance implica-
tions of targeting macro exposures, we also estimate
the trading costs of macro dedicated portfolios. Table
9 reports transaction costs as well as performance
measures net of transaction costs for different macro
strategies. We follow Corwin and Schultz (2012) and
Chung and Zhang (2014) to estimate the proxy of
effective bid-ask spread.16 These two approaches have
been shown to yield to reliable and accurate proxies
of intra-day effective spreads. The measures proposed
by Chung and Zhang (2014) is more accurate, but it
can only be applied starting in 1993, because the clos-
ing bid and ask prices that are used for computing the
effective bid-ask spread are only available from that
point. Hence, we rely on effective spreads following
Corwin and Schultz (2012) before 1993 and following
Chung and Zhang (2014) afterwards.
The results in Table 9 indicate that the annual trans-
action costs of macro dedicated portfolios are below
35 basis points. Therefore, the impact on the per-
formance is mild, and our conclusions from the ana-
lysis of gross returns remain intact. In particular,
macro strategies do not come with significantly lower
Sharpe ratios, and the multifactor alphas are statistic-
ally indistinguishable from zero for the portfolios tar-
geting either positive or negative exposures to the
short rates and term spread. The only significant
51
Financial Analysts Journal | A Publication of CFA Institute

Table 6. Unconditional Performance of Macro Dedicated Portfolios

United States
19 June
1970–30 Short Rates Short Rates Term Spread Term Spread Credit Spread Credit Spread
June 2022 Cap-Weighted (þ) (
) (þ) (
) (þ) (
)

Absolute 10.70% 11.35% 11.87% 11.28% 11.93% 12.45% 11.22%

return
Volatility 17.08% 18.61% 16.69% 18.78% 16.51% 15.97% 18.97%
Sharpe ratio 0.35 0.36 0.43 0.35 0.44 0.48 0.34
p value of the – 83.6% 22.6% 76.0% 10.8% 1.5% 60.2%
differences
in Sharpe
ratio
Relative return – 0.66% 1.18% 0.59% 1.23% 1.76% 0.53%
p value of – 25.9% 29.7% 40.9% 19.4% 9.3% 32.5%
relative
return
Tracking error – 5.45% 6.73% 7.16% 5.52% 5.85% 5.90%
Information – 0.12 0.17 0.08 0.22 0.30 0.09
ratio

Notes: The table reports performance and risk measures for macro dedicated portfolios. The analysis was done using daily returns
in USD, from 19 June 1970 to 30 June 2022. The annualised returns are computed as geometric average, whilst the statistical
tests of differences in Sharpe ratio and relative returns use arithmetic average. The p values of the differences in Sharpe ratios are
computed following the methodology proposed by Ledoit and Wolf (2008). Statistical significance of the differences in relative
returns and Sharpe ratios is indicated by a single asterisk () for results that are significant on a 5% basis and double asterisks ()
for results that are significant on a 1% basis.

Table 7. Unconditional Performance of Inflation Targeting Portfolios

United States
31 March 2002 to 30 June 2022 Cap-Weighted Inflation (þ) Inflation (
)

Absolute return 7.88% 6.31% 9.42%

Volatility 19.65% 26.12% 18.61%
Sharpe ratio 0.33 0.19 0.44
p value of the differences in Sharpe ratio – 16.3% 17.3%
Relative return –
1.56% 1.55%
p value of relative return – 89.1% 34.6%
Tracking error – 10.37% 5.56%
Information ratio –
0.15 0.28

Notes: The table reports performance and risk measures for macro dedicated portfolios. The analysis was done using
daily returns in USD, from 31 March 2002 to 30 June 2022. The annualised returns are computed as geometric aver-
age, whilst the statistical tests of differences in Sharpe Ratio and relative returns use arithmetic average. The p-values
of the differences in Sharpe ratios are computed following the methodology proposed by Ledoit and Wolf (2008).
Statistical significance of the differences in relative returns and Sharpe ratios is indicated by a single asterisk () for
results that are significant on a 5% basis, and double asterisks () for results that are significant on a 1% basis.

performance difference observed in our analysis is
for the credit spread (þ) portfolio, which outperforms
the market and the multifactor benchmark, even
after accounting for transaction costs.
The results are similar in Table 10, where we focus
on inflation portfolios over a more recent period. We
find no evidence that portfolios that come with posi-
tive or negative exposure to breakeven inflation
cause reduction in the performance relative to the
market portfolio or a multifactor benchmark. Note
that the transaction costs for inflation portfolios are
substantially smaller than those reported in Table 9.
This is mainly due to the differences in sample

52
 Targeting Macroeconomic Exposures in Equity Portfolios

Table 8. Factor Exposure Analysis of Macro Dedicated Portfolios

United States
19 June Credit Credit
1970a–30 Short Rates Short Rates Term Spread Term Spread Spread Spread Inflation Inflation
June 2022 (þ) (
) (þ) (
) (þ) (
) (þ) (
)

Unexplained 1.14% 1.53% 1.45% 1.36% 2.00% 1.13% 1.54% 0.37%

Market 1.05 0.87 0.94 0.99 0.88 1.04 1.06 0.99
Size 0.15 0.37 0.32 0.15 0.24 0.19 0.27 0.17
Value 0.08 0.06 0.26
0.05 0.19 0.07 0.37
0.10
Momentum 0.05
0.10 0.07
0.10
0.05 0.06 0.14
0.12
Volatility
0.01
0.05
0.28 0.30 0.00
0.11
0.16 0.22
Profitability
0.32 0.15
0.12
0.14 0.20
0.30
0.44 0.12
Investment
0.08
0.01 0.15
0.24 0.04
0.20
0.10
0.10
R-squared 90.95% 85.87% 89.79% 88.22% 87.24% 90.96% 92.26% 93.26%

Notes: The table reports factor exposures of macro dedicated portfolios. The analysis was done using weekly returns in USD, from
19 June 1970 to 30 June 2022. aThe only exception is portfolios targeting exposure to breakeven inflation, for which the analysis
starts in 31 March 2002. Factor intensity is the sum of all factor exposures except for the market factor. The unexplained returns
are annualised. Statistical significance is indicated by a single asterisk () for results that are significant on a 5% basis and double
asterisks () for results that are significant on a 1% basis.

Table 9. Performance of Macro Dedicated Portfolios Net of Transaction Costs

Short Rates Term Spread Credit Spread
31 December Broad U.S.
1977–30 June 2022 Cap-Weighted (þ) (
) (þ) (
) (þ) (
)

Gross annual return 11.41% 12.09% 11.84% 11.63% 12.53% 13.13% 11.60%
Net annual return 11.39% 11.79% 11.60% 11.31% 12.21% 12.85% 11.26%
Annual transaction cost 1 bps 30 bps 24 bps 32 bps 32 bps 27 bps 34 bps
Annual turnover 2.90% 51.77% 42.67% 53.49% 52.79% 48.45% 56.22%
Net Sharpe ratio 0.40 0.40 0.43 0.37 0.49 0.53 0.37
p value – 80.6% 63.2% 43.7% 13.3% 2.6% 38.1%
Net multifactor alpha 0.00% 0.80% 0.81% 0.64% 1.18% 1.84% 0.43%
p value – 36.3% 39.3% 48.0% 19.1% 3.5% 62.1%

Notes: Transaction cost is estimated as a difference in gross annualised performance and annualised performance net of transac-
tion costs. The stock-level transaction costs are estimated from daily prices. Prior to 1993, we use the methodology proposed by
Corwin and Schultz (2012), whilst after 1993, we rely on the methodology of Chung and Zhang (2014). The reported turnover is a
one-way turnover, computed quarterly and annualised by multiplied the average by 4. The p values of the differences in Sharpe
ratios are computed following the methodology proposed by Ledoit and Wolf (2008). The analysis was done from 31 December
1977 to 30 June 2022. The analysis starts only in 1977 due to data quality and availability of intra-day high and low prices.
Statistical significance of the differences in net Sharpe ratios and net multifactor alphas is indicated by a single asterisk () for
results that are significant on a 5% basis and double asterisks () for results that are significant on a 1% basis.

period.17 The transaction costs have declined dramat-
ically over recent decades, as the market structure
has improved, in particular through the reduction in
minimum tick size for trading (see Bruno et al.
2019). Therefore, historical estimates of trading
costs, such as the ones in Table 9, are inflated com-
pared to the costs that investors face today. Also
note that we did not apply further implementation
rules, such as buffers in stock selection or invest-
ability rules that reduce the amount of trading at
rebalancing. Such rules could be used to further
reduce transaction costs in practice (see Novy-Marx
and Velikov 2016).
Overall, our findings are consistent with empirical evi-
dence in a recent study by Herskovic, Moreira, and
Muir ([2019) which finds that hedging exposures to dif-
ferent macro risks, including the credit spread and term
spread, has little to no impact on expected returns.
Moreover, the empirical asset pricing literature has

Volume 79, Number 1 53

Financial Analysts Journal | A Publication of CFA Institute

Table 10. Performance of Inflation Portfolios Net of Transaction Costs

Breakeven Inflation
31 March 2002–31 December 2019 Broad U.S. Cap-Weighted (þ) (
)

Gross annual return 8.20% 7.71% 9.12%

Net annual return 8.20% 7.56% 8.99%
Annual Transaction cost <1 bps 14 bps 13 bps
Annual turnover 3.50% 47.58% 43.29%
Net Sharpe ratio 0.36 0.25 0.42
p value – 33.8% 32.6%
Net multifactor alpha 0.00% 1.41% 0.25%
p value – 35.6% 80.2%

Notes: Transaction cost is estimated as a difference in gross annualised performance and annualised performance net of transac-
tion costs. The stock-level transaction costs are estimated from daily prices following the methodology proposed by Chung and
Zhang (2014). Statistical significance of the differences in net Sharpe ratios and net multifactor alphas is indicated by a single
asterisk () for results that are significant on a 5% basis and double asterisks () for results that are significant on a 1% basis.

established a set of workhorse models to explain the
cross -section of stock returns that have evolved over
time (from the Fama and French [1993] three-factor
model to the Fama and French [2018] six-factor model
or the Hou, Mo, Xue, and Zhang [2021] five-factor
model). Whilst these models draw on slightly different
sets of equity-style factors, none of them includes
macroeconomic factors, such as the variables studied in
this paper. This is in line with the idea that such macro-
economic factors do not offer risk compensation over
and above what the typical equity-style factors capture.
In contrast, several papers argue that some macro-
economic factors may be separately priced in the
cross-section. Boons (2016) finds that the term
spread and credit spread are priced factors in a broad
cross-section of U.S. equities (including small- and
micro-cap stocks), but the short rate is not. Petkova
(2006) finds that exposures to short rates are priced
along with term spread, but the credit spread is not.
Such studies, however, have not led to a generally
accepted inclusion of macro factors in workhorse
models used to explain the cross-section of stock
returns. In our setting, where we analyse a universe
of liquid stocks and account for a broad set of style
factors, macro exposures are not associated with
separate risk premiums. An interesting avenue for
future research is to assess risk premia for macro
exposures whilst accounting for the fact that risk
premia may be dynamic and may change sign over
time (see, e.g., Boons et al. 2020, 2021).
Conclusion
We have proposed a methodology to estimate stock-
level exposures to different macroeconomic risks. We
rely on innovations in forward-looking variables that
quickly incorporate investors’ expectations about eco-
nomic conditions, such as short-term interest rates, the
term spread, the credit spread, and breakeven inflation.
To capture exposure of stocks to these macro sur-
prises, we apply statistical tools to improve robustness.
We then construct dedicated portfolios from the result-
ing firm-level measures of macro exposure. Such equity
strategies come with statistically and economically sig-
nificant exposures out of sample.
Our robust measurement procedure delivers a more
robust and reliable classification of stocks than stand-
ard estimation approaches. When constructing equity
portfolios that target high or low macroeconomic
exposure, our procedure leads to more than 40%
higher realised out-of-sample exposures and more
than 70% improvements in t-statistics associated with
these exposures, compared to naïve estimation.18
Our approach of building dedicated portfolios using
stock-level exposures also captures macro exposures
more reliably than using off-the-shelf building blocks,
such as sector or factor indices. We show that allo-
cating across sectors or factors leads to weaker
exposures than our dedicated macro portfolios and
sometimes even leads to exposures that are incon-
sistent with the objective. The reason behind this is
that sectors or factors are not designed to discrimin-
ate assets by macroeconomic risk exposures, whilst
the stock-level approach can fully exploit the hetero-
geneity of exposures in the cross-section of equities.
In addition, our proposed strategies are fully system-
atic and do not rely on common beliefs or the discre-
tionary choices of an asset manager.

54

Finally, we showed that our macro exposure strategies
do not suffer from lower performance compared to a
broad equity portfolio. The stand-alone returns of eight
macro exposure strategies as well as their Sharpe ratios
are not significantly different from the market portfolio
in our sample. They also do not come with negative
alphas in a multifactor model that includes the usual
style factors. Our conclusions hold when accounting
for trading costs of dedicated macro portfolios.
Editor's Note
Submitted 8 March 2022
Accepted 11 November 2022 by William N. Goetzmann
Notes
1. An incomplete list includes Novy-Marx (2013) and Bali,
Brown, and Caglayan (2014).
2. As we will see later, the statistical adjustments proposed
by Boons (2016) are similar to the ones we use.
However, there are two key differences between our
analyses. First, we use higher frequency for estimation
than Boons (2016). It has been shown that higher
frequency increases sample size and improves the
accuracy (see Levi and Welch 2017). Second, Boons
(2016) looks at all common stocks in the CRSP universe,
which is much larger than our investable universe that
consists of the largest 500 stocks in U.S. Our mission is
more difficult because distinguishing amongst high- and
low-exposure stocks in a smaller universe requires
more precision.
3. For clarity, we focus on a single objective of targeting
macro exposures. An interesting area for future research
would be to assess how investors can combine
objectives in terms of factor exposures aimed at
improving unconditional returns with objectives in terms
of macro exposures aimed at targeting
conditional returns.
4. We have also analysed the out-of-sample reliability of
exposures to various other macro variables using our
estimation approach, and the results are reported in the
Online Supplemental Material. First, we have estimated
exposures of stocks to macro variables that are based
on market prices that reflect investor expectations, such
as commodities, interest rates on U.S. Treasuries with
different maturities and inflation swaps. Our estimation
approach leads to highly reliable out-of-sample
exposures to these variables. In addition, we have
looked at aggregate economic indicators based on
macro fundamentals, such as the Aruoba-Diebold-Scotti
Business Conditions Index (ADS) and Bloomberg Weekly
Consumer Comfort Index (CCI). These two macro
variables aim to nowcast the economic activity, such as
growth (for ADS) and consumer confidence (CCI), and
are available at daily and weekly frequency, respectively.
Volume 79, Number 1
Targeting Macroeconomic Exposures in Equity Portfolios
Overall, we show that our macro exposure measures
can add value for investors with a dual objective,
which is to harvest the long-term equity premium
and protect their total portfolio from sudden changes
in economic conditions (or bet in favour of certain
economic conditions). In practice, such measures
could be used for a variety of applications such as
tilting long-only portfolios to target the desired
macro sensitivities.
Exposures we measure with respect to these variables
are not necessarily reliable out of sample. This suggests
that relying on market-based macro variables that
reflect shocks in investor expectations is important to
successfully detect a link with equity returns (see
Amenc et al. [2019] for a detailed discussion of
this argument).
5. Amenc et al. (2019) have not included expected inflation
in their analysis.
6. For a more detailed discussion on each state variable,
please refer to Amenc et al. (2019).
7. Various studies document that term spread predicts
returns on long-term asset classes. See, for example,
Fama and French (1989) for equities and bonds, Campbell
(1996) for human capital, Hong and Yogo (2012) for
commodities, and Ang, Nabar, and Wald (2013) for
real estate.
8. We have also tested innovations from VAR(1,1) model,
and our findings are unaffected by this change. Using
VAR model was proposed by Campbell (1996) in this
context and has been followed by various other studies
more recently (e.g., Petkova 2006, Boons 2016). Each
macro variable is modelled as a linear function of the past
values of all macro variables considered.
9. We also require that at least 208 weekly observations are
available for a given stock, which is equivalent to 80% of
total observations in 5 years. For other estimation
approaches, which use OLS or monthly frequency, we
also require that at least 80% of the total observations in
a five-year period are available.
10. Beyond information in returns, one might also
exploit information in company risk disclosures using
textual analysis. We have conducted additional
analysis augmenting our returns-based measures using
textual analysis. This approach also delivers robust
exposures out of sample but is not reported here
for brevity.
55
Financial Analysts Journal | A Publication of CFA Institute
11. Our universe excludes ADRs.
12. We also tested the out-of-sample reliability of our
estimation approach in an Asian universe. Results can be
found in the Online Supplemental Material.
13. Results of this long-term analysis can be found in the
Online Supplemental Material.
14. The sectors are (1) energy, (2) basic materials, (3)
industrials, (4) cyclical consumer, (5) non-cyclical
consumer, (6) financials, (7) healthcare, (8) technology, (9)
telecoms, and (10) utilities.
15. Note that we use exposure estimates before shrinkage,
because the standard errors of those exposures
correspond to “unshrunk” beta estimates. Such an
References
Amenc, N., M. Esakia, F. Goltz, and B. Luyten. 2019.
“Macroeconomic Risks in Equity Factor Investing.” The Journal
of Portfolio Management 45 (6): 39–60. doi:10.3905/jpm.2019.
1.092.
Andrews, D. W. K. 1991. “Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix Estimation.”
Econometrica: Journal of the Econometric Society 59 (3):
817–858. doi:10.2307/2938229.
re, and O. Signori. 2012. “Inflation and
Ang, A., M. Brie
Individual Equities.” Financial Analysts Journal 68 (4): 36–55.
doi:10.2469/faj.v68.n4.3.
Ang, A., N. Nabar, and S. Wald. 2013. “Searching for Common
Factor in Public Private Real Estate Returns.” The Journal of
Portfolio Management 39 (6): 120–133. doi:10.3905/jpm.2013.
39.6.120.
Bali, T. G., S. J. Brown, and M. O. Caglayan. 2014.
“Macroeconomic Risk and Hedge Fund Returns.” Journal of
Financial Economics 114 (1): 1–19. doi:10.1016/j.jfineco.2014.
06.008.
Boons, M. 2016. “State Variables, Macroeconomic Activity, and
the Cross Section of Individual Stocks.” Journal of Financial
Economics 119 (3): 489–511. doi:10.1016/j.jfineco.2015.05.
010.
Boons, Martijn, Fernando Duarte, Frans De Roon, and Marta
Szymanowska. 2020. “Time-Varying Inflation Risk and Stock
Returns.” Journal of Financial Economics 136 (2): 444–470. doi:
10.1016/j.jfineco.2019.09.012.
Bruno, G., M. Esakia, and F. Goltz. 2019. Towards Cost
Transparency: Estimating Transaction Costs for Smart Beta
Strategies. Nice, France: Scientific Beta.
Campbell, J. Y. 1987. “Stock Returns and the Term Structure.”
Journal of Financial Economics (18) (2): 373–399. doi:10.1016/
0304-405X(87)90045-6.
Campbell, J. Y. 1996. “Understanding Risk and Return.” Journal
of Political Economy 104 (2): 298–345. doi:10.1086/262026.
Chen, N.-F., R. Roll, and S. A. Ross. 1986. “Economic Forces
and the Stock Market.” The Journal of Business 59 (3): 383–403.
doi:10.1086/296344.
56
approach can be considered an alternative to shrinkage,
as both approaches try to account for the uncertainty
around point estimate.
16. Our analysis accounts for the effective bid-ask spread,
but ignores the costs such as brokerage commissions,
exchange taxes, and financial transaction taxes. Such
costs are difficult to account for because they may vary
widely depending on the bargaining power of investors.
17. We report more detailed analysis of turnover and
transaction costs over different time periods for all macro
dedicated portfolios in the Online Supplemental Material.
18. The average out-of-sample beta increases from 1.71 to
2.45, whilst the average t-statistic increases from 2.53 to
4.39, as shown in Table 2.
Chousakos, Kyriakos, and Daniel Giamouridis. 2020. “Harvesting
Macroeconomic Risk Premia.” The Journal of Portfolio
Management 46 (6): 93–109. doi:10.3905/jpm.2020.1.149.
Chung, Kee H., and Hao Zhang. 2014. “A Simple Approximation
of Intraday Spreads Using Daily Data.” Journal of Financial
Markets 17 : 94–120. doi:10.1016/j.finmar.2013.02.004.
Corwin, Shane A., and Paul Schultz. 2012. “A Simple Way to
Estimate Bid-Ask Spreads from Daily High and Low Prices.” The
Journal of Finance 67 (2): 719–760. doi:10.1111/j.1540-6261.
2012.01729.x.
Devarajan, M., V. Naik, A. Nowobilski, S. Page, and N.
Pedersen. 2016. Factor Investing and Asset Allocation – a
Business Cycle Perspective. Charlottesville, VA: The CFA
Institute Research Foundation. http://library.asue.am/open/
5334.pdf.
Estrella, A., and M. R. Trubin. 2006. “The Yield Curve as a
Leading Indicator: Some Practical Issues.” Current Issues in
Economics and Finance, Federal Reserve Bank of New York 12 (5).
https://msuweb.montclair.edu/lebelp/
EstrellaYieldCurveIndicatorFRBNY200608.pdf.
Fama, E. F. 1981. “Stock Returns, Real Activity, Inflation, and
Money.” The American Economic Review 71 (4): 545–565.
Fama, E. F., and K. R. French. 1989. “Business Conditions and
Expected Returns on Stocks and Bonds.” Journal of Financial
Economics 25 (1): 23–49. doi:10.1016/0304-405X(89)90095-0.
Fama, E. F., and K. R. French. 1993. “Common Risk Factors in
the Returns on Stocks and Bonds.” Journal of Financial
Economics 33 (1): 3–56. doi:10.1016/0304-405X(93)90023-5.
Fama, E. F., and K. R. French. 1997. “Industry Costs of Equity.”
Journal of Financial Economics 43 (2): 153–193. doi:10.1016/
S0304-405X(96)00896-3.
Fama, E. F., and K. R. French. 2018. “Choosing Factors.” Journal
of Financial Economics 128 (2): 234–252. doi:10.1016/j.jfineco.
2018.02.012.
Herskovic, B., A. Moreira, and T. Muir. 2019. Hedging Risk
Factors (Working paper). https://doi.org/10.2139/ssrn.3148693

Hong, H., and M. Yogo. 2012. “What Does Futures Market
Interest Tell Us About the Macroeconomy and Asset Prices?”
Journal of Financial Economics 105 (3): 473–490. doi:10.1016/j.
jfineco.2012.04.005.
Hou, K., H. Mo, C. Xue, and L. Zhang. 2021. “An Augmented q-
Factor Model With Expected Growth.” Review of Finance 25 (1):
1–41. doi:10.1093/rof/rfaa004.
Jurczenko, Emmanuel, and Je  ro
^ me Teiletche. 2020. Macro
Factor-Mimicking Portfolios (Working paper). https://doi.org/10.
2139/ssrn.3363598
Keim, D. B., and R. F. Stambaugh. 1986. “Predicting Returns in
the Stock and Bond Markets.” Journal of Financial Economics 17
(2): 357–390. doi:10.1016/0304-405X(86)90070-X.
Ledoit, O., and M. Wolf. 2008. “ Robust Performance
Hypothesis Testing With the Sharpe Ratio.” Journal of Empirical
Finance 15 (5): 850–859. doi:10.1016/j.jempfin.2008.03.002.
Levi, Y., and I. Welch. 2017. “Best Practice for Cost-of-Capital
Estimates.” Journal of Financial and Quantitative Analysis 52 (2):
427–463. doi:10.1017/S0022109017000114.
Longstaff, F. A. 2004. “The Flight-to-Liquidity Premium in U.S.
Treasury Bond Prices.” The Journal of Business 77 (3): 511–526.
doi:10.1086/386528.
Newey, W. K., and K. D. West. 1987. “A Simple, Positive Semi-
Definite, Heteroskedasticity and Autocorrelation Consistent
Volume 79, Number 1
Targeting Macroeconomic Exposures in Equity Portfolios
Covariance Matrix.” Econometrica 55 (3): 703–708. doi:10.
2307/1913610.
Novy-Marx, R. 2013. “The Other Side of Value: The Gross
Profitability Premium.” Journal of Financial Economics 108 (1):
1–28. doi:10.1016/j.jfineco.2013.01.003.
Novy-Marx, R., and M. Velikov. 2016. “A Taxonomy of
Anomalies and Their Trading Costs.” Review of Financial Studies
29 (1): 104–147. doi:10.1093/rfs/hhv063.
Petkova, R. 2006. “Do the Fama–French Factors Proxy
for Innovations in Predictive Variables?” The Journal of
Finance 61 (2): 581–612. doi:10.1111/j.1540-6261.2006.
00849.x.
Swade, Alexander, Harald Lohre, Mark Shackleton, Sandra
Nolte, Scott Hixon, and Jay Raol. 2021. “Macro Factor
Investing With Style.” The Journal of Portfolio Management 48
(2): 80–104. doi:10.3905/jpm.2021.1.306.
Ung, D., and P. Luk. 2016. “What is in Your Smart Beta
Portfolio? A Fundamental and Macroeconomic Analysis.” The
Journal of Index Investing 7 (1): 49–77. doi:10.3905/jii.2016.7.
1.049.
Vasicek, O. A. 1973. “A Note on Using Cross-Sectional
Information in Bayesian Estimation of Security Betas.” The
Journal of Finance 28 (5): 1233–1239. doi:10.1111/j.1540-
6261.1973.tb01452.x.
57
Research Financial Analysts Journal | A Publication of CFA Institute
https://doi.org/10.1080/0015198X.2022.2129946

Supply Chain
Climate Exposure
Greg Hall, Kate Liu, Lukasz Pomorski, and Laura Serban
Greg Hall is an executive director at AQR Capital Management, LLC, Greenwich, CT. Kate Liu is an executive director at AQR Capital
Management, LLC, Greenwich, CT. Lukasz Pomorski is a managing director at AQR Capital Management, LLC, Greenwich, CT. Laura
Serban is a managing director, AQR Capital Management, LLC, Greenwich, CT. Send correspondence to Lukasz Pomorski at
lukasz.pomorski@aqr.com.

We propose an intuitive measure Introduction

of supply chain climate risks,
reflecting the fact that even a
green company may have material
climate exposure if its customers
or suppliers face climate risks.
C most pressing priorities for investors. For most, climate-aware
Our measure captures price
movements around climate news
better than traditional climate
data and shows performance pat-
terns consistent with re-pricing of
climate risks. The metric is more
suitable for risk measurement
than scope 3 emissions and
requires raw data that is broadly
accessible for large cross-sections
of stocks and is of higher quality
than currently available scope 3
data. We discuss applications for
both portfolio and corporate deci-
sion-making.
Keywords: carbon emissions;
climate; emissions offshoring;
ESG; greenhouse gas emissions;
scope 3; supply chain
Disclosure: No potential conflict of
interest was reported by the author(s).
PL Credits: 2.0
limate-related risks and opportunities have become some of the
investing means managing a portfolio’s carbon emissions, or its
exposure to carbon reserves or green revenues. This is a natural starting
point but is unlikely to offer a complete, holistic view of a portfolio’s cli-
mate risk. Importantly, most traditional metrics will not capture expo-
sures along the supply chain. The few climate metrics that have a supply
chain component (e.g., scope 3 emissions) are usually motivated by car-
bon accounting rather than climate risk and are generally thought to be
extremely noisily estimated.1 We fill this gap by proposing a simple and
intuitive measure of supply chain climate exposure. The measure reflects
the strength of economic connections along the supply chain and is thus
more likely to capture risks; it can also be readily computed by investors
or their data providers for a broad cross-section of firms and using data
that is verifiable and of arguably high quality.
To motivate our approach, we consider an example of a real-world
payments services company from a sector (IT) and industry (IT serv-
ices) that are rarely associated with climate risks.2 In fact, based on
standard climate data such as carbon intensity or carbon footprint,
this company seems green even compared to its sector or industry
peers. For example, its scope 1 and 2 emissions are lower than the
industry and sector median, and its scope 3 emissions are in the
We thank two anonymous referees, Steffen Bixby, David Blitz, Nicole Boyson, Alfie
Brixton, Jeff Dunn, Shaun Fitzgibbons, Andrea Frazzini, Luis Garcia-Feijoo, Jess
Gaspar, Will Goetzmann, Antti Ilmanen, Lasse Pedersen, Roberto Rigobon, Luke
Taylor, Dan Villalon, Eddie Watts, and Jing Zhang for helpful discussion and com-
ments. AQR Capital Management is a global investment management firm which
may or may not apply similar investment techniques or methods of analysis as
described herein. The security mentioned herein was selected merely as an example
and is not a recommendation to buy, sell, or hold. The views expressed here are
those of the authors and not necessarily those of AQR.
This is an Open Access article distributed under the terms of the Creative Commons
Attribution-NonCommercial-NoDerivatives License (http://creativecommons.org/
licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and
reproduction in any medium, provided the original work is properly cited, and is not
altered, transformed, or built upon in any way.

58 © 2022 AQR Capital Management, LLC. Published with license by Taylor & Francis Group, LLC. Volume 79, Number 1

greenest quartile within its industry. Based on this
information alone, we may conclude that the com-
pany has little climate risk exposure. However, our
assessment may change once we look at its custom-
ers. For example, the company’s 2012 10Ks state
that its top three strategic relationships with major
oil companies represented in the aggregate approxi-
mately 20% of its consolidated revenue. The large
economic exposure to oil majors may indicate that
the company might be indirectly, but possibly materi-
ally, exposed to disruptions and potential loss of rev-
enue caused by climate-related risks.
This motivating example captures the economic intuition
behind our framework. In essence, we propose measur-
ing supply chain climate exposure as a business-
weighted average of standalone climate exposure of a
firm’s customers and suppliers. For example, the down-
stream (customer-oriented) measure we analyze empiric-
ally in this paper is a revenue-weighted average of
customers’ own scope 1 þ 2 carbon intensity. Our meas-
ure is higher for those firms that derive a large share of
revenue from carbon-heavy customers, with the idea
that climate risks harming those customers would also
harm the firm’s revenues. The lens of revenues makes it
clear why supply chain climate exposure may be rele-
vant for a firm’s fundamentals (as would be any other
exposure that potentially affects revenues: e.g., Kolay,
Lemmon, and Tashjian [2016] document large costs
incurred by suppliers to economically distressed firms).
While our empirical analyses focus on the downstream
supply chain, the framework we present here can be
employed to define an upstream metric: a cost-
weighted average of suppliers’ exposures. This meas-
ure would also capture an aspect of a firm’s funda-
mentals, with the idea that firms that rely on suppliers
exposed to climate risks may face disruptions to their
supply chains when such risks materialize. Beyond the
downstream and upstream measure, a third stand-
alone metric that may round out the analysis is the
portfolio company’s own climate exposure, for
example, its carbon intensity, a measure already very
popular in investment practice (e.g., CFA Institute
2020, and the case studies included in that report).
We next empirically validate our downstream supply
chain measure. This is a challenge rarely attempted in
prior literature due to a paucity of data and the reality
that the most severe climate risks may not have yet
materialized. Nonetheless, we believe investors should
insist on such analyses before they use a new metric in
practice. We propose a series of such data tests, with
the goal of not only increasing conviction in our supply
Volume 79, Number 1
Supply Chain Climate Exposure
chain climate exposure measure but also giving invest-
ors a blueprint for testing other climate indicators.
We show that our measure has intuitive, but rela-
tively low correlations with traditional climate data.
More importantly, it helps explain stock price reac-
tions to climate-related news sentiment, supporting
our interpretation of climate risk exposure. The cor-
relations we find (about 0.2) are similar in magnitude
to the correlations of measures explicitly designed to
track climate sentiment (e.g., Engle et al. 2020). In
addition, our measure subsumes climate-related
information that is contained in third-party measures
(e.g., scope 1, 2, or 3 emissions). We also demon-
strate that it exhibits intuitive patterns around those
few climate-related events that have already materi-
alized in our sample. For example, in the month fol-
lowing the G8 declaration to reduce emissions in
2009 (an example of positive climate news), the
stocks we identify as green on our measure outper-
form the brown stocks by about 5%. Conversely, in
the month following the US presidential election of
2016 (an example of negative climate news, given
President Trump’s vocal comments), green stocks
lagged brown stocks by about 2%.
In addition, we found strong historical performance
of long-short, market-neutral portfolios based on our
measure. Historically, such portfolios realized a
Sharpe ratio of about 0.85. This performance is not
driven by broad country or sector bets (the country-
adjusted within-industry version of the portfolio real-
ized a Sharpe ratio of more than 0.7) and is not
explained by standard academic factor models (the
portfolio has annualized alphas of 6.5% to 9%). This
is consistent with the market pricing in supply chain
climate risks, possibly because they are deemed to
be increasingly material or because investors have
less appetite for holding these risks. At the same
time, we show that downstream supply chain expos-
ure also predicts fundamentals such as earnings or
sales surprises, beyond what is captured by trad-
itional quality, fundamental, or price momentum sig-
nals. This is consistent with our measure capturing a
new facet of quality and potentially generating for-
ward-looking returns, similar to other quality signals.
Finally, and importantly, the supply chain measures
we propose are not meant to be a replacement for
scope 3 emissions data. They are related and can
answer similar questions but differ in their explicit
focus on financial risk exposure. For example, a com-
pany may have zero scope 3 emissions but still score
very poorly on our downstream measure because it
59
Financial Analysts Journal | A Publication of CFA Institute

derives material revenue from customers with high
climate exposure and vice versa. Another clear differ-
ence is that scope 3 isolates those supply chain emis-
sions that are attributable to the company’s products,
making it a preferred measure for carbon accounting
purposes. In contrast, our measure may double-count
a company’s emissions (e.g., a customer firm’s overall
emissions are counted separately for each of the
firm’s suppliers), which is not ideal for carbon
accounting but in our view is acceptable as it leads
to a more comprehensive climate exposure
risk measure.
Given the strong interest in scope 3 emissions,
Figure 1 compares scope 3 and our downstream sup-
ply chain measure and the key ingredients necessary
to compute them. The arrows illustrate the focus of
each metric: scope 3 captures emissions from suppli-
ers into the product and from the product to custom-
ers; our measure instead depends on the economic
revenue from customers to the company and from
the company to suppliers.
Practical Applications of the
Supply Chain Climate
Exposure Measures
The downstream supply chain measure we analyze in
detail in this paper and its upstream supply chain
analogue (cost-weighted average exposure of a firm’s
suppliers) lend themselves most naturally to portfolio
context applications, but they are also useful for cor-
porate decision-makers.
We motivate our framework by its potential to cap-
ture climate risks in the supply chain of a portfolio
company and in the sections below validate this
claim empirically. Hence, the most obvious applica-
tion is risk monitoring and measurement, for
example, highlighting situations when a portfolio
drifts toward stocks with increased supply chain
exposure and possibly triggering a further investiga-
tion of such stocks. The measures we propose can
also be easily incorporated as a linear constraint in
portfolio optimization, for example, targeting a

Figure 1. A Simplified Comparison between Scope 3 Emissions and Downstream Supply Chain
Climate Exposure

Scope 3: accounng for the emissions traceable to a company’s products and services
Company A
Company A’s suppliers Some of A’s own emissions Company A’s customers
product specific emissions e.g. business travel product specific emissions

Product 1
.
.
.
Product n

Supply chain climate exposure: accounng for the strength of economic links with, and overall
emissions/climate exposure of a company’s customers and suppliers

Company A’s suppliers’ Company A Company A’s customers’

overall emissions/climate exposure overall emissions/climate exposure

Cost of supplies Revenue

Note: The schematic shows which data inputs are relevant for each climate measure, highlighting the product-centric nature of scope 3
emissions and the economic linkage-centric nature of the downstream supply chain climate exposure measure proposed in this paper.
Source: AQR. For illustrative purposes only.

60

specific reduction in exposure versus a benchmark.
Some investors may also decide to formally incorpor-
ate them into a risk model, with the goal of hedging
out climate-type risks, possibly along the lines sug-
gested in Engle et al. (2020). In addition, our finding
of return patterns associated with the downstream
measure suggests that it may potentially improve an
investor’s investment view, for example, by enhanc-
ing the understanding of a company’s quality profile.
We anticipate that such applications may be of inter-
est to both discretionary and systematic investors.
Moreover, the data inputs underlying our approach
are broadly available to investors and are of relatively
high quality, resulting in excellent data coverage of
our measure. These features support the use of the
measure for reporting purposes, allowing compari-
sons across mandates or even asset classes (because
the measure is at the issuer level, it can inform both
equity and credit allocations).
Our indicator may also be useful in stewardship, for
example, by identifying companies with abnormally
high supply chain exposures as potential targets for
engagement or by measuring the progress that port-
folio companies make in engaging with their own
supply chain partners. This may be particularly rele-
vant for emissions offshoring, which occurs when a
company switches from making the most carbon-
intense components inhouse to buying them from
suppliers instead.3 When such components are pro-
duced using company-owned assets, they increase
the company’s scope 1 emissions; when they are
bought from suppliers instead, they only affect scope
3 emissions. Because most investors today focus on
scope 1 and 2, emissions offshoring may artificially
increase investors’ assessment of how green a port-
folio company is. In principle, this behavior would be
apparent in a firm’s scope 3 emissions, but as we
mention above, such data is unfortunately very noisy
(e.g., Andersson, Bolton, and Samama 2016; Callan
2020; Cheema-Fox et al. 2021; Ducoulombier 2021).
Our measure can at least partially fill this gap by
helping identify companies that dramatically reduce
their scope 1 emissions but at the same time look
increasingly brown on our measure.
Finally, we believe our measure will be useful not
only to investors but also to the portfolio companies.
Corporations face pressure to report environmental,
social, and governance (ESG)–type information,
increasingly including information related to their
supply chains, e.g., scope 3 emissions. Unfortunately,
estimating product-related emissions is a daunting
challenge. For example, in its 2020 Environment
Volume 79, Number 1
Supply Chain Climate Exposure
Report, IBM (2020) stated that “the assumptions that
must be made to estimate Scope 3 emissions in most
categories do not enable credible, factual numbers.”
Busch, Johnson, and Pioch (2022) not only find large
inconsistencies in scope 3 estimates but also find
that such inconsistencies actually increase over time,
suggesting that the growing investor and perhaps
also regulatory interest in this measure has not yet
led to any unified reporting. Ducoulombier (2021)
points out that reporting on scope 3 remains largely
voluntary today (even if some important regulators
such as the U.S. Securities and Exchange Commission
may require it in the near future), that existing
reporting standards do not support comparisons
between firms, and that data providers producing
scope 3 estimates typically take insufficient consider-
ation of firm-level circumstances. In contrast, our
measure may not only be more material to a com-
pany’s risks (as we argue above), but it will also be
far easier for a corporation to compute and will be
more comparable across companies. Corporations
know precisely who their customers and suppliers
are and how much revenue and cost they account
for. They still need the data on the overall emissions
of a customer or a supplier, but that data can be esti-
mated more precisely: Indeed, unlike scope 3 emis-
sions, they are self-reported by the majority of
companies in broad indexes such as MSCI World
(e.g., Bixby, Brixton, and Pomorski 2022; Bolton,
Kacperczyk, and Samama 2022). In addition, an
important advantage of our approach is that it can
be adjusted to address at least some of the data
weaknesses that are a perennial problem in ESG. For
example, when customers’ scope 1 þ 2 carbon emis-
sions data is not available (perhaps because they are
private firms that do not report emissions), one could
reasonably assume that their own climate exposure is
in line with that of their peers. For example, even a
simple regression of scope 1 þ 2 emissions on indus-
try dummies explains more than 40% of the variabil-
ity in emissions. This flexibility extends the resulting
data coverage for an investor and makes it easier for
a corporate entity to compute its own supply chain
climate exposure.
Data
For our empirical analyses below, we focus on large-
cap developed stocks, roughly equivalent to the hold-
ings of the MSCI World Index. Importantly, we do
not restrict the universe of customers or suppliers of
such stocks: For example, if a global developed firm
in our sample has emerging markets customers, we
61
Financial Analysts Journal | A Publication of CFA Institute
will utilize data on such customers to assess the
firm’s supply chain climate exposure.
To measure downstream supply chain climate expos-
ure, we need two core pieces of data. First, we need
data on the commercial linkages between a company
and its customers. This data has been popularized by
academic papers going back to Cohen and Frazzini
(2008) and Menzly and Ozbas (2010). The supply
chain data we utilize here is blended from multiple
third-party vendors. These vendors use data from
regulatory filings, earnings transcripts, company and
industry publications, and news sources to compre-
hensively track global supply chains and the strengths
of linkages. We merge vendor data to maximize
coverage and precision. The linkages data we utilize
provides the best coverage we are aware of for our
universe of large-cap developed stocks: For example,
as of December 2021, we can identify at least one
customer for 87% of the stocks in our sample.
Ideally, the data would reflect not only a commercial
link between two companies but also the strength of
this link, for example, the fraction of a firm’s revenue
from a particular customer. For instance, Cohen and
Frazzini (2008) collect this type of revenue data, lev-
eraging disclosure regulations in the U.S. market
(companies must report customers accounting for
more than 10% of revenues). We have at least partial
data on revenues for 58% of the companies with any
linkage data. While this coverage falls short of 100%,
we use state-of-the-art data, utilized in multiple other
research papers, and are not aware of a dataset with
better systematic coverage. Moreover, even with
imperfect coverage, we capture the more material
economic linkages (e.g., in the U.S., where these must
be reported to investors). Conditional on having rev-
enue figures for at least one customer, our data cap-
tures 19.4% of total revenues for the average firm.
The second piece of data we need is a measure of a
customer’s (or a supplier’s) own climate exposure. In
the analyses below, we opted for a simple measure
that we believe will resonate with investors: scope
1 þ 2 carbon intensity of customers and suppliers, as
reported in the Trucost database. We use this as an
intuitive indicator of which customers and suppliers
are relatively greener or browner. We do not claim
that this simple indicator is necessarily the “best”
measure of climate risk exposure. We expect that
investors interested in our approach will instead sub-
stitute their favored instruments, whether carbon
emissions from another vendor such as MSCI or cli-
mate VaRs, Implied Temperature Rise scores, and so
62
on. Our methodology easily accommodates
such measures.
Finally, we use a variety of standard returns and
financials data from XpressFeed and a variety of
ESG-related metrics from MSCI and Trucost, as indi-
cated below. Overall, our sample spans the period of
May 2009 through December 2021.
Missing Data and Limitations of the
Supply Chain Climate Exposure Measure.
Missing data is an inevitable part of research, and
how missing data is dealt with may meaningfully
influence the analysis. We assume that a company
with missing carbon intensity data has similar carbon
intensity as its same-sector peers. This is preferable
to ignoring customers or suppliers with missing data
and effectively giving them a free pass even when
they are in sectors that are very pollutive. As we dis-
cuss below, when dealing with missing data on the
fraction of revenue companies have from a given
customer, we impute the strength of the relationship
using both the number of customers with missing
revenue data and their market cap, so that same-
sized customers account for the same fraction
of revenue.
One potential weakness of our approach is that we
may not observe a customer–supplier linkage in the
first place (We note that if portfolio companies self-
report on such measures, they of course have full
information about their own economic links). To par-
tially address this, we normalize the data we have to
100%. For example, if the data at our disposal cap-
tures only 70% of a firm’s revenue, we rescale it so
that it sums up to 100%. This implicitly assumes that
the missing firms are as green as a typical customer
on which we do have coverage. This is not ideal
because, in some circumstances, the missing data
may be statistically biased. For instance, our data
does not include companies’ retail customers. One
might expect that retail customers have a different,
perhaps “greener” profile than that of a typical cor-
porate customer. Unfortunately, we do not know
how much of the missing data covers retail versus
corporate customers, making it challenging to adjust
the measure. Similarly, companies that may be off-
shoring emissions, as we discuss in the Practical
Applications section, may strategically seek suppliers
less likely to be covered by data providers (e.g., pri-
vate firms, or firms in emerging or frontier markets).4
In our assessment, data quality is sufficient for prac-
tical applications, in spite of the concerns above.
Data on customer linkages may not be perfect, but it

features prominently in academic and practitioner
research, so it seems fair to assume that much of the
industry is familiar and perhaps even comfortable
with it.5
Another potential weakness of our analyses below is
that we use scope 1 þ 2 emissions to gauge custom-
ers’ own climate exposure. This is only one of the
many climate metrics investors may consider, but our
approach easily extends to other metrics. To the
extent an investor is comfortable using a different
variable to gauge climate exposure of a portfolio
company, it stands to reason that the investor should
be comfortable using the same variable to gauge sup-
ply chain climate exposure.
Constructing the Downstream
Supply Chain Climate
Exposure Measure
To make our downstream measure more tangible,
we start with a hypothetical example, showing how
our measure is computed. Consider a firm with four
customers, whom we label with subscript j below
(so, j ¼ A, B, C, or D for the four customers). As
mentioned above, we do not restrict the customers
to be in the same investment universe: For example,
the firm we study may be a large-cap developed
stock, while its customers may be small-cap and
emerging markets companies. To compute our
measure, we need to know each customer’s climate
exposure, Cj : In our applications below, Cj is cus-
tomer j’s scope 1 þ 2 carbon emissions intensity. In
our example, we assume we have this data for each
of our hypothetical customers; in practice, missing
data could be replaced by industry or sector aver-
ages. We also need data on the fraction of revenue
each customer j represents. To make our example
Table 1. Computing Downstream Supply Chain Climate Exposure for a Hypothetical
Company, i, with Four Customers, j ¼ A, B, C, and D
Customer’s Carbon
  Intensity
Cj % Revenue from Each Customer
Customer A 200
Customer B 100
Customer C 100
Customer D 50
Source: AQR. For illustrative purposes only.
Volume 79, Number 1
Supply Chain Climate Exposure
realistic, we assume that we have that data for two
customers, A and B, and that the data is missing for
customers C and D. Table 1 summarizes the data for
our example.
We first impute the missing revenue data for custom-
ers C and D by normalizing the total revenue to
100%. This implicitly imposes the assumption that
the company only has four customers, that is, that
we are not missing any customer linkage. For most
firms in our sample, this is likely to be an approxima-
tion only, but we would expect that our data would
include at least the most important commercial rela-
tionships of each firm. For example, for U.S. compa-
nies, we know all relationships that account for more
than 10% of revenue, given the regulatory require-
ment to report them. In Table 1, we assume that cus-
tomers C and D are equally important; in our
empirical tests below, we impute the fraction of rev-
enue using both the number of customers and their
size, arriving at higher imputed revenue for custom-
ers with a higher market cap.
Once we have the data on both customers’ own cli-
mate exposure, Cj , and their relative revenue, wij ,
our downstream supply chain climate exposure meas-
ure for company i is simply the revenue-weighted cli-
mate exposure, as per Equation (1):
Downstream Supply Chain Climate Exposurei
X
¼ w Cj ,
j ij
(1)
where in our empirical tests below Cj is customer j’s
scope 1 þ 2 carbon intensity, defined as scope 1 þ 2
emissions divided by sales. We normalize by sales
rather than using overall emissions in tons, as the lat-
ter would skew the metric toward large companies (a
large company will generally have higher overall car-
bon emissions).
Imputed % Revenue
ðwij Þ
40% 40%
30% 30%
NA 15%
NA 15%
63
Financial Analysts Journal | A Publication of CFA Institute
We focus on the downstream (customer) version of
the supply chain measure in our subsequent discus-
sion. The upstream (supplier) version can be com-
puted analogously:
Upstream Supply Chain Climate Exposurei
X (indeed, we see a strong correlation between supply
¼ v Cj ,
j ij
(2)
where vij is the fraction of business that company i
does with supplier j and Cj is supplier j’s scope 1 þ 2
carbon intensity. The resulting metric is a cost-
weighted climate exposure of a company’s suppliers.
Summary Statistics of Downstream
Supply Chain Climate Exposure
We begin by comparing our supply chain measure to
the typical carbon emissions metrics: scope 1, 2, and
3 carbon intensity. Our analysis deals with large-cap
developed equities, so not surprisingly the resulting
data coverage is excellent. Our supply chain measure
can be computed for 93% of the index weight in the
average sample quarter and covers 96% of the index
toward the end of our sample; standard intensity
measures, as reported in the Trucost data, cover 87%
of the index weight on average and 95% toward the
end of our sample. While the resulting coverage hap-
pens to be similar for our measure and for standard
carbon intensity data, we note that this is not guar-
anteed. We can compute our metric even for those
stocks that are not covered by Trucost, as long as we
can identify at least some of their customers. In fact,
we can compute our measure even if we do not
have such customers’ own carbon intensity data,
because we can leverage sector or industry member-
ship of these companies and use sector or industry
averages as proxies. For example, an IT services com-
pany that received 10% of its revenue from an Oil
and Gas customer is likely exposed to supply chain
climate risks, whether we explicitly have the custom-
er’s emissions data or not. These features of our
approach may be particularly important in those
investment universes where the coverage of standard
emissions data is poor, for example, for small-cap
equities or for private issuers in credit.
Figure 2 presents supply chain and standard emis-
sions statistics for both a typical MSCI World con-
stituent and for the MSCI World Index overall. As
explained in the earlier sections, our downstream
supply chain climate exposure measure is a revenue-
level average of the carbon intensities of each firm’s
customers. The median stock’s downstream exposure
is smaller than, but relatively closer to, scope 3
64
emissions than to either scope 1 or scope 2.6 In con-
trast, at the index level, estimated supply chain cli-
mate exposure is meaningfully higher than that
implied by traditional emissions data. This is driven
by larger firms having higher supply chain exposure
chain exposure and market capitalization).
Figure 2 suggests that our measure is correlated with
standard carbon emissions metrics. For example, we
observe correlations of 0.1 to 0.15 between our
measure and scope 3 emissions. This may seem low
given that the two metrics both focus on the supply
chain, but—as we point out in the introduction—they
capture fundamentally different dimensions. More
surprisingly, perhaps, our measure correlates with
scope 1 emissions, meaning that firms that produce
more emissions from company-controlled sources
also have customers who emit a lot. In this case, the
correlations increase to a more meaningful but still
relatively low level of 0.3 to 0.4. This is partially
because our measure depends on customers’ scope 2
emissions: At least some of these will be the same as
scope 1 emissions of a given firm (in line with this
intuition, the correlation is about twice as high for
Utilities than for firms in other sectors).
We now turn to the distribution of our measure
across sectors and geographies, as presented in
Figure 3. For ease of presentation, we show the
results in terms of the contribution of each sector (in
Panel A) and each region (in Panel B) to the overall
index level. For comparison, we also show the index
weight of each sector and region. We compute the
contributions for each quarter in our sample and
then show the contribution for the average quarter
in Figure 3.
In Panel A, we see that our downstream measure
spikes for Utilities, Energy, and Materials, the three
sectors that are also associated with high scope 1
emissions, suggesting that their (revenue-weighted)
average customer is also carbon-intense. This may
not be surprising: Companies that are important cus-
tomers for these sectors, and therefore use a lot of
energy, oil, gas, or minerals, are likely companies that
also have substantial greenhouse gas emissions.
The distribution by sector also shows that the down-
stream carbon metric may capture different informa-
tion than scope 1, 2 and 3 carbon intensity. For
example, we see a relatively higher downstream
exposure for Industrials than one may expect from
traditional emissions data. Similarly, Financials have
relatively larger downstream supply chain climate
 Supply Chain Climate Exposure

Figure 2. Carbon Emissions Statistics over Time

A. Median Stock Carbon Emissions Exposures

140

median stock's emissions intensity 120

100
80
60
40
20
0

Jul-13

Jul-20
Jun-09
Jan-10

Mar-11

Feb-14

Jun-16
Jan-17

Mar-18
Dec-12

Nov-15

Dec-19

Feb-21
Aug-10

Oct-11

Sep-14
Apr-15

Aug-17

Oct-18

Sep-21
May-12

May-19
Downstream supply chain Scope 1 Scope 2 Scope 3

B. MSCI World Index–Weighted Carbon Emissions Exposures

400
MSCI World emissions intensity

350
300
250
200
150
100
50
0
Jan-10
Jun-09

Mar-11

Jul-13

Jun-16
Jan-17

Mar-18

Jul-20
Sep-14
Dec-12

Feb-14

Nov-15

Dec-19

Feb-21
Aug-10

Oct-11

Aug-17

Oct-18

Sep-21
May-12

Apr-15

May-19

Downstream supply chain Scope 1 Scope 2 Scope 3

Note: The downstream supply chain climate exposure and scope 1, scope 2, and scope 3 emissions intensity for the median MSCI
World constituent in Panel A and for the overall MSCI World Index in Panel B for the period of June 2009 to December 2021.
Source: AQR, MSCI, Trucost.

exposure relative to scope 1, 2, and 3 carbon inten- mentioned in our data section, our customer data
sities. This suggests that some industrials and finan- does not include data on retail customers who may
cial companies provide goods and services to be relatively more important for Consumer Staples;
customers who may be subject to significant climate however, for this to explain the patterns in Figure 3,
risk, despite these companies’ stand-alone processes retail customers would need to be more carbon-
being relatively green. In contrast, downstream heavy than corporate customers. Finally, we note
exposure is relatively less pronounced for Consumer that supply chain climate exposure is relatively less
Staples, at least compared to the scope 3 emissions concentrated at the sector level than scope
of that sector. This is perhaps surprising. As we 1 exposure.

Volume 79, Number 1 65

Financial Analysts Journal | A Publication of CFA Institute

Figure 3. Sector- and Country-Level Contributions to MSCI World Index–Level Carbon

Emissions Metrics

A. Sector Contribuon to Various Carbon Emissions Metrics

50%
%contribution to MSCI World

45%
40%
35%
30%
25%
20%
15%
10%
5%
0%

weight Scope 1 Scope 2 Scope 3 Downstream supply chain

B. Country–Level Contribuon to Various Carbon Emissions Metrics

70%
%contribution to MSCI World

60%

50%

40%

30%

20%

10%

0%
USA Europe ex UK FRA JPN CAN Other
UK&FRA

weight Scope 1 Scope 2 Scope 3 Downstream supply chain

Note: Each sample quarter, percentage contributions to MSCI World Index are calculated from each sector (Panel A) and each coun-
try or broader region (Panel B) as of December 31, 2021. The charts show the time series averages of these contributions. In Panel
A, the historically deprecated Telecommunication is omitted from the chart.
Source: AQR, Trucost, MSCI.

66

Panel B of Figure 3 repeats the analysis by geog-
raphy, looking at the countries with the highest
weight in the MSCI World Index. We observe some
variability related to size (e.g., USA stands out in
absolute terms, but not when compared to the index
weight of that region), but no significant concentra-
tion beyond that. A couple of countries stand out in
that their downstream supply chain exposure is
noticeably different than the traditional scope 1, 2,
or 3 exposure. For example, the UK contributes 4%
to 8% to MSCI World overall scope 1, 2, 3 weighted
average carbon intensity but as much as 10% to the
index’s downstream supply chain exposure (for com-
parison, the average index weight of the UK is about
8%). On the flipside, Japan contributes less to index-
level downstream exposure than it does to scope 1,
2, or 3 emissions, suggesting that it does business
with relatively greener customers than its own emis-
sions profile would suggest. Last, we see that com-
modity-based economies such as Canada contribute
more to the downstream supply chain exposure than
their weight in the index or their scope 3 emissions.
It seems intuitive that commodity producers have
carbon-heavy customers, but it is perhaps more sur-
prising that they score as “browner” on our supply
chain measure than they do on scope 3 emissions.
Canada’s contribution to index-level downstream
exposure is twice as high as its scope 3 contribution
(Australia, while lumped into “Other” in Panel B, also
contributes twice as much to the supply chain expos-
ure as it does to scope 3). One explanation for this
pattern is that the customers of commodity-produc-
ing countries are relatively browner not just because
of those countries’ own scope 3 emissions but also
because of other sources of emissions; another is
that scope 3 data may be missing some components
of customers’ scope 1 þ 2 emissions that we explicitly
account for here.
Validating the Measure
Our approach to measuring supply chain climate
exposure is based on a clear and intuitive economic
idea. However, before we can advocate practical use
of any measure, we want to go beyond the intuition
and to the extent possible to also verify it with data.
Unfortunately, this is a significant challenge for ESG
measures broadly speaking and for climate measures
in particular. This is because the outcome variables
we care about here—instances of climate risks affect-
ing company valuations—are rare in historical data;
the most severe climate events, whether related to
physical or transition climate risks, may only occur in
the coming years and decades. This may be true, but
Volume 79, Number 1
Supply Chain Climate Exposure
it is a very poor excuse not to look at data. Markets
do react to potential future risks, so climate risks
may well affect historical stock prices even before
they materialize. Realistically, climate may account
only for a small fraction of a stock’s total or even
residual volatility—clearly less than, say, market beta,
industry membership, and so on. For this reason, we
sharpen the empirical tests below by conditioning
them on climate-related variables, for example, look-
ing at returns around climate-related news, when cli-
mate risk is more likely to be priced in or re-priced.
We first relate our downstream supply chain measure
to more traditional climate data, including
Environmental (E-pillar) scores and carbon emissions.
A positive correlation with related third-party climate
data would lend credibility to our measure. If we
instead saw zero correlation with “standard” climate
data, it would be difficult for us, or for the investor
community, to accept this measure, no matter how
plausible the intuition behind it. Second, we test
whether downstream climate exposure correlates
with climate sentiment. That is, we check whether
the companies identified by our measure as relatively
green or relatively brown experience returns, with
intuitive signs, versus measures of market-wide senti-
ment following climate news. Third, we use an event
study methodology to gauge the price behavior of
stocks around meaningful macro-level climate news,
as a function of stocks’ downstream supply chain
exposure. These tests may be circumstantial, but we
believe that the preponderance of evidence justifies
the use of our proposed metric in invest-
ment practice.
Comparison to Third-Party Climate-
Related Metrics. To compare our downstream
supply chain climate risk measure with other climate
data, we regress it on scope 1, 2, and 3 emissions
intensity, as reported by Trucost, and on E-pillar and
climate change theme scores, as reported by MSCI.
Table 2 shows that downstream supply chain climate
exposure is strongly related to third-party measures
of emissions, in most cases with the intuitive signs.
For example, we expect a higher (riskier) downstream
exposure for those companies that also have high
carbon emissions intensity, at least in terms of scope
1 and 3 emissions.7 Similarly, we find intuitive signs
for climate-related measures from MSCI ESG data:
Higher downstream supply chain carbon exposure
correlates with lower (i.e., less attractive) overall
MSCI E score in regression (2) and lower Climate
Change theme scores, the component of E that is
most related to our measure, in (3).8 We pool the
67
Financial Analysts Journal | A Publication of CFA Institute

Table 2. Explaining Downstream Supply Chain Carbon Exposure with

Traditional Climate Data
(1) (2) (3) (4)

ln CO2 intensity, scope 1 0.29 0.08

(56.08) (14.43)
ln CO2 intensity, scope 2 0.09 0.01
(8.63) (1.76)
ln CO2 intensity, scope 3 0.15 0.09
(13.03) (6.83)
MSCI E-pillar score 0.05 0.00
(4.16) (0.90)
MSCI climate change score 0.04 0.02
(5.40) (3.93)
Sector þ region indicator NO NO NO YES
Observations 77,194 84,249 65,727 55,970
R2 0.280 0.011 0.010 0.461
Number of x-sections 51 51 36 36

Note: Estimates are from Fama-MacBeth regressions of the downstream supply chain carbon exposure (in logs)
on scope 1, 2, 3 carbon intensity (in logs), MSCI E-pillar, and MSCI Climate Change theme. Regressions (1)–(3)
do not include any additional explanatory variables; regression (4) also includes sector and regional indicator
variables. The regressions are run quarterly from June 2009 to December 2021 for the developed large-cap
universe (MSCI World); the table presents time series estimates t-statistics adjusted for heteroskedasticity and
auto-correlation with Newey-West procedure with 4 lags.
Indicates significance on a 5% basis and Indicates significance on a 1% basis.
Source: AQR, Trucost, MSCI.

explanatory variables and add sector and region indi-
cators in regression (4) and again find strong positive
correlations with scope 1 and 3 emissions. This time,
surprisingly, we find a positive rather than a negative
coefficient on the MSCI Climate Change theme,
counterintuitively suggesting that issuers with a
higher (i.e., more attractive) climate change score also
have higher downstream supply chain carbon expos-
ure. The sign flip is likely driven by the correlations
between the various climate variables; the interpret-
ation of the coefficient on MSCI Climate Change
score is more difficult when these correlated varia-
bles are included in the regression.9
Importantly, while we find strong and mostly intuitive
correlations with typical ESG data, these third-party
variables only capture a small amount of the variabil-
ity in our measure. R2 values in Table 2 range from
1% for high-level E and Climate Change scores to
20% to 30% when controlling for emissions-based
measures. Even in our “kitchen sink” regression (4),
where we also control for sector and region member-
ship, traditional climate data and sector/country fixed
effects capture less than half of the variability of
our measure.
This combination of findings is encouraging. If we
found no correlations whatsoever with third-party
climate data, this might mean that our measure sim-
ply does not capture any climate-related information.
The third-party measures we use here, and similar
measures we did not have access to for this paper,
come from well-known, reputable data providers
who work with some of the most sophisticated
investors and consultants. These measures may be
noisily estimated, but we posit that they capture at
least some climate-related information, meaning that
we should expect at least some correlation with our
own metric.
At the same time, it is comforting that the correlation
is imperfect. Had we found R2 values close to 1.0,
we would have concluded that our measure effect-
ively conveys the same information as what is
already captured by traditional providers. This would
reduce the potential usefulness of our idea to
those allocators who already subscribe to the stand-
ard climate data. Of course, finding correlations lower
than 1.0 is not a sufficient condition for success: The
low correlations might after all be attributable to
white noise that carries no incremental information
about climate. To test whether this may be the case,
and to isolate the incremental information in our
measure, we resort to sentiment analysis in the
next section.

68

Correlation with Climate News
Sentiment. While the most meaningful climate risk
events may materialize only in the future, their likeli-
hood of occurrence and potential impact are certainly
discussed in the news media today. We can leverage
such discussions to validate our measure: If it indeed
captures climate risk, then stocks with relatively
higher exposure should do poorly when negative cli-
mate news arrives; stocks with relatively lower
exposure should do well. The advantage of this
approach is that we rely on the overall market, rather
than on individual analysts, to assess which stocks
are more exposed to climate risks. Of course, we do
not observe such market assessment directly, but we
can infer it implicitly from price movements around
climate news events, as explained in Engle et al.
(2020) and AQR (2021). Moreover, we can use this
insight to check whether our measure captures infor-
mation not yet contained in standard climate data.
We do this by sorting stocks on a version of supply
chain exposure that is orthogonalized to other cli-
mate data.
To carry out this analysis, we need data on climate
sentiment. We utilize the Media Climate Change
Concern index (MCCC) from Ardia et al. (2021),
made publicly available from these authors, and
measure the arrival of (signed) climate news by look-
ing at the monthly changes to the index. This climate
news sentiment data was used in other papers as
well, for example, in Pastor, Stambaugh, and Taylor
(2021), and was available until June 2018.10
To measure the relationship between downstream
supply chain climate exposure and climate news sen-
timent, we form hypothetical factor portfolios that
long 50% of stocks with low downstream supply
chain climate exposure (we refer to these as “green
stocks”) and short the other 50% of stocks (“brown
stocks”). To avoid unintended bets, we form these
portfolios in a country-neutral manner, with or with-
out industry adjustments, and target zero beta
(equity market neutral) and 7% ex ante volatility. As
before, we use large-cap developed market stocks
for this analysis. Figure 4 presents the (monthly) cor-
relations between these factors and climate
news sentiment.
Figure 4 shows that factors based on downstream
supply chain climate exposure co-move with the cli-
mate news sentiment, with the positive correlation of
0.15 to 0.2. That is, the stocks we identify as greener
outperform the stocks we identify as brown at times
of increased climate concerns. Given the length of
our sample (109 months), the t-statistic on this
Volume 79, Number 1
Supply Chain Climate Exposure
correlation is 1.76. To put these correlations in con-
text, we note that our supply chain measure does
not utilize any news data. For comparison, portfolios
explicitly optimized to correlate with climate news
sentiment in Engle et al. (2020) also show out-of-
sample correlations of about 0.2.
We believe that these positive correlations are an
important validation of our measure, indicating that
downstream supply chain climate exposure captures
information that the market subsequently deems
important in the climate context.
We still need to check whether such information
might simply be a function of some relatively obvious
characteristics, such as sector or industry member-
ship (for example, as we saw above, Utilities stand
out on our measure). To address this, Figure 4 also
shows portfolios that are defined entirely within
industry. We still see a positive correlation here,
even if it is somewhat lower than before. Another
potential concern is that our measure is to some
extent correlated with traditional climate data (as we
saw in the prior section) and the return correlations
with climate news might be indirectly attributable to
such third-party data and not to our downstream
supply measure per se. Figure 4, Panel A shows that
this is not the case: We find similar levels of correla-
tions also for factors that were orthogonalized to
scope 1, 2, and 3 emission and to MSCI’s Low
Carbon Transition (LCT) scores.11
Strikingly, the reverse does not hold when we
instead analyze factors based on traditional climate
data in Figure 4, Panel B. On their own, factors based
on greenhouse gas emissions, or on MSCI Low
Carbon Transition scores, show correlations that are
relatively similar to those in Panel A. However, as
soon as we control for industry membership, these
correlations disappear for LCT scores or even turn
negative for scope 3 emissions. We still find low
positive correlations for scope 1 þ 2 emissions, but
they decrease to almost zero when we orthogonalize
the factors to our downstream supply chain measure.
We conclude that the correlations between trad-
itional climate data and climate news either are due
to industry membership or are subsumed by our
downstream supply chain metric.
Reaction to Macro-Level Climate Change
News. Our last validation test relies on a simple
event study methodology: We utilize the long-short
factors based on downstream supply chain carbon
exposure data, introduced above, and assess how
they behave around major climate events. This
69
Financial Analysts Journal | A Publication of CFA Institute

Figure 4. Correlation of Market-Wide Climate Sentiment with Factors Built Using Downstream
Supply Chain Carbon Exposure

A. Correlations of Downstream Supply Chain Factors with Climate News Sentiment

0.20
Correlation w/climate sentiment

0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
-0.20
Base Orthogonal to Orthogonal to Orthogonal to Orthogonal to
Scope 1+2 Scope 3 MSCI LCT all three

Overall Within Industry

B. Correlations of Traditional Climate Data with Climate News Sentiment
0.20
Correlation w/climate sentiment

0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
-0.20
Scope 1+2 Scope 3 MSCI LCT

Overall Within Industry

Overall, orth. to supply chain Within ind, orth. to supply chain

Note: The top panel shows monthly return correlations with the monthly changes in climate news sentiment index of Ardia et al.
(2021) for factors based on downstream supply chain climate exposure overall and within-industry exposure as well as after remov-
ing the commonality with other typical climate-related data. The bottom panel shows similar correlations for a hypothetical long-
short portfolio based on typical climate-related data, removing the commonality with the downstream supply chain measure, as indi-
cated in the chart. The hypothetical portfolio ranks stocks on (negated) downstream supply chain carbon exposure; neutralizes coun-
try and, as indicated, industry exposures; removes market exposures based on ex-ante beta; and targets 7% ex ante volatility. This
hypothetical portfolio is rebalanced monthly. Returns are gross of transaction costs, financing costs, and fees. Given the available cli-
mate sentiment data, the sample period is June 2009 through June 2018. Hypothetical performance results have many inherent lim-
itations, some but not all of which are described herein. No representation is being made that any fund or account will or is likely
to achieve profits or losses similar to those shown herein. Hypothetical performance results are presented for illustrative purposes
only. Hypothetical performance is gross of advisory fees and transaction costs and includes the reinvestment of dividends. If the
expenses were reflected, the performance shown would be lower.

70
 Supply Chain Climate Exposure

Figure 5. Stock Reaction to Climate News Events

A. Example Posive Climate News Event: G8 Agrees to Reduce GHG Emissions by 80% by 2050
6%
Beta-adjusted returns
4%

2%

0%

-2%

-4%

-6%
-30 -27 -24 -21 -18 -15 -12 -9 -6 -3 0 3 6 9 12 15 18 21 24 27 30
Days relative to the event

Long Short

B. Example Negave Climate News Event: President Trump Withdraws the U.S. from Paris
Agreement
2%
Beta-adjusted returns

1%

0%

-1%

-2%

-3%

-4%
-30 -27 -24 -21 -18 -15 -12 -9 -6 -3 0 3 6 9 12 15 18 21 24 27 30
Days relative to the event

Long Short

Note: Daily cumulative beta-adjusted returns to the longs and shorts from the downstream supply chain climate exposure factor, in
the 30-day window surrounding two climate events: G8 committing to reducing emissions by 80% by 2050 on 9th July 2009 (Panel
A) and President Trump withdrawing the U.S. from the Paris Agreement on 1st June 2017 (Panel B). The long and short side are
beta-adjusted using the MSCI World Index benchmark. Universe of large-cap developed markets stocks, similar to MSCI World
Index constituents. The long-short factor portfolio ranks stocks on their (negated) supply chain carbon exposure, neutralizes country
and industry exposures, removes market exposures based on ex ante beta, and targets 7% ex ante volatility. This hypothetical port-
folio is rebalanced monthly. Returns are gross of transaction costs, financing costs, and fees.
Source: AQR.

experiment is obviously related to the climate senti-
ment analysis above and may be noisier in that the
information about such macro events may have been
telegraphed to the media, and thus also the market,
before the event itself took place. At the same time,
this approach has the benefit of a straightforward
intuition and is arguably more transparent than the
more sophisticated natural language processing tech-
niques behind sentiment measures. For this reason,
we illustrate this approach with two sample events,
choosing one that is positive for climate (G8 commit-
ting on 9th July 2009 to reduce greenhouse gas
emissions by 80% by 2050) and one that is negative
(President Trump pulling the U.S. from the Paris
Agreement on 1st June 2017). Given the construc-
tion of our factors, we would predict them to do well
around positive climate news (i.e., we would expect
the long leg of the factor to appreciate relative to

Volume 79, Number 1 71

Financial Analysts Journal | A Publication of CFA Institute
the short leg of the factor) and do poorly around
negative climate news (longs depreciate relative to
shorts). Figure 5 tests whether this is the case, show-
ing market beta-adjusted cumulative performance of
the longs and shorts in the 30-day window around
the two events in question. The “0” date on the x-
axis corresponds to the event itself.
The results in Figure 5 are more mixed, but we still
observe some encouraging patterns. For positive
news in Panel A, the short side of the factor (brown
stocks) has negative market-adjusted performance
after the event happens. The long side (green stocks)
is more muted and noisier. For negative news, it is
the long side that reacts more strongly, with more
pronounced negative market-adjusted performance.
Of course, the two events we highlighted here are only
a fraction of possible climate events one may consider.
As we explained above, we showcase them here to
illustrate the overall approach but would argue that the
news sentiment analysis in the previous section is a
more informative way to validate the measure.
Performance
Our final section discusses the performance of stocks
sorted on the downstream supply chain climate
exposure measure, utilizing the long-short portfolio
we introduced in our climate sentiment analysis
above. Hypothetical performance, gross of fees and
transaction costs, is presented in Table 3.
We find strong performance for portfolios long in
green stocks and short in brown stocks, identified
using our measure. The portfolio formed ignoring
stocks’ industry membership realized a Sharpe ratio
of 0.85 over this period. This strong performance is
not subsumed by sector allocation (e.g., short Energy,
long IT) as we also find a relatively similar Sharpe
ratio for the industry-neutral version of the portfolio.
As we show in Table 4, the performance is not
explained by exposures to well-known investment
factors such as Value or Momentum.
When the baseline portfolio is regressed on standard
academic factors (Fama-French three- or five-factor
models, potentially augmented with momentum and
betting against beta [BAB]), it still shows economic-
ally and statistically significant abnormal returns of
6.5% to 9.4% p.a. As in Table 3, the within-industry
results are slightly weaker but still strong in regres-
sions (4)–(6). In regressions (7)–(9), we apply an even
more conservative test, redefining factors to match
the portfolio construction (within-country, within-
72
Table 3. Portfolio Performance for
Longshort Portfolios Based on
Downstream Supply Chain
Climate Exposure
Unadjusted Within-Industry
Average return 7.4% 6.5%
Standard deviation 8.7% 9.0%
of returns
Sharpe ratio 0.85 0.72
Beta to MSCI World 0.07 0.05
Hit ratio 56% 56%
Note: Universe of large-cap developed markets stocks, similar
to MSCI World constituents. The long-short portfolio ranks
stocks on (negated) downstream supply chain carbon exposure,
neutralizes country and, when indicated, industry exposures,
and removes ex-ante market exposures and targets 7% ex ante
volatility. This hypothetical portfolio is rebalanced monthly.
Returns are gross of transaction costs, financing costs, and
fees. The sample period is June 2009 to December 2021.
Source: AQR.
industry) and using the same universe of stocks as
that used for the portfolio. We believe that such fac-
tors lead to a more conservative assessment in that
they match the most important design choices of the
portfolio being tested. The estimated alphas slightly
decrease to 3.8% to 6% p.a. but remain economically
sizeable and, with the exception of regression (9), sig-
nificant at the 5% level.
Table 4 also indicates that the downstream supply
chain factor tends to have a positive growth exposure,
similarly to the green portfolio in Pastor, Stambaugh,
and Taylor (2021); estimated coefficients on other
investment factors are not consistent across specifica-
tions. We have verified that the value spread of the
portfolio does not meaningfully widen over this period,
as might be expected if our results were driven primar-
ily by ESG investment flows. We also see at least as
strong performance in the first half of our 2009 to
2021 sample as we do in the second half. In additional
robustness checks, we find that the performance is not
driven by a single country or region.
While this strong performance is arguably good
news, we do not view it as strictly necessary for the
validation of our measure. Our key research objective
is to study (downstream) supply chain climate expos-
ure, and it is at least theoretically possible that such
exposures may not have strong performance implica-
tions. Having said that, we note that this perform-
ance can be consistent with the market pricing in
Volume 79, Number 1
Table 4. Factor-Adjusted Performance of Downstream Supply Chain Climate Exposure (DSChCE) Long-Short Portfolios
Unadjusted DSChCE Portfolio Within-Industry DSChCE Portfolio Within-Industry DSChCE Portfolio
Unadjusted Academic Factors Unadjusted Academic Factors Within-Industry Academic Factors
(1) (2) (3) (4) (5) (6) (7) (8) (9)

Alpha 6.48% 9.41% 8.92% 5.31% 4.95% 6.55% 5.98% 4.85% 3.80%
(2.70) (3.92) (3.41) (2.15) (1.91) (2.35) (2.45) (1.97) (1.46)
Mkt-rf 0.02 0.05 0.05 0.01 0.01 0.02 0.02 0.03 0.03
(0.52) (1.12) (1.03) (0.22) (0.16) (0.31) (0.35) (0.52) (0.59)
Size 0.05 0.23 0.25 0.17 0.19 0.26 0.19 0.21 0.21
(0.38) (1.71) (1.75) (1.26) (1.32) (1.75) (2.24) (2.52) (2.40)
Value 0.39 0.56 0.53 0.44 0.57 0.51 0.40 0.27 0.18
(4.58) (4.34) (3.56) (4.99) (4.08) (3.17) (5.13) (3.03) (1.24)
Profitability 0.73 0.75 0.01 0.05 0.18 0.17
(4.10) (4.08) (0.04) (0.26) (2.24) (2.19)
Conservatism 0.02 0.05 0.31 0.33 0.14 0.15
(0.10) (0.23) (1.34) (1.37) (2.06) (2.15)
Momentum 0.02 0.13 0.07
(0.24) (1.27) (0.44)
BAB 0.04 0.23 0.11
(0.36) (1.88) (1.19)
Observations 151 151 151 151 151 151 151 151 151
R2 0.142 0.233 0.235 0.151 0.162 0.184 0.160 0.208 0.217

Note: Monthly returns of the downstream supply chain climate exposure long-short portfolio are regressed on the factors indicated in row labels. Regressions (1)–(3) use the port-
folio defined both across- and within-industry; in (4)–(9), the portfolio is based on within-industry supply chain exposures only. Regressions (1)–(6) use academic factors (Market,
Size, Value, Profitability, Conservatism, and Momentum are Mkt-rf, SMB, HML, RMW, CMA, and UMD, respectively from Ken French’s Data Library; BAB is from the AQR Data
Library), and regressions (7)–(9) use the within-industry versions of those factors, defined using equivalent stock characteristics (e.g., market cap for Size, book-to-price for Value,
etc.). Alphas are annualized, and t-statistics are shown in parentheses. The sample period is June 2009 to December 2021.
 indicates significance on a 5% basis and  indicates significance on a 1% basis.
Source: AQR, Ken French’s Data Library.

73
Supply Chain Climate Exposure
Financial Analysts Journal | A Publication of CFA Institute

Table 5. Regressions Relating Quarterly EPS and Sales Surprises to Downstream Supply
Chain Climate Exposure
Unadjusted Industry Demeaned
Additional Controls EPS Sales EPS Sales

No additional controls 9.30 5.64 3.83 1.72

Analyst revisions 7.77 5.02 3.04 0.79
Analyst revisions þ Profitability þ DProfitability þ Momentum 9.24 5.55 3.38 1.03

Note: Each quarter, EPS and Sales surprises, as indicated in column headers, are regressed on a given firm’s downstream supply
chain climate exposure and additional controls prior to the earnings announcement, as indicated by row labels. The table shows
the t-statistics of the coefficient on supply chain climate exposure, obtained from the time series of cross-sectional regressions
using the Fama-MacBeth method. The t-statistics are adjusted for heteroskedasticity and auto-correlation with Newey-West pro-
cedure with four lags. Profitability is defined as return on net operating assets; Fundamental Momentum is defined as 1-year
change in Profitability; and Price Momentum is based on the total return of stocks in the past 12 months excluding the most
recent month. The data in the right panel is industry-demeaned. The sample period is June 2009 to December 2021. Universe is
large-cap developed markets stocks, similar to MSCI World constituents.
Indicates significance on a 5% basis and  Indicates significance on a 1% basis.
Source: AQR.

climate risks. Arguably, markets have paid increas-
ingly more attention to climate, so as long as they
care about the downstream supply chain exposure,
the prices of issuers with high exposure may have
fallen relative to the prices of issuers with low expos-
ure, leading to the positive returns in Tables 3 and 4.
Pedersen, Fitzgibbons, and Pomorski (2021) show
how such repricing occurs in response to changes in
the preferences of, or the information available to,
market participants.
At the same time, there are alternative explanations:
For example, our climate exposure measure might be
indicative of higher production efficiency and higher
overall quality, as suggested in a related context in
Garvey, Mohanaraman, and Nash (2018). While trad-
itional measures of quality (e.g., profitability, or even
emissions, as in Garvey, Mohanaraman, and Nash
2018) relate to a specific portfolio company, it is pos-
sible that quality extends also across the supply
chain. For example, firms that have customers of
higher quality, including in the dimension of carbon
emissions, may be well positioned going forward
because their customers are less likely to experience
adverse risk events. To test such alternative explana-
tions of abnormal performance, we turn to Table 5,
where we assess whether downstream supply chain
climate exposure can help predict companies’ quar-
terly EPS and sales surprises. To do so, we run
Fama-MacBeth regressions of surprises at quarterly
earnings announcements on a given company’s
downstream supply chain exposure and a range of
controls, meant to capture analyst revisions, profit-
ability characteristics, fundamental momentum, and
price momentum ahead of the earnings
announcement.
Table 5 indicates that companies with “greener”
downstream exposure on average experience larger
EPS surprises than other companies. This effect
persists also when controlling for analyst revisions
and price and fundamental momentum, with or
without industry adjustment. We see similar results
for sales surprises, although we lose significance
when variables are industry-demeaned. While this
relationship with fundamentals might be explained
by risk exposures, we believe it is more likely to
reflect a facet of Quality, perhaps related to the
resource efficiency of a company along its supply
chain. If so, then the strong historical returns are
consistent with the premium we observe for other
Quality-type variables, say, profitability. In this case,
we may expect such premia to persist going for-
ward as well.
Conclusions
We propose a novel method to measure a firm’s cli-
mate risk exposure in the supply chain, leveraging
the firm’s economic linkages. We make a simple but,
to our knowledge, not-yet-proposed point that such
risks can be modeled as revenue-weighted climate
exposures of a company’s customers or a cost-
weighted climate exposure of its suppliers. Our
framework can be utilized with one’s preferred meas-
ure of exposure, potentially capturing physical cli-
mate risks, transition opportunities, green revenues,
and so on. In our empirical validations, we find that

74

downstream supply chain climate exposure captures
co-movement of stock returns around climate news
events, largely subsuming relevant information from
other traditional climate data, and that its perform-
ance is consistent with the market (re)pricing climate
risk exposures.
Our measure is similar to, but not a replacement for,
the often-discussed scope 3 emissions. It captures
somewhat different information that we argue may
be more material for issuers’ financial risks. It also
has a meaningful advantage of being easier to assess,
using data that is already broadly familiar to the
investment community. While our measure is not
available “off the shelf” from third-party vendors, the
straightforward methodology we explain in detail in
this article can be implemented in-house by at least
Editor's Note
Submitted 4 June 2022
Accepted 26 September 2022 by William N. Goetzmann
Notes
1. For example, Busch et al. (2022) find that scope 3
estimates are very rarely reported by portfolio companies
and third-party estimates are essentially uncorrelated
across data providers; Callan (2020) found no
respondents who use scope 3 data in investment practice.
2. To identify a company suitable for an illustrative example,
we sorted firms on their downstream supply chain
exposure and on their own emissions, focusing on those
that are at the top of the distribution of the former and
at the bottom of the latter. We then analyzed the public
statements the companies made to identify the ones that
disclosed the importance of customers who may have
large climate exposure (e.g., oil majors).
3. See, e.g., Ben-David et al. (2021) and Dai et al. (2022).
Offshoring may also be an issue for sovereign emissions.
4. These biases could perhaps be alleviated by leveraging
data on revenues from geographic segments and
assessing the climate exposure in each region in which a
company does business. We believe this approach is
coarser than what we propose; in addition, it would
require data on climate exposures of specific regions or
countries, which itself may pose challenges.
5. For example, a sampling of papers utilizing such data
include Cohen and Frazzini (2008), Menzly and Ozbas
(2010), Agarwal et al. (2017), or Lee et al. (2019).
6. The jump in early 2010 is driven by adding an additional
provider of customer-supplier data; from 2010 onward,
we have supply chain linkages data from all
underlying providers.
Volume 79, Number 1
Supply Chain Climate Exposure
some investors. Moreover, while we focus on public
equities in this article, investors can leverage our idea
also in other asset classes. This is most obvious in
credit, but also holds in sovereign fixed income,
where one could consider trade-weighted exposure
to countries that produce more emissions or are
otherwise exposed to climate risks.
The flexibility of our intuitive methodology, good
data coverage, and the additional validations we pre-
sent here bode well for the usefulness of this
approach to measure and report risks of individual
assets and broad portfolios. We believe our approach
will also interest corporate decision-makers, both for
their own risk analyses and also as an intuitive and
relatively easy-to-compute reporting item.
7. While scope 2 emissions show a negative sign in
regression (1), this is likely driven by the correlations
between scope 1, 2, and 3 emissions. When we regress
our measure on scope 2 alone, with or without sector
and region controls, we find reliably positive
coefficients.
8. We also confirmed a similar finding using MSCI Low
Carbon Transition scores (results available by request).
We do not report this regression given a shorter time
series for that variable (6 quarters of data, as opposed
to 36 or 51 quarters for other climate variables
we use).
9. (3) and (4) have somewhat different samples because of
data coverage, but when (3) is estimated using the same
sample as (4), the coefficient on MSCI Climate Change
theme is clearly negative, 0.05 (t-statistic ¼ 10.3).
Without the region/sector dummies but with all other
climate variables from (4), the coefficient is still positive,
0.03 (t-statistic ¼ 11.61).
10. We have repeated all analyses below also with a different
source of sentiment data: the Crimson Hexagon series,
used in Engle et al. (2020). The results are very similar to
what we presented here. We chose to highlight the Ardia
et al. data in our paper because the data is publicly
available and because it has a somewhat longer time
series (the Crimson Hexagon data at our disposal ended
in mid-2017).
11. LCT data has a relatively short sample, so we assume a
company’s historical LCT scores are the same as when
MSCI first reports them.
75
Financial Analysts Journal | A Publication of CFA Institute
References
Agarwal, Ashish, Alvin Chung Man Leung, Prabhudev Konana,
and Alok Kumar. 2017. “Co-Search Attention and Stock Return
Predictability in Supply-Chains.” Information Systems Research
28 (2): 265–88. doi:10.1287/isre.2016.0656.
Anderson, Mats, Bolton Patrick, and Frederic Samama. 2016.
“Hedging Climate Risk.” Financial Analysts Journal 72 (3): 13–32.
doi:10.2469/faj.v72.n3.4.
AQR. 2021. “Can Machine Learning Help Manage Climate
Risks?” AQR Alternative Thinking Q4: 2021.
Ardia, David, Keven Bluteau, Kris Boudt, and Koen
Inghelbrecht. 2021. Climate Change Concerns and the
Performance of Green versus Brown Stocks, working paper,
National Bank of Belgium.
Ben-David, Itzhak, Yeejin Jang, Stefanie Kleimeier, and Michael
Viehs. 2021. “Exporting Pollution: Where Do Multinational Firms
Emit CO2?” Economic Policy 36 (107): 377–437. doi:10.1093/
epolic/eiab009.
Bixby, Steffen, Alfie Brixton, and Lukasz Pomorski. 2022.
“Looking Forward with Historical Carbon Data.” In Climate
Investing: New Strategies and Implementation Challenges, edited
by E. Jurczenko. London: Wiley and ISTE.
Bolton, Patrick, Marcin Kacperczyk, and Frederic Samama.
2022. “Net-Zero Carbon Portfolio Alignment.” Financial Analysts
Journal 78 (2): 19–33. doi:10.1080/0015198X.2022.2033105.
Busch, Timo, Matthew Johnson, and Thomas Pioch. 2022.
“Corporate Carbon Performance Data: Quo Vadis?” Journal of
Industrial Ecology 26 (1): 350–63. doi:10.1111/jiec.13008.
Callan. 2020. Callan 2020 ESG Survey.
CFA Institute. 2020. “Climate Change Analysis in the Investment
Process.”
Cheema-Fox, Alexander, Bridget Realmuto LaPerla, George
Serafeim, David Turkington, and Hui (Stacie) Wang. 2021.
“Decarbonizing Everything.” Financial Analysts Journal 77 (3):
93–108. doi:10.1080/0015198X.2021.1909943.
76
Cohen, Lauren, and Andrea Frazzini. 2008. “Economic Links and
Predictable Returns.” The Journal of Finance 63 (4): 1977–2011.
doi:10.1111/j.1540-6261.2008.01379.x.
Dai, Rui, Rui Duan, Hao Liang, and Lilian Ng. 2022.
“Outsourcing Climate Change.” working paper, ECGI.
Ducoulombier, Frederic. 2021. “Understanding the Importance
of Scope 3 Emissions and the Implications of Data Limitations.”
The Journal of Impact and ESG Investing 2021: jesg.2021.1.018.
doi:10.3905/jesg.2021.1.018.
Engle, Robert, Stefano Giglio, Bryan Kelly, Heebum Lee, and
Johannes Stroebel. 2020. “Hedging Climate Change News.” The
Review of Financial Studies 33 (3): 1184–216. doi:10.1093/rfs/
hhz072.
Garvey, Gerald, Mohanaraman Iyer, and Joanna Nash. 2018.
“Carbon Footprint and Productivity: Does the 'E' in ESG
Capture Efficiency as Well as Environment?” Journal of
Investment Management 16 (1): 59–69.
IBM. 2020. “2020 IBM and the Environment Report.” https://
www.ibm.com/ibm/environment/annual/IBMEnvReport_2020.
pdf
Kolay, Madhuparna, Michael Lemmon, and Elizabeth Tashjian.
2016. “Spreading the Misery? Sources of Bankruptcy Spillover
in the Supply Chain.” Journal of Financial and Quantitative
Analysis 51 (6): 1955–90. doi:10.1017/S0022109016000855.
Lee, Charles, Stephen Sun, Rongfei Wang, and Ran Zhang. 2019.
“Technological Links and Predictable Returns.” Journal of Financial
Economics 132 (3): 76–96. doi:10.1016/j.jfineco.2018.11.008.
Menzly, Lior, and Oguzhan Ozbas. 2010. “Market Segmentation
and Cross-Predictability of Returns.” The Journal of Finance 65
(4): 1555–80. doi:10.1111/j.1540-6261.2010.01578.x.
Pastor, Lubos, Robert Stambaugh, and Lucian Taylor. 2021.
“Dissecting Green Returns.” Working Paper, University of
Chicago.
Pedersen, Lasse, Shaun Fitzgibbons, and Lukasz Pomorski.
2021. “The ESG-Efficient Frontier.” Journal of Financial
Economics 142 (2): 572–97. doi:10.1016/j.jfineco.2020.11.001.
Financial Analysts Journal | A Publication of CFA Institute
https://doi.org/10.1080/0015198X.2022.2126590
Trade Informativeness in
Modern Markets
Samarpan Nawn, CFA
Samarpan Nawn, CFA, is an assistant professor in the Department of Finance and Accounting at the Indian Institute of Management
Udaipur, Udaipur, Rajasthan, India. Gaurav Raizada is a founder at Irage Capital and visiting faculty at the Indian Institute of Management
Ahmedabad, Ahmedabad, Gujarat, India. Send correspondence to Samarpan Nawn at samarpan.nawn@iimu.ac.in.
Using transactions-based calendar
time (TBCT) portfolio analysis, we
investigate informativeness of trades
of investor categories, namely insti-
tutions, proprietary traders, and
retail clients. We find that trade
informativeness is positive for insti-
tutional and negative for retail-client
investors. The informativeness of
liquidity-demanding trades are less
than the informativeness of liquid-
ity-supplying trades for all trading
groups, over both long and short
horizons. We also find that institu-
tions are benefitted by algorithmic
executions compared to manual
executions and this benefit is ele-
vated on days of high volume and
volatility. Proprietary algorithmic
traders (high-frequency traders) gen-
erate positive alpha for their trades
only from their liquidity-supply-
ing trades.
Keywords: algorithmic trading; HFT;
informativeness; institutional
traders; limit orders; market orders
Disclosure: No potential conflict of interest
was reported by the author(s).
PL Credits: 2.0
Volume 79, Number 1
Research
, and Gaurav Raizada
hich group of traders are informed among institutional, pro-
W prietary, and retail client (i.e., non-institutional and non-pro-
prietary)? While such a question has widely interested
academicians and practitioners alike, the literature primarily has
chosen to concentrate on one of the groups—institutions
(Hendershott, Livdan, and Schu € rhoff 2015) or retail (Odean 1999;
Barrot, Kaniel, and Sraer 2016; Boehmer, Fong, and Wu 2021)—or at
most make a comparison between institutions and retail (Stoffman
2014). Additionally, as Boehmer, Fong, and Wu (2021) mention, most
of the existing studies are conducted on datasets that comprise only a
small part of the overall market. For example, Odean (1999) use data
from a single U.S. retail brokerage firm and Barrot et al. (2016) use
data from a single French brokerage firm. Only Barber et al. (2009)
provide an analysis that uses data from the entire equity market of
Taiwan; however, their analysis uses data (pre-2000) from a period
that did not see algorithmic trading (AT), and hence their results may
not be representative of the current trading landscape across the
world. While Boehmer, Fong, and Wu’s (2021) data is from a recent
period and represents a sizeable fraction of the U.S. market, their ana-
lysis focuses on only retail traders and uses a proxy to measure retail
marketable order flow (ignoring their limit orders). Finally, proprietary
traders seem to be almost entirely missing from informativeness ana-
lysis even though they contribute significantly to the overall volume.
In contrast to the extant literature, using a clean, rich, exact, and
exhaustive (exchange-provided) trading group identification, we pro-
vide a comprehensive informativeness analysis of all trading groups
using a 3-year (2013–2015) complete order and trade-level data from
the National Stock Exchange (NSE) of India. NSE is the tenth largest
stock exchange in the world in terms of market capitalization1 (ranked
just below the London Stock Exchange and the TMX Exchange of
Canada). The additional advantage of using NSE data is that the stock
markets in India can be considered unfragmented as NSE accounts for
more than 80% of the total trading volume, with the Bombay Stock
Exchange (BSE) accounting for the remaining. Thus, a comprehensive
informativeness analysis done on the NSE data is almost as good as
one performed on the entire Indian stock market.
© 2022 CFA Institute. All rights reserved. 77
Financial Analysts Journal | A Publication of CFA Institute
Apart from providing a clean identification of each
order and trade—on whether it is from the institu-
tional, proprietary, or retail-client traders—the data
also provides a flag to enable us to identify whether
it is coming from an algorithmic terminal. While we
start the analysis with the institutional, proprietary,
and retail-client trading group partition, we later use
the algorithmic flag and provide insights into the dif-
ference between algorithmic and non-algorithmic
traders within institutions and proprietary groups.
One particularly interesting subgroup is proprietary
traders using algorithmic services, often called high-
frequency traders (HFTs; SEC 2010). A stark reality is
that the architecture of modern markets has changed
after the emergence of AT technology. AT is growing
throughout globally (Boehmer, Fong, and Wu 2021)
and, currently, most of the trading in major stock
markets is happening through AT. In India, AT was
allowed in 2008, but it only started becoming popu-
lar after the advent of direct market access (allowing
traders to rent rack space in the premises of the
exchange itself) in 2010. Nawn and Banerjee (2019)
found in the NSE-listed top 50 stocks (accounting for
60% of the exchange’s market capitalization), 43% of
total trading volume and 85% of total order mes-
sages in 2013 emanated from AT. Thus, similar to
most active exchanges of the world, NSE sees a
heavy presence of AT.
Using the transactions-based calendar time (TBCT)
method (Seasholes and Zhu 2010; Fong, Gallagher,
and Lee 2014), our primary finding is that institutions
have a significantly positive alpha over all holding
periods, while the retail-client group has a signifi-
cantly negative alpha. For example, the risk-adjusted
alpha for institutions is 7 basis points (bp) for a 1-day
holding period. The proprietary group primarily has
an insignificant alpha throughout. Our results on
retail traders link more closely with the traditional lit-
erature suggesting they are naive (Odean 1999;
Barber et al. 2009) than with the recent literature
suggesting otherwise (Kaniel, Saar, and Titman 2008,
Kelley, and Tetlock 2013; Welch 2022) not only
because our data for the group is a more comprehen-
sive representation but also perhaps for the compos-
ition of the trading group. The retail-client trading
volume is quite significant in India (close to 50%),
with many individuals directly trading in the market.
This is similar to the Taiwanese market structure,
where more than 80% of the trading volume gener-
ates from retail order flow. In the United States, this
percentage is significantly lower. It is possible that
many are exposed in the equities markets there
through pension funds and mutual funds (Engen and
78
Lehnert 2000) and, thus, the proportion of “noise”
traders in the retail trading group is lower, contribu-
ting to the divergence in findings.
We work with all stocks listed in the NSE of India.
However, to ascertain that our results of institutional
informativeness are not driven by any particular
group of stocks, we segregate the stocks into three
groups by size—(1) the top or 50 largest stocks
(accounting for approximately 60% of total market
capitalization), (2) the following 450 stocks (account-
ing for 35%), and (3) the remaining stocks (accounting
for 5%)—and find that institutions (retail) have posi-
tive (negative) and significant alpha for all three
groups of stocks.
In spirit with much of the literature (Fong, Gallagher,
and Lee 2014), we equate the alpha from the four-
factor model with informativeness. However, the
alpha could also stem from other factors such as
investor ability to process public information and
behavioral biases. Specifically, Seasholes and Zhu
(2010) link the informativeness of trades to investors’
ability to exploit non-public information. Thus, for
the robustness of our main results, we repeat our
analysis on a group of stocks with zero analyst cover-
age. Very low–analyst coverage stocks are taken as a
proxy for stocks with high information asymmetry
(Hong, Lim, and Stein 2000; Chae 2005; Kovacs
2010). The alpha achieved in these stocks can be
mostly attributed to informativeness. We do find that
our main results of institutions having significantly
positive alphas and retail clients having significantly
negative alphas in both long and short horizons con-
tinue for the stocks with zero analyst coverage. Our
results from this section provide robustness to our
main analysis.
Traditional research assumed that informed traders
were more likely to use market orders (O’Hara 1995;
Easley et al. 1996), due to their execution certainty.
Limit orders have been known to face adverse selec-
tion and free option risk (Copeland and Galai 1983).
An overall informativeness analysis may not reveal
the true story where the relative usage of market
and limit orders differ among the trading groups. This
is especially relevant because retail traders are
known to supply immediacy to institutional traders
(Kaniel, Saar, and Titman 2008). In the spirit of Fong,
Gallagher, and Lee (2014), we next examine the
informativeness among our three groups of traders
separately on the trades they initiate (trading through
market orders) and on the trades they supply liquid-
ity to (trading through limit orders).

Our rich and detailed data allows us to correctly
identify the liquidity-demanding and liquidity-supply-
ing leg for each trade. Therefore, we do not need to
rely on any trade classification algorithms. Our results
in this section are reassuring and enticing in parallel.
For trades conducted using market orders, the insti-
tutions have positive and significant alphas at both
long and short horizons, while proprietary and retail-
client traders have negative and significant alphas
and the retail-client group has a more negative alpha.
For trades conducted using limit orders, both institu-
tions and proprietary traders have positive alphas,
with institutions having greater alphas compared to
proprietary traders. The retail-client group has insig-
nificant alphas over both long and short horizons
when using limit orders. The reassuring part of the
result is that the rank ordering of alphas among the
three trading groups remains valid whether we exam-
ine trades conducted through market orders or
trades conducted through limit orders. The enticing
part of the results is that for all three trading
groups—institutions, proprietary, and retail client—we
find that the liquidity-supplying trades have greater
alphas compared to liquidity-demanding trades at all
horizons (short and long).
While the alphas estimated from limit order trades
are high, the entire alpha could represent the com-
pensation for liquidity provision. The compensation
for supplying immediacy is earning the bid-ask spread
(Handa, Schwartz, and Tiwari 1998). We use the Roll
(1984) model, which utilizes the short-term return
reversal in trade prices, to estimate bid-ask spread
for each stock day. From these estimates, we find
that while half-spread can explain a part of the alpha
of the liquidity-supplying trades, there is still a signifi-
cant part of the alpha that remains unexplained and
thus can be attributed to informed traders’ relative
use of more limit orders than market orders in mod-
ern algorithmic markets.
In the final section of the paper, we use the algorith-
mic flags provided by the exchange. We first ask
whether the algorithmic executions provide value to
the institutions. We separate the institutional trades
as manual or algorithmic. The institutional trades,
which are algorithmic, are in essence executed
through broker execution algorithms. We do a differ-
ence in TBCT alphas2 of algorithmic and manual exe-
cutions for the day of the trading and find that this
difference is positive and significant. Thus, broker
execution algorithms are able to create a value to
their customers (the institutions), especially managing
their trades well and the institutional gain by hiring
these execution algorithms. The implication of the
Volume 79, Number 1
Trade Informativeness in Modern Markets
result is in line with previous research (Domowitz
and Yegerman 2005), which suggests that orders
executed through algorithms have lower implementa-
tion shortfall costs than manual orders, providing evi-
dence of how execution algorithms provide value to
their clients. We further restrict our analysis to (1)
only high-volume stock days and (2) only high-volatil-
ity stock days. We find that the difference in execu-
tion alphas between institutional algorithmic and
manual trades is positive and significant and is of
greater magnitude (8 bp) compared to the full-sample
(2 bp) results. Thus, our results suggest that algorith-
mic executions become more valuable on high-vola-
tility and high-volume days.
We attempt to identify channels through which algo-
rithmic executions are able to provide value to insti-
tutions compared to manual executions. We find two
channels: The first is able to continuously monitor
the market and trade at favorable junctures. For
example, institutional algorithms monitor stock liquid-
ity and trade using relatively more limit orders when
the liquidity is scarce (and hence expensive) and
trade using relatively more market orders when the
liquidity is plentiful (and hence cheap). The second
channel is that algorithms are less affected by behav-
ioral biases. The one example we find is that com-
pared to manual executions, algorithmic executions
have much less tendency to trade at round numbers,
a well-known behavioral bias that generates losses
(Bhattacharya, Holden, and Jacobsen 2012).
Last, we concentrate on the profitability of the HFT
trading group (proprietary algorithmic traders). We
simply ask whether HFT trades are informative (on
both the short and long horizon) and how informative
are their liquidity-demanding in relation to liquidity-
supplying trades. The TBCT results on HFT provide
important insights. First, we see that HFT trades are
informative, from the very short horizon (day of the
trade) to a longer horizon of up to 20 days. Second,
HFTs’ liquidity-demanding (-supplying) trades have a
negative (positive) and significant alpha in both long
and short horizons. Our results indicate that HFTs
make consistent profits from their liquidity provision
or market-making strategies. Part of this alpha from
the liquidity-supplying trades of HFT can be
explained by the estimated half-spread, which is the
compensation for providing liquidity; however, the
remaining alpha should be attributed to informed
HFTs’ use of limit orders (Brogaard, Hendershott, and
Riordan 2019).
HFTs are expected to rely on speed primarily, and
thus the TBCT analysis of the holding period from
79
Financial Analysts Journal | A Publication of CFA Institute
the end of the trading day to 6 months may not be
appropriate for them. We provide additional analysis
for this group of traders—an intraday TBCT where
the holding period ranges from the end of the trading
minute to 60 minutes. Here we see that HFT trades
are significantly informative across the intra-
day horizon.
There are several implications of our study. If a port-
folio mimicking the holdings of institutions is man-
aged, then it can provide an economically significant
alpha to the manager who can replicate the portfolio.
For example, institutions generate an alpha of 3.2 bp
for a 5-day holding period after accounting for Fama-
French-Carhart factors. The positive and negative
informativeness of trades by institutional and retail-
client trader category has practical implications for
brokerages and market makers. These intermediaries
in a way can differentiate between the informed and
uninformed order flow and act accordingly.
Our finding of the lack of informativeness of the
retail-client group requires further mention. As AT
has been competing against individual investors in
liquidity supplying role in recent years, the return
predictability by individual trading may have weak-
ened (Tian et al. 2015). Given that retail investors
have less of an information advantage in the market
and their aggressive orders further reduce trading
profits, an insight for retail traders is to be patient
when trading by using limit orders as much as pos-
sible. The institutions would do better if they hire
execution algorithms to trade on their behalf. In case
of algorithmic versus manual execution, algorithmic
execution provides an advantage of 2 bp on the exe-
cution day and 8.6 bp on high-turnover days.
We acknowledge that the data period used in the
paper (2013–2015) is slightly dated. However, post
2016, Indian markets saw a stupendous and continu-
ous bull run right up to February 2020. In compari-
son, 2013 to 2015 saw a more normal scenario, with
the market going up in the first 2 years and falling in
the third year. Thus, we believe an informativeness
analysis done in the 2013 to 2015 period is more
generalizable compared to an analysis done in a
period of only bull run or one that includes special
events such as the COVID-19 crisis. Additionally, by
2013, the AT activity had already stabilized in the
NSE.3 Thereafter, there has not been a significant
structural shift in the trading landscape in Indian mar-
kets. An unreported analysis by Banerjee and Roy
(2022) compares 2015 to 2019 and finds that the
that the trading volume distribution among different
80
groups of traders remained more or less stable
between these two time periods in the NSE.
Institutional Features of the NSE,
Data, and Descriptive Details
The NSE was established in 1992 and has played an
instrumental role in the evolution of Indian capital
markets in its present state. It is a completely elec-
tronic order-driven exchange without designated
market makers. Trading takes place by an order-
matching mechanism: matching the marketable sell
(buy) orders with the best available limit buy (sell)
orders. The orders are entered and matched at strict
price, visibility, and time priority. After a 15-minute
pre-opening session (call auction), the exchange gets
ready for continuous trading at 9:15 AM. The single
continuous session runs from 9:15 AM to 3:30 PM
with a T þ 2 rolling settlement cycle. The exchange
freely displays live, on its website, the five top bid
and ask quotes and the number of shares demanded
or offered at those quotes for all traded securities.
The NSE’s completely order-driven market structure
mimics some of the world’s most important non-U.S.
stock exchanges, such as the Tokyo Stock Exchange,
HKEX, and Euronext Paris. The NSE is one of the
most active stock exchanges in the world. It has con-
sistently ranked in the top five exchanges of the
world in terms of the number of trades for the last
several years (according to the World Federation of
Exchanges statistics). According to the latest report,
the NSE ranks fourth according to this parameter.4
Even in terms of market capitalization, the NSE has
been steadily growing over the last several years and
currently ranks tenth globally, with a market capital-
ization of $2.55 trillion. The only exchanges ahead of
the NSE are the NYSE ($26.23 trillion), NASDAQ
($19.06 trillion), Shanghai Stock Exchange ($6.97 tril-
lion), Japan Exchange Group ($6.71 trillion), HKEX
($6.13 trillion), Euronext ($5.44 trillion), Shenzhen
Stock Exchange ($5.23 trillion), LSE ($4.04 trillion),
and Toronto Stock Exchange ($2.6 trillion). Many of
the well-known stock exchanges of the world, such
as the ASX and the Deutsche Boerse, rank behind
the NSE in market capitalization. Finally, in terms of
dollar traded volume, the NSE ranks twelfth globally,
with $1.8 trillion worth of shares traded, just behind
the LSE ($2.1 trillion) and Deutsche Boerse
($2 trillion).
Similar to other important exchanges of the world,
the NSE sees a heavy presence in algorithmic trading.
Algorithmic trading was introduced in the NSE in

Figure 1. Share of Algorithmic Trading and Importance of the NSE
100.00%
90.00%
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
2010-11 2011-12 2012-13 2013-14 2014-15 2015-16 2016-17 2017-18 2018-19 2019-20 2020-21
Share of algo trading in NSE
The figure illustrates the proportion of traded value in cash equities generated through algorithmic trading in the National Stock
Exchange (NSE) of India from 2010–11 to 2020–21. It also provides the share of the NSE in the Indian stock markets’ trad-
ing volume.
2008, but the system became popular after coloca-
tion services5 were introduced in 2010. The
exchange provides (against a charge) live tick-by-tick
market data, generating a broadcast for every order
message to supplement the colocation service. Nawn
and Banerjee (2019) find that algorithmic traders are
responsible for 85% of the order messages in the
fifty NIFTY 50 index stocks in 2013.
The share of trading volume emanating from algorith-
mic terminals has been consistently on the rise in the
NSE over the last decade (Figure 1), approaching
50% in recent years. This has coincided with the
steady increase in the proportion of overall Indian
equities market trading volume to the NSE. An
important characteristic of the Indian equity trading
microstructure is its near unfragmented nature. There
are only two major stock exchanges in India (the
Bombay Stock Exchange [BSE] is the other). Almost
all the major stocks are listed in both the exchanges,
however, the NSE is by far the dominant exchange in
terms of trading volume. While the NSE’s share was
already quite high (75%) in 2010, it has increased to
nearly 90% by 2020 (Figure 1). While correlation is
not causation and we do not claim causality, verbal
conversations with algorithmic traders in India seem
Volume 79, Number 1
Trade Informativeness in Modern Markets
Share of NSE in overall Indian stock markets
to indicate that they prefer the NSE over the BSE,
and perhaps that has paved the way for further con-
solidation of trading volume to the NSE. The NSE
has always been a pioneer and frontrunner in intro-
ducing technological innovation, and despite the BSE
being an older exchange (it started in the 19th cen-
tury), the NSE has been able to leapfrog the BSE.
Biais and Foucault (2014) express concern that it is
sometimes difficult to generalize trading group–spe-
cific research conducted in a single exchange of a
fragmented market, as sophisticated traders may
have taken the opposite or hedging positions in other
exchanges. As discussed, our results would not face
such concerns.
The data contains the complete orders and transac-
tions history on the NSE from January 1, 2013,
through December 31, 2015. Each order contains
fields for the market session (regular/auction), order
number, order entry time, order side (buy/sell), order
type (new/revision/cancel), order price, and size. The
trades data includes the date and time of the trans-
action, a stock identifier, transaction price, and the
number of shares traded. An algorithmic flag and a
client flag earmark each order message and each
trade (both legs of the trade) in the data obtained
81
Financial Analysts Journal | A Publication of CFA Institute
Table 1. NSE Firms and Turnover Year-Wise from 2013 to 2015
No. of
No. of Trades Turnover
Year Securities (Millions) (Billion INR)
2013 1711 160003 28220
2014 1839 215933 38763
2015 1777 219551 43568
The table provides summary statistics for trades in the National Stock Exchange (NSE) of India. The number of securities available
for trading and total number of trades are shown in the first two columns, respectively. The annual turnover and average daily
turnover for the stock segment show the growth in exchange volumes. The average trade value is the turnover divided by the
number of trades. The turnover ratio is calculated as the average yearly turnover of domestic shares divided by their average
year-end total market capitalization.
from the NSE. The two flags help identify whether
an order was from an algorithmic terminal and
whether the order was for a proprietary account, a
custodian account, or a non-proprietary and non-cus-
todian (others) account.
The custodian group consists of traders who are not
allowed to conduct their own clearing and settlement
and must do that through custodians. This group
consists primarily of foreign institutional investors,
mutual funds, and financial institutions, and we refer
to this group as institutions for the rest of the study.
Proprietary traders consist of broker members
engaged in trading on proprietary trading capital.
These may fall in the category of speculators and day
traders who trade frequently. They may also fit in
the category of “pseudo” market makers providing
liquidity on both the buy and sell sides of the market.
The rest of the traders, who are non-institutional and
non-proprietary, are referred to as retail client. The
primary constituent in this category is the traditional
retail traders using manual or non-algorithmic plat-
forms. However, there are a small group of algorith-
mic traders who are included in the retail-client
group. These traders are using client money and
algorithms for primarily day trading (thus, they are
non-institutional and non-proprietary by definition).
They constitute only 5% of the trading volume,
whereas the non-algorithmic (or simply “retail”) group
constitutes 40% of the volume. Regardless of the
main results of the overall retail-client group, in unre-
ported tests, we find the results for the only “retail”
group mimic the results of the overall retail-cli-
ent category.
An important feature of the identification is that
trading group classification is at a message-level
granularity. For example, when a trader uses algo-
rithms to conduct proprietary trades, those are
82
Average Daily Market
Turnover Average Trade Capitalization
(Billion INR) Value (INR) (Billion INR) Turnover Ratio
112.88 19402 64248 0.44
158.86 22490 84119 0.46
175.68 23380 98458 0.44
categorized as being from HFTs, but if the same
trader uses algorithms to conduct client trades, those
will not be categorized as being from HFTs. This
advantage overcomes a known limitation of popular
HFT-identifying databases that classify HFTs at the
firm level (e.g., the Nasdaq HFT database does not
classify proprietary algorithmic trades from broker-
dealer firms such as Goldman Sachs or Morgan
Stanley as being from HFTs).
Table 1 provides the year-wise summary statistics of
the NSE from 2013 to 2015. It provides information
about the monthly and daily turnover, number of
trades monthly, as well as average trade size. The
average turnover velocity is around 0.45 times the
market capitalization. The turnover velocity is calcu-
lated as the average monthly turnover of domestic
shares, divided by their average month-end total
market capitalization.
Table 2 provides the investor-wise turnover across
the period from 2013 to 2015. We find that institu-
tional trading amounts to 32% of the total turnover,
while proprietary trading is around 22%. Retail-client
trading accounts for the largest share (46%) of the
total turnover. We further subdivide the turnover
across the six investor categories. Algorithmic execu-
tion accounts for most of the institutional, propri-
etary group, but manual execution dominates the
retail-client group.
We include all stocks listed and trading in the NSE in
our empirical results. However, for a better under-
standing of the performance of each trading group in
large-, mid-, and small-cap stocks, we divide the
stocks into three groups: (1) the top 50 stocks or the
constituents of the benchmark index NIFTY 50, (2)
the next 450 stocks or the constituents of index
NIFTY 500 that are not included in NIFTY 50, and (3)
all other stocks. The top 50, the next 450, and the
 Trade Informativeness in Modern Markets

Table 2. Trade Descriptive Statistics by Investor Category for the Period 2013–2015
Top 50 Next 450 Rest of Stocks
Investor Turnover Mode of Turnover Turnover Turnover Turnover
Category (Billion INR) Trading (Billion INR) (Billion INR) (Billion INR) (Billion INR)

Institutions 70432 (32%) Algo 44817.43 (20%) 29172.88 (25%) 13399.08 (13%) 0.17 (1%)
Non-algo 26244.69 (12%) 26244.69 (23%) 15343.54 (15%) 0.34 (2%)
Proprietary 47174 (22%) Algo 28500.13 (13%) 14926.41 (13%) 11853.17 (12%) 0.14 (1%)
Non-algo 18673.94 (8%) 6614.53 (6%) 9861.50 (10%) 3.1 (22%)
Retail clients 101653 (46%) Algo 14340.30 (7%) 7268.07 (6%) 6299.20 (6%) 0.18 (1%)
Non-algo 87312.42 (40%) 30845.67 (27%) 42939.38 (43%) 10.19 (72%)

The table shows total turnover and turnover percentage executed by each investor category: institutions, proprietary trading, and
retail clients. Further, the turnover is broken down by the mode of trading into “algo” and “non-algo” participants. The turnover is
subdivided into the stock categories based on market capitalization of “top 50,” “next 450,” and “the rest of stocks.” Our sample
includes all stocks traded at the National Stock Exchange (NSE) of India and the time period of our sample is 1st January 2013 to
31st December 2015.

rest of the stocks account for approximately 60%,
35%, and 5% of the total market capitalization of the
exchange, respectively. We check for the turnover
across NIFTY 50, the next 450, and the rest of the
stocks for each of the investor categories.
Institutional trading has a higher percentage turnover
from the NIFTY 50, while retail-client trading has
more turnover in the next 450 stocks. The rest of
the stocks have extremely low volumes. Most of the
volumes in the rest of the stocks are contributed by
retail-client trading.
price movements. Chakrabarty et al. (2020) track hid-
den orders of HFT using the same NSE data, and
their findings are qualitatively similar to their findings
from NASDAQ data. Also, their findings on hidden
orders of the manual traders match the findings of
Bessembinder et al. (2009), who use Euronext Paris
data of 2003 (hence, all trading was manual then).
The robust findings of Nawn and Banerjee (2019)
and Chakrabarty et al. (2020) lend credence that
results achieved with NSE data are more likely to be
generalizable than not.

In the latter part of the study, we put particular focus

on the trading group with the algorithmic flag as
affirmative and client flag as “proprietary” and call
these traders HFTs, per the SEC (2010). The litera-
ture indicates that the way HFTs operate would gen-
erate a very high order-to-trade ratio compared to
other traders (Hendershott, Jones, and Menkveld
2011; Jørgensen, Skjeltorp, and Ødegaard 2018). We
next verify whether we see the same pattern in our
data. We compare HFTs’ order-to-trade ratio with
that of other traders who use trading algorithms and
traders who do not use algorithms. We see that
HFTs’ order-to-trade ratios are ten times higher than
other algorithmic traders and about 100 times higher
than traders who operate manually.6
Using data from the same stock exchange, Nawn and
Banerjee (2019) show that the proprietary algorith-
mic traders can be taken as the HFTs. Additionally,
they find that these traders increase the supply of
liquidity when short-term volatility is high or when
price movement is extreme. These results corrobor-
ate the finding of Brogaard et al. (2018) using
NASDAQ data that HFTs increase their net supply of
liquidity during periods of extreme stock-specific
Related Literature and Main
Research Questions
The extant literature on trading group informative-
ness provides seemingly isolated findings. The early
research on the topic suggests that institutions are
informed (Sias and Starks 1997; Chakravarty 2001)
and retail traders are naive (Odean 1998; Odean
1999; Barber and Odean 2000). However, recent evi-
dence has shown that retail traders are indeed not
naive. Kaniel, Saar, and Titman (2008) find that
stocks with intense buying by individuals outperform
stocks with intense selling by individuals in a short
horizon of 20 days. Kaniel et al. (2012) study trading
before earnings announcements and document that
aggregate individual buying (selling) predicts large
positive (negative) abnormal returns on and after the
announcements in both a very short horizon (2 days)
and a relatively long horizon (60 days). Fong,
Gallagher, and Lee (2014) provide evidence that
trades via full-service retail brokers are significantly
informative over both the long and short horizon
compared to trades via discount retail brokers. Their
analysis indicates that there may be various

Volume 79, Number 1 83

Financial Analysts Journal | A Publication of CFA Institute
categories of retail investors implied by their choices
of brokerage houses, and such consideration should
be kept in mind while generalizing informativeness of
retail traders.7 Using marketable retail order flow,
Boehmer, Fong, and Wu (2021) show that individual
stocks with net buying by retail investors outperform
stocks with net selling by approximately ten basis
points over the following week. Finally, a very recent
strand of literature focuses on trades from an ultra-
low-cost brokerage platform, “Robinhood Markets
Inc” (expected to cater to less sophisticated invest-
ors) and suggest that “Robinhood” investors do not
appear to be naive (Angel 2021; van der Beck and
Jaunin 2021; Welch 2022). However, Barber et al.
(2009) use the complete trading history of traders in
Taiwan and conclude that individuals lose on average
4% per annum, while institutions garner 1.5% per
annum. Hendershott et al. (2015) suggest that intu-
itional order flow is informed about news announce-
ments. Taking all news from Reuters between 2003
and 2005, they conclude that institutions can trade
ahead of macroeconomic news, earnings announce-
ments, and any news related to the firm’s long-term
fundamental information.
The different conclusions to the aforementioned
studies can be partially attributed to the type of data
used. There are major differences in the samples in
terms of different exchanges and brokerage houses.
Most importantly, the majority of the studies use
data that is only a small part of the overall market
and, therefore, generalization of the results on any
group of traders may not be feasible.
Because institutions are known to have superior
information-gathering and -processing skills, have
economies of scale to monitor many sources of infor-
mation, employ professionals with superior skills and
unlike retail traders, and communicate directly with
firms (Hendershott et al. 2015), our prior expectation
is that institutional trades would be more informative
than retail trades if a comprehensive analysis is done
with all traders in the population. Similarly, because
proprietary traders are likely to be skilled, profession-
ally hired personnel, the expectation is that their
trades would be more informative than that of retail
traders. Thus, in our first research question, we ask
whether institutional and proprietary trades are more
informative than retail trades.
While the first research question focusses on overall
trades by each group, a trade can happen either
through a limit order or through a market order. An
informativeness analysis perhaps will not describe
the complete picture where the relative usage of
84
market and limit orders differ among the trading
groups, especially with traditional academic literature
having assumed that informed traders were more
likely to use market orders (O’Hara 1995; Easley
et al. 1996) due to their execution certainty.
The analysis assumes additional significance due to
the mixed evidence on the recent literature regarding
the informativeness of different trading groups while
using market and limit orders. Kaniel, Saar, and
Titman (2008) emphasizes that contrarian liquidity
provision by retail traders earn them positive returns.
However, Barrot et al. (2016) suggest that individuals
do not reap the rewards from liquidity provision as
they reverse their trades long after the excess
returns from liquidity provision are dissipated,
although their theoretical long-short portfolio delivers
a significant positive annualized return. Kelley and
Tetlock (2013) find that retail market orders can cor-
rectly predict upcoming news. Barber et al. (2009)
find that retail traders lose on trading through market
orders, while institutions’ aggressive and passive
trades generate positive alpha. Thus, we ask whether
the institutional and proprietary market (limit) order
trades are more informative than retail market (limit)
order trades.
Methodology
Transaction-Based Calendar Time
Methodology. Most studies use daily buy-sell
imbalances for investor categories to test for return
predictability. However, Fama (1998) recommends
the usage of calendar time portfolios in analysis of
the stock price performance, primarily for the rea-
sons of avoiding the bad model problem and also
due to the advantage that cross-correlations of
event-firm abnormal returns are automatically
accounted for in the portfolio variance. The distribu-
tion of this estimator is better approximated by nor-
mal distribution, allowing classical
statistical inference.
We utilize the TBCT methodology proposed by
Seasholes and Zhu (2010) to calculate the returns at
various time horizons for each trader category. TBCT
methodology allows aggregation of the buy and sell
trades across trading participants and stocks into a
single return series per trader category. This single
time series of returns also eliminates the problem of
cross-sectional dependence. In addition, it allows
value-weighting of trades and thus eliminates any
volume or stock bias. The transactions-based

approach allows measurement of abnormal returns
under different holding periods (1 day, 1 week, 1
month, etc.) and investigation of whether there is
any time-sensitive information in stock trades.
There can be several issues with aggregating cat-
egory data, and they are addressed below.
 Individualistic fallacy: An attempt to impute
aggregate relationships from individual trader/
firm relationships. There could be a significant
loss by a trading entity which is offset by another
entity in the same trading category. At an aggre-
gate level, this shall not be factored into
the model.
 Small-stock bias: As noted in Seasholes and Zhu
(2010), small stock performance can bias the
returns significantly; hence, value-weighting the
transactions portfolios addresses the small-stock
bias. We segment stocks into large, medium, and
small to account for this issue.
 Heterogeneity in investor categories: Barber and
Odean (2000) report that the quintile of the most
active traders underperforms the quintile of the
least active traders by about 6% per year after
trading costs are taken into account. Active trad-
ers trade much higher volumes than less active
traders. The holding periods and profitability are
also highly varying. However, from the perspec-
tive of the investor category and trading technol-
ogy, these are homogenous.
The TBCT portfolio methodology requires aggregat-
ing the number of shares in each buy trade to a cal-
endar time “buy” portfolio and in each sell trade to a
calendar time “sell” portfolio. These shares are held
in the portfolio for a fixed amount of time and are
then removed from the portfolio’s holdings. The port-
folio’s value and composition change over time as
shares are added and removed from the holdings.
Seasholes and Zhu (2010) compute daily value-
weighted returns for portfolios by marking the hold-
ings and new trades to daily close prices. We follow
Fong, Gallagher, and Lee (2014) by using the actual
transaction price of a trade instead of the end-of-
the-day price in adding a trade to the portfolio. The
informed nature of aggregated trades of an investor
category over a calendar time horizon is measured by
the mean daily difference between returns on the
buy portfolio and returns on the sell portfolio.
In prior studies, TBCT has been used for assessing
both short- and long-lived information. TBCT has
been used from a horizon of 5 minutes to 1 year
Volume 79, Number 1
Trade Informativeness in Modern Markets
(Linnainmaa and Saar 2012; Seasholes and Zhu
2010). For all investor categories, we compute the
TBCT portfolio using a variety of horizons from the
end of the trading day to 125 trading days (6 months)
such that we can capture information of different
longevity associated with trades of all the investor
categories. In order to compute abnormal returns,
TBCT returns are regressed against the Fama-
French-Carhart four-factor model to assess the
trader category alpha.
Measuring Abnormal Returns. We measure
the abnormal returns using the Fama-French-Carhart
four-factor model. We estimate the following time
series models using daily data:
rp, t ¼ a þ bp ðrm, t  rf, t Þ þ bSMB SMBt
þ bHML HMLt þ bUMD UMDt þ et (1)
where r(p,t) is the TBCT return on the buys minus
sells portfolio of an investor category and r(m,t) is
returns on market portfolio, estimated as the value-
weighted portfolio of all the stocks involved in the
estimation of SMB, HML, and WML portfolios.8 r(f,t)
is computed using the 91-day T-bill rate. The four-
factor data series has been taken from Agarwalla,
Jacob, and Varma (2014). Our primary interest is in
the estimation of alphas (intercept) on the buy minus
sell portfolios of investor categories.
Empirical Results: TBCT across
Three Investor Categories
Overall Informativeness across the
Institutions, Proprietary, and Retail
Client. The TBCT returns of the three mutually
exclusive and exhaustive groups of traders—institu-
tions, proprietary, and retail client—are first com-
puted by subtracting the “sell” portfolio returns from
the “buy” portfolio returns for different investment
horizons: end of the day of the trade, 1 day, 5 days,
10 days, 20 days, 60 days, and 125 days (trading
days), using the method described in the previous
section. Thereafter, we estimate Equation (1) for
each of the groups and report the risk-adjusted alpha
for the different investor groups for each horizon.
Table 3 provides the details.
First, we find that trades by institutions generate an
alpha that is high at the trade-day horizon. The aver-
age value of the alpha at the close of the first day is
9.0 bp. This implies that if the trades are offset at the
end of the trade day, the portfolio of the buy trades
85
Financial Analysts Journal | A Publication of CFA Institute

and sell trades generates a net return of 9.0 bp after

–0.053 (–22.624) –0.037 (–4.666) –0.017 (–3.176) –0.011 (–3.059) –0.008 (–3.105) –0.005 (–2.725) –0.004 (–2.267)
0.004 (2.091)

intercept in the regression of the buys-minus-sells portfolio return on the Fama-French-Carhart factors (i.e., market, size, value factors, and momentum over stock returns over the
Exchange (NSE) of India for the period 2013–2015. It reports the daily alpha in percentage from horizon of the end of trade day (EOD) to 125 days. The alpha is computed as the
–0.001 (–2.009) –0.000 (–0.853)
risk adjustment. It is equivalent to an annual alpha of
125 Days

25.2%, which is economically significant. Similarly,

This table presents the transactions-based calendar time (TBCT) buys-minus-sells portfolio alphas for institutions, proprietary traders, and retail clients in the National Stock
retail-client trades generate an alpha of 5.3 bp,
which is equivalent to 12.4% annualized returns.
The alpha of the proprietary traders is economically
insignificant. However, as Fong, Gallagher, and Lee
0.008 (3.172)

(2014) note, the large positive (negative) alpha of the

institutions (or of the retail client) is hypothetical in
60 Days

the sense that we do not observe the actual horizon

of each trade, and it is quite unlikely one category of
traders liquidates all trades at the end of the day of
day. Barrot et al. (2016) report that individual invest-
ors do not reap the rewards from liquidity provision
0.015 (3.799)

because they experience a negative return on the

–0.001 (–1.557)

day of their trade. Thus, it is also important to note

20 Days
Table 3. TBCT Alphas for Institutions, Proprietary Traders, and Retail Clients for All Trades

the alpha at longer horizons.

Alphas computed at longer holding periods show

slightly lower trading informativeness. The positive
alphas of the institutional trades continue to remain
0.022 (3.739)

significant at 1-day, 5-day, 10-day, 20-day, and 60-

0.000 (0.047)

day horizons. If the portfolios of institutions are liqui-

10 Days

dated at the end of the 60-day period, then a daily

alpha of 0.7 bp or annualized returns of 2.02% can
be expected after risk adjustment. In contrast, we
past 12 months).  and  denote statistical significance at the 1% and 5% levels, respectively.

see negative and significant alphas for trades by retail

clients until the 60-day horizon. At the 60-day hori-
0.032 (3.783)

zon, a daily alpha of 0.5 bp or annualized returns of

0.001 (0.869)

1% are seen. The alphas generated by trades of

5 Days

proprietary investors are not significant for any hori-

zon except for 60 days, where it is 1 bp.

Overall, the results in Table 3 additionally support

two points. First, the trade of proprietary traders,
0.070 (5.150)

despite constituting a quarter of the exchange turn-

0.002 (0.909)

over, does not indicate trading informativeness.

1 Day

Perhaps, it reinforces that proprietary traders are

largely performing the uninformed market makers’
role in modern markets. Second, there is a remark-
able similarity between the magnitude of the alphas
of the institutions and retail-client group across all
0.090 (23.976)

horizons. Retail-client traders alphas reflect their

information disadvantage when they are trading
0.001 (0.626)
EOD

against institutional investors.

We consider the four-factor alpha as a measure of

informativeness. However, the alpha could also stem
from other factors such as investor ability to process
Category/Horizon

public information. Seasholes and Zhu (2010) link

informativeness to investors’ ability to exploit non-
Retail clients
Proprietary
Institutions

public information. Thus, we reran our analysis on a

sample of stocks with greater information asymmetry
for robustness purposes. Alphas generated in stocks
with high information asymmetry can be attributed

86

to investor ability to exploit non-public information
with greater certainty. We consider stocks that are
covered by no analyst across all three of our sample
years9 as a sample of high–information asymmetry
stocks. Low analyst coverage has been well used as a
proxy for high information asymmetry in the litera-
ture (Hong, Lim, and Stein 2000; Chae 2005; Kovacs
2010). The unreported results10 validate our main
results: Institutions (retail clients) have positive (nega-
tive) and significant alphas.
Informativeness for Liquidity-Demanding
and -Supplying Trades. Trading through market
and limit orders has different implications in trad-
itional microstructure research. While the informed
are expected to use more market orders (Easley et al.
1996), limit order traders face adverse selection and
free option risk. An overall informativeness analysis
may not state the full story, as there could be differ-
ences in the relative proportion of the market and
limit order usage by different groups. In this section,
following Fong, Gallagher, and Lee (2014), we segre-
gate the trades as market order trades and limit
order trades and focus on the informativeness of
three trading groups on both. To do the same, we
follow the same method of “buy” portfolio minus
“sell” portfolio returns for different horizons but
restrict ourselves first to only liquidity-demanding
trades of each group and then to only liquidity-sup-
plying trades. The results are provided in Table 4
(Panels A and B).
The liquidity-demanding trades of institutions are
informative, but the informativeness is weaker at the
longer horizon. The 20-day alpha is not significantly
different from zero. The highest alpha occurs at a
horizon of 1 day. In contrast, their liquidity-supplying
trades show greater informativeness. All the alphas
are significant at the 1% level from the end-of-day
horizon to the 60-day horizon. Numerically, at all
horizons, the alpha from the liquidity-supplying
trades is greater than the corresponding alphas from
the liquidity-demanding trades. For example, the
alpha at the close of the trading day is 3.4 (14.7) bp
for liquidity-demanding (supplying) trades, which is
equivalent to 8.87% (44.37%) annually.
The informativeness in liquidity-demanding and -sup-
plying trades of proprietary traders show quite a sig-
nificant contrast. In both long and short horizons, the
liquidity-supplying trades have positive alphas,
whereas the liquidity-demanding trades have nega-
tive alphas. The positive alphas of liquidity-supplying
trades are significant from the day of the trade to
the 20-day horizon. The significance of the negative
Volume 79, Number 1
Trade Informativeness in Modern Markets
alpha of liquidity-demanding trades extends till the
60-day horizon. The alphas from both the legs of
proprietary traders balance out to a near-zero overall,
which we observe in Table 3.
For the retail-client group, we found negative and
significant alphas for all horizons in Table 3. Here we
see that all the lack of informativeness of this group
stems from the liquidity-demanding trades. The
alphas are negative and significant from the end of
the day of trading to the 60-day horizon. In parallel,
the liquidity-supplying trades of the group have zero
informativeness. This particular result almost mirrors
the findings of Barber et al. (2009) on the inform-
ativeness of individual traders’ trades conducted
using aggressive and non-aggressive orders.
Reassuringly, regardless of whether we look at trades
conducted through limit orders or at trades con-
ducted through market orders, we find that institu-
tions are most informed, followed by the proprietary
group. The informativeness of the retail-client group
is the lowest under both setups. Thus, the main
result of Table 3 is not an artifact of the trading
through market or limit orders.
Interestingly, we find that aggressive trades have
lower informativeness than passive trades across all
horizons for all three trading groups. While surprising
at first glance, the results align with very recent lit-
erature. Several papers have found that market order
imbalance no longer is related to information trading
in both the long and short run. In examining a sample
of corporate information events, Kim and Stoll (2014)
find that order imbalances do not reveal private
information about posterior events. Collin-Dufresne
and Fos (2015) find that activist traders (who move
prices) make very limited use of market orders and
use them only at times of high available liquidity.
Easley, de Prado, and O’Hara (2016) find that market
order imbalance is not related to the very short-term
future bid-ask spread. O’Hara (2015), in her seminal
paper on contemporary microstructure, explains
these phenomena by suggesting that sophisticated
traders rely primarily on limit orders in modern algo-
rithmic markets. She suggests that the architecture of
the markets has changed so fundamentally after the
advent of machine traders that many of the trad-
itional microstructure models and assumptions are no
longer valid.
One economic explanation could be that the higher
alpha of liquidity-supplying trades is perhaps entirely
a compensation for liquidity provision. The limit
orders earn the half bid-ask spread, and the market
87
88
Table 4. TBCT Alphas for Institutions, Proprietary Traders, and Retail Clients: Liquidity-Demanding and Liquidity-
Supplying Trades
Panel A: Liquidity-demanding trades
Category/Horizon EOD 1 Day 5 Days 10 Days 20 Days 60 Days 125 Days
Financial Analysts Journal | A Publication of CFA Institute

Institutions 0.034 (5.865) 0.041 (3.513) 0.018 (1.978) 0.012 (2.186) 0.007 (1.817) 0.004 (1.629) 0.003 (1.648)
Proprietary –0.029 (–11.672) –0.017 (–5.941) –0.010 (–6.163) –0.006 (–5.118) –0.004 (–4.144) –0.001 (–2.246) 0.000 (–0.440)
Retail clients –0.118 (–28.892) –0.081 (–10.050) –0.040 (–6.958) –0.027 (–6.771) –0.015 (–5.769) –0.006 (–3.131) –0.003 (–1.756)
Panel B: Liquidity-supplying trades
Category/Horizon EOD 1 Day 5 Days 10 Days 20 Days 60 Days 125 Days

Institutions 0.147 (22.166) 0.097 (6.977) 0.043 (4.897) 0.030 (4.729) 0.019 (4.591) 0.008 (2.774) 0.003 (0.981)
Proprietary 0.031 (13.093) 0.021 (7.714) 0.012 (6.805) 0.006 (5.807) 0.002 (2.331) 0.000 (0.025) 0.000 (–0.471)
Retail client 0.002 (0.479) –0.001 (–0.120) 0.001 (0.157) 0.000 (–0.036) –0.001 (–0.438) –0.001 (–0.793) 0.000 (–0.238)

This table presents the transactions-based calendar time (TBCT) buys-minus-sells portfolio alphas for institutions, proprietary traders, and retail clients for liquidity-demanding trades
(Panel A) and liquidity-supplying trades (Panel B) in the National Stock Exchange (NSE) of India for the period 2013–2015. It reports the daily alpha in percentage from horizon of
the end of trade day (EOD) to 125 days. We refer to the side of an electronic limit order book trade that initiated the trade as a liquidity-demanding trade and the other side of the
trade as a liquidity-supplying trade. The alpha is computed as the intercept in the regression of the buys-minus-sells portfolio return on the Fama-French-Carhart factors (i.e., mar-
ket, size, value factors, and momentum over stock returns over the past 12 months).  and  denote statistical significance at the 1% and 5% levels, respectively.

orders pay the half spread. We estimate the spread
for each stock each day using the Roll (1984) model,
which utilizes the short-term return reversal in trade
prices. The average estimated spread for the top 50
stocks is 4.4 bp. For the next 450 stocks, it is
24.3 bp, and for the rest of the stocks, it is 100.4 bp.
Thus, an approximate value-weighted average bid-ask
spread is 16 bp. The end-of-day alpha of institutions
from liquidity-supplying trades is 147 bp, and that
from liquidity-demanding trades is 34 bp. From these
estimates, we find that while half-spread can explain
a part of the extra alpha of the liquidity-supplying
trades, the majority part of the alpha remains unex-
plained and thus can be attributed to informed trad-
ers’ relative use of more limit orders than market
orders in modern algorithmic markets.
Informativeness across Different Levels
of Market Capitalization. A stock’s informa-
tional environment can change with the size of the
stock, and thus the trading environment can also
consequently change (Fong, Gallagher, and Lee
2014). The NIFTY 50 stocks see higher coverage
from brokerage analysts and media. We saw in Table
2 that stocks that are part of the NIFTY 50 index
have higher participation from institutions, whereas
the smaller stocks have relatively greater participa-
tion from the retail-client group.
We next examine whether the informativeness of the
three groups of traders is different across the differ-
ent market capitalization groups. The overall TBCT
results are inherently value-weighted and thus may
not reflect the performance in smaller stocks well. As
discussed before, we divide the stocks into three
groups—the 50 stocks of NIFTY 50, the next 450
stocks, and the rest of the stocks—and for each
group of stocks, we compute the TBCT returns by
subtracting the sell portfolio returns from the buy
portfolio returns for all the time horizons. Table 5
provides the details. Panels A, B and C provide the
results of the first 50, the next 450, and the rest of
the stocks.
The institutions outperform in all market capitaliza-
tion categories. Their positive alpha is significant
across all the board. In the case of small stocks, the
alpha is higher and lasts up to the 125-day horizon.
The retail client loses in all market capitalization cate-
gories. The proprietary traders have positive and sig-
nificant alpha only in the small-stock category.
Trading Informativeness–Flow of
Information among Groups. To complement
the informativeness analysis using the TBCT method,
Volume 79, Number 1
Trade Informativeness in Modern Markets
we provide an additional analysis that allows us to
understand the flow of information in the market
among the three trading groups more clearly. For the
same, we use an analysis proposed by Stoffman
(2014), who used only two groups of traders—institu-
tions and households (or individuals)—and examined
how prices moved when one group buys from the
other and vice versa. To recognize how prices are
affected by the interaction of different groups of
investors in the market, we examine the relation
between stock returns and the proportion of trading
within and between each group of investors. We
estimate the regression:
ri, t ¼ bI, I ðI=IÞi, t þ bI, P ðI=PÞi, t þ bI, R ðI=RÞi, t
þ bP, I ðP=IÞi, t þ bP, P ðP=PÞi, t þ bP, R ðP=RÞi, t
þ bR, I ðR=IÞi, t þ bR, P ðR=PÞi, t þ bR, R ðR=RÞi, t
X
3 X
3
þ fk ri, tk þ  k RMi, tk
k¼1 k¼0
X
3
þ nk ABNVOLi, tk þ ei, t
k¼0
(2)
where ri, t denotes return for stock i on day t. The
explanatory variables represent the fraction of trad-
ing in stock i on date t that is accounted for by trad-
ers of type X buying from traders of type Z, and the
types “I,” “P,” and “R” indicate institutions, propri-
etary, and retail clients, respectively. Three lags of
the stocks’ daily return and the lags and contempor-
aneous values of stocks’ abnormal volume (the daily
volume divided by its trailing 20-day average) and
the market return are included as controls.
We run the regression for the three market capital-
ization groups of stocks separately, and the results
are reported in Table 6. As expected, within-group
trading does not explain returns, but between-group
trading does. For the top 50 stocks, we find that bI,R
is positive and significant (at the 1% level) and bR,I is
negative. Thus, when institutions buy from retail cli-
ents, prices move up, and when they sell to retail cli-
ents, prices move down (weakly). Similarly, when
proprietary traders buy from retail clients, prices
move up, and when they sell to retail clients, prices
go down (both significant at the 1% level). None of
the other main explanatory variables is significant,
including the trade variables between institutions and
proprietary.
For the next 450 stocks, we find that prices move up
when institutions buy from either retail-client or pro-
prietary groups, as bI,R and bI,P are both positive and
significant. We also find that when the proprietary
89
90
Table 5. TBCT Alphas for Investor Categories of Institutions, Proprietary Traders, and Retail Clients across Stock Categories of
Top 50, Next 450, and the Rest of Stocks
Category/
Horizon EOD 1 Day 5 Days 10 Days 20 Days 60 Days 125 Days

A: top
50 stocks
Institutional 0.048 (11.310) 0.040 (3.972) 0.021 (2.336) 0.017 (2.739) 0.011 (2.717) 0.005 (1.721) 0.002 (0.692)
Financial Analysts Journal | A Publication of CFA Institute

Proprietary 0.003 (1.109) 0.006 (1.514) 0.004 (2.267) 0.002 (1.549) 0.000 (–0.030) 0.000 (–0.633) 0.000 (0.138)
Retail client –0.045 (–14.762) –0.036 (–3.915) –0.021 (–2.327) –0.019 (–2.878) –0.012 (–2.785) –0.006 (1.900) –0.003 (–0.883)
B: next
450 stocks
Institutional 0.161 (25.956) 0.111 (5.717) 0.044 (4.076) 0.029 (4.004) 0.016 (3.566) 0.005 (2.297) 0.003 (1.656)
Proprietary –0.004 (–2.024) –0.004 (–2.055) –0.003 (–2.557) –0.003 (–3.287) –0.002 (–3.745) –0.001 (–3.212) 0.000 (–2.224)
Retail client –0.069 (–25.351) –0.046 (–5.163) –0.019 (–3.366) –0.012 (–3.445) –0.007 (–3.090) –0.002 (–2.059) –0.001 (–1.346)
C: the rest
of stocks
Institutional 0.213 (13.372) 0.117 (8.482) 0.044 (4.073) 0.028 (3.158) 0.020 (2.724) 0.017 (2.984) 0.016 (2.911)
Proprietary 0.025  (5.971) 0.011  (2.795) –0.002 (–0.666) 0.001 (0.372) 0.002 (0.833) 0.001 (0.538) –0.001 (1.358)
Retail client –0.048 (–15.928) –0.027 (–9.463) –0.011 (–4.350) –0.007 (–3.235) –0.004 (–2.207) –0.003 (–2.404) –0.003 (–2.284)

This table presents the transactions-based calendar-time (TBCT) buys-minus-sells portfolio alphas for institutions, proprietary traders, and retail clients across stock categories “top
50,” “next 450,” and “the rest” of stocks in the National Stock Exchange (NSE) of India for the period 2013–2015. It reports the daily alpha in percentage from horizon of the end
of trade day (EOD) to 125 days. We compute the alpha as the intercept in the regression of the buys-minus-sells portfolio return on the Fama-French-Carhart factors (i.e., market,
size, value factors, and momentum over stock returns over the past 12 months).  and  denote statistical significance at the 1% and 5% levels, respectively.
 Trade Informativeness in Modern Markets

Table 6. Flow of Information across Three Trading Groups

Top 50 Stocks Next 450 Stocks Rest of Stocks

Inst/Inst 0.021 0.047 0.100

Inst/Prop 0.008 0.040 0.019
Inst/Ret-Cl 0.068 0.014 0.000
Prop/Inst –0.024 –0.107 0.031
Prop/Prop 0.013 –0.055 0.046
Prop/Ret-Cl 0.176 0.095 0.009
Ret-Cl/Inst –0.037 0.021 0.000
Ret-Cl/Prop –0.148 0.033 –0.015
Ret-Cl/Ret-Cl 0.016 –0.020 0.000
Return(t-1) –0.002 –0.001 –0.002
Return(t-2) –0.002 –0.001 –0.002
Return(t-3) –0.001 –0.001 –0.001
AbVol 0.002 –0.025 –0.029
AbVol(t-1) 0.006 0.009 0.009
AbVol(t-2) –0.000 0.003 0.003
AbVol(t-3) –0.001 0.001 0.003
Market_ret 0.008 –0.002 0.008
Market_ret(t-1) 0.000 –0.000 0.001
Market_ret(t-2) –0.000 –0.002 –0.002
Market_ret(t-3) –0.002 0.006 0.008

The table provides an analysis of the flow of information among and within the three groups of trad-
ers: institutions (Inst), proprietary (Prop), and retail client (Ret-cl). The cells represent the regression
coefficient in estimating Equation (2). The dependent variable is the return of stock i on day t. The
explanatory variables represent the fraction of trading in stock i on date t that is accounted for by
traders of type X buying from traders of type Z (X/Z). Three lags of the stocks’ daily return and the
lags and contemporaneous value of stocks’ abnormal volume (the daily volume divided by the trailing
20-day average) and market return are included as controls. Our sample is formed using all stocks at
the National Stock Exchange (NSE) of India and stocks are divided according to market capitalization
into “top 50,” “next 450,” and “the rest” of stocks, The time period is 2013–2015.

group buys from the retail client, prices move up,
and when they buy from institutions, prices go down.
The fact that institutions are most informed, followed
by proprietary groups, is validated by the information
flow analysis of Stoffman (2014).11
Trading Informativeness of Institutions
Using Algo and Non-Algo Execution
Systems. The institutions, through their brokers, can
choose to use algorithmic or non-algorithmic execution.
In this section, we aim to see whether it beneficial for
the institutional traders to go for algorithmic execu-
tions. Institutional traders can have various methodolo-
gies and time horizons for executing and holding the
positions, for example, they could be large market cap
funds, diversified funds, or small-cap funds. These trad-
ers empanel brokers to execute their trades. These
brokers can use either algorithmic trading systems or
manual terminals manned by dealers to handle the exe-
cution. The funds may deem it appropriate to use dif-
ferent brokers for different sets of stocks based on
liquidity, size and value of trade, and brokerage fees. It
could very well be the case that an institution may split
an order for the same stock with two or more brokers,
one of whom uses an algorithmic system and the other
broker executes manually. We aim to see whether
there is any difference in alpha on the day of the trade
between algorithmic trades and manual trades of the
institution. We do not go on and check the difference
in informativeness across longer horizons since that will
not provide insight on the value added by the execu-
tion services (i.e., algorithmic versus manual).
Methodologically, we use again the Fama-French-
Carhart four-factor model. We estimate the following
time series models using daily data:
rpalgo, t  rpnonalgo, t ¼ a þ bp ðrm, t  rf, t Þ
þ bSMB SMBt þ bHML HMLt
þ bUMD UMDt þ et
(3)
where r(p-algo,t) and r(p-non-algo,t) are the returns on the
buys minus sells portfolios of the institutional algo
and non-algo investor categories.

Volume 79, Number 1 91

Financial Analysts Journal | A Publication of CFA Institute

Table 7. TBCT Difference between Alphas for Algo and Non-Algo Institutions on Trade Day
EOD EOD Liquidity-Demanding EOD Liquidity-Supplying

All stocks 0.020 (2.128) 0.076 (6.765) 0.010 (0.866)

NIFTY 50 stocks 0.016 (2.033) 0.033 (3.837) 0.009 (0.906)
Next 450 stocks 0.073 (5.899) 0.016 (8.455) 0.029 (1.498)
Rest of stocks 0.066 (1.607) 0.149 (3.504) 0.033 (0.792)

This table presents the transactions-based calendar-time (TBCT) alpha difference for algo and non-algo institutional trades in the
National Stock Exchange (NSE) of India for the period 2013–2015. The dependent variable in the “algo minus non-algo” alpha cal-
culation is the buys-minus-sells portfolio return of trades via algo minus the return of trades via non-algo. The alphas for TBCT
portfolios based on liquidity-demanding and liquidity-supplying trades are reported. We refer to the side of an electronic limit
order book trade that initiated the trade as a liquidity-demanding trade and the other side of the trade as a liquidity-supplying
trade. The daily alpha in percentage at the end of trade day (EOD) is reported.  and  denote statistical significance at the 1%
and 5% levels, respectively.

In Table 7, we provide the estimates of the alphas
and show that the end-of-day alphas for algorithmic
executions are greater than manual executions for
institutional trades.
To further probe the usefulness of the trading algo-
rithms, we examine their differential performance on
high-volume and high-volatility days. Those days, in
particular, are important for institutional investors as
there may be some news that requires them to
unwind a large position or some index reconstitution
that forces them to buy and sell large quantities
of stocks.
We thus create two sets of TBCT portfolios. One set
of portfolios is created with only stock-days trades
that have more than two times the 21-day rolling
volume. The second set of portfolios is created with
trades of stock days that have a difference between
the high and low price that is greater than 3% of the
open price. We then re-estimate Equation (3), but
only for these high-volume and high-volatility port-
folios. We see that the end-of-day difference in
alphas between algorithmic and manual execution is
positive and statistically significant on high-volume
and high-volatility days at the 5% and 1% levels,
respectively. What is more important, the magnitude
of the difference in the alpha is 8.6 bp (4.8 bp) for
high-volume (-volatility) days is more than four (two)
times the corresponding alpha in the full sample
(Table 7). Thus, we find that there is more value to
algorithmic executions for the institutional traders in
days of high volume and high volatility.
How do algorithms help create value over manual
executions? First, algorithms have the technology to
monitor markets in real time and do it without any
explicit cost. This advantage enables the algorithms
to find suitable window of trading opportunities and
choose the route of trade accordingly. Hendershott
and Riordan (2013) find that algorithms take liquidity
(trade via market orders) when it is cheap and supply
liquidity (trade via limit orders) when it is expensive,
while human traders do the opposite. We compare
days of extremely high volatility (top 20th percentile)
with days of extremely low volatility (bottom 20th
percentile). The bid-ask spreads are expected to be
very high during the former and very low during the
latter (Wyart et al. 2008). We define passivity ratio
as the number of shares traded as liquidity suppliers
over the total number of shares traded and find that
the passivity ratio of institutions using algorithms is
higher by 1.5% (significant at the 1% level) on days
of high volatility compared to days of low volatility.
When we see the manual executions of institutions,
the situation is completely the opposite. The passivity
ratio is 3.6% less (significant at a 1% level) on days
of high volatility compared to days of low volatility.
Thus, the ability to monitor the markets continuously
and switching between market and limit orders
depending on the market condition (depending on
whether liquidity is cheap or expensive) is one of the
value additions of executing through algorithms.
The other value-providing channel of algorithms is
that they are much less prone to cognitive biases.
One of the well-known characteristics of financial
markets is price clustering or trading at round num-
bers (Ahn, Cai, and Cheung 2005; Aitken et al. 1996).
The clustering tendency can create losses, and
sophisticated traders can devise profitable trading
strategies such as cluster undercutting to take advan-
tage of this phenomenon (Bhattacharya et al. 2012).
Davis, Van Ness, and Van Ness (2014) argue that
trades done through algorithms should see much less
clustering tendency than trades done manually. The
NSE has a uniform tick size of 5 paise or (` 0.05).
We define a trade as a clustered trade if the trade
price is of the form ` X.00 or ` X.50.12 We find that

92
 Trade Informativeness in Modern Markets

institutional manual trades have a clustering ten-

0.000 (7.060) 0.000 (8.063)

limit order book trade that initiated the trade as a liquidity-demanding trade and the other side of the trade as a liquidity-supplying trade. The daily alpha in percentage from horizon
India for the period 2013–2015. The alphas for TBCT portfolios based on liquidity-demanding and liquidity-supplying trades are also reported. We refer to the side of an electronic
0.001 (2.078)
Liquidity-demanding –0.017 (–6.942) –0.008 (–2.830) –0.005 (–3.374) –0.003 (–2.900) –0.002 (–3.373) –0.001 (–1.952) 0.000 (–0.091)

This table presents the transactions-based calendar-time (TBCT) alphas for proprietary algorithmic traders (high-frequency traders [HFTs]) in the National Stock Exchange (NSE) of
dency of 33%, whereas the institutional algorithmic

125 Days
trades have a clustering tendency of only 22% (with
the difference significant at 1% level). Thus, we see
that algorithmic executions provide value in terms of
reducing one of the behavioral biases of humans: the
tendency to cluster. There are several other cognitive
biases operating in financial markets that we are not

0.001 (1.191)
60 Days
able to examine in our study, and we sense that exe-
cuting through trading algorithms mitigates
those biases.

Trading Informativeness of Proprietary

Algorithmic Traders or HFTs. One of the

0.001 (7.213)

0.003 (2.890)
more intriguing facts in modern algorithmic markets

of the end of trade day (EOD) to 125 days is reported.  and  denote statistical significance at the 1% and 5% levels, respectively.
20 Days
is the presence of HFTs. With superior speed, HFTs
are often accused of having an unfair advantage in
the stock market over other participants. Academic
literature has argued that HFT adversely selects
other traders due to its speed advantage (Biais,

0.002 (7.401)

0.007 (5.512)
Foucault, and Moinas 2015) and its liquidity-demand-
ing orders look out for stale quotes (Van Kervel and

10 Days
Menkveld 2019) from other market participants while
its liquidity-supplying orders avoid the adverse selec-
tion themselves by quick modification or cancellation
(Hoffmann 2014).

In this section, we simply ask whether the HFT earns

Table 8. TBCT Alphas for Proprietary Algorithmic Traders (HFTs)

0.003 (7.376)

0.012 (6.969)
risk-adjusted trading alphas and whether their liquid-
5 Days

ity-demanding and liquidity-supplying orders primarily

contribute to the same. To accomplish the task, we
re-estimate Equation (1) again for all the horizons,
restricting ourselves to only HFTs’ trades (proprietary
traders with algorithmic trading). We already saw in
0.007 (6.077)

0.037 (15.232) 0.025 (9.083)

Table 3 that overall proprietary traders’ alphas are

not significantly different from zero. However, in
1 Day

Table 8, we find that HFTs’ trades show positive

risk-adjusted alpha that is significant until a horizon
of 10 days. Thus, HFTs’ trades are informative.

Next, we segregate the trades as either liquidity-

demanding or liquidity-supplying and compute the
0.007 (3.417)

HFT alphas for all horizons. We find that for all hori-
zons from the end of the trading day to 60 days,
EOD

HFTs’ liquidity-demanding trades have negative and

significant alphas. On the other hand, for all horizons
from the end of the trading day to 20 days, HFT
liquidity-supplying trades have positive and signifi-
cant alphas. Comparing the liquidity-supplying alpha
Liquidity-supplying
Category/Horizon

with the estimate of half-spread derived earlier, we

see that a significant part of HFTs’ liquidity-supplying
alpha cannot be explained by compensation for
All trades

trades

trades

liquidity provision. Our result here again goes back to

O’Hara’s (2015) argument that sophisticated traders
primarily use limit orders to trade.

Volume 79, Number 1 93

Financial Analysts Journal | A Publication of CFA Institute

Table 9. Intraday TBCT Alphas for Proprietary Algorithmic Traders (HFTs)

Category/Horizon EOM 1 Minute 5 Minutes 20 Minutes 60 Minutes

All Trades 0.0247 (404.544) 0.0113 (208.570) 0.0031 (84.337) 0.0006 (26.34) 0.0001 (8.887)

This table presents the intraday transactions-based calendar-time (TBCT) alphas for proprietary algorithmic traders (high-frequency
traders [HFTs]) in the National Stock Exchange (NSE) of India for the period 2013–2015. The alphas are calculated using the one-
factor market model. The alphas (expressed in percentages) are calculated at several horizons, starting at of the end of trade
minute (EOM) to 60 minutes post-trade.  and  denote statistical significance at the 1% and 5% levels, respectively.

To cap off our informativeness analysis of HFT, we
perform an intraday TBCT analysis for this group.
Since the essence of high-frequency trading is speed,
HFTs could be reversing their trades many times
within the day. Thus, a higher-frequency TBCT port-
folio may be more appropriate to track the inform-
ativeness of HFTs. What is not trivial is the
frequency at which the HFT activity should be
tracked. Given that the NSE is ranked fourth in the
world in terms of the number of transactions, HFT
activity in the NSE is quite intense. We choose a 1-
minute benchmark for this analysis, keeping the same
in mind. The method we follow is the same as
defined earlier, with the unit of analysis shifted from
trading day to trading minute; the holding periods
are now the end of the trading minute, 1 minute,
5 minutes, 10 minutes, 20 minutes, and 60 minutes.
Additionally, we compute the alpha from the one-fac-
tor market model, as the size and the book-to-market
and momentum factors are not expected to be sig-
nificant at high-frequency intervals. We also make
sure that no holding period continues over to the
next day.
Table 9 provides details. We find that the TBCT
alpha for the end-of-trading minute is positive and
highly significant at the 1% level. The TBCT alphas
go down in magnitude as we increase the holding
period; however, they remain highly significant at the
1% level throughout. The economic magnitude of the
end-of-minute TBCT alpha (2.4 bp) of HFTs is quite
high and should be interpreted with caution. Because
the actual horizon of the HFT holding period could
be a few milliseconds to few hours, it may be difficult
to state whether the HFT is actually able to realize
this alpha. However, the results no doubt reflect that
HFT trades are strongly informative intraday at vari-
ous horizons. Overall, our results in this section con-
tribute to the widely growing literature on HFT
informativeness (Baron et al. 2019; Brogaard,
Hendershott, and Riordan 2019).
Finally, we provide evidence on HFTs’ daily net clos-
ing position, which should indicate whether the HFT
plays the role of the modern market maker
(Menkveld 2013; Hagstro € mer and Norde
n 2013). If
the HFT acts as a voluntary market maker or liquidity
provider, then we should see HFT trading on both
sides of the market. To analyze the same, we focus
on the following quantity on a stock-day level:
ABS Scaled net volume
 
Shares boughtShares sold
¼ ABS
Shares bought þ Shares sold
If a trading group is trading on both sides of the mar-
ket and reversing their positions, then we should see
this quantity close to 0. However, if a trading group
is trading on one side of the market on a trading day,
then we should see this quantity close to 1. For each
trading group and mode of execution (algorithmic
and manual), we compute the stock-day average of
this quantity. Figure 2 provides the details. HFTs’
value of this variable is closest to 0 at 6.2%. Among
all trading groups, HFTs reverse their trades most
intraday and end their day at near-flat inventory,
atypical of market makers. Whether trading manually
or through algorithms, institutions primarily trade on
one side of the market on a day; their absolute
scaled net volume are respectively 70.5% and 67.9%,
very close to 1.
Robustness
Our analysis for this paper uses 3 years of data, from
2013 to 2015. Over this period, the Indian stock
market initially rose for the first 2 years and then fell
in the final year. In this section, we provide evidence
on whether our results could be influenced by the
market conditions during the sample period. We div-
ide the sample into two parts—1) the bull period
from 1st January 2013 to 28th February 2015, and
2) the bear period from 1st March 2013 to 31st

94

Figure 2. Average Absolute Scaled Volume
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
Instuons algo Instuons non-
algo
Absolute Scaled Net Volume
The figure illustrates the average values of the abs(shares bought  shares sold)/(shares bought þ shares sold) for each investor
type-execution type. Our sample includes all stocks traded at the National Stock Exchange (NSE) of India, and our sample period is
2013 to 2015.
December 2015. The NIFTY index increased from
5950.9 to 8901.9 in the first period and went down
to 7946.4 in the second period. We run the TBCT
analysis in both periods separately and find that our
results are qualitatively similar in both subperiods.13
Finally, we provide evidence that our results remain
unaffected if we use a one-factor market model or
the Fama-French three-factor model instead of the
four-factor model we use in our main results.14
Conclusion
We provide a comprehensive analysis of informative-
ness for three mutually exclusive and exhaustive
trading groups—institutions, proprietary, and non-
institutional non-proprietary (or retail-client) groups—
using 3-year complete order- and trade-level data for
all stocks trading on the NSE of India. Past studies of
informativeness were primarily done on data that
was non-comprehensive for the market, have mostly
focused on one or two of the trading groups, and
were done in non-algorithmic markets.
Our main finding is that institutions are informed,
and the retail-client group is uninformed. For all hori-
zons starting from the end of the day to 60 days, the
Volume 79, Number 1
Trade Informativeness in Modern Markets
HFT Prop non-algo Retail-Client algo Retail non-algo
institutional trades show a positive and significant
risk-adjusted alpha. Similarly, the retail-client group
shows a negative and significant risk-adjusted alpha
for similar horizons. The proprietary traders have an
insignificant alpha. Further, we find that the rank
ordering of alphas of the three groups of traders
remains valid when either we look at only limit-order
trades or we look at only market-order trades. More
interestingly, the alpha from the limit-order trades
exceeds the alpha for market-order trades for all
trading groups (even after controlling for the bid-ask
spread), suggesting that informed traders make use
of limit orders more than market orders.
Finally, we make use of the exchange-provided algo-
rithmic flags to do two additional investigations. First,
we see whether algorithmic executions carry any
value over manual execution for institutional traders.
We do a difference in TBCT portfolio returns on the
day of the trade between institutional algorithmic
trades and institutional manual trades and find that
the alpha is positive and significant (even more so on
days of high volume and high volatility), suggesting a
non-trivial value provided by algorithmic executions.
Second, we separately examine the profitability of
proprietary algorithmic traders or HFTs. We find that
over all horizons starting from the day of the trade
95
Financial Analysts Journal | A Publication of CFA Institute
to 10 days, HFT alphas are positive and significant.
We also find that the HFT alpha coming from the
liquidity-supplying trades is positive and significant,
while that coming from liquidity-demanding trades is
negative and significant.
An important implication of our results for retail trad-
ers is that they should be patient and trade using
limit orders. On the whole, they have an information
disadvantage in the market, and this disadvantage
gets attenuated when they use limit orders to trade
as they receive compensation for liquidity provision.
An implication for institutions is that they should use
Editor's Note
Submitted 22 October 2021
Accepted 15 September 2022 by William N. Goetzmann
Notes
1. World Federation of Exchanges 2020 report.
2. Alphas are risk adjusted (Fama-French-Carhart
four factor).
3. The direct market access was provided in 2010.
4. https://www.world-exchanges.org/our-work/articles/
2020-annual-statistics-guide.
5. Colocation services amount to renting rack space in the
premises of the exchange itself to reduce latency, i.e., to
reduce the time gap between an order being placed and it
actually reaching the exchange.
6. The results are available on request from the authors.
7. Odean (1998), Odean (1999), and Barber and Odean
(2000) use trading account data from a discount
brokerage house.
8. SMB is the Fama-French size factor constructed by
longing a portfolio of small firms and shorting a portfolio
of big firms. HML is the Fama-French value factor
constructed by longing a portfolio of high-value firms and
shorting a portfolio of low-value firms. UMD is the
References
Agarwalla, S. K., J. Jacob, and J. R. Varma. 2014. “Four
Factor Model in Indian Equities Market.” Indian
Institute of Management, Ahmedabad Working Paper
(2013-09), 05.
Ahn, H. J., J. Cai, and Y. L. Cheung. 2005. “Price Clustering on
the Limit-Order Book: Evidence from the Stock Exchange of
Hong Kong.” Journal of Financial Markets 8 (4): 421–51. doi:10.
1016/j.finmar.2005.07.001.
96
broker execution algorithms. We find that algorithmic
executions provide value to institutions, especially
during high-volume and high-volatility days. The
source of this alpha comes from the trading algo-
rithm’s ability to monitor the market continuously at
a near-zero cost and the fact that algorithms are
much less affected by behavioral bias compared to
humans. As a limitation, we believe that the data
period in our study is perhaps on the shorter side.
Future research should attempt a more robust study
that uses data for a much longer period, subject to
the availability of the data.
momentum factor constructed by subtracting the returns
of a portfolio of losing stocks from the returns of a
portfolio of winning stocks.
9. As an alternative, we also create a sample of stocks with
fewer than three analysts covering in all the 3 years. The
results are qualitatively similar.
10. Available from the authors.
11. Finally, for the group of the rest of stocks, we do not
see any major information flow from this method.
Table 2 suggests that the retail client contributes
almost 75% of the turnover in this group of smallest
stocks. Such a skew in the distribution perhaps makes
this analysis on the group of the rest of stocks
incomprehensible.
12. Thus, the trades of the form (` X.05, ` X.10, ` X.15, … , `
X.45, ` X.55, ` X.60, … ., ` X.90, ` X.95) are termed as
non-clustered.
13. Unreported, available from the authors.
14. Unreported, available from the authors.
Aitken, M., P. Brown, C. Buckland, H. Y. Izan, and T. Walter.
1996. “Price Clustering on the Australian Stock Exchange.”
Pacific-Basin Finance Journal 4 (2–3): 297–314. doi:10.1016/
0927-538X(96)00016-9.
Angel, J. 2021. “Gamestonk: What Happened and What to do
About It.” Available at SSRN 3782195.
Banerjee, A., and P. Roy. 2022. "Are High Frequency Traders
Really Market-Makers." SSRN working paper.

Barber, B. M., Y. T. Lee, Y. J. Liu, and T. Odean. 2009. “Just
How Much Do Individual Investors Lose by Trading?” Review of
Financial Studies 22 (2): 609–32. doi:10.1093/rfs/hhn046.
Barber, B. M., and T. Odean. 2000. “Trading is Hazardous to
Your Wealth: The Common Stock Investment Performance of
Individual Investors.” The Journal of Finance 55 (2): 773–806.
doi:10.1111/0022-1082.00226.
Baron, M., J. Brogaard, B. Hagstro€ mer, and A. Kirilenko. 2019.
“Risk and Return in High-Frequency Trading.” Journal of
Financial and Quantitative Analysis 54 (3): 993–1024. doi:10.
1017/S0022109018001096.
Barrot, J. N., R. Kaniel, and D. Sraer. 2016. “Are Retail Traders
Compensated for Providing Liquidity?” Journal of Financial
Economics 120 (1): 146–68. doi:10.1016/j.jfineco.2016.01.005.
Bessembinder, H., M. Panayides, and K. Venkataraman. 2009.
“Hidden Liquidity: An Analysis of Order Exposure Strategies in
Electronic Stock Markets.” Journal of Financial Economics 94 (3):
361–83. doi:10.1016/j.jfineco.2009.02.001.
Bhattacharya, U., C. W. Holden, and S. Jacobsen. 2012. “Penny
Wise, Dollar Foolish: Buy–Sell Imbalances on and around
Round Numbers.” Management Science 58 (2): 413–31. doi:10.
1287/mnsc.1110.1364.
Biais, B., and T. Foucault. 2014. “HFT and Market Quality.”
Bankers, Markets & Investors 2014 (128): 5–19.
Biais, B., T. Foucault, and S. Moinas. 2015. “Equilibrium Fast
Trading.” Journal of Financial Economics 116 (2): 292–313. doi:
10.1016/j.jfineco.2015.03.004.
Boehmer, E., K. Fong, and J. J. Wu. 2021. “Algorithmic Trading
and Market Quality: International Evidence.” Journal of Financial
and Quantitative Analysis 56 (8): 2659–88. doi:10.1017/
S0022109020000782.
Boehmer, E., C. M. Jones, X. Zhang, and X. Zhang. 2021.
“Tracking Retail Investor Activity.” The Journal of Finance 76 (5):
2249–305. doi:10.1111/jofi.13033.
Brogaard, J., A. Carrion, T. Moyaert, R. Riordan, A. Shkilko, and
K. Sokolov. 2018. “High Frequency Trading and Extreme Price
Movements.” Journal of Financial Economics 128 (2): 253–65.
doi:10.1016/j.jfineco.2018.02.002.
Brogaard, J., T. Hendershott, and R. Riordan. 2019. “Price
Discovery without Trading: Evidence from Limit Orders.” The
Journal of Finance 74 (4): 1621–58. doi:10.1111/jofi.12769.
Chae, J. 2005. “Trading Volume, Information Asymmetry, and
Timing Information.” The Journal of Finance 60 (1): 413–42. doi:
10.1111/j.1540-6261.2005.00734.x.
Chakrabarty, B., T. Hendershott, S. Nawn, and R. Pascual.
2020. “Order Exposure in High Frequency Markets.” Available
at SSRN 3074049.
Chakravarty, S. 2001. “Stealth-Trading: Which Traders’ Trades
Move Stock Prices?” Journal of Financial Economics 61 (2):
289–307. doi:10.1016/S0304-405X(01)00063-0.
Collin-Dufresne, P., and V. Fos. 2015. “Do Prices Reveal the
Presence of Informed Trading?” The Journal of Finance 70 (4):
1555–82. doi:10.1111/jofi.12260.
Volume 79, Number 1
Trade Informativeness in Modern Markets
Copeland, T. E., and D. Galai. 1983. “Information Effects on the
Bid-Ask Spread.” The Journal of Finance 38 (5): 1457–69. doi:
10.1111/j.1540-6261.1983.tb03834.x.
Davis, R. L., B. F. Van Ness, and R. A. Van Ness. 2014.
“Clustering of Trade Prices by High-Frequency and Non–High-
Frequency Trading Firms.” Financial Review 49 (2): 421–33. doi:
10.1111/fire.12042.
Domowitz, I., and H. Yegerman. 2005. “The Cost of Algorithmic
Trading: A First Look at Comparative Performance.” The Journal
of Trading 1 (1): 33–42. doi:10.3905/jot.2006.609174.
Easley, D., M. L. de Prado, and M. O’Hara. 2016. “Discerning
Information from Trade Data.” Journal of Financial Economics
120 (2): 269–85. doi:10.1016/j.jfineco.2016.01.018.
Easley, D., N. M. Kiefer, M. O’Hara, and J. B. Paperman. 1996.
“Liquidity, Information, and Infrequently Traded Stocks.” The
Journal of Finance 51 (4): 1405–36. doi:10.1111/j.1540-6261.
1996.tb04074.x.
Engen, E. M., and A. Lehnert. 2000. “Mutual Funds and the US
Equity Market.” Fed. Res. Bull 86:797.
Fama, E. F. 1998. “Market Efficiency, Long-Term Returns, and
Behavioral Finance.” Journal of Financial Economics 49 (3):
283–306. doi:10.1016/S0304-405X(98)00026-9.
Fong, K. Y., D. R. Gallagher, and A. D. Lee. 2014. “Individual
Investors and Broker Types.” Journal of Financial and
Quantitative Analysis 49 (2): 431–51. doi:10.1017/
S0022109014000349.
€ mer, B., and L. Norde
Hagstro n. 2013. “The Diversity of High-
Frequency Traders.” Journal of Financial Markets 16 (4):
741–70. doi:10.1016/j.finmar.2013.05.009.
Handa, P., R. A. Schwartz, and A. Tiwari. 1998. “The Ecology of
an Order-Driven Market.” The Journal of Portfolio Management
24 (2): 47–55. doi:10.3905/jpm.24.2.47.
Hendershott, T., C. M. Jones, and A. J. Menkveld. 2011. “Does
Algorithmic Trading Improve Liquidity?” The Journal of Finance
66 (1): 1–33. doi:10.1111/j.1540-6261.2010.01624.x.
€rhoff. 2015. “Are
Hendershott, T., D. Livdan, and N. Schu
Institutions Informed about News?” Journal of Financial
Economics 117 (2): 249–87. doi:10.1016/j.jfineco.2015.03.007.
Hendershott, T., and R. Riordan. 2013. “Algorithmic Trading
and the Market for Liquidity.” Journal of Financial and
Quantitative Analysis 48 (4): 1001–24. doi:10.1017/
S0022109013000471.
Hoffmann, P. 2014. “A Dynamic Limit Order Market with Fast
and Slow Traders.” Journal of Financial Economics 113 (1):
156–69. doi:10.1016/j.jfineco.2014.04.002.
Hong, H., T. Lim, and J. C. Stein. 2000. “Bad News Travels
Slowly: Size, Analyst Coverage, and the Profitability of
Momentum Strategies.” The Journal of Finance 55 (1): 265–95.
doi:10.1111/0022-1082.00206.
Jørgensen, K., J. Skjeltorp, and B. A. Ødegaard. 2018.
“Throttling Hyperactive Robots–Order-to-Trade Ratios at the
Oslo Stock Exchange.” Journal of Financial Markets 37:1–16.
doi:10.1016/j.finmar.2017.09.001.
Kaniel, R., S. Liu, G. Saar, and S. Titman. 2012. “Individual
Investor Trading and Return Patterns around Earnings
97
Financial Analysts Journal | A Publication of CFA Institute
Announcements.” The Journal of Finance 67 (2): 639–80. doi:10.
1111/j.1540-6261.2012.01727.x.
Kaniel, R., G. Saar, and S. Titman. 2008. “Individual Investor
Trading and Stock Returns.” The Journal of Finance 63 (1):
273–310. doi:10.1111/j.1540-6261.2008.01316.x.
Kelley, E. K., and P. C. Tetlock. 2013. “How Wise Are Crowds?
Insights from Retail Orders and Stock Returns.” The Journal of
Finance 68 (3): 1229–65. doi:10.1111/jofi.12028.
Kim, S. T., and H. R. Stoll. 2014. “Are Trading Imbalances
Indicative of Private Information?” Journal of Financial Markets
20:151–74. doi:10.1016/j.finmar.2014.03.003.
Kovacs, T. 2010. “Equity Issues and Temporal Variation in
Information Asymmetry.” Journal of Banking & Finance 34 (1):
12–23.
Linnainmaa, J. T., and G. Saar. 2012. “Lack of Anonymity and
the Inference from Order Flow.” The Review of Financial Studies
25 (5): 1414–56.
Menkveld, A. J. 2013. “High Frequency Trading and the New
Market Makers.” Journal of Financial Markets 16 (4): 712–40.
doi:10.1016/j.finmar.2013.06.006.
Nawn, S., and A. Banerjee. 2019. “Do Proprietary
Algorithmic Traders Withdraw Liquidity during Market
Stress?” Financial Management 48 (2): 641–76. doi:10.1111/
fima.12238.
Odean, T. 1998. “Are Investors Reluctant to Realize Their
Losses?” The Journal of Finance 53 (5): 1775–98. doi:10.1111/
0022-1082.00072.
Odean, T. 1999. “Do Investors Trade Too Much?” American
Economic Review 89 (5): 1279–98. doi:10.1257/aer.89.5.
1279.
O’Hara, M. 1995. Market Microstructure Theory. New York:
Wiley.
98
O’Hara, M. 2015. “High Frequency Market Microstructure.”
Journal of Financial Economics 116 (2): 257–70. doi:10.1016/j.
jfineco.2015.01.003.
Roll, R. 1984. “A Simple Implicit Measure of the Effective Bid-
Ask Spread in an Efficient Market.” The Journal of Finance 39
(4): 1127–39. doi:10.1111/j.1540-6261.1984.tb03897.x.
Seasholes, M. S., and N. Zhu. 2010. “Individual Investors and
Local Bias.” The Journal of Finance 65 (5): 1987–2010. doi:10.
1111/j.1540-6261.2010.01600.x.
Securities and Exchange Commission. 2010. “Concept Release
on Market Structure.”
Sias, R. W., and L. T. Starks. 1997. “Return Autocorrelation and
Institutional Investors.” Journal of Financial Economics 46 (1):
103–31. doi:10.1016/S0304-405X(97)00026-3.
Stoffman, N. 2014. “Who Trades with Whom? Individuals,
Institutions, and Returns.” Journal of Financial Markets 21:
50–75. doi:10.1016/j.finmar.2014.08.002.
Tian, X., B. Do, H. N. Duong, and P. S. Kalev. 2015. “Liquidity
Provision and Informed Trading by Individual Investors.” Pacific-
Basin Finance Journal 35:143–62. doi:10.1016/j.pacfin.2014.11.005.
van der Beck, P., and C. Jaunin. 2021. “The Equity Market
Implications of The Retail Investment Boom.” Available at SSRN
3776421.
Van Kervel, V., and A. J. Menkveld. 2019. “High-Frequency
Trading around Large Institutional Orders.” The Journal of
Finance 74 (3): 1091–137. doi:10.1111/jofi.12759.
Welch, I. 2022. “The Wisdom of the Robinhood Crowd.”
Journal of Finance 77 (3): 1489–1527. doi:10.1111/jofi.13128.
Wyart, M., J. P. Bouchaud, J. Kockelkoren, M. Potters, and M.
Vettorazzo. 2008. “Relation between Bid–Ask Spread, Impact
and Volatility in Order-Driven Markets.” Quantitative Finance 8
(1): 41–57. doi:10.1080/14697680701344515.
Financial Analysts Journal | A Publication of CFA Institute
https://doi.org/10.1080/0015198X.2022.2100234
Option Pricing via
Breakeven Volatility
Blair Hull, Anlong Li, and Xiao Qiao
Blair Hull is a founder and chairman of Hull Tactical Asset Allocation, LLC, Chicago, IL. Anlong Li is a senior financial engineer at Hull
Tactical Asset Allocation, LLC, Chicago, IL. Xiao Qiao is an assistant professor at the City University of Hong Kong and Hong Kong Institute
for Data Science, Hong Kong, China. Send correspondence to Xiao Qiao at xiaoqiao@cityu.edu.hk.
The fair value of an option is given
by breakeven volatility, the value of
implied volatility that sets the profit
and loss of a delta-hedged option to
zero. We calculate breakeven vola-
tility for 400,000 options on the
T Cox, Ross, and Rubinstein 1979) launched a revolution in modern
S&P 500 and build a predictive
model for these volatilities. A two-
stage regression approach captures
the majority of the observed vari-
ation. By providing a link between
option characteristics and breakeven
volatility, we establish a non-para-
metric approach to pricing options
without the need to specify the
underlying price process. We illus-
trate the economic value of our
approach with a simulated trading
strategy based on breakeven volatil-
ity predictions.
Keywords: options; prediction;
trading strategy; volatility
Disclosure: The authors report no funding
sources. The authors alone are responsible
for the content and writing of the paper.
Hull Tactical Asset Allocation, LLC is a
registered investment adviser that deploys
strategies, such as the one described in
this paper in investment products including
separately managed accounts and an
exchange-traded fund. All errors are our
own.
PL Credits: 2.0
Volume 79, Number 1
Research
Introduction
he advent of option pricing models (Black and Scholes 1973;
finance that saw explosive growth in options markets and mas-
sive interest in option pricing in the subsequent decades. Although
the option pricing literature contains a multitude of models, the
approaches often share a common idea: The value of an option is equal
to that of a replication portfolio with the same eventual payoffs
(Merton 1973). The most natural implementation of this simple and ele-
gant idea is to build a distribution of outcomes from a large set of repli-
cation portfolios using historical data, which can then be used to price
additional options. However, the implementation of this approach turns
out to be fraught with issues. Zou and Derman (1999) point out that
such replication can be time-consuming, and hedging errors due to
inaccurate volatility forecasts or infrequent hedging make the exercise
difficult. As a result, most researchers opt for a more structural
approach: Specify a return process for the underlying asset, derive the
pricing relationships, then calibrate the model to data.
We set out to build a non-parametric option pricing model based on
Merton’s (1973) original insight. In the past 20 years, computing power
has grown rapidly and option databases have become more compre-
hensive. These developments enable us to overcome the limitations
cited in Zou and Derman (1999). The first step towards an option pric-
ing model is to build a measure for the appropriate value of options.
Using historical prices, we calculate the fair value of an option, the
value of implied volatility (IV) that sets the profit and loss of a delta-
hedged option position to zero. This value is called breakeven volatil-
ity (BEV).
The profit and loss from a dynamically hedged option position recover
the difference between implied volatility and realized volatility, which
indicates a
The authors thank William Goetzmann (Executive Editor), Luis Garcia-Feijoo
(Managing Editor), Daniel Giamouridis (Co-Editor), two anonymous referees, Petra
Bakosova, Yeguang Chi, Jake DeSantis, Ruslan Goyenko, Steve Heston, Ray
Iwanowski, Craig Iseli, Kris Jacobs, Jason Wei, Lei Wu, Qi Wu, and Zonghao Yang
for helpful comments. We are especially grateful to Ben Albert for extensive discus-
sions and data processing.
© 2022 CFA Institute. All rights reserved. 99
Financial Analysts Journal | A Publication of CFA Institute
positive or negative risk premium (Bakshi, Cao, and
Chen 1997; Christoffersen et al. 2018). In computing
breakeven volatility, we set the profit and loss to
zero, removing risk premia embedded in market pri-
ces. To the extent market implied volatility of an
option differs from its breakeven volatility, the mar-
ket price contains either a risk premium or a dis-
count. If the IV were higher than the BEV, a delta-
hedged short position in this option would result in a
positive profit. If the IV were lower than the BEV, a
delta-hedged long position would show a positive
profit. In this sense, breakeven volatility provides the
fair value of an option to both sides of the contract
(Dupire 2006).
We build a large set of breakeven volatility values
for options on the S&P 500 Index (SPX), and we
construct a predictive model that connects option
characteristics to BEV. By providing a link
between moneyness, time to expiration, and
other observable characteristics and the fair
value of the option, the predictive relation can be
interpreted as a non-parametric option pricing
model.1 Given a new option with its own set of
characteristics, we can quickly determine its
fair value.
Parametric approaches are favored over non-para-
metric ones when the underlying asset’s return
dynamics are known, but this is rarely the case in
practice. The non-parametric approach to option
pricing can be data-intensive, but it offers a promis-
ing alternative to standard parametric pricing models
when parametric restrictions are violated. Since
non-parametric models do not rely on restrictive
assumptions, such as log-normality or sample-path
continuity, they are robust to the specification
errors that plague parametric models (Ait-Sahalia
and Lo 1998).
We construct delta-hedged positions for nearly
400,000 S&P 500 options with a variety of strike
prices and time to expiration ranging from 5 to 74
trading days. The empirical distribution of breakeven
volatility is similar to, but distinct from, that of
implied volatility. We construct volatility smirks for
both measures of volatility, and we uncover the fol-
lowing pattern: Breakeven and implied volatility are
almost identical in the at-the-money (ATM) region,
but as we move further away from the ATM strike
prices, the two measures diverge. Implied volatility
is higher than breakeven volatility for strike prices
lower than the underlying price, as well as for
strikes exceeding the underlying. The shapes of the
volatility smirks indicate that delta-hedged short
100
positions tend to earn positive profits for out-of-
the-money (OTM) puts, an empirical regularity con-
firmed in the literature (Coval and Shumway 2001;
Bakshi and Kapadia 2003).
We build a predictive model of breakeven volatility
using 12 variables and their transformations. In
addition to typical parameters used in option
pricing, such as moneyness and time to expiration,
we also include variables that have the potential to
bring added predictive power, such as the CBOE
VIX Volatility Index (VVIX), the difference between
the delta of the option and the ATM delta, and the
sensitivity of the option price with respect to vola-
tility. We use a two-stage regression approach to
build a statistical model that connects the predictor
variables and breakeven volatility. The first stage
produces forecasts of logged breakeven volatility,
and the second stage makes an adjustment such
that the predicted breakeven volatility values
are unbiased.
Our statistical model provides a reliable fit to break-
even volatility. The two-stage model captures 90%
of the variation in breakeven volatility. Model diag-
nostics reveal that the model performs best
between volatility values of 5 and 50%, but has dif-
ficulties if breakeven volatility is below 5%. Our
model predictions exhibit convexity in volatility
space, and option prices are consistent with an
underlying return distribution that exhibits negative
skew and excess kurtosis. While the above features
can be carefully engineered into a parametric model,
they arise naturally in our approach as we learn
from data.2
We test the economic value of our statistical model
through a simulated trading strategy that exploits the
difference between implied volatility and our forecast
of breakeven volatility. For a particular option, if its
implied volatility were lower than the predicted BEV,
the market price of the option is too low, and we
take a delta-hedged long position. If the implied vola-
tility were higher than the predicted BEV, the market
price is too high, so we take a delta-hedged short
position. We trade options whose breakeven volatil-
ity values differ more than $1 compared to the mid-
point price.
In our testing period from January 2015 to
November 2020, our statistical model achieves an
impressive out-of-sample R-squared of 0.71. The
BEV-based strategy identifies 6,783 trading oppor-
tunities, earning 6.8% per year with an annual vola-
tility of 4.5%. With a Sharpe ratio of 1.50, this

strategy outperforms three alternative options strat-
egies in terms of expected returns per unit of risk.
The BEV strategy also provides a more attractive
risk-return tradeoff compared to a buy-and-hold
strategy in the S&P 500, which has a Sharpe ratio of
0.79 in our sample period.
We are not the first to use historical data to com-
pute terminal distributions and option prices.
Stutzer (1996) and Zou and Derman (1999) also use
historical data to compute option prices. Rather
than looking at the replicated value of an option,
these papers compute a risk-neutral distribution
from the empirical distribution of the underlying
asset. Stutzer (1996) notes the shortcoming of not
incorporating market information related to option
prices into his approach, stating “there is need for a
more detailed comparison of … non-parametric
model values to actual transaction prices.” We
answer Stutzer’s call to action by using actual
transaction prices to inform us about the fair
values of options. Furthermore, Zou and Derman’s
(1999) goal is relative pricing: Take exchange-
traded option prices as given, build a pricing
model, then calculate the prices of less liquid, over-
the-counter options. In contrast, our goal is
absolute pricing—to determine the fair value of indi-
vidual options based on their observable
characteristics.
Some papers have used breakeven volatility as a
diagnostic tool for market prices. Dupire (2006)
builds breakeven volatility surfaces and explains that
BEV is a fair volatility for both sides of an option
contract. He highlights using BEV surfaces as a tool
to understand market volatility surfaces. Mitoulis
(2019) carries out breakeven volatility calculations
using simulations, comparing delta-hedging results
using Black and Scholes (1973) and Heston (1993)
models. Neither paper attempts to forecast BEV val-
ues or explore trading implications of the differences
between implied and breakeven volatilities. As far as
we know, we are the first to build a non-parametric
option pricing model via a predictive model of break-
even volatility.
A crucial distinction between our paper and the
existing literature is the calculation of breakeven
volatility. Zou and Derman (1999), Dupire (2006),
and Mitoulis (2019) all construct a constant BEV
throughout the life of an option of a given strike
price and time to maturity. This calculation requires
perfect foresight of the path of the option and is
not feasible in real-time. Additionally, the resulting
BEV surfaces often admit arbitrage. Our approach
Volume 79, Number 1
Option Pricing via Breakeven Volatility
relaxes the constant breakeven volatility assumption
and allows for a term structure of BEV, which is
then solved by backward recursion. In doing so, we
use the implied volatility path as well as the stock
price path for the calculation. Breakeven volatility
values calculated from our approach do not allow
for arbitrage.
Our paper is also related to the literature on non-
parametric approaches to option pricing. A strand of
literature first developed in the 1990s. Hutchinson,
Lo, and Poggio (1994) use neural networks to
approximate the Black-Scholes formula. Ait-Sahalia
and Lo (1998) propose a kernel density estimator of
the risk-neutral distribution, effectively providing a
link between option characteristics and the second
derivative of the volatility smirk. These pioneering
papers led to more recent literature on using
machine learning methods to approximate the rela-
tionship between model inputs and outputs. Liu,
Oosterlee, and Bohte (2019) propose a neural net-
work model to approximate the Black-Scholes and
Heston (1993) models. Manzo and Qiao (2021) use
neural networks to approximate nine different credit
risk models. Compared to these papers, we do not
try to approximate existing option pricing models—
we let the data dictate how option characteristics
and prices are related. Furthermore, our target option
price is based on a fair valuation calculation rather
than the market price.
Breakeven Volatility
Data. We obtain option prices, implied volatility,
and deltas from SpiderRock Platform Services LLC.
SpiderRock is an options trading platform provider
and data vendor based in Chicago that serves trad-
ing desks at large banks, hedge funds, and propri-
etary trading firms. SpiderRock’s historical data and
option analytics include implied volatility, option
Greeks, risk metrics, and volatility surfaces derived
from the live data from SpiderRock’s trad-
ing systems.
Since OptionMetrics is commonly used in academic
studies, we would like to point out two differences
between SpiderRock and OptionMetrics data (Please
refer to Appendix A for more details). First, daily
options and underlying prices are recorded differ-
ently across the two databases. OptionMetrics uses
option prices at 3:59 p.m. Eastern Time and stock
settlement prices at market close at 4 p.m., whereas
SpiderRock uses the same option price at 3:59 p.m.
101
Financial Analysts Journal | A Publication of CFA Institute
but also uses stock prices at 3:59 p.m. Second,
SpiderRock data imposes put-call parity whereas
OptionMetrics maintain separate volatility surfaces
for puts and calls. Even if OptionMetrics were to
synchronize the option and stock quotes, one may
still prefer using forward prices derived from put-call
parity in implied volatility calculation.3 For these rea-
sons, we choose to use SpiderRock data. These data
are used for the computation and prediction of
breakeven volatility values in an attempt to use real-
time market data that traders rely on to create trad-
ing signals.
Variables used to predict breakeven volatility are
obtained from SpiderRock and Bloomberg. In particu-
lar, the strike price, days until expiration, implied
volatility, and other information related to the under-
lying are from SpiderRock. The CBOE Volatility Index
(VIX), the CBOE VIX Volatility Index, and variables
used to construct a realized volatility forecast of the
S&P 500 are from Bloomberg. Our sample is from
January 2013 to November 2020.
Calculation of Breakeven Volatility. The
breakeven volatility is the value that sets the profit
and loss from a delta-hedged option position to be
zero. Suppose we want to calculate the breakeven
volatility for a call option. The delta-hedged profit
and loss are given as follows:4
X
T
PnLt ðrÞ ¼ cT  ct  Ds1 ðSs  Ss1 Þ
s¼1
where cT is the payoff of the call option at maturity,
ct is the call price at the time of initiating the
position, Ds is the option delta at time s, and Ss is
the price of the underlying asset at time s: The
profit and loss of the delta-hedged option
depend on the volatility of the underlying. By
changing the volatility, we can trace out PnLt ðrÞ as
a function of volatility. The breakeven volatility is
the value that sets PnLt ðrÞ to zero. We can also
compute breakeven volatility in a similar manner
using put options.
To operationalize the above expression, we compute
the breakeven volatility as follows:
PnLT1 ðrT1 Þ ¼ cT  cT1 ðrT1 Þ  DT1 ðST  ST1 Þ
…
PnLt ðrt Þ ¼ ctþ1  ct ðrt Þ  Dt ðStþ1  St Þ
For day s, we solve for the breakeven volatility rs
such that PnLs is set equal to zero. For a call option,
102
we start our computation when the call has five days
until expiration, which is taken to be T in the above
expressions. The call value at time T is taken to be
the market value. For the previous day, at time T 
1, we compute the fair value of the call option that
sets the profit and loss of the delta-hedged position
on that day to be zero. We then solve for the fair
value of the call option until we reach a time to
expiration of 74 trading days, so we can build a term
structure of breakeven volatility for a variety of time
to expiration ranging from 5 to 74 days.5 We com-
pute the breakeven volatility for call and put options
available from SpiderRock that have 5–74 days until
expiration, including every option with a non-zero
bid price.6
We transform the fair values into volatility space
using the Black and Scholes (1973) model. A trans-
formation of prices into volatilities aids a compari-
son between breakeven volatility and implied
volatility. Volatilities are also more well-behaved
compared to dollar amounts, which facilitates build-
ing a predictive model. Because the Black-Scholes
equation is only used to transform option prices
into a more convenient space, much like an affine
transformation or taking logarithms, this transform-
ation does not require that the Black and Scholes
(1973) model is the correct pricing model
(Shimko 1993).
To delta-hedge the call options, we take the deltas
(1)
from SpiderRock, calculated using the Cox-Ross-
Rubinstein (CRR, Cox, Ross, and Rubinstein 1979)
binomial tree model.7 We compute the fair value of
the option until five trading days before maturity.
We do not hold the option until expiration due to
two reasons. First, due to the high gamma exposure
in the few days immediately before expiration,
volatility calculations can be quite noisy, and the
associated volatility curves are not smooth func-
tions of strike prices.8 Second, option traders often
do not roll over their positions on the expiration,
but rather a several days before expiration. Our
range of 5–74 trading days until maturity covers
the typical time to expiration options traders
engage in.9
Figure 1 provides a time series plot of the number of
options we use to calculate breakeven volatility. On
average, we have 278 strikes each day. This value is
somewhat biased upwards, by the extreme days with
(2) more than 750 strikes; the median number of strikes
we use is 244. Although the equity index options
market is already quite liquid at the beginning of our
sample in 2013, the number of strikes steadily
 Option Pricing via Breakeven Volatility

Figure 1. Number of Options Used to Calculate Breakeven Volatility

Table 1. Summary Statistics

S&P 500
Variable Mean SD 10th 25th Median 75th 90th

BEV 25.4% 16.2% 9.5% 14.1% 22.1% 32.2% 45.0%

IV 27.2% 15.6% 10.9% 15.4% 24.1% 34.8% 46.5%
BEV-IV 1.8% 9.8% 7.2% 4.2% 2.3% 0.3% 3.0%

This table presents summary statistics for breakeven volatility and implied volatility of the S&P 500 Index. 10th, 25th, 75th, and
90th represent the quantile values. Our sample is from January 2013 to November 2020.

increased towards the later part of our sample. In
total, we compute breakeven volatility for
396,899 options.
Panel A of Table 1 presents the summary statistics
for breakeven volatility. From January 2013 to
November 2020, the average value of all the com-
puted breakeven volatilities is 25.4%, and the
median is 22.1%. We observe a broad range of val-
ues—the 10th percentile of the empirical distribu-
tion is just 9.5%, whereas the 90th percentile is
45.0%. The standard deviation of breakeven volatil-
ities is 16.2%. The distributional statistics for
breakeven volatility are similar, but not identical to,
those for implied volatility. In our sample
period, the implied volatility of options on the
S&P 500 index average 27.2%, with a median value
of 24.1%. The standard deviation of implied volatil-
ity, 15.6%, is somewhat lower compared to break-
even volatility. The percentiles of implied volatility
are also similar compared to those of breakeven
volatility. We also present the difference between
BEV and IV to understand their joint distribution.
On average, BEV is lower than IV by 1.8%. Their
difference has a standard deviation of 9.8%. BEVs
are generally lower than IVs, as indicated by the
negative percentile values up to the 75th, which
reflects that in our sample, many options appear to

Volume 79, Number 1 103

Financial Analysts Journal | A Publication of CFA Institute
Figure 2. Volatility Smirks of BEV and IV
This figure shows the volatility smirks constructed from break-
even volatility and implied volatility for the S&P 500 Index,
averaged across the full sample period and smoothed. The sam-
ple period is from January 2013 to November 2020.
be trading at a premium relative to their
fair values.
Existing papers focusing on isolating a variance risk
premium (VRP) often do so in the form of excess
returns of delta-hedged positions (Bakshi, Cao, and
Chen 1997; Goyenko and Zhang 2020; Bali et al.
2021). A delta-hedged short position that earns an
average positive return indicates implied volatility is
on average higher than subsequent realized volatility.
By setting the delta-hedged profit and loss to zero,
the breakeven volatility is a measure that sets the
VRP to zero for each individual option. In this way,
BEV calculations are more flexible than a single esti-
mate of realized volatility, since we retain the flexibil-
ity of having different volatility values for options
with the same maturity but different strike prices.
Therefore, breakeven volatilities provide richer infor-
mation set on the realized distribution.
Volatility Smirks. The idea of constructing repli-
cating portfolios using historical data has been dis-
cussed in the literature, but implementation has been
limited. Zou and Derman (1999) note that theoretic-
ally, the appropriate implied volatility for a given
option is determined by the cost of replicating that
option throughout its lifetime. However, the authors
do not pursue this idea because “such replication can
be time-consuming, and the hedging errors due to
inaccurate volatility forecasting and infrequency of
104
hedging make the exercise difficult.” Greatly
improved computing power in the past 20 years, as
well as more comprehensive option datasets, over-
come much of Zou and Derman’s (1999) concerns.
Mitoulis (2019) carries out breakeven volatility calcu-
lations using simulations but stops short of a more
complete empirical analysis. Having constructed BEV
values for a large set of S&P 500 options, we com-
pare the behavior of breakeven volatility and
implied volatility.
Figure 2 compares volatility smirks using breakeven
and implied volatilities. We average across all dates
and maturities to trace out the unconditional distri-
butions for these volatility measures. We plot the
volatility curves as a function of the normalized
strike, the ratio of the log moneyness of the option
divided by its scaled volatility.10 Defined this way,
the normalized strike provides a better measure to
compare options that have very different volatility
levels or time to expiration. There is a limited number
of listed call options with high strike prices, which
causes the volatility calculations to be truncated on
the right side, whereas the left side extends much
further. This asymmetry around the number of strikes
around the at-the-money value is commonly found in
the literature (e.g., Figlewski and Malik 2014). Since
the 1987 stock market crash, equity index options
have shown a persistent skew to OTM puts. The
most common explanation for this pattern is that
market participants want to protect against the possi-
bility of a large market crash, but they do not neces-
sarily purchase deep OTM calls in anticipation of a
large upward move. This sort of market participant
preference can explain why we observe many more
strike values below the current price of
the underlying.
In the vicinity of the at-the-money region, breakeven
volatility and implied volatility are largely similar.
When we get farther away from ATM strike prices,
the two volatility measures diverge. In the regions
where the strike price exceeds the underlying price—
the normalized strike is negative—implied volatility is
higher than breakeven volatility, indicating that delta-
hedged short positions in puts or calls with these
strikes tend to have positive profits on average.
Similarly, in the regions where the strike price is
lower than the underlying price, the breakeven vola-
tility is also lower than the implied volatility. The
implications for the profitability of delta-hedged posi-
tions from Figure 2 are consistent with documented
empirical facts in the literature. Coval and Shumway
(2001) and Bakshi and Kapadia (2003) find that short

delta-hedged positions are profitable for out-of-the-
money puts.
A Data-Driven Valuation Process
In the previous section, we used historical data to
build a database of nearly 400,000 breakeven volatil-
ities for options of different moneyness and time to
expiration. While this exercise presents us with the
fair values of options, the breakeven volatility calcu-
lation uses future information, so it cannot be com-
puted in real-time. Suppose an investor wants to
know what an option with a time to expiration of
70 days should be worth. She cannot calculate the
breakeven volatility for this option without knowing
its future price path. To compute the fair value of an
option in real-time, we require a model that estab-
lishes a link between the current observable charac-
teristics of an option to its breakeven volatility. To
this end, we build a predictive model for breakeven
volatility using a variety of input variables. The map-
ping from input variables to the breakeven volatility
learned by our model constitutes an option pric-
ing model.
Volatility Predictors. We consider the
following set of predictors for breakeven volatility.
Whereas some predictors are motivated by
financial theory, others are included for their likely
potential to contain strong predictive power. Since
our primary goal is to make a robust prediction for
breakeven volatility, we are not restricted to only
using variables that have a strong economic motiv-
ation. All variables are calculated at the
daily frequency.
1. VIX captures the market’s expectation for the
implied volatility of the S&P 500 Index over the
next 30 days. It is commonly used by investors
to gauge market sentiment. We use the log of
the CBOE Volatility Index at the market close.
2. VVIX is a volatility of volatility measure that
captures the expected volatility of the 30-day
forward price of the VIX. We use the closing
value of the CBOE VVIX Index.
3. rv is a GARCH forecast of the volatility of the S&P
500 returns over the same horizon as the time to
expiration of the option, including an adjustment
for the overnight component of volatility.
4. rt provides a measure for the time to expir-
ation. It is calculated as the square root of the
number of trading days to expiration divided
by 252.
Volume 79, Number 1
Option Pricing via Breakeven Volatility
5. lsk is a measure of the moneyness of the
option calculated as the natural log of the
price of the underlying over the strike price.
This variable is motivated by the empirical
finding that volatility is not constant across
moneyness (Zou and Derman 1999), and it
allows the breakeven volatility to vary as a
function of moneyness.
6. ATM.IV is the implied volatility of the option
closest to at-the-money, calculated using the
Black and Scholes (1973) model.
7. ImpliedVol is the market implied volatility of
the option calculated using the Black and
Scholes (1973) model.
8. RR is the difference between the delta of the
option and the ATM delta. This variable pro-
vides another way to capture higher volatilities
in the wings compared to ATM options.
9. vega is the sensitivity of the option price
with respect to volatility. We use the ffi Black-
pffiffiffiffiffiffiffiffiffiffi
Scholes vega, equal to St N0 ðd1 Þ T  t, where
St is the price of the underlying, N0 ðÞ is the
probability density function of standard nor-
mal distribution, d1 comes from the Black-
Scholes formula,11 and T  t is the time
to expiration.
10. gamma is the rate of change in the delta
with respect to changes in the underlying
price, or the second derivative of the option
price with respect to the price of the underly-
ing. We calculate the Black-Scholes gamma,
N0p
ð d1 Þ
ffiffiffiffiffiffi , where r is the volatility of
St r Tt
the underlying.
11. vrt is defined as the VIX multiplied by the
square root of time to p expiration.
ffiffiffiffiffiffiffiffiffiffiffi This variable is
a construction of r T  t from the Black-
Scholes model. Breaking up the Black-Scholes
model into its constituent parts allows for add-
itional flexibility in predicting breakeven volatility.
12. vvt2 is defined as the VIX squared multiplied
by the time to expiration. This variable is
another piece in the Black-Scholes
model, r2 ðT  tÞ:
2
We also include some transformations of the
above variables. lsk2 is the squared value of lsk,
motivated by the empirical observation that volatil-
ity measures are a convex function of moneyness.
ImpliedVol2 is equal to ImpliedVol squared. RR2 and
RR3 are the squared and cubed values of RR, moti-
vated by interpolation methods used in the litera-
ture to build volatility curves that try to capture the
curvature in the volatility surface through a cubic
105
Financial Analysts Journal | A Publication of CFA Institute

Table 2. Summary Statistics of Predictors

Variable Mean SD 10th 25th Median 75th 90th

VVIX 97.67 18.31 79.53 85.83 93.97 106.01 118.94

VIX 18.08 8.93 11.53 12.85 15.00 20.10 27.96
rt 0.39 0.10 0.24 0.32 0.40 0.47 0.52
lsk 0.15 0.21 0.07 0.00 0.12 0.26 0.40
lsk2 (10) 0.68 1.73 0.00 0.03 0.17 0.70 1.63
ATM.IV 16% 8% 10% 12% 14% 18% 25%
ImpliedVol 27% 16% 11% 15% 24% 35% 46%
ImpliedVol2 10% 14% 1% 2% 6% 12% 22%
RR 0.21 0.35 0.45 0.01 0.40 0.47 0.49
RR2 0.17 0.08 0.02 0.11 0.20 0.23 0.24
RR3 0.04 0.07 0.09 0.00 0.07 0.11 0.11
vega 1.50 1.55 0.09 0.25 0.88 2.40 3.88
vega2 4.63 7.87 0.01 0.06 0.78 5.77 15.08
gamma (x100) 1.14 1.72 0.10 0.17 0.39 1.28 3.43
vrt (x10) 0.63 0.34 0.29 0.41 0.57 0.75 1.01
vvt2 (x100) 0.25 0.36 0.04 0.08 0.16 0.28 0.51
rv 0.24 0.18 0.14 0.16 0.19 0.27 0.39

This table presents summary statistics for the different variables used to predict breakeven volatility. The sample period is from
January 2013 to November 2020.

spline (Malz 2014; Neuberger 2012). vega2 is
vega squared.
Table 2 shows the summary statistics for the pre-
dictor variables, and Table 3 provides the pairwise
correlations. This table highlights the diversity of
our predictors; most variables are weakly correlated
with others. For example, the realized volatility fore-
cast rv is basically uncorrelated with the time to
expiration rt or the moneyness measures lsk and
lsk2, but it is positively correlated to volatility varia-
bles, such as the at-the-money implied volatility,
VIX, VVIX, as well as vrt and vvt2. Variable transfor-
mations are often highly correlated with the original
variables: lsk and lsk2 have a correlation of 0.77,
ImpliedVol and ImpliedVol2 have a correlation of
0.93, and vega and vega2 have a correlation of
0.94. Some variable pairs show large negative corre-
lations. Gamma has a 0.79 correlation with RR,
whereas RR2 has correlations of 0.90 and 0.79
with vega and vega2.
We include several volatility measures as predic-
tors, including VIX, ATM.IV, and rv. As one would
expect, these variables are highly correlated with
one another. To limit multicollinearity, we reduce
the correlations among these predictors through
orthogonalization, via a Gram-Schmidt process
(Cheney and Kincaid 2009). We run an ordinary
least squares regression of VIX on ATM.IV, then
we take the residual of the regression, VIX.res, as
the new predictor in place of VIX. We repeat the
same process for rv to obtain rv.res. By construc-
tion, VIX.res and rv.res are uncorrelated with
ATM.IV. Additionally, they are also only 0.33 corre-
lated with each other. We do not orthogonalize
ImpliedVol or ImpliedVol2, since these have greater
variation depending on the moneyness of the
option and tend to bring significantly different
information compared to the previous three volatil-
ity variables.
ATM.IV, VIX.res, and rv.res are all included as predic-
tors in our model. Presumably, the idiosyncratic com-
ponent of each volatility measure contributes to
predicting breakeven volatility, rather than just the
common component. If only the common component
of the volatility measures mattered for prediction, it
would not be necessary to retain all three volatility
measures—some form of average would suffice. For
example, we could construct a derived predictor as
the largest principal component (Hull and Qiao
2017). Because the three volatility measures capture
different information related to breakeven volatility,
we choose to include all three in the predic-
tion model.
The first row of Table 3 contains correlations among
the predictor variables and breakeven volatility, our
prediction target. Correlation is one measure of how

106

Table 3. Correlation Matrix of Predictors
This table presents the pairwise correlations for the predictor variables and breakeven volatility. Positive correlations are shown in
shades of red, whereas negative correlations are shown in shades of blue. The shade of the cell indicates the strength of the cor-
relation. The sample period is from January 2013 to November 2020.
well each variable would do in univariate predictive
regressions. Almost all the predictors exhibit consid-
erable correlations with our prediction target,
although the relationship can be positive or negative.
VVIX, lsk, RR, and vrt all have a positive relationship
with breakeven volatility; vega, vega2, and gamma
have a negative relationship with BEV. Suggestive
economic interpretations of a positive or negative
correlation with BEV can be found in our description
of the predictor variables.
Statistical Model. Our goal is to make good pre-
dictions of breakeven volatility. To the extent linear
models can offer good predictions for the target vari-
able, they are preferred over more complex models
given their simplicity and interpretability. Our
Volume 79, Number 1
Option Pricing via Breakeven Volatility
prediction target is BEVi, K, t, T for option i with
strike price K and time to expiration T  t: Because
the BEV calculation requires knowledge of the path
of the option and the underlying, it is only known on
the expiration date T: We use predictor variables at
time t, Xi, m, t , to form a conditional expectation of
the actual BEV value at time T: To simplify notation,
we will denote the target and predictor variables as
BEVi and Xi, m with the understanding that they are
known at different times.
The raw BEV values are truncated at zero and show
significant right skew. In the first stage, we transform
the dependent variable, BEV, into its logarithms. By
working in the log space of BEV, we obtain a distri-
bution well-approximated by a normal distribution,
and the predicted breakeven volatility value will be
107
Financial Analysts Journal | A Publication of CFA Institute
guaranteed to be positive. Our first-stage regression
models logged BEV as a linear function of the pre-
dictor variables:
X
M
logðBEVi Þ ¼ b0 þ Xi, m bm þ ei
m¼1
where BEVi is the breakeven volatility linked to a par-
ticular set of parameters including time to expiration
and moneyness, Xi, m are the M predictors associated
with BEVi , and ei is the prediction error.
For a given set of predictor variables, Equation (3)
provides unbiased estimates for the expectation of
logged BEV, E½logðBEVÞ: However, our goal is to find
the best prediction for the expectation of BEV,
E½BEV: The approach to retransform a predicted
quantity back to its original scale is known as a
“smearing adjustment” in statistics (Duan 1983;
Taylor 1986). We cannot simply exponentiate our
prediction from Equation (3), because eE½logðBEVÞ 
E½BEV by Jensen’s inequality.12 This point has been
emphasized by Duan (1983), who asserts that
unbiased and consistent predictions on a transformed
scale (using a monotonic function) do not transform
into unbiased or consistent quantities on the untrans-
formed scale.
If the residuals from Equation (3) were normally
distributed, there is a closed-form transformation
of the prediction of Equation (3), E½logðBEVÞ, into
E½BEV :
E½BEV ¼ eE½logðBEVÞ e2g
1 2
where g is the standard deviation of the residuals
ei of Equation (3). For residuals that are approxi-
mately normally distributed, the above expression
provides a fairly precise adjustment from the
logged values to the original values (Duan 1983).
However, the adjustment does not work well if the
residuals were not normally distributed. In the case
of non-normal residuals, Duan (1983) proposes the
following:
1X N
E½BEV ¼ eE½logðBEVÞ eej
N j¼1
where N is the number of data points used to esti-
mate Equation (3), and ej s are the residuals. Duan
(1983) posits that like Equation (4), our desired quan-
tity E½BEV and the exponentiated prediction from
Equation (3) eE½logðBEVÞ differ by a constant multiple.
Whereas normal theory implies a multiplicative factor
P
of e2g , Duan’s is N1 Nj¼1 eej :
1 2
108
Empirically, the residuals from the first-stage
regression are not normally distributed. If we were
to apply Equation (4), the resultant breakeven
volatility values differ significantly from the actual
(3) values. Equation (5) provides a better adjustment
factor, but the resulting BEV values still make
poor predictions for the actual values. To
overcome this retransformation issue, we use
a more general, supervised approach to
estimate the adjustment factor in a second-stage
regression:13
d
BEVi ¼ ceE½logðBEVi Þ þ ei (6)
where logðd BEVi Þ are the predicted values from the
first-stage regression in Equation (3). Equation (6)
is estimated without an intercept, adhering to the
constant multiple relationships derived in Duan
(1983), and c is an estimate of the multiplicative fac-
tor that wedges between E½BEV and eE½logðBEVÞ : The
final prediction of the breakeven volatility comes
from the predicted component of Equation (6),
d
^c eE½logðBEVi Þ : Combining the two stages, the following
expression links the predictor variables with the final
prediction of BEV:
P
M
bb0 þ Xi, m bbm
E½BEVi  ¼ ^c e m¼1 (7)
The in-sample results of the first-stage regression are
shown in Table 4. The predictors are all statistically
significant at the 1% level, indicating strong
predictive power of these variables for logged
(4) values of breakeven volatility. Taken together, the
predictor variables can explain 53% of the variation
in BEV. The adjusted R-squared is nearly identical to
the unadjusted R-squared because we have many
observations compared to the number of
regressors.14
In-sample R-squared can overstate the predictive
power of our model. To more accurately assess how
well our model can predict breakeven volatility, we
use a 10-fold cross-validation approach to calculate
the coefficient of determination. We partition our
(5) data into 10 subsamples of roughly equal size. For
each subsample, we fit our statistical model on the
remaining 90% of data and use the fitted model to
make predictions on the subsample. We repeat this
procedure to obtain 10 validation R-squareds and
take their average to be the cross-validated R-
squared. The cross-validated R-squared is very close
to the full-sample R-squared, indicating that our pre-
diction model performs with a high degree of stability
on this dataset. Indeed, the 10 validation R-squareds
 Option Pricing via Breakeven Volatility

Table 4. Predicting Logged Table 5. Smearing Adjustment of Breakeven

Breakeven Volatility Volatility Prediction
Estimate SE t-Stat Estimate SE t-Stat

(Intercept) 1.85 0.02 116.7 FirstStagePrediction 1.03 0.001 1310

VVIX 0.187 0.009 20.6 R squared 0.90
VIX.res 0.002 0.001 2.6 Adj. R squared 0.90
rt 0.59 0.03 19.7 10-Fold CV R-squared 0.66
lsk 0.69 0.02 30.0 Number of observations 396,899
lsk2 0.11 0.01 7.6
ATM.IV 1.71 0.06 28.1 This table shows the second stage of the predictive model for
ImpliedVol (coef  10) 2.08 0.04 52.7 breakeven volatility. We regress breakeven volatility values on
ImpliedVol2 0.01 0.00 44.8 the predictions from the first stage.
RR 0.15 0.01 15.5 d
RR2 0.90 0.03 26.8 BEVi ¼ ceE½logðBEVi Þ þ ei
RR3 1.29 0.06 22.3 where BEVi is the breakeven volatility and logðdBEVi Þ is the pre-
vega 2.05 0.37 5.5 dicted value from the first-stage regression. c is an estimate of
vega2 0.54 0.04 12.3 the smearing adjustment. Standard errors and t-statistics are
adjusted for heteroskedasticity.
gamma 67.53 1.44 46.8
vrt 4.33 0.20 21.7
vvt2 33.74 0.85 39.5 penalization parameter we select using 10-fold
rv.res 0.43 0.01 34.1
cross-validation. We find that the optimal
R-squared 0.53
penalization parameter is zero, indicating that the
Adjusted R-squared 0.53
10-Fold CV R-squared 0.53 optimally-selected ridge model is equivalent to
Number of observations 396,899 ordinary least squares. We further explore a range
of possible values for the penalization parameter,
This table shows the first stage of the predictive model for and we confirm that the model with the lowest
breakeven volatility. We use linear regression to predict the mean squared error is the one with the penalization
logged values of breakeven volatility.
parameter set to zero. These results suggest that
X
M multicollinearity does not pose a major issue for
logðBEVi Þ ¼ b0 þ Xi, m bm þ ei our model.15
m¼1
where BEVi is the breakeven volatility linked to a particular set Table 5 presents the smearing adjustment according
of parameters including time to expiration and moneyness, Xi, m to Equation (6). As implied by the equation, this
are the M predictors associated with BEVi , bm are regression
coefficients. The coefficients associated with VVIX, ATM.IV, second-stage regression does not include an inter-
ImpliedVol, ImpliedVol2, vega, vega2, and rv.res are multiplied cept term. The smearing adjustment indicates
E½BEV ¼ 1:03 eE½logðBEVÞ , consistent with the impli-
by 100 for ease of exposition. Standard errors and t-statistics
adjusted for heteroskedasticity are shown.
cations of Jensen’s Inequality eE½logðBEVÞ  E½BEV:
This adjustment procedure shows a close
show a narrow range between 0.51 to 0.56, with a
relationship between the predicted log breakeven
standard deviation of just 0.012. The regression coef- volatility values and actual BEV—the full-sample R-
ficients show a high level of stability across valid- squared is 0.90, and so is the adjusted R-squared.
ation folds. The 10-fold cross-validated R-squared is somewhat
If high correlations among the predictors lead to lower but still suggests that the majority of the
multicollinearity issues, the statistical integrity of variation in breakeven volatility is captured by
our model.
the model may be compromised. While ordinary
least squares estimates remain unbiased, their Our approach attempts to estimate breakeven volatil-
variances can be large. Ridge regression helps to ities as a fixed function of certain characteristics of
resolve multicollinearity by allowing for better options, such as the current level of the underlying
convergence of the variance-covariance matrix and moneyness. If our model captures the data well,
and introducing a small degree of bias to produce its functional form should be relatively stable over
more reliable estimates. We check the robustness time, although the actual breakeven volatility predic-
of our prediction using ridge regression, whose tions can change over time or across strike prices if

Volume 79, Number 1 109

Financial Analysts Journal | A Publication of CFA Institute
the characteristics on which the prediction is based
on change.
Our predictive model takes a data-driven approach
to quantifying the value of an option, which stands
in contrast to commonly used structural approaches
to option pricing. Existing approaches typically spe-
cify key parameters necessary for calculating
options prices—volatility, time to expiration, volatil-
ity of volatility, etc., then use these parameters to
derive the value of an option. In comparison, a data-
driven approach learns the relationship between the
variables that are important for option valuation,
rather than specifying the model parameters
in advance.
Any pricing model must provide a link between a
set of input variables and output variables, typically
taken to be the option prices or volatilities. For
example, the Black-Scholes model provides a map-
ping between the model parameters (the price of
the underlying, volatility of the underlying, time to
expiration, strike price, and interest rate) and the
dollar value of an option. The Heston (1993) model
stipulates a different mapping for a set of model
parameters and the price of an option. Our
approach also establishes a link between the
predictor variables and the breakeven volatility,
Figure 3. Probability Densities of Actual and Predicted Values
The figures below show density plots of the first and second-stage regressions. In each plot, kernel densities of the prediction target
and their predicted values are shown. In the first-stage regression, the prediction target is logged breakeven volatilities. In the
second stage, the target is the level of breakeven volatility.
110
a measure of the fair value of an option.
However, the relationship between the predictor
variables and BEV is learned from a large database
of historical BEV values and predictors, which
allows for more flexibility compared to the exam-
ples above.
Researchers have explored non-parametric option
pricing models in the past. Hutchinson, Lo, and
Poggio (1994) use neural networks to approximate
the Black-Scholes formula, but the paper does not
consider a non-parametric estimation of the relation-
ship between option characteristics and option pri-
ces. Ait-Sahalia and Lo (1998) propose a kernel
regression estimator of the risk-neutral distribution
conditional on option characteristics, such as the
price of the underlying, strike price, and time to
expiration. In the same spirit, we also try to capture
the relationship between predictor variables and
option values in a data-driven manner. Our approach
is different in that our target is breakeven volatility,
which serves as an estimate for the fair value of
an option.
Model Diagnostics. Although economic intu-
ition may guide us in thinking about the univariate
relationships among predictors and breakeven
volatility, it is much more difficult to determine

the signs of these relationships in a multivariate
setting. Rather than focus on the contribution of
individual predictors, we conduct model diagnostics
to identify the strengths and weaknesses of our
predictive model. We investigate the statistical val-
idity of the model, as well as the economic implica-
tions of the volatility predictions. To start, we
inspect the differences between the prediction tar-
get and the predicted values. In Figure 3, we plot
the kernel densities of the prediction target and the
predicted values in the same figure. A comparison
of the two densities can reveal where our model is
doing well and where it is failing. In the first-stage
regression, the prediction target is the logarithm of
breakeven volatilities. The predictions match the
target values reasonably well in the middle part of
the distribution but deviate from the target in the
tails. In both the left and right tails, there are too
few predicted values compared to the target distri-
bution, and there are too many predicted values
hovering in the center of the distribution. The mod-
el’s predictive power appears to be the strongest
when the breakeven volatility is between 5
and 50%.
In the second-stage regression, the left-hand vari-
able is breakeven volatility and the right-hand vari-
able is a transformation of logged BEV values. Like
the first-stage regression results, the centers of the
two distributions mostly overlap, although it is clear
that the predicted values are more concentrated in
the middle of its distribution than the tails. The
model predictions have the greatest difficulty with
the left tail; there are too few predicted values com-
pared to the actual breakeven volatility values,
especially if BEV is <5%. The predicted distribution
and the actual distribution of breakeven volatilities
are more closely matched in their right tails, albeit
we still observe too few predictions compared to
the target density.
We verify that the predicted volatilities do not
violate arbitrage conditions. Orosi (2015) derives
static no-arbitrage conditions for call options:
Prices must be a decreasing and convex function
of the strike price. We check the parallel
conditions for puts: Prices are an increasing and
convex function of the strike price. These
conditions can be verified by checking whether
put prices violate arbitrage conditions implied by
putting spreads and butterfly spreads. A put
spread combined buying and writing put options
with the same expiration but different strike prices.
Suppose we have two put options p1 and p2 , with
the same expiration date but different strike
Volume 79, Number 1
Option Pricing via Breakeven Volatility
prices, K1 and K2 : To prevent risk-free profits,
we need the following condition to hold:
p2  p1 for K2 > K1 (8)
If the above condition were violated, risk-free profit
would be available. If p2  p1 , a portfolio that com-
bines a long position in p2 and short position in p1 is
costless to initiate, and will offer a non-negative pay-
off in all possible states of the world. The no-arbi-
trage condition for put spreads compels put prices to
be an increasing function of the strike price, an intui-
tive condition that keeps the volatility curve
well-behaved.
For put prices to be a convex function of the strike
price, the no-arbitrage condition for butterfly
spreads must be satisfied. A butterfly spread uses
three contracts: Buy a put at the lowest strike, buy
a put at the highest strike, and write two puts at an
intermediate strike. These positions form a portfolio
that earns the maximum payoff if the underlying
does not move—a short volatility position. The no-
arbitrage condition for butterfly spreads is the fol-
lowing:
K3 K2 K2 K1
p2  p1 þ p3 (9)
K3  K1 K3  K1
for
K3 > K2 > K1
In the above expressions, K1 , K2 , and K3 are strike
prices of increasing value. p1 , p2 , and p3 are the put
prices associated with the respective strikes. This
condition states that the put price associated with
the intermediate strike, p2 , must be less than a
weighted average of the put prices associated with
the highest and lowest strikes.16 The absence of arbi-
trage from butterfly spreads implies that put prices
must be a convex function of the strike price, such
that the shape of the volatility curve matches its
well-known empirical profile (Bollen and
Whaley 2004).
For each unique combination of date and time to
expiration, options with different strike prices
trace out a volatility curve. We examine the pre-
dicted BEV values using the above conditions, and
we find that none of the volatility curves
admit arbitrage.
Volatility Arbitrage
In the previous section, we built a predictive model
for breakeven volatility, the fair value of an option
111
Financial Analysts Journal | A Publication of CFA Institute
that does not admit any risk premia. Implied volatility
from market prices includes the risk attitudes of mar-
ket participants and therefore embeds risk premia.
To the extent our predictive model can make accur-
ate forecasts of breakeven volatility, we can poten-
tially convert the difference between implied and
breakeven volatility into trading profits. In this way,
the economic value of our model can be understood
through a simulated trading strategy. The purpose of
this section is to evaluate whether the difference
between implied volatility and predicted breakeven
volatility can generate profitable trading
opportunities.
Simulated option trading strategies have often been
used in the literature to determine the difference
between implied and realized volatility. Coval and
Shumway (2001) and Bakshi and Kapadia (2003)
find large returns to delta-hedged options and inter-
pret this result as a volatility risk premium. Bollen
and Whaley (2004) attempt to match the profits of
delta-hedged positions with net buying pressure in a
demand-based asset pricing framework. Zou and
Derman (1999) ask, “How is an investor to know
which strike and expiration provide the best
value? What metric can option investors use to
gauge their estimated excess return?” Existing
approaches to option strategies do not offer
answers to these questions, as they are not able to
determine which options offer relatively more
attractive investment opportunities or what
expected returns we should expect from delta-
hedged positions.
Our non-parametric option pricing model based on
the prediction of breakeven volatility readily
answers the above questions. The investor can
make breakeven volatility predictions for several
options and compare her predictions against implied
volatilities. If the predicted BEV and IV for an option
are different, the investor may initiate a position
with positive expected returns. If implied volatility
were lower than the predicted breakeven volatility,
the market price of the option is too low, and the
investor can “buy cheap” by taking a long delta-
hedged position. Conversely, if implied volatility
were higher than the predicted breakeven volatility,
the market price of the option is too high, and the
investor can “sell dear” by taking a short delta-
hedged position.
Simulated Trading Strategy. Trading the
difference between a volatility prediction and
implied volatility is commonly called “volatility
arbitrage” (Ammann and Herriger 2002). It is not
112
arbitrage in the strict sense, but rather refers to
whether the investor can make a sufficiently
accurate forecast for future realized volatility to
lock in a trading profit. To illustrate the
economic value of our predictive model, we
construct a trading strategy that exploits the
difference between our predicted breakeven
volatility values and implied volatility from mar-
ket prices.
We keep the trading strategy simple because we
want to test the accuracy of our statistical predic-
tions. If we were to proceed with a more complex
trading strategy—for example, adding in a layer of
portfolio optimization—it becomes more difficult to
evaluate the accuracy of our breakeven volatility
forecasts. In a strategy that combines BEV forecasts
with sophisticated portfolio optimization, we would
be testing the transformation of our trading signals
into portfolio positions as well as the accuracy of the
BEV forecasts, resulting in a “joint hypothesis prob-
lem.” By keeping the simulated trading strategy sim-
ple, we can test the accuracy of the BEV forecasts
with limited confounding factors. Therefore, our
empirical choices in the simulated trading strategy
closely mirror those we made in the construction of
our predictive model.
We use an initial training period from January
2013 to December 2014 to fit our model. We fix
the model parameters and make predictions for
breakeven volatilities throughout 2015, and we
refit the model at the end of 2015 using all the
data from 2013 through 2015. We then make new
predictions using the updated model parameters.
Each subsequent year, the model is refit using an
expanding window. These predictions only use the
information available at the time of the forecast, so
they can be made in real-time. The final set of pre-
dictions goes from January 2015 to
November 2020.
In the modeling section, we showed that our statis-
tical model achieves high in-sample R-squared.
However, good in-sample performance does not
always imply generalizable results. This point is
emphasized in return predictability studies where
the goal is to produce the best predictions of future
market returns (e.g., Welch and Goyal 2008;
Campbell and Thompson 2008). Our expanding win-
dow predictions constitute out-of-sample (OOS)
forecasts of breakeven volatility, and we evaluate
the out-of-sample predictive performance of our
model compared to the historical mean. We assess
the volatility forecasts using the mean squared

forecast error (MSFE), calculated for our prediction
and the historical mean:
^e 0, i ¼ BEVi  BEV i (10)
di
^e 1, i ¼ BEVi  BEV (11)
d i is
where BEVi is the actual breakeven volatility, BEV
the predicted value of breakeven volatility, and BEV i
is the historical mean estimated using the same infor-
d i : The mean squared forecast error
mation set as BEV
is calculated as the sum of squared forecast errors
across the full set of N forecasts:
XN
dk¼1
MSFE ^e 2 , k ¼ 0, 1 (12)
N i¼1 k, i
Based on the Clark and West (2007) procedure,
Dong et al. (2022) develop a test for a difference in
the population of the MSFEs:
H0 : MSFE0  MSFE1 , H1 : MSFE0 > MSFE1 (13)
Campbell and Thompson (2008) propose an alterna-
tive method to evaluate competing forecasts via an
out-of-sample R-squared statistic:
d1
MSFE
R2OS ¼ 1  (14)
d0
MSFE
Dong et al. (2022) show that the Clark and West
(2007) statistics to test Equation (13) is equivalent to
testing whether the out-of-sample R-squared of
Campbell and Thompson (2008) is positive or nega-
tive. We follow Dong et al. (2022) to evaluate the
d i captures more
predictive power of our model. If BEV
variation of actual breakeven volatility than the his-
torical mean, the sum of squared errors in the
numerator would be smaller than the denominator,
indicating our predictive model forecasts BEV better
than its historical average. In this case, the out-of-
sample R-squared R2OS would be positive, correspond-
ing to the alternative hypothesis in Equation (13). If
d i captures less variation of BEVi than BEV i , the
BEV
out-of-sample R-squared would be negative, corre-
sponding to the null hypothesis in Equation (13).
Every OOS prediction of breakeven volatility allows
for a comparison with implied volatility. Starting in
January 2015, each day, we scan through all available
OTM options with time to expiration between five
and 74 days, and we identify those with a predicted
profit remaining >$1.17 Profit remaining is calculated
as the difference between the breakeven dollar value
and the midpoint between the bid and ask prices.
For a particular option, if the breakeven price were
Volume 79, Number 1
Option Pricing via Breakeven Volatility
lower than the mid-price, we sell one contract. If the
breakeven price were higher than the mid-price, we
purchase one contract. All trades are delta-hedged
and held until five days to expiration, at which time
the positions are closed out. Portfolio returns are cal-
culated by summing up the daily profit and loss and
dividing by the aggregate notional value of the
contracts.18
Results
In the simulated trading strategy, the predictive
model makes out-of-sample forecasts of breakeven
volatility. We first assess these OOS forecasts before
evaluating the trading strategy. The OOS forecasts
provide a valuable perspective for understanding the
validity of our predictive model. The out-of-sample
R-squared is 0.71, indicating that our predictive
model for breakeven volatility significantly outper-
forms the historical average estimate—the mean
squared forecast error is 71% smaller for our fore-
casting model compared to the historical mean.
Table 6 presents a confusion matrix of the implied
trading strategy from the breakeven volatility fore-
casts. The confusion matrix compares the number of
long and short investment decisions implied by the
model predictions with the decisions that an investor
would make if she could perfectly observe the actual
BEV values. For example, suppose an option with
30 days to expiration has an IV of 25% and BEV of
28%. With 30 days to go, we would not observe the
true BEV value. Suppose our predicted BEV value is
24%. In this case, we would take a short position in
the option, whereas if we could observe the true
BEV value, we would have taken a long position.
Since the options literature has documented a robust
Table 6. Confusion Matrix of BEV
Predictions
Predicted
Short Long
Actual
Short 3002 1467
Long 0 2314
This table shows the out-of-sample BEV predictions for those
observations that show favorable trades. Each year, we retrain
our forecasting model for breakeven volatility using as much
historical data as possible, and we use this model to make pre-
dictions throughout the year. If the breakeven price were lower
than the mid-price, we sell one contract. If the breakeven price
were higher than the mid-price, we purchase one contract.
113
Financial Analysts Journal | A Publication of CFA Institute

Table 7. Strategy Performance

A. Raw values
BEV Short Options Short Puts Short Calls Buy-and-Hold S&P

Average returns 6.8% 4.2% 3.1% 1.1% 10.4%

Volatility 4.5% 4.3% 3.0% 1.6% 13.0%
Sharpe ratio 1.50 0.99 1.04 0.68 0.79
Max drawdown 3.2% 5.1% 4.7% 2.1% 19.8%

B. Option strategies scaled to the same volatility as the S&P 500

BEV Short Options Short Puts Short Calls Buy-and-Hold S&P

Average returns 19.6% 12.9% 13.5% 8.9% 10.4%

Volatility 13.0% 13.0% 13.0% 13.0% 13.0%
Sharpe ratio 1.50 0.99 1.04 0.68 0.79
Max drawdown 9.3% 15.7% 20.3% 17.7% 19.8%

C. Including transaction costs

One Tick Two Ticks Three Ticks

Average returns 4.4% 3.2% 2.1%

Volatility 4.5% 4.5% 4.5%
Sharpe ratio 0.98 0.72 0.46
Max drawdown 4.9% 7.2% 11.5%

This table presents the performance statistics of our simulated trading strategy and several other investment strategies. “BEV” is a
strategy that uses breakeven volatility forecasts to select and invest in options. “Short Options” is a strategy that takes short posi-
tions in all available options. “Short Puts” and “Short Calls” only invest in puts or calls separately. Panel A shows the statistics for
the raw returns of the strategies. Panel B scales the options strategies to have the same volatility as that of the S&P 500. Panel C
includes transaction costs for the BEV strategy (each tick is five cents, see text for details).

variance risk premium (Carr and Wu 2009; Bollerslev,
Tauchen, and Zhou 2009), our trading strategy will
have a default bias towards short positions. In this
sense, the short positions are the “positive” class in
the confusion matrix, and the long positions are the
“negative” class.
Our out-of-sample predictions cover 6,783 observa-
tions that display an investing edge. Of these, 3,002
are true positives—the model short positions match
the actual BEV implied short positions, and 2,314
are true negatives. Therefore, the accuracy of the
model, measured as the total correct predictions
divided by the total number of observations, is 78%.
Precision measures the ratio of correctly predicted
positive observations to the total predicted positive
observations. This number comes out to be 100%—
in this particular sample, every predicted short pos-
ition with an expected profit remaining of more
than $1 turned out to be correct. Recall, that the
ratio of correctly predicted positive observations to
all the actual positive observations, is 67%, indicat-
ing that the number of correctly predicted short
positions is 67% of the total actual number of short
positions an investor would take if she could per-
fectly observe breakeven volatilities. F1 score, the
harmonic mean of precision and recall, provides a
holistic evaluation of the predictive power of the
model that balances false positives and false nega-
tives. The F1 score is 0.80 in our case. By looking at
the predictions from our regression model through
the lens of classification, we gain an additional per-
spective on its performance.
We provide a summary of the performance of our
BEV-based trading strategy in Table 7. We include
three additional options strategies for comparison: A
portfolio of delta-hedged short positions in all avail-
able options, a portfolio of delta-hedged short posi-
tions in all puts, and a portfolio of delta-hedged short
positions in all calls. Finally, we include a buy-and-
hold S&P 500 strategy. The BEV-based strategy
earns an annual return of 6.8% with a volatility of
4.5%, achieving the highest Sharpe ratio and the
highest ratio of expected to maximum drawdown
among all the options strategies. A buy-and-hold

114

strategy in the S&P 500 Index performed well in our
sample period, earning 10.4% per year with 13%
volatility. Its Sharpe ratio of 0.79 shows an attractive
risk-return tradeoff between 2015 and 2020, higher
than its historic longer-term value.
Due to the differences in volatility, it is difficult to com-
pare average returns and maximum drawdowns across
strategies. In Panel B of Table 7, we scale all the option
strategies such that their annual volatilities are all equal
to that of the S&P 500. Putting all the strategies on
the same scale makes clear the distinct investment
opportunities available to investors. At annual volatility
of 13%, the BEV strategy earns 19.6% per year with a
maximum drawdown of 9.3%, providing the highest
average return and the lowest tail risk. In comparison,
the other options strategies earn between 8.9% and
13.5%, with maximum drawdowns between 15.7 and
20.3%. The buy-and-hold S&P strategy has a max-
imum drawdown of 19.8%.
Margins and transaction costs are major factors that
can impact the return of option strategies (Zhan
et al. 2022). While our return calculation makes a
conservative assumption about margin requirement—
fully collateralized positions—we did not include
transaction costs thus far. We explore the impact of
transaction costs for our BEV strategy in Panel C of
Table 7. The tick size for SPX options is five cents.
For far out-of-the-money options, the bid-ask spread
is often five cents. For at-the-money options, the
typical bid-ask spread ranges from 20 to 30 cents. To
incorporate costs, we assume we can trade the E-
mini futures contract (which we use to hedge) one
tick worse than the midpoint price, and we assume
three values for trading SPX options: One tick, two
ticks, and three ticks.19 According to Muravyev and
Pearson (2020), the average quoted bid-ask spread is
8.1 cents for stock index options, and the average
effective spread is 6.2 cents. For sophisticated trades
that take into account the expected future price
movement of the options, the average effective
spread is just 1.3 cents. The authors point out that
proprietary traders or institutional investors who
have algorithms or use brokerage firm execution
algorithms pay closer to 1.3 cents than 6.2 cents. In
light of the findings of Muravyev and Pearson (2020),
our transaction cost assumptions are on the conser-
vative side.
Panel C shows that if we trade one tick worse than
the midpoint price for SPX options, the average return
reduces to 4.4%, the Sharpe ratio reduces to 0.98,
and the maximum drawdown rises to 4.9%. Incurring
larger transaction costs further reduces the average
Volume 79, Number 1
Option Pricing via Breakeven Volatility
return and Sharpe ratio, and increases the maximum
drawdown. Unless the investor can limit her transac-
tion costs, a strategy based on breakeven volatility
will not offer attractive risk-adjusted returns.
Conclusion
In this paper, we build a non-parametric option pric-
ing model whose output is the fair value of an option
as measured by breakeven volatility, the value of
implied volatility that sets the profit and loss of a
delta-hedged option to zero. We compute breakeven
volatilities for a large set of S&P 500 index options,
and we use these historical values to construct a pre-
dictive model. A two-stage regression approach cap-
tures the majority of the variation in breakeven
volatility, and the resulting predictions satisfy no-
arbitrage conditions. The predictions from our model
can be used to formulate a volatility arbitrage strat-
egy that exploits the difference between implied
volatility and the predicted breakeven volatility.
There are several interesting future directions to
explore. A natural and immediate extension would be
a careful comparison of the different ways of delta
hedging. We used deltas associated with individual
options, but alternative choices may include using the
at-the-money delta or a delta calculated from a real-
ized volatility forecast. A deeper analysis of the rela-
tionship between breakeven volatility, realized
volatility, and realized skewness would contribute to
the existing literature on modeling and forecasting
higher moments of asset distributions. Garleanu,
Pedersen, and Poteshman (2009) show that option
prices are associated with demand pressure from
option end-users, which may serve as an additional
predictor variable in an option pricing model. An
investigation of the connection between breakeven
volatility and the Recovery Theorem of Ross (2015)
could shed light on the viability of Ross’s assumptions.
For a fixed time to expiration, the second derivative
of option prices with respect to strike prices gives
the risk-neutral distribution (Breeden and
Litzenberger 1978). To the extent breakeven volatil-
ities are different for options with different strike pri-
ces, the second derivative of the prices associated
with BEVs provides another distribution. Because pri-
ces associated with breakeven volatilities do not
admit any risk premia, this distribution does not
embed the risk attitudes of market participants. A
breakeven volatility surface that changes as a func-
tion of moneyness and time to expiration may be of
interest for additional investigation.
115
Financial Analysts Journal | A Publication of CFA Institute

Another interesting research direction is to expand Appendix A

beyond equity indices to individual equity options. It
The table below outlines some differences between
would be worth exploring whether one can use the
the closing price methodology for the SpiderRock
same predictive model for BEVs in different sectors,
platform and OptionMetrics. We refer to the IvyDB
and what sort of adjustments may be necessary.
Europe Reference Manual v2.4 for OptionMetrics,
Researchers have documented distinct patterns for
and we refer to the SpiderRock technical notes for
variance risk premium and option skewness for
calculating volatility surfaces and forward rates.
equity index options compared to individual equity
While both firms capture market data and use volatil-
options. Presumably, these discrepancies would
ity surface calculations to report these values, they
translate to differences between breakeven volatility
have different approaches. OptionMetrics has long
and implied volatility. One may also consider expand-
been the standard for academic and institutional
ing the analysis to options traded on other asset
research, whereas SpiderRock is widely used by prac-
classes, such as bond indices or commodity futures.
titioners to support valuation in live trading environ-
Consistent empirical findings across asset classes can
ments and recently started to provide full datasets
prevent overfitting and promote a common explan-
going back at 10 years.
ation for the behavior of option prices.

Data Item OptionMetrics Approach SpiderRock Approach

Marking closing prices and
characteristics
Closing theoretical surface
price for options
Surface-fitting approach to
determine the surface price
Alignment of call and
put surfaces
Forward calculation and
dividend rate calibration
Calculation of time
to expiration
Stock and option close: settlement/last/
bid/ask mid waterfall
Closing prices use the last trade or
closing prices for the close. Implied
volatility calculations use underlying
prices that are time synchronized with
options bid, ask, or last quotes
Option prices used to calculate the
implied volatility are selected as the
first available price in the following: (1)
settlement, (2) last trade, (3) bid-ask
average, (4) bid, (5) ask
Vega-weighted kernel fit to options
implied volatility space. Standardized
options are interpolated from all
options (all terms and strikes) for
that day.
Surface fits for calls and puts are fit
independently. Calls and puts are
interpolated with a weighted scheme
that penalizes when option types
are different.
Forward fit to normalized maturities
using a forward rate curve.
Continuous calendar time
Stock close: bid/ask mid logic
Option: bid/ask/surface
SpiderRock closing prices include the
surface price are taken 1 min before
the option market close. Implied
volatility is synchronized with the
underlying market
Call and put bid and ask markets are
fitted directly as a component of the
volatility surface shape fitting process.
The instantaneous surface price is
determined by aligning the surface
shape with the underlying mid-price
and ATM volatility in real-time.
Price fits have an adaptive number of
spline points. The result is a multi-
point spline that describes a single
volatility curve for both calls and
puts for each expiry. Curves for
normalized maturities are also
generated as part of this process.
Call and put prices are fit
independently, and the price data
are transformed into a single
volatility curve for both calls
and puts.
Forward prices are calculated using a
continuous forward dividend rate
curve with an adjustment based on
aligning call and put volatility as part
of the curve fitting process.
A hybrid time convention that takes
both trading day and continuous
calendar time into account

116
 Option Pricing via Breakeven Volatility

Appendix B
Left-Hand Variable: BEV-IV
The figure shows a histogram of the difference
between breakeven volatility and implied volatility Estimate SE t-Stat
for options on the S&P 500 Index. The sample is
(Intercept) 7.46 0.16 47.0
from January 2013 to November 2020. VVIX 0.043 0.001 48.1
VIX.res 0.188 0.006 30.3
rt 11.22 0.30 37.3
lsk 24.99 0.24 106.0
lsk2 8.82 0.14 61.4
ATM.IV 0.53 0.01 88.4
ImpliedVol (coef  10) 6.03 0.04 156.8
ImpliedVol2 (coef  100) 1.22 0.03 45.0
RR 0.75 0.10 7.6
RR2 36.38 0.34 105.9
RR3 16.68 0.59 28.1
Vega 2.48 0.04 64.8
vega2 (coef  10) 1.03 0.05 22.4
gamma 137.00 14.23 9.6
vrt 15.81 2.01 7.9
vvt2 52.55 0.88 60.0
Appendix C rv.res (coef  10) 2.11 0.01 146.0
This table presents a predictive model for the differ- R-squared 0.13
ence between breakeven volatility and implied Adjusted R-squared 0.13
volatility:
X
M
BEVi  IVi ¼ b0 þ Xi, m bm þ ei
k¼m overvalued or undervalued relative to its
In this specification, rather than first predicting breakeven volatility. We include the same
log(BEV) followed by a transformation, we predictor variables used in the first stage
directly try to predict whether an option is regression as shown in Table 4.

Editor's Note
Submitted 2 February 2022
Accepted 7 July 2022 by William N. Goetzmann

Notes
1. Traditional option pricing models typically try to match
the market price of options, whereas our pricing model
provides the fair value. In this sense, traditional models
are positive models that capture how the world is,
whereas our model is normative.
2. Parametric models, such as Black-Scholes can be
constrained in matching certain empirical patterns.
Stochastic volatility models may have difficulty eliminating
pricing errors in short-term option prices because the
distribution of the underlying does not have enough
kurtosis. Jump models may have issues with longer-term
options because they revert too quickly to the Black-
Scholes prices as the time to expiration increases.
Nonparametric models can be more flexible in capturing
the features of the data, which must be pre-specified and
built into parametric models.
3. For stock indices, deviation from put-call parity is
typically small. For individual stocks that are hard-to-
borrow, deviation from put-call parity could be
much larger.
4. This expression is approximate. To be precise, the
summation is missing a borrowing or lending term equal
to (Dt1 St1  ct1 Þ: At the daily frequency, this term is
very close to zero, especially in a low-interest
environment that our data span.
5. The iterative approach we take was used for a different
purpose by Dumas et al. (1998), who try to determine
whether the delta-hedging component matched the
changes in option prices. They show that the Black-
Scholes model provides the best approximation for option
price changes.

Volume 79, Number 1 117

Financial Analysts Journal | A Publication of CFA Institute
6. Our calculation of breakeven volatility requires knowing
the future path of the option, so it cannot be
implemented in real time. To compute BEV in a real-time
implementation, we would require a predictive model that
uses currently available information to forecast the
eventual BEV values.
7. There are a number of alternatives ways of hedging. One
could choose to hedge at market open rather than at
market close, use delta calculated from the at-the-money
volatility or forecasted volatility, hedge only when the
net delta exceeds a certain threshold, or adjust the
hedge ratio depending on whether the order flow is
believed to be informed. Different hedging methods
correspond to different objectives, and there is no single
best method. On a practical note, the BEV calculation
should match the specific trading desks’ preferred
hedging policy.
8. More frequent delta hedging (i.e., intraday) can
alleviate the issue of large changes in gamma in the
volatility calculations near maturity, especially for
ATM options.
9. Extending our calculation to the expiration date will not
significantly change our results. We do find that
numerical computation of breakeven volatilities in the
last several days just before expiration to be more
difficult, as more volatility calculations do not converge
in the days immediately before expiration. More
frequent delta hedging can improve this
convergence issue.
 rðTtÞ 
log Se K
10. Normalized Strike ¼ ATM:IV p ffiffiffiffiffiffi for underlying price S,
Tt
strike price K, the at-the-money implied volatility ATM:IV,
and time to expiration T  t: The annualized interest rate
r is set to 0.5%.
11. d1 in the Black-Scholes model is given by the following
expression: "    #
 19.
1 St r2
d1 ¼ pffiffiffiffiffiffiffiffiffiffiffi log þ rþ ðTt Þ
r Tt K 2
References
Ait-Sahalia, Y., and A. W. Lo. 1998. “Nonparametric Estimation
of State-Price Densities Implicit in Financial Asset Prices.” The
Journal of Finance 53 (2): 499–547. doi:10.1111/0022-1082.
215228.
Ammann, M., and S. Herriger. 2002. “Relative Implied-Volatility
Arbitrage with Index Options.” Financial Analysts Journal 58 (6):
42–55. doi:10.2469/faj.v58.n6.2485.
Bakshi, G., C. Cao, and Z. Chen. 1997. “Empirical Performance
of Alternative Option Pricing Models.” The Journal of Finance
52 (5): 2003–2049. doi:10.1111/j.1540-6261.1997.tb02749.x.
Bakshi, G., and N. Kapadia. 2003. “Delta-Hedged Gains and the
Negative Market Volatility Risk Premium.” Review of Financial
Studies 16 (2): 527–566. doi:10.1093/rfs/hhg002.
118
where St is the price of the underlying, K is the strike
price, r is the interest rate, r is the volatility of the
underlying, and T  t is the time to expiration.
12. The exponential function, eðÞ , is convex. Jensen’s
inequality states that for a convex function
f, f ðE½XÞ  E½f ðXÞ:
13. We compare our supervised approach with the parametric
methods of Duan (1983) and we find that our approach
achieves the lowest mean squared error.
14. With 396,899 observations and 17 predictors, the
adjusted R-squared is 1.00004 times larger compared to
the unadjusted R-squared.
15. We thank an anonymous referee for this suggestion.
16. If Equation (9) were violated, a costless arbitrage profit
may be earned by writing a put with strike price K2 ,
buying KK33 K
K1 units of the put with strike price K1 , and
2
buying K2 K1
units of the put with strike price K3 :
K3 K1
17. We also explore thresholds of $0.50, $2, and $3, and we
find our results are qualitatively unchanged. A higher
trading threshold has two offsetting effects: We have
more conviction on each trade, but we may miss some
profitable opportunities that do not exceed the threshold.
In our sample, these two effects apparently have similar
magnitudes such that changing the threshold has only a
marginal effect on the strategy performance.
18. Our return calculation is on the conservative side. By
taking the notional value as the denominator, we have
made the implicit assumption that the option positions
are fully collateralized with 100% margin requirement. In
practice, margin requirements typically range between 5
and 30% for puts and calls on the S&P 500 Index.
Private correspondence with option traders indicates that
unless one wants immediate execution, transacting one
tick from the midpoint price is achievable with very high
probability.
Bali, T. G., H. Beckmeyer, M. Moerke, and F. Weigert. 2021.
“Option Return Predictability with Machine Learning and Big
Data.” Available at SSRN 3895984.
Black, F., and M. Scholes. 1973. “The Pricing of Options and
Corporate Liabilities.” Journal of Political Economy 81 (3):
637–654. doi:10.1086/260062.
Bollen, N. P., and R. E. Whaley. 2004. “Does Net Buying
Pressure Affect the Shape of Implied Volatility Functions?” The
Journal of Finance 59 (2): 711–753. doi:10.1111/j.1540-6261.
2004.00647.x.
Bollerslev, T., G. Tauchen, and H. Zhou. 2009. “Expected Stock
Returns and Variance Risk Premia.” Review of Financial Studies
22 (11): 4463–4492. doi:10.1093/rfs/hhp008.

Breeden, D. T., and R. H. Litzenberger. 1978. “Prices of State-
Contingent Claims Implicit in Option Prices.” The Journal of
Business 51 (4): 621–651. doi:10.1086/296025.
Campbell, J. Y., and S. B. Thompson. 2008. “Predicting Excess
Stock Returns out of Sample: Can Anything Beat the Historical
Average?” Review of Financial Studies 21 (4): 1509–1531. doi:
10.1093/rfs/hhm055.
Carr, P., and L. Wu. 2009. “Variance Risk Premiums.” Review of
Financial Studies 22 (3): 1311–1341. doi:10.1093/rfs/hhn038.
Cheney, E. W., and D. R. Kincaid. 2009. Linear Algebra: Theory
and Applications. Jones & Bartlett Learning.
Christoffersen, P., R. Goyenko, K. Jacobs, and M. Karoui. 2018.
“Illiquidity Premia in the Equity Options Market.” The Review of
Financial Studies 31 (3): 811–851. doi:10.1093/rfs/hhx113.
Clark, T. E., and K. D. West. 2007. “Approximately Normal
Tests for Equal Predictive Accuracy in Nested Models.” Journal
of Econometrics 138 (1): 291–311. doi:10.1016/j.jeconom.2006.
05.023.
Coval, J. D., and T. Shumway. 2001. “Expected Option
Returns.” The Journal of Finance 56 (3): 983–1009. doi:10.
1111/0022-1082.00352.
Cox, J. C., R. A. Ross, and M. Rubinstein. 1979. “Option Pricing:
A Simplified Approach.” Journal of Financial Economics 7 (3):
229–263. doi:10.1016/0304-405X(79)90015-1.
Dong, X., Y. Li, D. E. Rapach, and G. Zhou. 2022. “Anomalies
and the Expected Market Returns.” The Journal of Finance 77
(1): 639–681. doi:10.1111/jofi.13099.
Duan, N. 1983. “Smearing Estimate: A Nonparametric
Retransformation Method.” Journal of the American Statistical
Association 78 (383): 605–610. doi:10.1080/01621459.1983.
10478017.
Dumas, B., J. Fleming, and R. E. Whaley. 1998. “Implied
Volatility Functions: Empirical Tests.” Journal of Finance 53 (6):
2059–2106. doi:10.1111/0022-1082.00083.
Dupire, B. 2006. “Fair Skew: Break-Even Volatility Surface.”
Technical Report, Bloomberg L.P.
Figlewski, S., and F. Malik. 2014. “Options on Leveraged ETFs:
A Window on Investor Heterogeneity.” Available at SSRN
2477004.
Garleanu, N., L. H. Pedersen, and A. M. Poteshman. 2009.
“Pricing Options in an Extended Black-Scholes Economy with
Illiquidity: Theory and Empirical Evidence.” Review of Financial
Studies 22 (10): 4259–4299.
Goyenko, R., and C. Zhang. 2020. “The Joint Cross Section of
Option and Stock Returns Predictability with Big Data and
Machine Learning.” Available at SSRN.
Heston, S. L. 1993. “A Closed-Form Solution for Options with
Stochastic Volatility with Applications to Bond and Current
Options.” Review of Financial Studies 6 (2): 327–343. doi:10.
1093/rfs/6.2.327.
Volume 79, Number 1
Option Pricing via Breakeven Volatility
Hull, B., and X. Qiao. 2017. “A Practitioner’s Defense of Return
Predictability.” The Journal of Portfolio Management 43 (3):
60–76. doi:10.3905/jpm.2017.43.3.060.
Hutchinson, J. M., A. W. Lo, and T. Poggio. 1994. “A
Nonparametric Approach to Pricing and Hedging Derivative
Securities via Learning Networks.” The Journal of Finance 49 (3):
851–889. doi:10.1111/j.1540-6261.1994.tb00081.x.
Liu, S., C. W. Oosterlee, and S. M. Bohte. 2019. “Pricing
Options and Computing Implied Volatilities Using Neural
Networks.” Risks 7 (1): 16. doi:10.3390/risks7010016.
Malz, A. M. 2014. “A Simple and Reliable Way to Compute
Option-Based Risk-Neutral Distributions.” FRB of New York
Staff Report 677.
Manzo, G., and X. Qiao. 2021. “Deep Learning Credit Risk
Modeling.” The Journal of Fixed Income 31 (2): 101–127. doi:10.
3905/jfi.2021.1.121.
Merton, R. C. 1973. “Theory of Rational Option Pricing.” The
Bell Journal of Economics and Management Science 4 (1):
141–183. doi:10.2307/3003143.
Mitoulis, N. 2019. “Breakeven Volatility.” Technical Report,
University of Cape Town.
Muravyev, D., and N. D. Pearson. 2020. “Option Trading Costs
Are Lower than You Think.” The Review of Financial Studies 33
(11): 4973–5014. doi:10.1093/rfs/hhaa010.
Neuberger, A. 2012. “Realized Skewness.” Review of Financial
Studies 25 (11): 3423–3455. doi:10.1093/rfs/hhs101.
Orosi, G. 2015. “Estimating Option-Implied Risk-Neutral
Densities: A Novel Parametric Approach.” The Journal of
Derivatives 23 (1): 41–61. doi:10.3905/jod.2015.23.1.041.
Ross, S. 2015. “The Recovery Theorem.” The Journal of Finance
70 (2): 615–648. doi:10.1111/jofi.12092.
Shimko, D. 1993. “The Bounds of Probability.” Risk 6 (4):
33–37.
Stutzer, M. 1996. “A Simple Nonparametric Approach to
Derivative Security Valuation.” The Journal of Finance 51 (5):
1633–1652. doi:10.1111/j.1540-6261.1996.tb05220.x.
Taylor, J. M. 1986. “The Retransformed Mean after a Fitted
Power Transformation.” Journal of the American Statistical
Association 81 (393): 114–118. doi:10.1080/01621459.1986.
10478246.
Welch, I., and A. Goyal. 2008. “A Comprehensive Look at the
Empirical Performance of Equity Premium Prediction.” Review
of Financial Studies 21 (4): 1455–1508. doi:10.1093/rfs/
hhm014.
Zhan, X., B. Han, J. Cao, and Q. Tong. 2022. “Option Return
Predictability.” The Review of Financial Studies 35 (3):
1394–1442. doi:10.1093/rfs/hhab067.
Zou, J., and E. Derman. 1999. Strike-Adjusted Spread: A New
Metric for Estimating the Value of Equity Options. SSRN. https://
ssrn.com/abstract=170629
119
Taylor & Francis takes seriously its contribution to protecting our environment.
In addition to all paper used in our journals being FSC-certified, this journal has gone
plastic-free and no longer uses plastic cover lamination or polywrap for mailing.