Hedge Fund Replication PDF

Hedge Fund Replication
This page intentionally left blank

Edited by
Greg N. Gregoriou
Professor of Finance, State University of New York (Plattsburgh), USA
EDHEC Business School Research Associate, Nice, France
and
Maher Kooli
Professor of Finance, Université du Québec à Montréal, Canada
Editorial matter and selection © Greg N. Gregoriou and Maher Kooli 2012
All remaining chapters © respective authors 2012
Softcover reprint of the hardcover 1st edition 2012 978-0-230-33681-0
All rights reserved. No reproduction, copy or transmission of this
publication may be made without written permission.
No portion of this publication may be reproduced, copied or transmitted
save with written permission or in accordance with the provisions of the
Copyright, Designs and Patents Act 1988, or under the terms of any licence
permitting limited copying issued by the Copyright Licensing Agency,
Saffron House, 6–10 Kirby Street, London EC1N 8TS.
Any person who does any unauthorized act in relation to this publication
may be liable to criminal prosecution and civil claims for damages.
The authors have asserted their rights to be identified as the authors of this
work in accordance with the Copyright, Designs and Patents Act 1988.
First published 2012 by
PALGRAVE MACMILLAN
Palgrave Macmillan in the UK is an imprint of Macmillan Publishers Limited,
registered in England, company number 785998, of Houndmills, Basingstoke,
Hampshire RG21 6XS.
Palgrave Macmillan in the US is a division of St Martin’s Press LLC,
175 Fifth Avenue, New York, NY 10010.
Palgrave Macmillan is the global academic imprint of the above companies
and has companies and representatives throughout the world.
Palgrave® and Macmillan® are registered trademarks in the United States,
the United Kingdom, Europe and other countries.
ISBN 978-1-349-34059-0 ISBN 978-0-230-35831-7 (eBook)
DOI 10.1057/9780230358317
This book is printed on paper suitable for recycling and made from fully
managed and sustained forest sources. Logging, pulping and manufacturing
processes are expected to conform to the environmental regulations of the
country of origin.
A catalogue record for this book is available from the British Library.
A catalog record for this book is available from the Library of Congress.
10 9 8 7 6 5 4 3 2 1
21 20 19 18 17 16 15 14 13 12
Contents
List of Tables vii

List of Figures ix
Preface x
Acknowledgments xii
Editor Biographies xiii
Contributor Biographies xiv
Chapter Overview xx
1 Can We Really “Clone” Hedge Fund Returns? Further

Evidence 1
Maher Kooli and Sameer Sharma
2 Hedge Fund Replication: Does Model Combination Help? 15
Jérôme Teïletche
3 Factor-Based Hedge Fund Replication with Risk Constraints 30
Richard D. F. Harris and Murat Mazibas
4 Takeover Probabilities and the Opportunities for Hedge Funds
and Hedge Fund Replication to Produce Abnormal Gains 48
Anthony Ravi, Peter Mayall, and John Simpson
5 Benchmarking of Replicated Hedge Funds 61
Martin D. Wiethuechter and Lajos Németh
6 Insight—Distributional Hedge Fund Replication via State
Contingent Stochastic Dominance 76
Clemens H. Glaffig
7 Nonparametric Hedge Funds and Replication Indices
Performance Analysis: A Robust Directional Application 90
Laurent Germain, Nicolas Nalpas, and Anne Vanhems
8 Hedge Fund Cloning through State Space Models 106
Roberto Savona
9 Hedge Fund Return Replication via Learning Models 119
R. McFall-Lamm Jr.
v
vi Contents
10 Linear Model for Passive Hedge Fund Replication 133

Giovanni Barone-Adesi and Simone Siragusa
11 Can Hedge Fund-Like Returns be Replicated in a Regulated
Environment? 146
Iliya Markov and Nils S. Tuchschmid
12 A Factor-Based Application to Hedge Fund Replication 159
Marco Rossi and Sergio L. Rodríguez
Index 191
List of Tables
1.1 Selected hedge fund indices and statistics (January

1998–September 2009) 4
1.2 Selected descriptive statistics (monthly) for clone indices:
forecast period February 2004–September 2009 8
1.3 Set of risk factors for fixed-weight and extended Kalman
filter clones 9
2.1 Individual models’ replication properties 24
2.2 Combined models’ replication properties 27
3.1 Summary statistics and time series properties of hedge
fund series 34
3.2 List of assets used in replicating portfolio construction 35
3.3a Out-of-sample evaluation criteria of monthly rebalancing
hedge fund return replicating portfolios 40
3.3b Out-of-sample evaluation criteria of monthly rebalancing
hedge fund return replicating portfolios 42
4.1 Risk–return characteristics: merger arbitrage portfolio 53
2
5.1 In-sample R of five fixed-income strategies 70
2
5.2 In-sample R of eight equally weighted HFR indexes 71
5.3 In-sample R2 of 11 HFR indexes 72
5.4 Out-of-sample evaluation of the replicated clones of Jaeger
and Wagner 73
6.1 Goodness of fit measures, using daily and monthly out-
of-sample return data for the period January 1, 2010 to
December 31, 2010 87
6.2 Moments for the out-of-sample daily return series for the
period January 1, 2010 to December 31, 2010 87
7.1 Monthly statistics 95
7.2 Monthly statistics 96
7.3 Spearman rank correlations with DEA measures 100
7.4 Spearman rank correlations 100
vii
viii List of Tables
7.5 Spearman rank correlations 100

7.6 IQ replication indices ranks within the TASS HF database
(in deciles) 102
8.1 Time-varying beta parameter estimates from January 1998
to December 2006 112
8.2 Clones vs. hedge fund indices from January 1998 to
December 2006 113
8.3 Clones vs. hedge fund indices from January 2007 to
September 2008 115
9.1 Clone vs. actual fund-of-funds returns 126
9.2 Clone vs. actual fund-of-funds returns by year 128
10.1 Univariate statistics of S&P 500 and CSFB Tremont Indexes;
data from July 1996 to December 2010 135
10.2 Regression of the Dow Jones Tremont Hedge Fund Indexes
on S&P 500 index 136
10.3 Performance comparison of 36 rolling window TEV and
MAD clones with six months rebalancing 139
10.4 Performance comparison of 48 rolling window TEV and
MAD clones with six months rebalancing 141
11.1 Strategy breakdown of alternative UCITS funds at the end
of March 2011 153
11.2 Performance of alternative UCITS funds and hedge funds at
the end of March 2011 154
12.1 Variable definitions 161
12.2 Hedge funds’ annualized monthly returns, percent 164
12.3 Risk factors: annualized monthly returns, percent 166
12.4 Simple correlations between hedge funds’ returns and
risk factors 170
12.5 Panel regression: fixed effects and random effects 1 173
12.6 Panel regression: fixed effects by hedge fund investment
strategy 176
12.7 Individual OLS regressions: summary indicators for
estimated coefficients and adjusted R-Sq 179
List of Figures
1.1 Comparison of average Sharpe ratios of clones vs. indices

from February 2004 to September 2009 11
2.1 Predicted and realized hedge fund returns 25
2.2 Forecasting errors 26
3.1 Net asset values of replicating model portfolios and
replicated hedge fund strategies 45
4.1 The profitability of the merger arbitrage strategy 49
4.2 Short-term wealth effects: target shareholders 53
4.3 Short-term wealth effects: acquirer shareholders 54
4.4 The merger arbitrage spread 54
4.5 Short-term profitability: plain vanilla strategy 55
4.6 Short-term profitability: long-only strategy 55
4.7 Value of $1 invested: merger arbitrage strategy 56
6.1 Performance graph for the daily performance from
January 2, 2008 to March 31, 2011 78
6.2 Out-of-sample performance graphs for the period January 1,
2010 to December 31, 2010 88
9.1 The value of past information: rolling windows vs.
constant gain 123
9.2. Clone returns vs. fund-of-funds 128
9.3 Clone net positions 129
9.4 Out-of-sample replication via ETFs 130
12.1 IMF credit outstanding (in SDR billions) 163
ix
Preface
Much has been written in recent years about hedge funds and how they
can generate positive abnormal returns or alpha. However, recently
several academicians have shown that many hedge fund returns are
mostly made up of repackaged beta (or risk factors). If this is true, investors
who are already paying high fees with hedge funds will be tempted to
invest in alternative beta strategies or hedge fund clones. In addition to
lower fees and ease of trading, hedge fund clones also offer transparency.
Investors generally know what is inside a clone’s portfolio, whereas tradi-
tional hedge funds are considered as black boxes. However, while there
may be a consensus in the industry that hedge fund clones will bring
better liquidity and lower fees, it is still debatable whether replication
products should serve as a complement in the hedge fund allocation
decision or as a replacement. Many financial experts also consider
hedge fund clones to be unproven and risky for investors.
The hedge fund clone industry remains very much in the embryo
stage and more academic research is needed for the market to gain more
confidence in such products. Interestingly, hedge fund clones, while
heterogeneous in nature, have performed relatively well during the recent
financial crisis and merit further attention.
This book will be helpful to financial professionals in general, consult-
ants, regulators, advisors, academicians, and students. It is suitable both as
an introduction to hedge fund replication issues and as a reference book
for the empirical analysis of hedge fund clones. Specifically, the book
sheds light on various issues regarding the construction of hedge fund
clones: how we should consider them, their pros and cons; and whether
or not they really expand the efficient frontier for investors.
Chapters 1 and 2 present different approaches to replicating hedge
fund returns. Chapter 3 introduces advanced optimization methodolo-
gies to the hedge fund universe. Chapter 4 investigates the profitability of
the merger arbitrage trading strategy in the Australian market. Chapter 5
examines how a hedge fund actually generates alphas, exploits opportuni-
ties over other market participants, manages and profits from exposure to
market risks (beta), and enhances returns through exposures to alterna-
tive systematic risk. Chapter 6 highlights some of the pitfalls that occur
in real life for both the linear factor and distributional approaches
and introduces an alternative replication scheme, combining aspects of
x
Preface xi
factor replication with an innovative version of distributional replication.

Chapter 7 presents robust directional performance measures and applies
these measures to hedge fund clones. Chapter 8 introduces a Bayesian
State Space Model for hedge fund replication. Chapter 9 demonstrates
that the use of the Kalman filter, a method for incorporating learning that
discounts past information geometrically, produces robust clones that
represent a viable alternative to actually investing in hedge funds. Chapter
10 investigates linear tracking error minimization models for hedge fund
replication. Chapter 11 describes the European Union’s (EU) directive for
Undertakings for Collective Investment in Transferable Securities (UCITS),
which is a regulatory framework that permits the replication of certain
hedge fund-like investment strategies. It presents a summary of the main
aspects of this European framework, offers an analysis of their advantages
and disadvantages, and examines whether the regulations in Europe
impose significant restrictions on hedge fund clones. Chapter 12 estimates
a factor-based model using data from the HFR database, around a time of
heightened market volatility. It shows that limited liquidity, high manage-
ment fees, and poor transparency provide the motivation for replicating
the risk–return profile of hedge funds with liquid assets.
Thus the book offers the reader valuable insights into the thinking
behind hedge fund replication. The numerous international contributors
to this book give the reader a clear and objective overview of the topic.
Acknowledgments
We would like thank the handful of anonymous referees for selecting the
chapters for this book. In addition, we would like thank Lisa von Fircks
Senior Commissioning Editor-Finance and Gemma d’Arcy Hughes,
Senior Editorial Assistant-Finance at Palgrave Macmillan UK. Neither
the editors nor the publisher is responsible for the chapters in this book.
Each author is responsible for his or her own work.
xii
Editor Biographies
Greg N. Gregoriou has published 42 books, 60 refereed publications in

peer-reviewed journals, and 20 book chapters since his arrival at SUNY
(Plattsburgh) in August 2003. Professor Gregoriou’s books have been
published by McGraw-Hill, John Wiley & Sons, Elsevier-Butterworth/
Heinemann, Taylor and Francis/CRC Press, Palgrave Macmillan, and Risk
Books. His articles have appeared in the Review of Asset Pricing Studies,
Journal of Portfolio Management, Journal of Futures Markets, European Journal
of Operational Research, Annals of Operations Research, Computers and
Operations Research, etc. He has also been quoted several times in the New
York Times and the Financial Times of London. Professor Gregoriou is
hedge fund editor and editorial board member for the Journal of Derivatives
and Hedge Funds, as well as editorial board member for the Journal of Wealth
Management, the Journal of Risk Management in Financial Institutions, Market
Integrity, IEB International Journal of Finance, and the Brazilian Business
Review. Professor Gregoriou’s interests focus on hedge funds, funds of
funds, and CTAs. He is an EDHEC Research Associate in Nice, France.
Maher Kooli is Professor of Finance at the School of Management,

Université du Québec à Montreal (UQAM). He is also the finance graduate
programs director and in charge of the Trading room at UQAM. He holds
a PhD in finance from Laval University (Quebec) and was a postdoctoral
researcher in finance at the Center of Interuniversity Research and Analysis
on Organisations. Professor Kooli also worked as a Senior Research Advisor
for la Caisse de Depot et Placement de Québec (CDP Capital). Professor
Kooli has also published articles in a wide variety of books and journals
including the Journal of Future Markets, the Financial Management, the
Journal of Multinational and Financial Management, The Financial Review,
The International Finance Review, The Journal of Private Equity, The Journal
of Wealth Management, The Canadian Investment Review, Derivatives Use and
Trading Regulations, and Journal of Derivatives and Hedge Funds. He has
coauthored three books in financial management and venture capital.
xiii
Contributor Biographies
Giovanni Barone-Adesi is Director of the Swiss Finance Institute at the

University of Lugano. He holds MBA and PhD degrees from the Graduate
School of Business, University of Chicago, under Myron Scholes. In 1981
he joined the Faculty of the University of Alberta, Canada, as Assistant
Professor (1981–3), Associate Professor (1983–7), and Professor (1987–98).
Since 1998 he has been Professor of Financial Theory at the Institute
of Finance of the University of Lugano, Switzerland. During 2001–5 he
was Dean of the Faculty of Economics at the University of Lugano. He is
a referee for the following journals: Journal of Finance, Journal of Financial
Economics, Journal of Financial and Quantitative Analysis, Management
Science, SSHRC, Journal of Banking and Finance, Journal of Financial Research,
Journal of Economics and Business, Journal of Business and Economic Statistics,
Canadian Journal of Administrative Sciences, and Journal of Mathematical
Finance. He is known for his work with Robert Whaley on American
options pricing.
Laurent Germain is Professor of Finance and the Head of the Finance

Group at Toulouse Business School (TBS). He also teaches at ISAE. His
research interests are Market Microstructure, Behavioral Finance, and
Corporate Finance. He is a graduate of TBS, Toulouse School of Economics,
New York University, and Université Paris Dauphine. After post–doctoral
study at the London Business School (LBS) in 1996, financed by the
European Commission, he took the position of Assistant Professor of
Finance at LBS. He left LBS in 2000 to join TBS. He was one of the Directors
of the European Financial Management Association and has published
articles in leading journals such as Review of Financial Studies, Journal of
Financial and Quantitative Analysis, Journal of Financial Intermediation,
and European Financial Management.
Clemens H. Glaffig is principal partner of Panathea Capital Partners

GmbH & Co. KG (an asset management and trading company) and
managing director of DCG GmbH, an analytics company with offices
in Freiburg, Germany. Prior to this, he was head of structured finance
Europe and head of capital markets Central and Northern Europe at CIBC
World Markets, London. Before joining CIBC, Mr Glaffig worked for AIG
Financial Products/Banque AIG in London/Paris and Commerzbank,
Frankfurt. Mr Glaffig held academic positions as visiting professor in
xiv
Contributor Biographies xv
Mathematics at the University of California and research fellow at the

University of Bochum, Germany, after obtaining a PhD in mathemati-
cal physics from the California Institute of Technology, Pasadena, USA,
in 1988.
Richard D. F. Harris is a Professor of Finance in the XFI Centre for

Finance and Investment at the University of Exeter. He has a first class
degree in Economics from University College London, a Master’s degree
with distinction in Economics from Birkbeck College, University of
London, another Master’s degree with distinction in Chinese Language,
Business and International Relations from the University of Sheffield, and
a PhD in Finance from the University of Exeter. Richard is a Fellow of the
Higher Education Academy and a Chartered Member of the Chartered
Institute for Securities and Investment. He has held visiting academic
positions in Belgium, China, New Zealand, Norway, Russia, and Sweden.
Richard’s research interests lie in the areas of financial econometrics and
risk management. He has publications in leading international journals
in economics and finance, including the Economic Journal, Journal of
Econometrics, Journal of Derivatives, Journal of Futures Markets, and Journal
of Banking and Finance. He also has extensive consultancy experience in
the investment banking and fund management sectors, in the fields of
currency and commodity valuation, volatility modelling, and directional
trading strategies. Prior to becoming an academic, Richard worked as a
technical consultant in information technology.
Peter Mayall is a lecturer in finance at the Curtin University of

Technology in Perth, Western Australia. His primary qualification was in
chartered accountancy and he worked in this capacity in his early career
in Africa, the Middle East, and the UK. He then moved to Australia and
switched to the finance industry, being involved in the assessment and
funding of capital projects. He joined academia in 1993 and lectures
in corporate finance, mergers and acquisitions, and financial decision
making. His research interests include the topics of mergers, agency
issues, and the teaching of finance. He has published in the area of the
teaching of finance.
Iliya Markov has an MSc in Operational Research with Finance from

the University of Edinburgh and a BA in Mathematics and Economics
from the American University in Bulgaria. His research interests include
the financial and commodity markets, financial modeling and optimi-
zation, and risk management. He is a recipient of numerous awards and
distinctions, including an Outstanding Achievement in Mathematics
xvi Contributor Biographies
at the American University in Bulgaria and a full scholarship at the

University of Edinburgh. Mr Markov is currently working as a research
assistant at Haute Ecole de Gestion, University of Applied Sciences in
Geneva, Switzerland.
Murat Mazibas is a PhD student in the XFI Centre for Finance and
Investment at the University of Exeter. He has a first class degree in
Economics from Istanbul University, a Master’s degree with distinction
in Econometrics from Gazi University, and another Master’s degree
with distinction in Financial Analysis and Fund Management from
the University of Exeter. Currently he is in the final year of his PhD in
Finance in the University of Exeter and is also carrying on PhD studies
in Financial Mathematics in Turkey. He is a member of CFA Institute
and GARP, and holds CFA and FRM charters. Murat’s research interests lie
in the areas of financial econometrics, risk management, and investment
management. He has numerous professional and academic publica-
tions on hedge funds, operational risk measurement, financial applications
of artificial neural network models, and volatility models. Prior to his
doctoral research, Murat worked for the Turkish Treasury 10 years and for
BRSA as a banking specialist.
R. McFall-Lamm Jr. is Chief Investment Officer at Stelac Advisory Services,

an investment management and consulting boutique in New York. He
was formerly the Chief Investment Officer of the Global Hedge Fund
Group and Chief Investment Strategist at Deutsche Bank in London.
Prior to that, Dr Lamm was Head of Global Portfolio Management at
Bankers Trust in New York.
Nicolas Nalpas is Professor of Finance at Toulouse Business School

(France). He obtained his PhD in financial economics in 2003 at the
University of Paris I Panthéon-Sorbonne (France). He was assistant
professor of finance at HEC Montreal from 2001 to 2004. He obtained
a grant from the French Minister of Research to visit the Department
of Economics of the Ohio State University in 2000. His research
focuses on decision making under risk, asset pricing, and empirical
finance.
Lajos Németh holds an MSc in Management from EBS Business School

at EBS Universität für Wirtschaft und Recht in Wiesbaden and an MBA
from Bond University in Queensland, Australia. The main topics of his
research are corporate finance valuation techniques and fund perform-
ance benchmarking.
Contributor Biographies xvii
Anthony Ravi is a past honors student at the School of Economics and

Finance, Curtin Business School, Curtin University in Western Australia.
His dissertation on using takeover probabilities as a strategy to achieve
abnormal returns in the Australian share market was awarded a high
distinction. He is a high distinction degree student of Curtin University
in financial economics. Anthony is now pursuing a career in corporate
and international finance in the investment banking area.
Sergio L. Rodríguez received his PhD in Economics from UCLA, and

is currently Economist at the International Monetary Fund. Previously,
he worked in Mexico as General Director for Operational Supervision
at CONSAR and as Acting CFO at BANOBRAS. He has been consultant
for the IADB and lecturer at CIDE, Universidad Iberoamericana and
Universidad Panamericana in Mexico City.
Marco Rossi received his PhD in Quantitative Economics jointly from

the Catholic University of Louvain and the London School of Economics.
He is currently Senior Economist at the International Monetary Fund
(IMF). Prior to joining the IMF, he worked at the Bank of England in
its Research Department. He has published on a wide range of topics in
monetary and financial economics and international finance. He is the
author of Payment Systems in the Financial Markets (Macmillan Press
and St. Martin’s Press).
Roberto Savona is Associate Professor of Financial Markets and

Institutions at the University of Brescia, Department of Business Studies,
Italy. After receiving his PhD in finance from University of Udine, Italy
(2002) he visited at Hass School of Business at the University of
California, at Carroll School of Management at Boston College, and at
the Department of Statistics at the University of California. He is a member
of the Board of Directors on the European of Financial Management
Association. His works have been published in Applied Financial Economics,
Economic Notes and some international books.
Sameer Sharma is the Head of International Funds at Investment

Professionals Ltd (IPRO) Fund Management in Mauritius. Sameer
holds a Masters degree in Financial Engineering from the School of
Management, Université du Quebec à Montréal (UQAM) and is also a
Chartered Alternative Investment Analyst. Over the past four years he
has worked in the financial sector in Mauritius and Canada and brings
his quantitative experience in alternative investment strategies and
risk management to IPRO. Sameer has participated and presented at
multiple conferences in the USA and in Canada. He specializes in hedge
xviii Contributor Biographies
fund-related research and consulting while maintaining research links

at École des sciences de la gestion (ESG) at UQAM.
John Simpson is a Professor at Curtin University in Western Australia in

the School of Economics and Finance. He is well published in respect-
able internationally refereed financial economics journals and books. His
research areas include international business risk and international busi-
ness risk management in international finance, banking, and economics.
A more recent research area is the financial economics of energy.
Simone Siragusa obtained his Bachelor’s degree from the University

of Brescia in 2002 and his Master’s in Finance from the University of
Lugano, Switzerland in 2005. Simone is a PhD Student at the University
of Lugano (2006–present), where his supervisor is Professor Barone Adesi
Giovanni. His interests include clustered covariance and copulae, hedge
fund replication models, stochastic volatility models, and numerical
methods. His professional experience is in Risk Management of Market
and Credit Risk. He currently works as Head of Risk Management at
Dinamis Advisors (Switzerland). Dinamis Advisors is an advisory com-
pany managing private and institutional money.
Jérôme Teïletche is the Head of Systematic Strategies and Funds of Hedge

Funds at Lombard Odier in Geneva. He has more than 12 years’ experience
in the financial sector in both investment banking and asset management,
where he has specialized in alternative and quantitative strategies. Jerome
holds a PhD in economics with a specialization in financial econometrics,
and has been publishing in leading peer-reviewed journals in the domain,
such as Journal of Portfolio Management, Journal of Alternative Investment,
Journal of Empirical Finance and Journal of Asset Management. He is adjunct
Professor of Finance at Université Paris Dauphine.
Nils Tuchschmid is Professor of Banking and Finance at Haute Ecole de
Gestion, University of Applied Sciences in Geneva, Switzerland. He is a
member of the Swiss Society of Financial Market Research, an invited
professor at HEC Lausanne University and a lecturer at the University of
Zürich. Nils was previously a Managing Director and a senior member
of the Investment Committee of Alternative Funds Advisory (AFA) at
UBS. Prior to joining UBS, Nils was a Managing Director in the Asset
Management division at Credit Suisse (CS), where he was Head of Multi
Manager Portfolios in Alternative Investments. Nils previously was at
Banque Cantonale Vaudoise (BCV). He worked as Head of Quantitative
Research and Alternative Investments and served as a member of the
Investment Committee of the Swiss regulated AMC Alternative Funds. Prior
Contributor Biographies xix
to this, Nils was Senior Vice-President of Synchrony Asset Management, a

Geneva based company. Nils was Professor of Finance at HEC University
of Lausanne, Switzerland.
Anne Vanhems is Professor of Statistics and Econometrics at Toulouse

Business School (France), and affiliated researcher at Toulouse School of
Economics (TSE). She graduated from ENSAE, Paris, in 1998, obtained
her PhD in applied mathematics in 2001 at the University of Toulouse,
France, and obtained a grant from the Fulbright research program in
2002. Her research interests are structural econometrics, nonparametric
functional estimation, demand analysis, inverse problems, and non-
parametric measures of performances.
Martin D. Wiethuechter is a doctoral research assistant at EBS Business

School at EBS Universität für Wirtschaft und Recht in Wiesbaden,
Germany. He holds a diploma degree in business administration from
the University of Mannheim. In 2008 he also studied at the University
of Michigan. His main research interests are asset liability management,
quantitative investment models, and performance measurement for alter-
native investments. He has authored several articles in finance journals as
well as book chapters.
Chapter Overview
Chapter 1
Can We Really “Clone” Hedge Fund Returns? Further
Evidence
We examine the possibility of creating hedge fund “clones” using liquid

exchange traded instruments. Specifically, we analyze the performance
of fixed weight and Extended Kalman filter (EKF) generated clone
portfolios for 14 hedge fund strategies from February 2004 to September
2009. For five strategies out of 14 we find that our EKF clones outperform
their corresponding indices. Thus, for certain strategies, the possibility
of cloning hedge fund returns is indeed real. Results should be con-
sidered with caution. While our Kalman filter approach does seem to
decrease the tracking error of the clone portfolios when compared to the
fixed weight clone portfolios, we find that index clones are, by construc-
tion, more highly correlated to various asset classes than their actively
managed counterparts. We also find that the rolling correlation between
index and clone portfolios of most strategies may be quite volatile over
time. Our empirical findings suggest that the most important benefits of
clones are to serve as benchmarks and to help investors to better under-
stand the various risk factors that impact hedge fund returns.
Chapter 2
Hedge Fund Replication: Does Model Combination Help?
Jérôme Teïletche
Hedge fund replication is a growing interest in the financial industry.

Most products use factor-based models where one fits a model of hedge
fund returns in terms of investable market factors such as the S&P 500
Index. We investigate whether the combination of methodologies for
estimating factor exposures helps when designing better replicators.
Chapter 3
Factor-Based Hedge Fund Replication with Risk Constraints
We propose a method for hedge fund replication using a factor-based

model supplemented with a series of risk and return constraints that
xx
Chapter Overview xxi
implicitly target all the moments of the hedge fund return distribution.
We use the approach to replicate the monthly returns of ten broad
hedge fund strategy indices, using long-only positions in ten equity,
bond, foreign exchange, and commodity indices, all of which can be
traded using liquid investible instruments such as futures, options, and
exchange-traded funds. In out-of-sample tests, our approach provides an
improvement over the pure factor-based model, offering a closer match
to both the return performance and risk characteristics of the hedge fund
strategy indices.
Chapter 4
Takeover Probabilities and the Opportunities for Hedge Funds
and Hedge Fund Replication to Produce Abnormal Gains
Hedge funds are always seeking to maximize their returns, and one
way is to buy into potential takeover targets. This study calculates
the probability of companies becoming takeover targets and uses these
probabilities to maximize the returns from buying into potential target
companies. Hedge funds and investors involved in hedge fund replica-
tion would do well to invest in target companies where the probabilities
of further takeover plays are higher. This study highlights the areas where
short-term plays should be made to created value maximization based on
Australian evidence.
Chapter 5
Benchmarking of Replicated Hedge Funds
To make a truly informed investment decision, an investor needs to

understand how a hedge fund actually generates returns and exploits
opportunities over other market participants, namely generating alpha.
They also need to manage and profit from exposure to market risks
(beta) and enhance returns through exposures to alternative systematic
risks that do not exist in traditional buy-and-hold portfolios. Several
benchmark concepts have been developed to assist hedge fund investors
to make better decisions, but lots of them are vulnerable to biases which
generate distorting impacts on the suitability of benchmarks. In this
chapter we provide an analysis of several benchmark concepts and will
explain why factor-based benchmarking would help investors to manage
their hedge fund investments effectively.
xxii Chapter Overview
Chapter 6
Insight—Distributional Hedge Fund Replication via State
Contingent Stochastic Dominance
Clemens H. Glaffig
We propose a new hybrid hedge fund replication technique, which com-

bines aspects of portfolio construction from factor-based replication with
an innovative version of distributional replication. It uses a parameter-
ized replicating strategy for which we match a version of state-contingent
integral stochastic dominance. The dominance will be over a set of
distributions reflecting preferred state-contingent distributional perform-
ance characteristics, granting insight into return features to arbitrary
fine detail. It emphasizes the replication of desired aspects rather than
the replication of the performance path. A further application of this
approach is to replicate certain features of a target fund and at the same
time dominate any less desired aspects. Before the new approach to
replication is presented, a brief recollection of its evolution, the various
different approaches, and some of the pitfalls of hedge fund replication
are highlighted.
Chapter 7
Nonparametric Hedge Funds and Replication Indices
Performance Analysis: A Robust Directional Application
The objective of this chapter is to evaluate the performance of hedge

funds and replication indices using recent production frontier methods.
The classical nonparametric DEA (Data Envelopment Analysis) method
suffers from several drawbacks, such as the assumption of strictly posi-
tive inputs and/or outputs, or sensitivity to outliers, and is not adapted
in our context. Using TASS Hedge Funds data from 2004 to 2009 and IQ
replication indices, we analyze rankings with robust directional measures
and compare them with traditional (parametric and nonparametric)
performance measures.
Chapter 8
Hedge Fund Cloning through State Space Models
Roberto Savona
Using a Bayesian State Space Model we clone hedge fund return indices
based on liquid underlying assets easy to replicate through common
financial instruments. Our empirical findings using CSFB/Tremont
Chapter Overview xxiii
indices over the period January 1994 to September 2008 prove that
hedge fund returns can be both cloned and outperformed. In- and out-
of-sample analysis provide convincing evidence about the profitability of
our cloning technology, leading to hedge fund clones with perform-
ances that significantly exceed their hedge fund counterparts on a
risk-adjusted basis.
Chapter 9
Hedge Fund Return Replication via Learning Models
R. McFall-Lamm Jr.
The majority of hedge fund managers modify their trading techniques

over time as market conditions change. They exhibit “learning,” a
process which should be taken into account in clone construction. This
chapter demonstrates that the use of the Kalman filter, a method for
incorporating learning that discounts past information geometrically,
produces robust clones that represent a viable alternative to actually
investing in hedge funds.
Chapter 10
Linear Model for Passive Hedge Fund Replication
Traditional models of factor replication are based on squared error mini-

mization. This is realized through constrained Quadratic Programming
(also called Style Analysis). This optimization technique is widely used
to understand the risk exposure of a hedge fund manager or replicate
his behavior. Starting from a different definition of Tracking Error we
investigate the Mean Absolute Deviation minimization model for hedge
fund replication.
Chapter 11
Can Hedge Fund-Like Returns be Replicated in a Regulated
Environment?
Iliya Markov and Nils Tuchschmid
UCITS are a series of directives issued by the European Commission

whose aim is the facilitation of cross-border marketing and selling of
collective investment schemes. This regulatory framework has recently
gained popularity as an investment vehicle for the replication of hedge
fund-like strategies. This trend can be explained in part by customers
shying away from direct investments into hedge funds and their lack of
xxiv Chapter Overview
transparency. For some, the importance of UCITS is inevitable and the

competition will grow over time. For others, the constraints imposed
by the UCITS regulatory environment will prevent these investment
vehicles from seriously competing with hedge funds. In this chapter,
we analyze the advantages and disadvantages of the UCITS framework.
In particular, we study the different solutions that are proposed and
analyze how adequate they are to offer hedge fund-like returns. For each
strategy or hedge fund style we then emphasize the strengths and weak-
nesses of the Alternative UCITS solution.
Chapter 12
A Factor-Based Application to Hedge Fund Replication
Limited liquidity, high management fees, and poor transparency provide

the motivation for replicating the risk–return profile of hedge funds with
liquid assets. While these replication strategies may not generate the
same alpha as the original hedge fund strategy, they may achieve pay-
offs that are close enough in net terms, that is, when the drawbacks of
investing in hedge funds are taken into account. This chapter estimates
a factor-based model using data from the HFR database. The focus is on
the replication performance across types of hedge fund and across time,
specifically, around times of heightened market volatility.
1
Can We Really “Clone” Hedge
Fund Returns? Further Evidence
1.1 Introduction
While investors generally consider hedge fund investments as pure alpha

products, academic research has shown that hedge funds earn most of
their returns from systematic exposures. Jaeger and Wagner (2005),
among others, argue that hedge fund returns are derived from a mix of
traditional and alternative beta exposures and skill-based returns. Alpha
is simply defined as the part of the returns that cannot be explained by
exposure to systematic risk factors and is a measure of the manager’s
skill. Traditional beta is generated as part of the returns derived from
long-only investing, while alternative beta is the return that can be
specified in a systematic way, but which involves techniques often used
by hedge funds, such as leverage and short-selling (Anson, 2006).
Jaeger and Wagner (2005) find that 80 percent of hedge fund returns
originate as a result of beta exposure (systematic risk factors) and that
only 20 percent is accounted for by the manager’s skill or risk factors
that have yet to be determined. They argue that much of the alpha in
hedge funds is actually repackaged alternative beta. These findings are
consistent with those of Fung and Hsieh (2006), who argue that much
of the alpha of hedge fund returns could be explained by the various
biases that are known to plague those indices. The popular academic
jargon that hedge fund returns are simply beta in alpha clothing is an
important case for passive replication of hedge fund returns. If much of
the return from hedge funds is not true alpha, but rather beta, it may
make more sense to replicate them rather than to invest directly in hedge
funds. Furthermore, hedge funds typically follow the two and twenty
rule when it comes to fees, where the investor pays a 2 percent annual
management fee and 20 percent of the profits that fall above a certain
1
2 Can We Really “Clone” Hedge Fund Returns?
pre-specified hurdle rate. Passive hedge fund replication can, however,

provide a cheaper alternative to investors assuming that true alpha does
not exist on average.1 Hedge fund replicas, or clones as they are often
called, are also compared to investable hedge fund indices. Given the fact
that investable hedge fund indices have inferior performance compared
with non-investable hedge fund indices (because of different biases that
are specific to this family), one would expect that a well-functioning
clone would have superior performance to the former.
With many hedge fund managers now requiring longer lock-up periods,
hedge fund clones have also been marketed as a more liquid alternative
or as a temporary investment in a passively managed hedge fund until a
suitable fund manager can be found. Furthermore, a properly functioning
clone that can be shorted can offer interesting hedging characteristics to
portfolios that allocate to actively managed funds. Indeed, in theory,
making the case for passive hedge fund replication should not be too
difficult if active hedge fund managers can mostly generate beta in alpha
clothing. However, as noted by Wallerstein, Tuchschmid and Zaker (2010),
while there may be a consensus in the industry that clones will bring
better liquidity and lower fees, it is still debatable whether replication
products should serve as a complement in the hedge fund allocation
decision or as a replacement. In general, the authors conclude that the
$2 billion hedge fund replication business is far from dead. Hedge fund
clones, while heterogeneous in nature have performed relatively well
in 2008–9 and merit further attention. However, because the hedge fund
industry remains relatively new, it can be difficult to properly analyze the
return and risk characteristics of such products without relying too heavily
on the back-tested data provided on the replication product brochures.
The purpose of this chapter is to examine the question of whether
hedge fund returns can be replicated. Rather than using fixed-weight
and rolling windows approaches (Hasanhodzic and Lo, 2007), we con-
sider an extended version of the Kalman filter, a computational algorithm
that better captures the time-varying dynamics of hedge fund returns.
Furthermore, in order to be practical, we consider investable factors and
that the models themselves may not be constant over time. In other words,
our only inclusion/exclusion rule when it comes to factor selection is that
they must remain statistically significant throughout the “increasing
window” period which includes both the in-sample and out-of-sample
periods. In different states of the world, hedge funds are not bound
to follow the same strategies over and over again, and hence in order
to ameliorate the out-of-sample performance of the extended Kalman
clone we continuously update the model.
Maher Kooli and Sameer Sharma 3
The rest of this research proceeds as follows: Section 1.2 discusses

the data and the methodology. Section 1.3 presents and discusses the
results. Section 1.4 summarizes and concludes.
1.2 Data and methodology
1.2.1 Data
We use monthly returns data from the EDHEC database from December
1998 to September 2009. We also consider the HFRI (Hedge Fund
Research, Inc.) Fund Weighted Composite Index from the HFR database.
For equity factors, we consider the MSCI-Barra database, while for bond
factors we consider the Credit Suisse and Barclays Capital datasets. We
should note, however, that not all of the factors considered began in
December 1998, as some funds were launched more recently. All factors
are denominated in US dollars and need to be investible, and preferably
exchange traded funds (ETFs) should exist. The forecast period starts in
February 2004 and ends in September 2009. Table 1.1 shows descriptive
statistics.
1.3 Methodology
The factor-based approach is applied to the first generation of hedge

fund replication products. It involves finding appropriate risk factors and
their associated sensitivities to hedge fund returns and aims to construct
a portfolio of long and short positions in a set of risk factors that help to
explain the performance of hedge funds compared with the replicated
portfolio. The clone portfolio can be estimated via the traditional ordi-
nary least squares regression method within an in-sample period and can
be held passively during an out-of-sample period. The performance of the
clone portfolio during the out-of-sample period can then be compared to
the actual hedge fund or a hedge fund index. The factor-based approach
assumes that the return of a clone portfolio is theoretically equal to the
return of the hedge fund if their values agree with probability one.
The factor-based approach is based on two important steps: (1.1)
Calibrating an appropriate factor model within the sample of the form:
K
∑ βjk Fkt εjk
HFj
Rt (1.1)
k=1
where RtHF is the return of hedge fund or a hedge fund index j at time t,
j
b̂jk is the estimated exposure of the return of hedge fund j to factor k,

4
Table 1.1 Selected hedge fund indices and statistics (January 1998–September 2009)
Mean Standard deviation Sharpe Skewness Kurtosis Jacques– Q-stat

annual annual ratio annual Bera p-value (1 lag)
Directional strategy
Long short equity 8.21% 7.68% 0.62 –0.27 4.26 0.01 12.54
Short selling 1.95% 17.58% –0.09 0.09 3.69 0.25 1.62
Emerging markets hedge 12.57% 11.73% 0.78 –0.59 5.64 0.00 17.85
CTA global 7.08% 8.57% 0.42 0.09 2.82 0.84 0.91
Corporate restructuring
Merger arbitrage 7.61% 3.49% 1.19 –0.93 4.50 0.00 16.22
Distressed securities arbitrage 10.39% 6.08% 1.14 –1.19 7.61 0.00 40.23
Event-driven 9.31% 5.93% 0.99 –1.18 6.21 0.00 24.91
Convergence trading
Convertible arbitrage 7.88% 7.31% 0.61 –2.62 18.27 0.00 48.31
Fixed income arbitrage 5.94% 4.38% 0.57 –3.66 26.14 0.00 37.65
Market neutral 6.42% 3.12% 0.95 –3.13 22.51 0.00 9.43
Relative value arbitrage 7.85% 4.64% 0.95 –2.09 12.56 0.00 32.74
Opportunistic strategies
Funds of funds 6.75% 6.05% 0.55 –0.30 6.56 0.00 18.60
Global macro 8.39% 5.22% 0.95 0.49 3.93 0.01 2.12
HFRI composite fund 8.60% 7.39% 0.70 –0.25 4.57 0.00 12.31
weighted index
Fkt is the return of factor k at time t, and êjk is the estimated specific risk
in the return of hedge fund j; and (1.2). Identification of the clone as:
K
∑ βjk Fkt
Clone j
Rt (1.2)
k1
The main challenge of this factor-based process is the selection of

appropriate factors and retaining a proper factor model. If the explanatory
power of the in-sample regression is low, the clone portfolio will not be
expected to perform well during the out-of-sample tests. Since hedge
funds are exposed to a number of risks, a multifactor model is generally
used. Hedge fund returns are typically regressed against asset-based style
(ABS) factors, using the following relation:
In sample return α ∑ (βi ABS factori ) Error term (1.3)
The ABS factors are generally selected in three ways:
1. Arbitrary factor specification. This approach could increase the risk

of under or over specifying the model.
2. Optimization approach. This preferred approach involves the use
of backward or forward step regressions during the factor selection
process.
3. Statistical approach. In order to address concerns about specification
risk, the ABS factors can be obtained by using principal component
analysis (PCA).
Once the appropriate factors are selected, hedge fund returns can be
regressed against the most significant factors and estimates of the beta
coefficients are obtained. The replicating portfolio returns can then be
estimated using the factors and associated beta coefficients with out-of-
sample data, as follows:
Out-of-sample return = ∑ ( βi factori ) (1.4)
We consider two main methods to replicate hedge fund returns:

Fixed weight clone: Depending on data availability, a first trial ordinary
least squares (OLS) regression is run from December 1998 (or February
2000) to January 2004. A stepwise process is followed during the factor
selection process while the factors themselves are chosen based on the
intuition gained from previous academic research. Factor weights from
the calibration period are kept constant throughout the forecast period
when generating the fixed weight clone.
Kalman filter clone: Since hedge funds do not have a constant exposure
over time, it is important to incorporate these time-varying changes
in a multifactor model that aims to replicate hedge fund returns. The
Kalman filter consists of two steps: the prediction step and the updating
step. Its equations can be defined as follows:
k
Rt ∑ βi ,t Fit εt (1.5)
i1
βit1 βi ,t bi ,t1 (1.6)
The bi,t1 and et terms are error terms which are assumed to be inde-
pendently normally distributed with mean zero and constant variance.
This implies that the beta terms evolve over time according to a random
walk. The first equation is called a measurement equation while the
other is called a transition equation. The model assumes that the beta
terms depend on their past values and an error term at any point of
time. One can see that when the error terms of the transition equations
are equal to zero, the time invariant estimation of alpha and beta can be
obtained. As in Swinkels and Van der Sluis (2006), while the filter only
uses past data, it can react rather quickly to changes in the environment,
which makes it highly applicable to the ever-changing hedge fund
industry. Another advantage of the Kalman filter is that it is less sensi-
tive to multicollinearity vs. the traditional OLS approach.
In this chapter, we consider a specific version of the Kalman filter
called the extended Kalman filter (EKF). With this approach, we do not
impose any normality constraints on the error terms, which allows the
filter to find the optimal recursive process by itself. Such models could
adjust even faster to sudden shifts in market conditions vs. a standard
Kalman filter. The measurement and transition equations are allowed to
take any functional forms allowing for a more flexible solution to passive
hedge fund replication. It also provides us with a flexible model to build
on during future research. Further, we do not simply assume that only
the factor weights will change over time but that the factors them-
selves can change throughout the forecast period. Indeed, we periodi-
cally add or remove factors as they become statistically significant or
insignificant, respectively, as long as the remaining/added factors are
significant during the entire increasing window period (December 1998
or February 2000).
1.4 Empirical results
Table 1.2 provides a comparison between the performance of fixed

weight and Kalman clones as well as the original fund indices from
which the clones are derived. In addition, Table 1.3 shows a set of risk
factors for fixed-weight and EKF clones. We notice that for five out of 14
strategies the average mean return of the clones is higher than that of
their index.
EMN for Equity Market Neutral, GM for Global Macro, FIA for Fixed
Income Arbitrage, CTA for Commodities Trading Advisor, RVA for relative
Value Arbitrage, MA for Merger Arbitrage, SS for Short Selling, ED for
Event Driven, EM for emerging Markets, CBA for Convertible Bond
Arbitrage, DSA for Distressed Securities Arbitrage, FOF for fund of Fund,
LSE for Long Short Equity, HFRI for HFR Index.
MSCI Large Cap US: Morgan Stanley Capital Markets Index for large
cap US stocks; MSCI Small Cap US: Morgan Stanley Capital Markets Index
for small cap US stocks; Corporate Bond Index: Barclays US Credit Bond
Index (Investible Grade); DXY: US Dollar Index; CS High Yield Corpo:
Credit Suisse High Yield Bond Fund; MSCI EM: Morgan Stanley Capital
Markets Emerging Markets Standard Index; MSCI EM Value: Morgan
Stanley Capital Markets Emerging Markets Core Value Index; MSCI
Growth: Morgan Stanley Capital Markets US Growth Index; MSCI Value:
Morgan Stanley Capital Markets US Value Index; EAFE Spread: Spread
between the MSCI EAFE index and the MSCI U.S. Large Cap Index;
EM Spread: Spread between the MSCI Emerging Markets Index and the
MSCI Large Cap US Index; Growth-Value Spread: Spread between the
MSCI US Growth Index and the MSCI US Value Index; GSCI: Goldman
Sachs Commodities Index; MBS Index: Barclays Capital Mortgage Backed
Securities Index; Money Market Fund: Fidelity Investments Money
Market Fund; MSCI EAFE: Morgan Stanley Capital Markets Index primarily
invests in developed market stocks ex US (Europe, Australasia, Far East).
EMBI: JP Morgan Emerging Markets Bond Index; MSCI EM Growth:
Morgan Stanley Capital Markets Emerging Markets Growth Index; Size
Spread: The spread between the MSCI Small Cap US and the MSCI Large
Cap US; Ten Year Bond Index: Barclays Capital US 10 Year Treasury Bond
Index; USTIPS: Barclays Capital US Treasury Inflation Protected Index;
Yield Curve: Spread between the 10 Year US Government Treasury Index
and Barclays Capital US Treasury 1–3 Year Index.
Figure 1.1 compares the average Sharpe ratios of clones versus indices.
We find that the average Sharpe ratios of the fixed weight clone are higher
than those of the corresponding indices in six out of 14 cases, while the
8
Table 1.2 Selected descriptive statistics (monthly) for clone indices: forecast period February 2004–September 2009
Hedge fund Fixed weight clone Extended Kalman filter clone

index
Mean Std. Mean Std. Correlation Tracking Information Mean Std. Correlation Tracking Information
dev. dev. to index error Ratio Dev. to Index Error Ratio
EMN 0.31% 1.07% 0.29% 0.57% 0.01 1.21% –0.02 0.39% 0.49% 0.58 0.89% 0.09
GM 0.56% 1.38% 0.44% 1.85% 0.6 1.51% –0.08 0.42% 1.41% 0.76 0.97% –0.14
FIA 0.32% 1.64% 0.44% 0.60% 0.39 1.52% 0.08 0.37% 0.87% 0.63 1.31% 0.04
CTA 0.49% 2.35% 0.21% 1.35% 0.29 2.31% –0.12 0.58% 1.26% 0.32 2.33% 0.04
RVA 0.45% 1.59% 0.15% 1.14% 0.76 1.09% –0.28 0.37% 1.01% 0.87 0.95% –0.08
MA 0.53% 1.04% 0.23% 0.33% 0.51 1.45% –0.2 0.48% 0.82% 0.68 0.77% –0.06
SS 0.20% 3.48% –0.11% 5.24% 0.91 2.56% –0.12 –0.07% 4.82% 0.92 2.15% –0.12
ED 0.59% 1.92% 0.41% 1.05% 0.83 1.16% –0.15 0.36% 1.45% 0.83 1.10% –0.2
EM 0.78% 3.48% 0.67% 3.41% 0.95 1.08% –0.11 0.67% 3.35% 0.95 1.08% –0.11
CBA 0.29% 2.72% 0.46% 1.57% 0.42 2.61% 0.07 0.36% 1.25% 0.67 2.09% 0.03
DSA 0.61% 1.99% 0.42% 1.01% 0.71 1.41% –0.13 0.41% 1.06% 0.73 1.39% –0.14
FOF 0.30% 1.77% 0.00% 1.62% 0.76 1.18% –0.25 0.27% 1.51% 0.82 0.91% –0.03
LSE 0.52% 2.20% 0.28% 1.99% 0.79 1.34% –0.18 0.85% 2.21% 0.93 0.98% 0.34
HFRI 0.50% 2.04% 0.36% 1.96% 0.88 0.99% –0.14 0.47% 2.04% 0.95 0.73% –0.04
Table 1.3 Set of risk factors for fixed-weight and extended Kalman filter clones
Factors used in the Factors (final period) in the

fixed weight clone Extended Kalman Filter
Equity Market EM Spread (MSCI) GSCI
Neutral CS High Yield Corpo CREDITCORPO
EMBI MSCIEMGROWTH
Size Spread Size Spread
Global macro MSCI EM Growth MSCI EM Growth
MSCI Small Cap U.S. Size Spread
MSCI Growth DXY
Ten Year Bond Index
(Barclays)
Fixed Income MSCI EM Value MSCI EM
Arbitrage MSCI EAFE CREDITCORPO
MBS Index USTIPS
DXY
Corporate Bond Index
(Barclays)
MSCI Small Cap U.S.
CTA DXY DXY
GSCI GSCI
Ten Year Bond Index YIELD CURVE
(Barclays) MSCI Large Cap U.S.
MSCI Large Cap U.S.
Relative Value CREDITCORPO MSCI EM Growth
Arbitrage MSCI Large Cap U.S. CREDITCORPO
MSCI Small Cap U.S. Yield Curve
USTIPS
Merger arbitrage Size Spread MSCI EAFE
US TIPS DXY
MSCI Value US TIPS
Yield Curve Size Spread
Short selling MSCI Large Cap U.S. MSCI Large Cap U.S.
MSCI Small Cap U.S. MSCI Small Cap U.S.
Event driven CS High Yield Corpo MSCI Small Cap US
Corporate Bond Index MSCI EM
(Barclays) CREDITCORPO
MSCI Small Cap U.S.
MSCIEM EMBI
Emerging markets MSCI EM Growth MSCI EM Growth
hedge
Convertible bond USTIPS CREDITCORPO
arbitrage MBS INDEX MBS INDEX
YIELDCURVE MSCI SMALL CAP US
(continued)
Table 1.3 Continued
Factors used in the Factors (final period) in the

fixed weight clone Extended Kalman Filter
Distressed EMBI CREDITCORPO
securities Size Spread MSCI EM Growth
arbitrage CREDITCORPO EMBI
MBS INDEX
Fund of hedge Yield Curve CREDITCORPO
funds USTIPS Size Spread
MSCI Small Cap U.S. MSCI EM Growth
MSCI EM Growth GSCI
GSCI Growth Value Spread
MBS Index
Long short equity MSCI Large Cap U.S. Growth Value Spread (MSCI)
MSCI Small Cap U.S. MSCI EM Value
Growth-Value Size Spread
Spread (MSCI) MSCI EAFE
HFRI fund MSCI Growth CREDITCORPO
weighted Size Spread MSCI EM Growth
composite Index CS High Yield Corpo Size Spread
MBS Index MSCI Growth
GSCI
average Sharpe ratios of the Kalman clone are higher than those of the cor-
responding indices in seven out of 14 cases. Overall, we confirm that the
possiblity of cloning hedge fund returns does exist for certain strategies. In
the following section, we discuss our results for each strategy separately.
Our empirical findings suggest that clones generated from multifactor
models look promising for some strategies but remain less successful for
other strategies. Interestingly, for the long/short strategies that account
for more than 40 percent of the asset allocated to hedge funds, we find
that our Kalman clone outperforms the long/short index during the
forecast period.
Furthermore, during the Kalman filter replication process, factors are
removed or added as long as they remain statistically significant through-
out the increasing window. However, even this approach does not guar-
antee that model misspecification disappears. Indeed, by simply looking
at the rolling correlations of many strategies, we have no guarantee that
some factors remain significant over time. Hence, during the factor selec-
tion process, it is important not to deviate too much from factors suggested
by academic research.
Sharpe ratio
–0.02
–0.08
0.18
0.28
0.38
0.48
0.58
0.68
0.78
Figure 1.1
Market Neutral Fixed

Market Neutral Kalman
Market Neutral Index
Rel Value Fixed
Rel Value Kalman
Rel Value Index
Fixed Income Arbitrage Fixed
Fixed Income Arbitrage Kalman
Fixed Income Arbitrage Index
Convertible Arbitrage Fixed
Convertible Arbitrage Kalman
Convertible Arbitrage Index
Event Driven Fixed
Event Driven Kalman
Event Driven Index
Merger Arbitrage Fixed
Merger Arbitrage Kalman
Merger Arbitrage Index
Distressed Securities Fixed
Distressed Securities Kalman
Distressed Securities Index
Short Selling Fixed
Short Selling Kalman
Short Selling Index
Hedge funds strategies

Long/Short Equity Fixed
Long/Short Equity Kalman
Long/Short Equity Index
Emerging Hedge Fixed
Emerging Hedge Kalman
Emerging Hedge Index
Global Macro Fixed
Global Macro Kalman
Global Macro Index
CTA Global Fixed
CTA Global Kalman
Comparison of average Sharpe ratios of clones vs. indices from February 2004 to September 2009
CTA Global Index
FOF Fixed
FOF Kalman
FOF Index
HFRI Fund Weighted Fixed
HFRI Fund Weighted Kalman
HFRI Fund Weighted Index
11
Generally, while we find that it is possible to replicate hedge fund

returns for some strategies, investors should be wary of big promises.
Hedge fund clones could indeed cost less than their actively managed
counterparts, this lower fee does come however at the cost of higher cor-
relations (to other asset classes). In relation to the information ratio, the
tracking error of clones in the hedge fund world may not resemble those
of traditional equity ETFs that are characterized by relatively low tracking
error over time.
Finally, if the tracking error of individual ETFs vs. their respective
benchmarks is not stable over time, then these errors could influence
the performance and tracking error characteristics of the clones to their
respective benchmarks. An ETF on oil, for example, could not always
generate the same returns as the underlying asset when the futures curve
switches from backwardation to contango due to the ETF rebalancing
process. Hence the clone industry will need to be careful and more
dynamic when switching from a model that is based on the factor indices
to investing in the underlying ETFs themselves.
For practitioners who attempt to clone hedge fund returns, we recom-
mend the following points:
1. The factors used during replication need to be investible and in general

replication should be cheap enough in order to be a viable alternative
to active hedge fund management.
2. The clone should have minimal tracking error over time, and where
there is tracking error, the information ratio should be above zero.
3. Correlation with the benchmark index should be relatively constant
over time.
4. Correlation between clones and other asset classes should be similar to
those between the hedge fund indices and these same asset classes.
5. The tracking errors of ETFs used as factors themselves need to be
minimal (vs. their respective benchmarks).
1.5 Conclusion
Much has been made in recent years about the “high fee” structure of
hedge funds. It has been argued that since hedge fund returns are mostly
made up of repackaged beta being sold at alpha prices, such funds can
either be passively replicated at a cheaper cost or/and fund managers
need to offer more competitive pricing. In theory, if the alternative
betas of hedge funds can be passively replicated, hedge fund clones
would be a cheaper and preferred alternative. By using the Kalman filter,
a computational algorithm that better captures the time-varying dynamics

of hedge fund returns, we find that our clones generated from multi-
factor models look promising for some strategies, but a great deal of
work remains to be done for other strategies. Specifically, for seven strat-
egies out of 14 we find that for the 2004–9 period our Kalman clones
outperform their corresponding indices, and that for six strategies out
of 14 our fixed weight clones outperform their corresponding indices.
Thus, for certain strategies, the possibility of cloning hedge fund returns
is indeed real. The results should, however, be considered with caution.
Despite the use of an increasing window method during the factor selec-
tion process of the Kalman clones, the risk of model misspecification
remains relatively high in general, especially during structural breaks in
the systematic risk exposures of hedge funds. For instance, we find that
rolling correlations increased during the 2007–9 subprime crisis period.
Further, there is an increasing interest in the clone industry to replicate
hedge fund returns by following naïve rule-based strategies and reverse
engineering. Replication costs could, however, increase as hedge fund
clone managers attempt to make their models more complex. We leave
these topics for future research.
Overall, despite the fact that the hedge fund clone industry remains
very much in the embryo stage, clones could be used for benchmarking
purposes or in a “core–satellite” framework. For examples, clones could
in some cases serve as alarm bells warning investors of the potential risks
of investing in strategies that are admittedly mostly made up of alterna-
tive beta. Investors could also use clones during the hedge fund selection
and fee negotiation process or in a “core–satellite” framework as an
attempt to separate “pure” alpha from beta. Hence the myths and limits
of passive hedge fund replication should not discourage the growth of
such an industry, but caveat emptor.
Note
1. Marc Hogan notes that “indexes (clones) may help put downward pressures
on fees in the hedge fund arena as they once did for mutual funds” (Business
Week, December 4, 2006).
References
Anson, M. J. (2006). Handbook of Alternative Assets. Hoboken, NJ: John Wiley
and Sons.
Fung, W. and Hsieh, D. (2006). “Hedge Funds: An Industry in its Adolescence.”
Federal Reserve Bank of Atlanta Economic Review, 91(4): 1–34.
Hasanhodzic, J. and Lo, A. W. (2007). “Can Hedge Fund Returns be Replicated?

The Linear Case.” Journal of Investment Management, 5(2): 5–45.
Swinkels, L., and Van Der Sluis P. J. (2006). “Return-Based Style Analysis with
Time-Varying Exposures.” The European Journal of Finance, 12(6–7): 529–52.
Wallerstein, E., Tuchschmid N., and Zaker, S. S. (2010). “How Do Hedge Fund
Clones Manage the Real World?” The Journal of Alternative Investments, 12(3):
37–50.
Jaeger, L. and Wagner C. (2005). “Factor Modeling and Benchmarking of Hedge
Funds: Can Passive Investments in Hedge Fund Strategies Deliver?” The Journal
of Alternative Investments, 8(3): 9–36.
2
Hedge Fund Replication: Does
Model Combination Help?
Jérôme Teïletche
2.1 Introduction
Hedge fund replication is one of the best-known innovations of the

asset management industry in the recent years. Despite some early skepti-
cism from both practitioners and the academic world (see Amenc et al.,
2008), hedge fund clones appear to have been successful in meeting
their replication objectives, though they are still struggling to raise signifi-
cant assets.1 In practice, even if at least three alternative approaches are
in competition (Jaeger, 2007; Kat, 2007), the industry remains mainly
organized around factor-based models.2
The motivation for such models finds its roots in the extensive literature,
which has identified systematic, potentially nonlinear, exposures to tradi-
tional and exotic asset classes in hedge fund returns. On this basis, Jaeger
and Wagner (2005), Hasanhodzic and Lo (2007), and Fung and Hsieh
(2007) have advocated that basic linear models incorporating investable
market indices such as S&P allow one to replicate hedge fund returns
properly. A striking feature is that the vast majority in both the literature
and the industry use a simple ordinary least squares (OLS) approach to
estimate the factor loadings, whereas, hedge fund return samples are not
necessarily well-suited for this. They are notoriously noisy, with frequent
outliers, and they also do not deal with interesting aspects in replication,
such as factor selection or alternative weights of observations (notably
attributing more weight to recent observations).
In this chapter we investigate whether model combination can help
design better hedge fund replicators. We start from the growing literature
on Bayesian model averaging, which has proven in various domains that
models’ forecasting capabilities are generally improved when combined.
Factor-based hedge fund replication is inherently a forecasting exercise,
15
16 Hedge Fund Replication
as one aims to replicate future returns on the basis of models fitted to

historical data. We use three different combination methodologies (thick
modeling, thin modeling, and dynamic mixture), which we apply to
seven different linear regression models (OLS, weighted least squares
with two different weighting schemes, ridge regressions, quantile regres-
sions, stepwise regressions, and robust regressions). Each of these estima-
tion models aims to address specific features of hedge fund returns or
hedge fund replication. We show that model combinations help to design
better models in terms of replication properties. However, there is little
improvement to be gained from using dynamic averaging rather than
simple averaging. The results are robust to alternative modeling choices.
The rest of this chapter is as follows. In Section 2.2 we describe the
methodological framework. In Section 2.3 we present the data and the
results. Finally, we draw some conclusions.
2.2 Methodology
2.2.1 Building and evaluating the replicator

We start with a factor-based regression model of the form:
yt xt′ βt εt , t 1,..., T . (2.1)
yt denotes the monthly hedge fund returns, xt′ are the k factors values
and bt the exposures of hedge funds to those factors at time t. The list
of factors may or may not include an intercept, which would represent in
this context an estimate of the average alpha of hedge funds.3 Following
Hasanhodzic and Lo (2007), we omit the intercept to force the regression
to fit the hedge fund returns mean with factors.
In order to select the market-timing ability of hedge funds, models are
estimated on rolling samples of t months. In conformity with industry
practice and academic literature (see, e.g., Hasanhodzic and Lo, 2007),
we choose t 24. Furthermore, the factors and the hedge fund returns
are all considered as unfunded or self-financed positions. Thus, for
example, yt HFt rft where HFt denotes the hedge fund USD monthly
returns and rft is the risk-free rate for period t. Factors are expressed in
excess of rft or as spreads (for instance, for equity indices). The total
return version of the replicator is then built as:
HF _ Clonet xt′ βt1 rft , t τ 1,..., T . (2.2)
In this model, the replicator is built with the exposures as estimated for
the previous period. The idea is to incorporate the lag between the
Jérôme Teïletche 17
observation of the performance of hedge funds and the implementation

of the positions. This one-month lag implies that we consider exposures
based on the previous month applied at the beginning of the subsequent
month, which assumes that the hedge fund return is available on the first
day of the month.4 By adding that lag, we mimic the constraint that one
faces in building a replicator in real life, which is an inherently out-of-
sample forecasting exercise. One should notice that, following Amenc
et al. (2008), we do not consider the constraints applied by Hasanhodzic
and Lo (2007) that the sum of factor loadings should equal 1 and that the
in-sample volatility of the clone should be similar to that of the cloned
hedge fund index. The budget constraint is respected by the fact that
we assume that the replicator is based on unfunded positions, typically
through futures where the cash deposit is rewarded at the risk-free rate.
Moreover, this constraint seems at odds with hedge fund practice. The
volatility constraint is not necessary either, in the sense that the volatility
of the index is fairly well reproduced by the replicator in-sample (after all,
this is what a regression is supposed to do). Furthermore, the fact that one
targets a volatility in-sample offers no guarantee out-of-sample, that is, in
“real” life.
Starting from the hedge fund index HFt and the replicator HF _ Clonet
time series, we compute various statistics. First, we use strict replication
(
metrics through the linear Pearson correlation r corr HFt , HF _ Clonet , )
( )
the annualized tracking error TE 12 s HFt HF _ Clonet , where s(.)
is the standard deviation, and the average absolute error AAE (T t )
1
Σ Tt t1 HFt HF _ Clonet . Next, we look at hit ratios HIT (T t )

1

Σ Tt t1I (sgn (HF ) sgn (HF _ Clone )) , where I(.) is an indicator function
t t
which takes value 1 when the condition between the brackets is fulfilled
and 0 elsewhere. Finally, we also compare the first four moments (mean,
volatility, skewness, and kurtosis) of the distribution of replicators, their
maximum drawdown, and their Sharpe ratios with those of the original
hedge fund series.
2.2.2 Details of the individual modeling approaches

We describe the list of the various methodologies utilized to estimate
equation (2.1). Following industry practice and academic literature,
the starting model is OLS. All other models are extensions of this
simple model. There are some exceptions in the literature where authors
choose alternative methodologies. Roncalli and Teïletche (2008) advocate
the use of Kalman filters. Amenc et al. (2008) demonstrate that the use
of Markov-switching and Kalman filter models does not significantly
improve the replication strategy. In this chapter, we do not make use

of such time series models, as they can lead to instable results in out-of-
sample contexts because of the limited sample size of hedge fund returns,
while the calibration of these models necessitates rather large samples
and is sensitive to the hypothesis about the distribution of returns.
We start from extensions of OLS, which attempt to deal with several
issues characteristic of hedge fund returns, notably that they are observed
with quite a high amount of noise and that outliers are frequently
observed. Therefore one might be interested in giving more weight to
recent observations so as to be more in line with current hedge fund
portfolios or the use of automatic factor selection methods. In each case,
the dynamic behavior of hedge fund exposures is modeled through
rolling samples, as is traditional. See, for instance, Hastie, Tibshirani, and
Friedman (2008) for more details about the models.
2.2.2.1 Ordinary Least Squares

The first method we use is ordinary least squares (OLS), which despite
its simplicity seems to be widely used in the context of replication. OLS
consist in minimizing the sum of squared residuals (SSR):
t
SSRols ( bt ) ∑ (y ′ bt ) ,
2
m xm (2.3)
mt t1
where bt denotes fitted values of the factors’ loadings at time t for the
rolling sample spanning from t t 1 to t , which is used for the repli-
cator at the time t 1.
2.2.2.2 Weighted Least Squares

A common recommendation is that, in the context of hedge fund
replication, data should be weighted in order to give more emphasis
to recent observations as a way to limit the inertia of the model and to
cope with the implications of the risk of correlation breakdown among
assets. Weighted least squares offers a simple way to deal with that issue
by differently weighting the observations. The solution is obtained by
modifying the SSR accordingly, that is:
t
SSRwls ( bt ) ∑ ′ bt ) ,
wm ( y m xm
2
(2.4)
mt − t1
Here we adopt two weighting schemes. The first one (LW) is based
on a linear decrease in the weight as time elapses. More specifically,
the weight attributed to observation t i , i 0,..., t 1, is equal to

( )
wt i (t i ) / t (t 1) / 2 . The second one (EW) is based on an exponen-
( )
tial decrease, wt i δ i (1 δ ) / 1 δ τ1 with d 2 / (t 1) .
2.2.2.3 Stepwise regression

The preceding models make use of all the factors at each point in time.
Subset selection models aim at reducing the variance of OLS estimator
by taking some bias. The objective is to improve prediction accuracy. The
most common approach is stepwise selection (SW). We use backward–
forward selection starting from an initial model: at each point in time,
the algorithm compares the explanatory power of incrementally larger
and smaller models through F-tests of significance (at the 10 per cent risk
level). The algorithm ends when no single step improves the model.
2.2.2.4 Ridge regression

As they reduce the number of variables, subset methods à la stepwise
regression often imply lower prediction errors than the full model. But
as they are discrete (variables are either retained or discarded), they often
exhibit high variance. One way to deal with this issue is to use shrinkage
methods. Ridge regression (RRG) is a common approach in this perspec-
tive and shrinks the coefficients by imposing a penalty on their size.
More specifically, the ridge coefficients are obtained by minimizing a
modified SSR:
t K
SSRrrg ( bt ) = ∑
mt − t1
( ym xm′ bt )2 l∑ bjt2 ,
j1
(2.5)
The variable l controls for the amount of shrinkage; the larger it is,
the greater the amount of shrinkage. The ridge regression helps to deal
with correlated variables, as large positive coefficients will be canceled
by the large negative coefficients of their counterparts. More generally,
ridge regression will down weight the impact of large coefficients in
absolute terms. In the special case of orthonormal inputs, the ridge esti-
mates are only scaled versions of OLS ones with bridge bols / (1 l ) .
2.2.2.5 Quantile regression

Another consideration for robustness is the sensitivity of the OLS
regression to potential outliers. OLS regression gives an estimate of
the conditional mean of the dependent variable given certain values
of the predictor variables. A quantile regression is a generalization to
any quantile of the distribution of the dependent variable. A common
choice is to model the median, which is similar to the least absolute

deviation (LAD), where one minimizes the sum of absolute values of
deviations:
t
∑
mt t1
′ bt .
y m xm (2.6)
While biased, this estimator is assumed to be more robust, as extreme

observations are less heavily weighted than through the square trans-
formation of OLS.
2.2.2.6 Robust estimators

An alternative approach to robustness is given by M-estimators (ROB).
A general M-estimator minimizes the objective function:
t t
∑
mt t1
f ( em ) ∑
mt t1
′ bt ).
f ( y m xm (2.7)
A special case is given by OLS where f ( et ) et2 . It can be shown that

the optimization of the objective function is similar to solving the system
of equations:
t
∑
mt t + 1
′ bt ) xm
wm ( y m xm ′ 0. (2.8)
This system is solved through iteratively reweighted least squares with

bisquare weights.
2.2.3 Mixing the fitted models

We have seven different estimation models (OLS, LW, EW, SW, RRG,
LAD, ROB), each of them being potentially different in its estimated
factor exposures at each point in time. Let Ml denote model l, l 1,..., L
and blt be the associated factor exposures. We can either combine or
choose between models, that is, we can perform model selection or model
averaging. In each case, the mixed factor loadings will be obtained as a
weighted average ⎯b t Σ lL1wlt blt , where wlt denote the weight accorded to
model l at time t. At each point in time, models can be ranked according
to either a statistical criterion (AIC, R2, ...) or an economic one (tracking
error, Sharpe ratio, ...), and the weights attributed to each model can be
based on this criterion. In this perspective, three main approaches can
be distinguished.
Thin modeling consists in selecting one single model. Thus, at any

point in time, the weights wlt are equal to 0 for all but one whose weight
is equal to 1. This category encompasses real-time selection approaches,
which have been studied many times in the literature (Bossaerts and
Hillion, 1999; Pesaran and Timmermann, 2000).
Thick modeling (Aiolfi and Favero, 2005; Granger and Jeon, 2004)
consists of simply averaging among all the models or a subset of those
which are ranked as best according to a criterion. The idea is to reduce
model uncertainty associated with the choice.
An intermediate approach consists in weighting the models proportion-
ally to their ranks or the values associated with each criterion. The thin
and thick approaches are obtained as special cases.
The Bayesian Model Averaging (BMA) consists in updating the forecaster
priors on each model probability (weight), conditional on observation of
sample data. BMA came to prominence in statistics in the mid-1990s (see
Hoeting et al., 1999, for a survey) and has expanded into various different
fields, including economics and finance, where it has proved useful in
out-of-sample forecasting (see Avramov (2002) or Cremers (2002)). Given
observed data D, BMA computes the weight of model l as the posterior
probability that the lth model is the true model:
(
wlt P Ml D ) (
P D Ml P ( Ml )) , (2.9)
∑ ( )
L
P D Ml P ( Ml )
l1
where
P ( D Ml ) ∫ P ( D θ , Ml ) P ( θ Ml ) d θ (2.10)
( )
is the marginal likelihood of the lth model, P q Ml is the prior density of
( )
the parameter vector in this model and P D q , Ml is the likelihood. The
main shortcoming of BMA is that it is mainly based on in-sample behavior,
while hedge fund replication is inherently an out-of-sample methodology.
The AFTER (Aggregate Forecasting Through Exponential Re-weighting)
algorithm of Yang (2004) deals directly with this issue as weights are
modified recursively according to past forecasting errors; see Hagmann and
Loebb (2006) for an application to stock return predictability. We follow a
similar principle and update our weights according to the following parsi-
monious formula (we refer to “dynamic mix” to qualify this version):
exp(γCl ,t1/2 )
wlt . (2.11)
∑l1
L
exp(γCl ,t1/2 )
Cl , t 1 denotes the metrics used to implement the mix. We concentrate

on tracking error but also use other metrics for comparison. The tracking
error is computed over 12 months, but we find that our results are insen-
sitive to that hypothesis (see below). The variable g is a positive constant
whose role is explained below; while the expression might seem ad hoc,
it has some statistical foundation. In particular, if one uses the BIC
criterion as a metric, the weights in (2.11) correspond to the posterior
probability of each model (Hastie, Tibshirani, and Friedman, 2008).
In the end, the difference between the three models comes mainly
from the amount of model uncertainty one wants to incorporate. In thin
modeling, full confidence is given to the data, whereas in thick modeling,
no confidence is given to the metrics to discriminate in out-of-sample
exercises. The dynamic mix is located between the two, and it is the
parameter g which sets this. In particular, when g 0 , the dynamic mix
is equivalent to thick modeling, as the weight attributed to each model
is the same whatever the values of the various Cl , t 1 are. At the opposite
extreme, when g → ∞ the dynamic mix is equivalent to thin modeling,
as only the model with the lowest Cl , t 1 will have a weight different from
zero. In the results below, we choose g 100 as a reasonable compromise
between thin and thick modeling for the dynamic mix.
Several remarks can be made. First, one may note that if the true model
is part of the set of models, dynamic mix and thin modeling might find
it out, which is not the case for thick modeling. However, the latter
approach might have attractive properties in small samples, thanks to its
robustness. Second, one should observe that the issues raised by hedge
fund replication are very different from those raised by former applica-
tions of model selection/model combination in financial variables, where
one typically aims to design models by selecting appropriate forecasting
variables which will offer better forecasting abilities. Here, we try to select
appropriate econometric specifications in order to achieve good replica-
tion results, either judged in terms of out-of-sample statistical abilities
(correlation, tracking error) or in terms of risk-adjusted performances.5
2.3.1 Data
The hedge fund returns are drawn from the HFR Fund Weighted
Composite Index (HFRI hereafter), published by Hedge Fund Research,
Inc. The index is based on an equally weighted average of funds which
are present in the HFR database (over 2000 funds), and is acting as target
index for many hedge fund replicators.
Like any other non-investable hedge fund indices, the HFRI index is
affected by various forms of biases, notably survivorship and selection
bias, as managers have the option to publish or not their NAV. Regarding
replication, we do not see this issue as detrimental. The main implication
is that it might be difficult to reproduce fully the performance of the index
on an out-of-sample basis. Conversely, replicators might capture part of
the biases as there is ample evidence of hedge fund performance persist-
ence (see, e.g., Boyson, 2008). In that case, the sample of funds used by
the replicator is probably biased toward the best managers, both in the
past and in the near future.
Regarding the set of factors, we select a list which is both character-
istic of hedge fund exposures as stated by previous academic literature
and investable through liquid and cost-efficient listed instruments.
Moreover, all our factors are built as self-financed positions, through
spreads between indices or as excess returns over risk-free rate, the latter
being measured through the 1-month USD Libor rate. More specifically,
we retain the following set of factors:
• SPX: the S&P 500 excess return

• RTY: the spread between the Russell 2000 Index and the S&P 500
total returns
• EAFE: the spread between the MSCI EAFE Index and the S&P 500
total returns
• EMG: the spread between the MSCI Emerging Index and the S&P 500
total returns
• UST: the US Treasury 10-year note excess return
• HY: the spread between the Bank of America–Merrill Lynch US High
Yield Master II index and the US Treasury 10-year note returns
• USD: the US Dollar effective exchange rate excess return
• GSCI: the Goldman Sachs Commodity Index excess return
• VIX: the CBOE VIX excess return
The sample spans the period from January 1990 to December 2010,
with 252 monthly observations. With a window of 24 months for roll-
ing samples in regressions and 12 months for calculations of tracking
error statistics for the dynamic mix, we are left with a sample of 216
out-of-sample forecasts.
2.3.2 Results
The results for individual models are summarized in Table 2.1. For
comparison, we also report the statistics for the HFRI index, which is
24
Table 2.1 Individual models’ replication properties
HFRI OLS LW EW LAD SW RRG ROB

Correlation 100.00% 89.94% 89.78% 89.96% 88.50% 89.06% 89.32% 89.84%
Tracking error 0.00% 3.33% 3.35% 3.32% 3.67% 3.39% 3.35% 3.38%
Hit ratio 100.00% 87.50% 87.04% 87.50% 87.04% 88.43% 87.96% 87.50%
Average monthly absolute error 0.00% 0.79% 0.77% 0.76% 0.84% 0.79% 0.83% 0.80%
Average annual return 10.67% 6.72% 7.14% 7.21% 6.91% 6.07% 5.72% 6.84%
Volatility 7.16% 7.57% 7.57% 7.55% 7.87% 7.34% 7.32% 7.67%
Skewness –0.73 –0.78 –0.73 –0.72 –1.00 –0.81 –0.77 –0.87
Kurtosis 5.731 4.728 4.817 4.784 5.667 4.757 5.009 5.039
Maximum drawdown 21.42% 22.68% 20.94% 20.85% 22.47% 21.85% 22.86% 23.66%
Sharpe ratio 1.07 0.49 0.55 0.56 0.50 0.42 0.37 0.50
the target of replication. When looking at these numbers, the reader

should be reminded that the entire replication methodology is applied
here in an out-of-sample framework, which is more realistic but also
tends to lower the replication statistics. The first rows in Table 2.1 cover
the direct statistics related to replication, that is correlation, tracking
error, average absolute errors, and hit ratios. The first striking feature is
that all models are rather close whatever the replication measure. This
demonstrates that the replication is probably a robust process, in the
sense that it is not highly dependent on the estimation methodology.
Furthermore, the replication process seems quite efficient, as for all
models the correlation and hit ratios seem fairly high. Tracking errors
and average absolute errors are more complex to judge in absolute
terms, but they indicate that hedge fund replication is a complex matter
and that the amount of error is much larger than in equity index replica-
tion, for instance. Looking into the details, it seems that the best model
in terms of replication is the weighted least squares with exponential
weights (EW) while the worst seems to be the median regression (LAD);
the traditional approach adopted in the industry—ordinary least squares
(OLS)—is quite highly ranked.
To illustrate the potential importance of these errors in a more
dynamic manner, we compare the predicted and realized hedge fund
returns in Figure 2.1. In Figure 2.2 the associated errors are displayed.
OLS LW EW
10.0%
LAD SW RRG
8.0% ROB Observed
6.0%
4.0%
2.0%
0.0%
–2.0%
–4.0%
–6.0%
–8.0%
–10.0%
–12.0%
1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
Figure 2.1 Predicted and realized hedge fund returns

OLS LW EW LAD
4.0% SW RRG ROB
2.0%
0.0%
–2.0%
–4.0%
–6.0%
1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
Figure 2.2 Forecasting errors
Errors greater than 2 percent per month are fairly rare and it also seems
that the replication was particularly accurate in the bull market from
2003 to 2007. The crisis period has led to larger errors but ultimately
these are limited when compared with what they would have been in
the 1990s.
The last rows of Table 2.1 report other metrics of comparison among
models. The first striking feature is the difference between the average
returns of the original index and the replicators: more than 3 per cent per
annum for all models. However, it is well known that single hedge fund
indices suffer from notorious biases which are of the same order of
magnitude as the difference we have here (see Fung and Hsieh, 2006).
As a result, hedge fund replicators can definitely appear to be serious com-
petitors for more realistic measures of hedge fund average returns, such as
a fund of hedge fund average returns or investable indices, and in prac-
tice they seem to be (Tuchschmid, Wallerstein, and Zaker, 2010). Higher
moments of the single hedge fund index seem fairly well replicated, even
if volatility is slightly higher. Comparing across models again shows that
exponentially weighted least squares constitute the best model.
Table 2.2 shows the results one get when combining models based on
past tracking errors. The results are similar when applying other met-
rics, such as correlation or average absolute errors, or lags other than 12
months (available upon request). A dynamic mix based on (2.11) offers
Table 2.2 Combined models’ replication properties
All models Four best models

Thick Thin Dynamic Thick Thin Dynamic
Correlation 90.60% 90.27% 90.75% 90.66% 89.91% 90.79%
Tracking error 3.18% 3.30% 3.15% 3.16% 3.30% 3.14%
Hit ratio 87.96% 88.89% 87.50% 88.89% 88.89% 88.89%
Average error 0.77% 0.77% 0.76% 0.75% 0.77% 0.75%
Average return (p.a) 6.66% 6.53% 6.56% 6.79% 6.35% 6.73%
Volatility (p.a) 7.46% 7.64% 7.44% 7.43% 7.47% 7.43%
Skewness –0.84 –0.89 –0.86 –0.79 –0.75 –0.80
Kurtosis 4.956 5.434 5.067 4.766 4.766 4.837
Maximum 22.18% 22.11% 22.32% 21.57% 22.13% 21.68%
Drawdown
Sharpe ratio 0.49 0.46 0.48 0.51 0.45 0.50
the best results, but the difference from thick modeling is small. Larger
differences are obtained with thin modeling, which concentrates on
one single model at each different point in time. The last three columns
illustrate that some improvement in the results can be obtained when
combining only the best individual models (OLS, LW, EW, and SW), but
that the hierarchy of models is unchanged.
2.4 Conclusion
Hedge fund replication is one of the most recent innovations in the asset
management industry. We have shown that model combinations help in
designing better models in terms of replication properties, thus poten-
tially improving existing industry approaches. However, there is little
improvement to be gained by using dynamic averaging rather than simple
averaging. The results are robust to alternative modeling choices.
Acknowledgments
I would like to thank my colleagues Laurent Joué, Marc Pellaud, and

Anne-Valère Amo for their help and suggestions.
Notes
1. See “Hedge Fund Clones Beat Index but Inflows Low,” Financial Times Fund
Management Supplement, 18 April 2011 for industry assessment. See Wallerstein
et al. (2010) and Tuchschmid et al. (2010) for detailed analysis.
2. For instance, among the 18 components of the independent hedge fund

replication index set by Israel Cohen (see http://www.hedgefundreplication.
com/), we estimate that 15 are—at least partially—applying a factor-based
approach.
3. Notice that this alpha might also be considered as an estimate of the various
biases which are known to be very typical of hedge fund indices (notably
self-selection, survivorship, and instant history biases; see Fung and Hsieh,
2006).
4. Notice that a flash estimate of the non-investable HFR index that we use
below is available as soon as the fifth business day, while the performance for
the investable HFRX indices is available every day with a two-day lag. More
generally, thanks notably to the development of managed account platforms,
hedge fund returns are available with shorter notice than in the past.
5. Notice that one could also treat the issue of variable selection in hedge fund
replication. We leave this for future work.
References
Aiolfi, M. and Favero, C. A. (2005). “Model Uncertainty, Thick Modelling and the
Predictability of Stock Returns.” Journal of Forecasting, 24(4): 233–54.
Amenc, N., Géhin, W., Martellini, L., and Meyfredi, J. C. (2008). “Passive Hedge
Fund Replication: A Critical Assessment of Existing Techniques.” Journal of
Alternative Investments, 11(2): 69–83.
Avramov, D. (2002). “Stock Return Predictability and Model Uncertainty.” Journal
of Financial Economics, 64(3): 423–58.
Bossaerts, P. and Hillion, P. (1999). “Implementing Statistical Criteria to Select
Return Forecasting Models: What Do We Learn?” Review of Financial Studies,
12(2): 405–28.
Boyson, N. (2008). “Hedge Fund Performance Persistence: A New Approach.”
Financial Analysts Journal, 64(6): 27–44.
Cremers, K. (2002). “Stock Return Predictability: A Bayesian Model Selection
Perspective.” The Review of Financial Studies, 15(4): 1223–49.
Fung, W. and Hsieh, D. A. (2006). “Hedge Funds: An Industry in its Adolescence.”
Fung, W. and Hsieh, D. A. (2007). “Will Hedge Funds Regress Towards Index-like
Products?” Journal of Investment Management, 5(2): 46–65.
Granger, C. and Jeon, Y. (2004). “Thick Modelling.” Economic Modelling, 21(2):
323–43.
Hagmann, M., and Loebb, J. (2006). “Model Combination and Stock Return
Predictability.” Working Paper, Swiss Finance Institute, Zurich.
Hasanhodzic, J. and Lo, A. (2007), “Can Hedge-Fund Returns be Replicated?:
Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical
Learning: Data Mining, Inference, and Prediction. New York, NY: Springer-Verlag.
Hoeting, J. A., Madigan, D., Raftery, A. E., and Volinsky, C. T. (1999). “Bayesian
Model Averaging: A Tutorial.” Statistical Science, 14(4): 382–417.
Jaeger, L. (2007). “Can Hedge Fund Returns Be Replicated Inexpensively?” CFA
Institute Conference Proceedings Quarterly, 24(3): 28–40.
Jaeger, L. and Wagner, C. (2005). “Factor Modelling and Benchmarking of Hedge

Funds: Can Passive Investments in Hedge Fund Strategies Deliver?” Journal of
Kat, H. (2007). “Alternative Routes to Hedge Fund Return Replication.” Journal of
Wealth Management, 10(3): 25–39.
Pesaran, M. and Timmermann, A. (2000). “A Recursive Modelling Approach to
Predicting UK Stock Returns.” The Economic Journal, 110(460): 159–91.
Roncalli, T. and Teïletche, J. (2008). “An Alternative Approach to Alternative
Beta.” Journal of Financial Transformation, 24(1): 41–50.
Tuchschmid N., Wallerstein E., and Zaker, S. (2010). “Hedge Fund Clones Are Still
to Count On.” Working paper, Geneva Business School, Geneva, Switzerland.
Wallerstein, E., Tuchschmid, N., and Zaker, S. (2010). “How Do Hedge Fund
Clones Manage the Real World?” Journal of Alternative Investments, 12(3):
37–50.
Yang, Y. (2004). “Combining Forecasting Procedures: Some Theoretical Results.”
Econometric Theory, 20(1): 176–222.
3
Factor-Based Hedge Fund
Replication with Risk Constraints
3.1 Introduction
The hedge fund industry has witnessed rapid growth over the last two
decades, from as few as 300 funds in 1990 to about 9,000 funds today.
Although there was a reduction both in the number of hedge funds and
in their average level of leverage during the credit crisis of 2007–8, today
total hedge fund investment amounts to $2.4 trillion (Stowell, 2010).
In parallel with this rapid growth in the hedge fund industry, there has
been increased demand from investors for products that deliver the
returns of hedge funds at lower cost, and without the risks that are typi-
cally associated with hedge fund investment, such as illiquidity, lack
of transparency, and management-specific risks. To meet this demand,
investment banks and asset management firms have developed invest-
ment products, commonly known as “clones,” that seek to replicate
hedge fund returns by employing statistical models or algorithmic
trading strategies.
The dynamics of hedge fund returns are relatively complex owing to
the non-traditional investment strategies and tools that are commonly
used by hedge fund managers, such as leverage, short selling, deriva-
tives, and dynamic trading. This results in a nonlinear relationship
between hedge fund returns and the returns of the major asset classes.
Moreover, it is well established that hedge fund returns are not normally
distributed, with most strategies exhibiting high levels of negative skew-
ness and excess kurtosis, and displaying positive autocorrelation as a
result of holding illiquid assets (see, for example, Brooks and Kat (2002)
and Fung and Hsieh (1997), among others). It is therefore considerably
more challenging to replicate hedge fund returns than it is to repli-
cate, for example, mutual fund returns. Attempts to model hedge fund
30
Richard D. F. Harris and Murat Mazibas 31
returns have searched for assets, styles, or trading rules that can mimic
the strategies that hedge fund managers employ (see, for example, Fung
and Hsieh (2001) and Hasanhodzic and Lo (2007), among others).
Although most hedge fund managers claim that they achieve superior
risk-adjusted performance, and are hence able to justify the high fees
that they commonly charge, some studies estimate that up to 60–80
percent of hedge fund returns can be captured by systematic risk factors
(Jaeger and Wager 2005; Fung and Hsieh, 2006, 2007). The purpose of
replication, therefore, is not to achieve exactly the same level of return
performance, but to capture a significant part of it with lower fees and
better liquidity.
There are three broad approaches to hedge fund replication: the
factor approach, the distribution-matching approach, and the rule-based
approach. The factor approach projects hedge fund returns on to a set
of investible factors, and uses linear regression (Jaeger and Wagner,
2005; Hasanhodzic and Lo, 2007) or nonlinear optimization (Amenc
et al., 2010) to minimize the tracking error between the hedge fund
return and the weighted average return of the factors. Factor approaches
are often able to generate a good fit to hedge fund returns in-sample,
depending on the choice of factors and the time period considered, but
are often found to have poor out-of-sample performance. In particular,
it is commonly found that the replicating portfolio has lower average
return and higher standard deviation (and hence higher risk) than the
hedge fund portfolio that it is designed to track. This is potentially
due to the dynamic nature of the investment strategies that hedge
funds typically employ, which cannot be captured by the essentially
backward-looking factor approach. For a detailed summary of research
in this area, see Amenc et al. (2008), Fung and Hsieh (2004), Mitchell
and Pulvino (2001), Tancar and Viebig (2008), and Takahashi and
Yamamoto (2008), among others. The distribution-matching approach
seeks to replicate the unconditional distribution of the payoffs of the
hedge fund using an equivalent investment in the replicating assets (Kat
and Palaro, 2005, 2006; Papageorgiou, Remillard, and Hocquard, 2008;
Takahashi and Yamamoto, 2010). In contrast with the factor approach,
the distribution-matching approach is relatively robust out-of-sample
in the sense that the higher moments (such as variance, skewness, and
kurtosis) of the replicating portfolio are similar to those of the hedge
fund portfolio. However, as noted by Amenc et al. (2008), by focusing on
the higher moments of hedge fund returns, rather than their time-series
properties, there is nothing to guarantee the out-of-sample perform-
ance of the replicating portfolio, since the first moment of returns
32 Factor-Based Hedge Fund Replication with Risk Constraints
is ignored. Indeed, the time-series correlation between the replicating

portfolio and the hedge fund portfolio is often found to be extremely
low. As such, the distribution-matching approach is more relevant to
fund design than to performance replication (see, for example, Amenc
et al., 2008; Wallerstein, Tuchschmid, and Zaker, 2010). The rule-based
approach seeks to mimic well-known hedge fund strategies by imple-
menting relatively simple trading algorithms that invest in liquid assets
in a way that generates a similar risk–return profile to the hedge fund
being replicated. These algorithms are mainly proprietary in nature, and
therefore there is little academic research concerning their performance.
Mitchell and Pulvino (2001) and Duarte, Longstaff, and Lu (2007)
investigate the risk–return characteristics of merger and fixed income
arbitrage strategies and search for trading rules to mimic these strategies.
In practice, the rule-based approach is often combined with factor-based
replication.
In this chapter, we investigate a composite approach that combines
the factor and distribution-matching methodologies. In particular, we
specify a linear factor model for hedge fund returns to capture their
time-series properties, but impose a range of constraints to ensure that
the replicating portfolio matches various risk measures of the hedge fund,
including conditional value at risk, conditional drawdown at risk, and
the partial moments of returns. These risk measures are nonlinear func-
tions of the higher moments of returns, so our approach can be thought
of as incorporating the distribution-matching approach. We also impose
a return constraint to ensure that the clone portfolio delivers the same
absolute performance as that of the hedge fund. We use this approach
to replicate the monthly returns of ten hedge fund strategy indices
using long-only positions in ten equity, interest rate, exchange rate, and
commodity indices, all of which can be traded using liquid, investible
instruments such as futures, options, and exchange traded funds. Using
out-of-sample evaluation, we show that our composite approach yields
replicating portfolios that better mimic both the risk-adjusted perform-
ance and distributional characteristics of the hedge fund indices that
they are designed to track. On balance, our approach appears to represent
an improvement over the out-of-sample performance of the factor and
payoff distribution approaches reported by Jaeger and Wagner (2005) and
Amenc et al. (2010).
The outline of the remainder of this chapter is as follows. In the fol-
lowing section, we describe the characteristics of the data and outline
the replication methodology. In Section 3.3, we report the results of our
out-of-sample tests. Section 3.4 concludes.
3.2 Data and methodology
3.2.1 Data
We use monthly data on ten hedge fund strategy indices obtained from
Hedge Fund Research, Inc. The hedge fund strategy indices include con-
vertible arbitrage (CA), distressed securities (DS), event-driven (ED), equity
hedge (EH), emerging markets (EM), equity market neutral (EMN),
mergers arbitrage (MA), macro (MAC), relative value (RV), and fund of
funds (FOF) strategies. The full sample covers the period June 1994 to
January 2011 (200 observations), and includes a number of crises (e.g.
Asian financial crisis, the default of the Russian government, the collapse
of Long Term Capital Management, the collapse of the dotcom bubble
and the most recent credit crisis). The initial estimation period is
June 1994 to September 2002 (100 observations), and the out-of-sample
evaluation period is October 2002 to January 2011 (100 observations).
Summary statistics for the hedge fund return series over the full sample
of 200 observations are reported in Table 3.1.
Panel A of Table 3.1 reports various descriptive statistics for the monthly
hedge fund strategy returns. Some strategies (such as emerging markets)
exhibit relatively higher volatility than others (such as equity market
neutral and mergers arbitrage). The returns for all ten hedge fund strate-
gies are leptokurtic, and with the exception of the macro strategy they all
exhibit negative skewness. The null hypothesis of normality is strongly
rejected in all cases. Panel B reports the first five autocorrelation
coefficients of hedge fund strategy returns, the Ljung–Box portmanteau
test for serial correlation up to 10 lags and the ARCH test of Engle (1982).
With the exception of the macro strategy, all ten hedge fund strategies
exhibit highly significant autocorrelations. The ARCH test suggests that
there is evidence of volatility clustering in six of the ten strategies. The
significant autocorrelation in hedge fund returns is largely due to the
artificial smoothing of monthly returns that arises from time lags in
the valuation of the securities held by the hedge fund, especially in less
liquid strategies such as distressed securities. To correct for this autocor-
relation, we use the method of Geltner (1991), originally proposed for
smoothing appraisal-based returns of commercial real estate assets. In
order to replicate the performance of the hedge fund strategy indices,
we use ten equity, bond, commodity, and foreign exchange indices,
which are taken from Datastream. These indices are listed in Table 3.2.
Although not directly investible, they can be traded via a range of low
cost, highly liquid instruments, such as futures, options, and exchange
traded funds.
Table 3.1 Summary statistics and time series properties of hedge fund series 34
Panel A: Summary statistics
Index Mean Median SD Min. Max. Skew E. Kurt. JB Stats pval

Convertible arbitrage 0.008 0.010 0.021 –0.160 0.097 –3.07 26.05 5971.0 0.00
Distressed securities 0.008 0.011 0.018 –0.085 0.056 –1.66 6.31 423.3 0.00
Event-driven 0.009 0.013 0.020 –0.089 0.051 –1.39 4.58 238.6 0.00
Equity hedge 0.010 0.012 0.027 –0.095 0.109 –0.23 2.12 39.0 0.00
Emerging markets 0.009 0.015 0.041 –0.210 0.148 –1.00 4.29 187.0 0.00
Equity market neutral 0.005 0.005 0.009 –0.029 0.036 –0.11 1.49 19.0 0.00
Mergers arbitrage 0.007 0.008 0.011 –0.057 0.031 –1.71 6.31 428.8 0.00
Macro 0.008 0.006 0.019 –0.038 0.068 0.42 0.53 8.2 0.02
Relative value 0.007 0.008 0.013 –0.080 0.039 –3.08 17.03 2731.1 0.00
Fund of funds 0.005 0.007 0.018 –0.075 0.069 –0.75 4.05 155.2 0.00
Panel B: Basic time series properties
Index ARCH pval LB-Q pval ACF(1) ACF(2) ACF(3) ACF(4) ACF(5)
Convertible arbitrage 39.73 0.00 108.48 0.00 0.59 0.29 0.17 0.13 –0.03
Distressed securities 21.96 0.00 90.53 0.00 0.54 0.30 0.19 0.15 0.05
Event-driven 5.12 0.28 42.05 0.00 0.39 0.17 0.12 0.09 0.04
Equity hedge 17.92 0.00 24.13 0.01 0.27 0.15 0.10 0.04 –0.06
Emerging markets 3.15 0.53 34.01 0.00 0.35 0.15 0.09 0.06 0.00
Equity market neutral 16.26 0.00 62.68 0.00 0.17 0.20 0.17 0.19 0.10
Mergers arbitrage 3.00 0.56 42.99 0.00 0.28 0.17 0.17 0.05 0.10
Macro 3.78 0.44 5.45 0.86 0.07 –0.04 –0.02 –0.01 0.04
Relative value 31.00 0.00 72.51 0.00 0.49 0.27 0.14 0.07 –0.03
Fund of funds 13.46 0.01 37.62 0.00 0.37 0.19 0.07 0.00 –0.05
Notes: The table (Panel A) reports summary statistics in percentages for the replicated monthly Hedge Fund Research (HFRI) strategy indices over the
period of June 1994 to January 2011. Panel B reports the autoregressive conditional heteroskedasticity (ARCH) and autocorrelation test results for the
full period. The Ljung–Box Q test for autocorrelation of order upto 10 asymptotically distributed as a central chi-square with 10 d.o.f. under the null
hypothesis, with 5 percent critical value 18.307. ARCH(4) is Engle’s LM test for autoregressive conditional heteroskedasticity, which is asymptotically
distributed as a central chi-square with four d.o.f. under the null hypothesis with 5 percent critical value 9.488. p-values are also reported in the
adjacent columns.
Table 3.2 List of assets used in replicating portfolio construction
Ticker Asset
Equity indices
MSIEMF MSCI Emerging Markets: Investable TR Index
WILDJMI DJ US Micro Cap. Total Stock Market TR Index
WILDJMG DJ US Medium Cap. Growth Total Stock Market TR Index
WILDJMV DJ US Medium Cap. Value Total Stock Market TR Index
WILDJSV DJ US Small Cap. Value Total Stock Market TR Index
Foreign exchange futures
ICDCS CME-Canadian Dollar Cont. Settlement Price
Bonds index
LHTBW3M BARCLAYS US Treasury Bellwethers 3M
Commodity indices
GSCI S&P GSCI Commodity TR Index
GSEN S&P GSCI Energy TR Index
GSPM S&P GSCI Precious Metal TR Index
Notes: The table lists stock, fixed income, commodity, and foreign exchange rate assets used in
replicating portfolio construction. All indices are total return index and there are tradable highly
liquid instruments on these indices (futures, ETFs, etc.). All assets are traded in US dollars.
3.2.2 Methodology
Our starting point is the factor-based approach to replicating hedge fund
returns, but we supplement this with a number of constraints on the
return and risk of the replicating portfolio, and hence indirectly on its dis-
tributional characteristics. Specifically, the objective function is given by:
min f ( x) var( rhf ,t rp ,t ) (3.1)

x
subject to
m
∑ x 1,
i1
i i 1,..., m (3.2)
xi 0 (3.3)
m
∑r x r
i1
i i hf (3.4)
CVaR p CVaR hf (3.5)
CDaR p CDaR hf (3.6)
UPM p = UPMhf (3.7)

LPM p LPMhf (3.8)
where rhf , t is the return of the hedge fund index at time t, xi , i 1,..., m , is
the weight of instrument i in the replicating portfolio, and rp ,t Σ im1ri ,t xi
is the return of the replicating portfolio at time t. The budget constraint
(3.2) represents full investment without leverage, while the positivity
constraint (3.3) ensures long-only portfolio positions. The constraint
(3.4) matches the mean return of the replicating portfolio with the mean
return of the hedge fund index, which addresses the return component
of the risk-adjusted performance of the replicating portfolio. The remain-
ing constraints concern the risk of the replicating portfolio, and are
described in detail in the following sub-sections.
3.2.2.1 CVaR constraint

Conditional value at risk (CVaR) is a risk measure derived from value
at risk (VaR) and can be defined as the expected value of losses exceed-
ing VaR over a specified time horizon at a specified confidence level
(Rockfeller and Uryasev, 2002). Let z be VaR with confidence level a. We
define CVaR for the replicating portfolio as
⎪⎧ ⎫
N m
∑ ∑ r x z⎪⎬⎪⎭
1 1
CVaR p ,a ( x) z max ⎨0, ij i (3.9)
1−a N ⎪⎩
j1 i1
and for the replicated hedge fund strategy index as

N
∑ max { 0, r }
1 1
CVaR hf ,a z hf , j z (3.10)
1−a N
j1
where z is estimated from the returns of the hedge fund strategy and
CVaRp,a ( x) is a convex function of portfolio positions with respect to a.
The constraint in (3.5) therefore aims to match the CVaR of the repli-
cating portfolio in (3.9) with that of the individual hedge fund strategy
in (3.10).
3.2.2.2 CDaR constraint

Drawdown, also known as the underwater portfolio level, is a com-
monly used performance indicator in portfolio management, and is
defined as the reduction in portfolio value from a previous maximum.
The drawdown concept helps investors construct portfolios in a way
that avoids losses that exceed a fixed percentage of the maximum
value of their wealth achieved up to that point in time. Chekhlov,
Uryasev, and Zabarankin (2000) propose the conditional drawdown at
risk (CDaR) measure, which combines the drawdown concept with the
CVaR approach. Analogous to CVaR, CDaR can be defined as the expec-
tation of drawdowns that exceed a certain threshold drawdown level,
z , which is defined at an a-confidence level similar to the way VaR is
defined in the specification of CVaR. However, unlike CVaR, CDaR is
a risk measure that accounts not only for the aggregate of losses over
some period, but also for the sequence of those losses. For portfolio
implementation of CDaR, see, for example, Chekhlov, Uryasev, and
Zabarankin (2005).
Let z be threshold drawdown level estimated at confidence level a.
We define CDaR for the replicating portfolio as
⎡ ⎡ m ⎛ k ⎞ ⎤ ⎤
⎢0, max ⎢
⎢ 0 k j ⎢ ∑∑
⎜
⎜
ris ⎟ xi ⎥
⎟ ⎥
⎥
⎥
⎣ i1 ⎝ s1 ⎠ ⎦
N
∑
1 1
CDaR p ,a ( x) z max ⎢ ⎥ (3.11)
1−a N ⎢ m ⎛ j ⎞ ⎥
⎢
∑∑ ⎥
j1
⎜ ris ⎟ xi z
⎢ ⎜ ⎟⎠ ⎥
⎢⎣ i1 ⎝ s1 ⎥⎦
and for the hedge fund strategy index as
⎡ ⎛ k ⎞ ⎤
N
⎢0, max ⎜
⎢ 0k j ⎜⎝ s1 ∑
rhf ,s ⎟
⎟⎠
⎥
⎥
∑
1 1
CDaR hf ,a ( x) z max ⎢ ⎥ (3.12)
1−a N ⎢ j ⎥
∑
j1
⎢ rhf ,s z ⎥
⎢⎣ s1
⎥⎦
where z is estimated from the drawdowns of the hedge fund strategy

and CDaR p ,a ( x) is a convex function of portfolio positions with respect
to a. The constraint in (3.6) therefore aims to match the CDaR of the
replicating portfolio in (3.11) with that of the individual hedge fund
strategy in (3.12).
3.2.2.3 Partial moments constraint

The constraints on the partial moments in (3.7) and (3.8) are moti-
vated by the Omega performance measure first introduced by Keating
and Shadwick (2002). Omega is a generalized measure of risk-adjusted
return that implicitly utilizes all moments of the distribution of portfo-
lio returns, rather than focusing merely on the mean and the variance,
and is defined as the upper partial moment of returns with respect to
some threshold, divided by the lower partial moment of returns. Here
we consider the components of Omega—the upper and lower partial
moments—separately. In particular, we define upper and lower partial

moments for the return rp as the probability weighted ratio of portfolio
gains and losses relative to a threshold return defined by the investor:
∫ (1 F (y )) dx E
rmax
UPM p (rb ) ⎡max (0, rp rb )⎤ (3.13)
rb
P ⎣ ⎦
rb
F ( y ) dx EP ⎡⎣max (0, rb rp )⎤⎦
LPM p (rb )
∫
rmin
(3.14)
where F(.) is the cumulative probability distribution function of the

portfolio returns, F ( y ) P ⎡⎣rp y ⎤⎦. Given the threshold return level rb, we
define the UPM and LPM functions of the replicating portfolio in (3.7)
and (3.8) as
UPM p E ⎡⎣rp | rp rb ⎤⎦ rb (3.15)
LPM p rb E ⎡⎣rp | rp rb ⎤⎦ (3.16)
and the UPM and LPM functions of the replicated hedge fund strategy
in (3.7) and (3.8) as follows:
UPMhf E ⎡⎣rhf | rhf rb ⎤⎦ rb (3.17)
LPM hf rb E ⎡⎣rhf | rhf rb ⎤⎦ (3.18)
where rhf is the returns of individual hedge fund strategy.
3.2.3 Estimation and evaluation

We test the out-of-sample performance of the replicating portfolios with
different constraints over the period October 2002 to January 2011. We
report portfolio performance for a number of different specifications
of the model. FM is the pure factor model, which imposes only the
full investment constraint. RC is the pure factor model supplemented
with the long-only and average return constraints. CVaRC, CDaRC,
and PMC each have one additional constraint (on CVaR, CDaR, or the
upper and lower partial moments, respectively). ALLC imposes all of
the constraints simultaneously. We initially estimate the model using
the first 100 months, June 1994 to September 2002, to generate out-of-
sample forecasts of the replicating portfolio weights for October 2002.
The estimation sample is then rolled forward one month to forecast
the portfolio weights for November 2002, and so on until the end of
the sample is reached. The estimation window length is kept constant at

100 months. The model is estimated using the Matlab fmincon optimiza-
tion function. In estimating the CVaR and CDaR constraints a 95 percent
confidence level is used.
We evaluate the out-of-sample performance of the replicating port-
folios using a number of statistical and economic measures. Firstly, we
estimate a regression of realized hedge fund portfolio returns on the
realized replicating portfolio returns over the out-of-sample period. For
brevity, we provide only the adjusted R-squared statistic and beta coeffi-
cient of this regression. Secondly, we report the mean, standard deviation,
and skewness and kurtosis coefficients of the return distribution for both
the hedge fund portfolio and the replicating portfolio. Thirdly, we report
the Sharpe ratio, maximum drawdown, and annualized CVaR and CDaR
statistics for both the hedge fund portfolio and the replicating portfolio.
In computing the Sharpe ratio, we use an annualized risk-free rate of 2.03
percent, representing the average yield on US Treasury securities at a
constant 3-month maturity over the out-of-sample period.
The regression results are reported in columns 1–3 of Table 3.3a and 3.3b.
Generally, the estimated beta coefficient is significantly greater than
zero, and the adjusted R-squared values are relatively high. The factor-
based model generates a beta value closer to one than all other models.
For some strategies (such as emerging markets), the FM model is able to
explain up to 85 percent of the variance in hedge fund returns. However,
the highest R-squared statistic is generated by the PMC model in six out
of ten cases, but by the FM model in only three cases. The CVaRC model
generates the second highest adjusted R-squared statistic in seven out of
ten strategies. In terms of statistical performance, therefore, the replicat-
ing portfolios in many cases display a significant improvement over the
out-of-sample performance of the factor-based model. However, for some
strategies, such as equity market neutral, the relatively low level of the
systematic component is detrimental to the performance of the replicat-
ing portfolios.
The annualized mean and standard deviation of returns and the
skewness and kurtosis coefficients are reported in columns 4–7 of Table
3.3a and 3.3b. In general, the composite model replicates the statistical
properties of hedge fund returns in terms of their first four moments
reasonably well. Composite models tend to offer slightly higher aver-
age returns relative to the hedge fund strategies, but also slightly higher
Table 3.3a Out-of-sample evaluation criteria of monthly rebalancing hedge fund return replicating portfolios
40
Contraint Beta tstat adjR² AR SD Skew Kurt SR MDD CVaR CDaR

Panel 1: Convertible arbitrage 0.07 0.19 –1.44 11.42 0.07 0.73 1.16 0.05
CVaRC 0.75 6.86 31.8% 0.12 0.14 –0.82 5.64 0.19 0.51 0.61 0.04
CDaRC 0.85 6.36 28.5% 0.12 0.12 –0.07 3.28 0.23 0.30 0.34 0.03
PMC 0.58 7.15 33.6% 0.14 0.19 –0.95 5.82 0.18 0.76 0.79 0.05
ALLC 0.57 4.91 19.0% 0.05 0.15 –0.38 5.02 0.06 0.63 0.44 0.05
RC 0.72 4.20 14.4% 0.11 0.10 –0.02 3.64 0.25 0.23 0.31 0.02
FM 1.11 5.70 24.1% 0.05 0.09 –2.09 15.61 0.09 0.31 0.56 0.03
Panel 2: Distressed securities 0.11 0.12 –0.87 4.99 0.21 0.39 0.49 0.03
CVaRC 0.60 9.61 48.0% 0.11 0.14 –1.73 11.43 0.18 0.47 0.74 0.04
CDaRC 0.58 8.67 42.8% 0.12 0.13 –1.28 9.03 0.22 0.38 0.67 0.03
PMC 0.45 11.71 57.9% 0.12 0.20 –1.37 8.81 0.15 0.79 1.01 0.05
ALLC 0.56 7.98 38.8% 0.09 0.13 –0.76 5.21 0.16 0.42 0.53 0.04
RC 0.61 9.81 49.0% 0.11 0.13 –1.53 10.60 0.19 0.43 0.71 0.04
FM 0.90 9.84 49.2% 0.06 0.09 –1.32 8.46 0.12 0.40 0.45 0.03
Panel 3: Event-driven 0.10 0.10 –1.14 5.30 0.24 0.31 0.46 0.03
CVaRC 0.58 14.22 67.0% 0.11 0.14 –1.67 10.38 0.19 0.52 0.71 0.04
CDaRC 0.60 13.78 65.6% 0.10 0.13 –1.10 6.88 0.19 0.37 0.60 0.03
PMC 0.35 10.77 53.7% 0.12 0.20 –1.25 7.00 0.14 0.79 0.95 0.05
ALLC 0.62 12.00 59.1% 0.09 0.12 –1.07 7.13 0.17 0.38 0.54 0.03
RC 0.51 11.08 55.1% 0.11 0.14 –1.00 7.35 0.19 0.47 0.66 0.04
FM 0.84 16.37 72.9% 0.06 0.10 –1.60 9.57 0.11 0.44 0.52 0.04
Panel 4: Equity hedge 0.08 0.11 –0.84 4.29 0.15 0.39 0.46 0.03
CVaRC 0.68 15.77 71.4% 0.09 0.14 –1.33 7.63 0.14 0.54 0.66 0.04
CDaRC 0.67 15.11 69.7% 0.10 0.14 –0.74 5.10 0.17 0.40 0.56 0.03
PMC 0.39 11.84 58.5% 0.10 0.22 –1.05 5.97 0.11 0.89 1.06 0.06
ALLC 0.74 15.37 70.4% 0.09 0.13 –1.11 6.18 0.17 0.46 0.59 0.04
RC 0.39 7.83 37.8% 0.08 0.18 –0.59 7.45 0.10 0.47 0.74 0.04
FM 0.92 19.38 79.1% 0.06 0.11 –2.09 13.45 0.12 0.46 0.66 0.04
Panel 5: Emerging markets 0.15 0.17 –1.00 4.55 0.22 0.56 0.69 0.04
CVaRC 0.60 11.79 58.2% 0.13 0.22 –1.73 11.11 0.14 0.90 1.29 0.06
CDaRC 0.80 12.67 61.7% 0.16 0.17 –1.45 12.10 0.25 0.45 0.89 0.04
PMC 0.43 9.39 46.8% 0.05 0.27 –1.68 9.83 0.03 1.57 1.69 0.07
ALLC 0.73 10.47 52.3% 0.11 0.17 –1.13 8.17 0.16 0.60 0.84 0.05
RC 0.74 15.94 71.9% 0.17 0.19 –1.20 8.19 0.23 0.60 0.92 0.05
FM 0.88 23.44 84.7% 0.16 0.17 –1.45 7.86 0.23 0.70 0.88 0.05
Notes: The table reports evaluation criteria for the out-of-sample monthly rebalancing replicating portfolios of HFRI indices in the period October
2002 to January 2011 (100 months). Evaluation criteria include regression results (i.e. beta, t statistics (tstat) of beta coefficient and adjusted R square
(adjR2)), first four moments (i.e. annualized average return (AR), annualized standard deviation (SD), skewness (Skew), kurtosis (Kurt)), Sharp Ratio
(SR), and risk measures (i.e. maximum drawdown (MDD), annualized conditional value at risk (CVaR), and annual conditional drawdown at risk
(CDaR)). CVaR and CDaR statistics are estimated at a 99 percent confidence level.
41
Table 3.3b Out-of-sample evaluation criteria of monthly rebalancing hedge fund return replicating portfolios
42
Contraint Beta tstat adjR² AR SD Skew Kurt SR MDD CVaR CDaR

Panel 6: Equity market neutral 0.03 0.03 –1.15 5.28 0.07 0.10 0.15 0.01
CVaRC 0.14 3.52 10.3% 0.06 0.07 –1.33 10.43 0.15 0.18 0.39 0.02
CDaRC 0.10 2.42 4.7% 0.08 0.08 –0.28 3.75 0.23 0.12 0.24 0.01
PMC 0.10 4.70 17.5% 0.10 0.13 –1.79 10.88 0.16 0.57 0.73 0.04
ALLC 0.13 2.64 5.7% 0.03 0.06 –0.83 5.91 0.05 0.18 0.27 0.02
RC 0.13 3.00 7.5% 0.07 0.07 –1.09 8.43 0.19 0.16 0.33 0.02
FM 0.19 1.32 0.7% 0.00 0.02 –3.07 22.95 –0.30 0.09 0.15 0.01
Panel 7: Mergers arbitrage 0.06 0.05 –0.82 4.37 0.26 0.10 0.14 0.01
CVaRC 0.34 8.29 40.6% 0.07 0.09 –1.53 10.15 0.18 0.22 0.44 0.02
CDaRC 0.30 6.76 31.1% 0.08 0.09 –0.69 5.86 0.21 0.17 0.36 0.02
PMC 0.19 9.04 44.9% 0.10 0.16 –1.60 8.58 0.15 0.68 0.82 0.05
ALLC 0.37 7.46 35.6% 0.05 0.07 –0.91 6.12 0.12 0.20 0.31 0.02
RC 0.36 8.73 43.2% 0.08 0.08 –1.02 7.54 0.21 0.20 0.39 0.02
FM 0.62 8.29 40.6% 0.02 0.05 –2.21 13.74 –0.02 0.20 0.31 0.02
Panel 8: Macro 0.08 0.06 0.26 3.16 0.30 0.05 0.18 0.01
CVaRC 0.28 6.23 27.6% 0.08 0.11 –1.82 11.67 0.16 0.31 0.61 0.03
CDaRC 0.21 4.52 16.4% 0.12 0.12 –0.27 5.20 0.25 0.18 0.43 0.02
PMC 0.18 6.17 27.3% 0.11 0.17 –2.03 12.48 0.15 0.59 0.96 0.05
ALLC 0.31 5.30 21.5% 0.08 0.09 –1.09 7.06 0.18 0.19 0.41 0.02
RC 0.29 6.61 30.1% 0.11 0.11 –1.43 10.16 0.23 0.28 0.58 0.03
FM 0.60 5.96 25.8% 0.04 0.05 –2.29 16.57 0.09 0.16 0.30 0.02
Panel 9: Relative value 0.07 0.08 –2.20 11.97 0.18 0.28 0.58 0.03
CVaRC 0.54 8.78 43.4% 0.09 0.10 –1.55 10.09 0.20 0.31 0.53 0.03
CDaRC 0.41 5.78 24.7% 0.10 0.10 –0.67 4.64 0.22 0.26 0.40 0.02
PMC 0.37 10.58 52.9% 0.12 0.16 –1.37 7.52 0.18 0.68 0.79 0.05
ALLC 0.46 6.46 29.2% 0.07 0.10 –0.80 4.84 0.14 0.36 0.38 0.03
RC 0.56 8.51 41.9% 0.09 0.10 –1.31 9.25 0.21 0.27 0.48 0.03
FM 0.97 8.69 42.9% 0.04 0.06 –1.87 11.85 0.08 0.23 0.33 0.02
Panel 10: Fund of funds 0.05 0.08 –1.22 5.52 0.10 0.28 0.40 0.03
CVaRC 0.59 12.40 60.7% 0.08 0.11 –1.76 10.87 0.15 0.43 0.58 0.04
CDaRC 0.62 9.86 49.3% 0.08 0.09 –0.96 7.19 0.19 0.27 0.42 0.03
PMC 0.39 12.88 62.5% 0.12 0.16 –1.64 9.37 0.17 0.69 0.92 0.05
ALLC 0.47 7.71 37.1% 0.06 0.11 –0.93 4.61 0.11 0.40 0.46 0.03
RC 0.73 12.00 59.1% 0.07 0.09 –1.37 9.37 0.16 0.30 0.44 0.03
FM 0.86 11.68 57.8% 0.04 0.07 –2.54 16.84 0.10 0.32 0.46 0.03
Notes: The table reports evaluation criteria for the out-of-sample monthly rebalancing replicating portfolios of HFRI indices in the period October
2002 to January 2011 (100 months). Evaluation criteria include regression results (i.e. beta, t statistics (tstat) of beta coefficient and adjusted R square
(adjR2)), first four moments (i.e. annualized average return (AR), annualized standard deviation (SD), skewness (Skew), kurtosis (Kurt)), Sharp Ratio
(SR), and risk measures (i.e. maximum drawdown (MDD), annualized conditional value at risk (CVaR), and annual conditional drawdown at risk
(CDaR)). CVaR and CDaR statistics are estimated at a 99 percent confidence level.
43
standard deviation. In contrast, the benchmark FM model tends to offer

lower average returns than the hedge fund strategies, and lower standard
deviation. Both the benchmark FM model and the composite models
approximately match the skewness and kurtosis of hedge fund returns.
The risk-adjusted performance of the models is given by the Sharpe
ratio, reported in column 8 of Table 3.3a and 3.3b. The composite
models, ALLC, CVaRC, and CDaRC, offer the best replication of risk-
adjusted performance. The FM model underperforms the hedge fund
strategies in seven out of ten cases, and in two cases actually yields a
negative Sharpe ratio. In contrast, the CDaRC model outperforms the
hedge fund strategies in seven out of ten cases, and where it underper-
forms, the differences are relatively small. The success of the composite
model over the pure factor-based model can be attributed to the different
objective functions of the two models: the composite model implicitly
considers all the moments of the return distribution through the return
and risk constraints, while the factor-based model considers only the
second moment. The risk characteristics of the replicating portfolios and
hedge fund strategies are reported in columns 9–11 of Table 3.3a and 3.3b.
In terms of CVaR and CDaR, the replication performance of the composite
models is similar to the factor-based model, except for the PMC model.
Increasing the number of risk constraints reduces the out-of-sample
explanatory power of the replicating portfolios, with the R-squared
dropping significantly. However, this fall in explanatory power is not
matched by a reduction in portfolio performance. Indeed, the ALLC
model generates higher returns, better risk-adjusted return, and better
replication of the higher moments than the factor-based model, and with
similar (or better) risk, as measured by CVaR and CDaR. Among the com-
posite models, although the PMC and CVaRC models explain more of the
hedge fund return variance (i.e. they have a higher R-squared), the ALLC
model produces the best overall replication performance. In particular,
compared with the other composite models, the ALLC model provides a
better match for the first four moments of hedge fund returns, better risk-
adjusted return, better CVaR replication, and similar CDaR replication.
The course of net asset values of the replicating portfolios and the hedge
fund strategies over the out-of-sample test period is displayed in Figure 3.1.
In general, the composite model portfolios provide a reasonable fit to the
time series of hedge fund returns. For example, the ALLC model slightly
underperforms the EM and DS strategies and closely follows and outper-
forms other strategies. The PMC model generates the highest portfolio
values and significantly outperforms all hedge fund strategies except EM
and EH. On the other hand, the factor-based model closely follows the EM
and FOF strategies but clearly underperforms other hedge fund strategies.
45
Convertible Arbitrage Distressed Securities

3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0
Sep-02 Sep-03 Sep-04 Sep-05 Sep-06 Sep-07 Sep-08 Sep-09 Sep-10 0
Sep-02 Sep-03 Sep-04 Sep-05 Sep-06 Sep-07 Sep-08 Sep-09 Sep-10
CA CVaRC CDaRC PMC ALL RC FM DS CVaRC CDaRC PMC ALL RC FM
Event Driven Equity Hedge

3 2.5
2.5
2
2
1.5
1.5
1
1
0.5 0.5
0 0
Sep-02 Sep-03 Sep-04 Sep-05 Sep-06 Sep-07 Sep-08 Sep-09 Sep-10 Sep-02 Sep-03 Sep-04 Sep-05 Sep-06 Sep-07 Sep-08 Sep-09 Sep-10
ED CVaRC CDaRC PMC ALL RC FM EH CVaRC CDaRC PMC ALL RC FM
Emerging Markets Equity Market Neutral

4.5 3
4
2.5
3.5
3 2
2.5
1.5
2
1.5 1
1
0.5
0.5
0 0
EM CVaRC CDaRC PMC ALL RC FM EMN CVaRC CDaRC PMC ALL RC FM
Merger Arbitrage Global Macro

3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
MA CVaRC CDaRC PMC ALL RC FM MAC CVaRC CDaRC PMC ALL RC FM
Relative Value Arbitrage Fund of Funds

3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
RV CVaRC CDaRC PMC ALL RC FM FOF CVaRC CDaRC PMC ALL RC FM
Figure 3.1 Net asset values of replicating model portfolios and replicated hedge
fund strategies
3.4 Conclusion
In principle, the ability to replicate hedge fund performance represents

an attractive opportunity for investors to benefit from the high returns
that hedge fund strategies are potentially able to offer, while avoiding
the risks that such strategies involve. In practice, however, the effective-
ness of existing hedge fund replication methods appears to be limited.
Factor-based approaches work well in-sample, but are typically unable to
maintain this performance out-of-sample. In contrast, payoff distribution
matching approaches successfully replicate the unconditional distribu-
tion of hedge fund returns out-of-sample, but ignore the first moment
of returns and hence are not able to deliver the absolute return perform-
ance associated with hedge fund strategies. In this chapter, we investigate
an approach to hedge fund replication that combines the factor-based
methodology with a series of risk and performance constraints. We use
this approach to replicate the monthly returns of ten hedge fund strategy
indices using long-only positions in a broad set of equity, interest rate,
exchange rate, and commodity indices, all of which can be traded using
liquid, investible instruments such as futures, options, and exchange
traded funds. In out-of-sample tests, we show that our composite
approach yields replicating portfolios that are potentially able to mimic
both the risk-adjusted performance and distributional characteristics of
the hedge fund indices that they are designed to track. On balance, our
approach appears to represent an improvement over the out-of-sample
performance of the factor and payoff distribution approaches reported by
Jaeger and Wagner (2005) and Amenc et al. (2010).
References
Amenc, N., Géhin, W., Martellini, L., and Meyfredi, J.-C. (2008). “Passive Hedge
Fund Replication: A Critical Assessment of Existing Techniques.” Journal of
Amenc, N., Géhin, W., Martellini, L., and Meyfredi, J.-C. (2010). “Passive Hedge
Fund Replication: Beyond Linear Case.” European Financial Management, 16(2):
191–210.
Brooks, C. and Kat, H. M. (2002). “The Statistical Properties of Hedge Fund Index
Returns and Their Implications for Investors.” Journal of Alternative Investments,
5(3): 26–44.
Chekhlov, A., Uryasev, S., and Zabarankin, M. (2000). “Portfolio Optimization
with Drawdown Constraints.” Research Report 2000–5. ISE Dept., University
of Florida, Gainesville, Florida.
Chekhlov, A., Uryasev, S., and Zabarankin, M. (2005). “Drawdown Measure in
Portfolio Optimization.” International Journal of Theoretical and Applied Finance,
8(1): 13–58.
Duarte, J., Longstaff, F. A., and Lu, F. (2007). “Risk and Return in Fixed Income
Arbitrage: Nickels in Front of a Steamroller?” The Review of Financial Studies
20(3): 769–811.
Engle, R. F. (1982). “Autoregressive Conditional Heteroscedasticity with Estimates
of Variance of United Kingdom Inflation.” Econometrica, 50: 987–1008.
Fung, W. and Hsieh, D.A. (1997) “Empirical Characteristics of Dynamic Trading
Strategies: The Case of Hedge Funds.” Review of Financial Studies 10(2): 275–302.
Fung, W. and Hsieh D. A. (2001). “The Risk in Hedge Fund Strategies: Theory and
Evidence from Trend Followers.” Review of Financial Studies 14(2): 313–41.
Fung, W. and Hsieh D. A. (2004). “Hedge Fund Benchmarks: A Risk-Based
Approach.” Financial Analysts Journal, 60(5): 65–80.
Fung, W. and Hsieh, D. A. (2006). “Hedge Funds: An Industry in Its Adolescence.”
Fung, W. and Hsieh, D. A. (2007). “Hedge Fund Replication Strategies: Implications
for Investors and Regulators.” Banque de France, Financial Stability Review, Special
Issue on Hedge Funds, 10(4): 55–66.
Geltner, D. M. (1991). “Smoothing in Appraisal-Based Returns.” Journal of Real
Estate Finance and Economics, 4(3): 327–45l.
Hasanhodzic, J. and Lo, A. W. (2007). “Can Hedge Fund Returns Be Replicated?
Jaeger, L. and Wagner, C. (2005). “Factor Modeling and Benchmarking of Hedge
Kat, H. and Palaro, H. (2005). “Who Needs Hedge Funds? A Copula-Based Approach
to Hedge Fund Return Replication.” Working Paper 27, Alternative Investment
Research Centre, Cass Business School, Reading, UK.
Kat, H. and Palaro, H. (2006). “Hedge Fund Indexation the Fund Creator Way:
Efficient Hedge Fund Indexation Without Hedge Funds.” Working Paper 38,
Alternative Investment Research Centre, Cass Business School, Reading, UK.
Keating, C. and Shadwick, F. (2002). “A Universal Performance Measure.” Journal
of Performance Measurement 6(3): 59–84.
Mitchell, M. and Pulvino, T. (2001). “Characteristics of Risk and Return in Risk
Arbitrage.” Journal of Finance, 56(6): 2135–75.
Papageorgiou, N., Remillard, B., and Hocquard, A. (2008). “Replicating the
Properties of Hedge Fund Returns.” Journal of Alternative Investments, 11(2): 8–38.
Rockafellar, R.T. and Uryasev, S. (2002). “Conditional Value-At-Risk for General
Loss Distributions.” Journal of Banking & Finance, 26(7): 1443–71.
Stowell, D. P. (2010). An Introduction to Investment Banks, Hedge Funds, and Private
Equity: the New Paradigm. San Diego, CA: Elsevier Inc.
Takahashi, A. and Yamamoto, K. (2008). “Hedge Fund Replication.” CARF
Working Papers. Tokyo, Japan.
Takahashi, A. and Yamamoto, K. (2010). “A New Hedge Fund Replication Method
with the Dynamic Optimal Portfolio.” Global Journal of Business Research, 4(4):
23–35.
Tancar, R. and Viebig, J. (2008). “Alternative Beta Applied—An Introduction to
Hedge Fund Replication.” Financial Markets Portfolio Management, 22(3): 259–79.
Wallerstein, E., Tuchschmid, N., and Zaker, S. (2010). “How Do Hedge Fund Clones
Manage the Real World?” Global Journal of Business Research, 4(4): 23–34.
4
Takeover Probabilities and the
Opportunities for Hedge Funds
and Hedge Fund Replication to
Produce Abnormal Gains
4.1 Introduction
It is important for hedge fund managers and those investors involved

in hedge fund replication to recognize that, following the announce-
ment of a takeover offer, empirical marketplace evidence shows that the
target’s share price does not always trade at the offered price. If investors
are to replicate hedge fund strategies in relation to the pricing of a target
in a takeover offer where the target is part of the hedge fund portfolio,
it follows that the hedge fund managers must be seen to get the pricing
strategy right. This chapter confirms a methodology that has been avail-
able to hedge fund managers to fine tune strategy in relation to takeover
targets in a hedge fund portfolio, using evidence from Australia.
The target price anomaly usually lasts for the duration of the announce-
ment period. The difference between the offer price and the market
price, known as the “offer spread” represents the “money left on the
table” following the initial takeover announcement. This “spread” broadly
reflects the market’s expectation regarding the eventual outcome of the
offer. Where the market expects the offer to succeed the target’s price
trades close to, the offered price.
Conversely, when the market is anticipating deal failure the spread
widens. Empirical evidence suggests that as successful offers approach
the final outcome the spread narrows, converging upon the offered price
at the consummation of the offer (Brown and Raymond, 1986; Hutson,
2000; Hutson and Partington, 1994; Samuelson and Rosenthal, 1986).
The “Merger Arbitrage Strategy”1 is an event-driven trading strategy that
has been successfully employed by hedge fund managers to exploit
the convergence in this spread. Figure 4.1 illustrates that the strategy
48
Anthony Ravi, Peter Mayall, and John Simpson 49
40%
30%
20%
10%
Annual return (%)
0%
1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
–10%
–20%
–30% Barclay merger arbitrage index

Credit suisse merger arbitrage index
–40%
The world index
–50%
Figure 4.1 The profitability of the merger arbitrage strategy
has generally been profitable for hedge funds over the last decade,
outperforming the MSCI Word index, particularly during economic
downturns such as the 2008 global financial crisis (GFC).
Over much of the past two decades the merger arbitrage trading strategy
was believed to be the exclusive domain of large hedge fund managers.
It was generally considered both unprofitable and too risky for the average
retail investor. Larcker and Lys (1987) suggest that the abnormal returns
generated by the strategy are due to the fact that hedge fund managers are
better informed than the market about the probability of deal success and
therefore retail investors could not profit from the strategy. However, the
combination of attractive gains and lower transaction costs has begun to
attract the attention of researchers and retail investors in recent years. All
of the tables and figures in this chapter are sourced and originated in the
empirical work of the writers and specifically in the work of Ravi (2011).
Despite “arbitrage” being in the title, the strategy is far from a risk-free
trade.2 In fact, the strategy can be viewed as a bet on the ultimate success
of the offer. Therefore, the biggest risk to the strategy is deal failure (Branch
and Yang, 2003), in which case the target’s share price tends to fall, result-
ing in losses to the investor (For example, Bishop, Dodd, and Officer, 1987;
Bradley, Desai, and Kim, 1983; Dodd, 1976). The risk arbitrage strategy
itself is quite simple, involving the purchase of the target’s shares in a cash-
only deal, with the additional short sale on the acquirer’s shares if the scrip
is the method of payment.
Following the GFC of 2008, hedge funds suffered significant losses.
Hedge Fund Research (HFR)3 reports that the average loss for 2008 was
–35.98 percent across all American hedge funds. As a point of comparison,
the Dow Jones Industrial average returned negative 33.80 percent over the
50 Takeover Probabilities
same period, the third worst period on record. However, hedge funds that
employed the event-driven merger arbitrage strategy seem to have avoided
this demise. The HFRX Merger Arbitrage Index™ reports a positive return
of 3.69 percent for 2008. Additionally, Figure 4.1 indicates that the Barclay
and Credit Suisse merger arbitrage indices reported only slightly negative
returns.
Figure 4.1 illustrates the returns of two prominent merger arbitrage
indexes in the US. The Barclay merger arbitrage index is an equally
weighted index that includes the returns from 33 professional hedge
fund managers specializing in merger arbitrage. The Credit Suisse index
is value weighted and includes the returns from ten managers. This
index did not begin until late 2001. The MSCI World Index is also shown
on the graph for comparative purposes, and indicates that the strategy
has generally outperformed, especially in 2001 and 2008 recessions.
These results suggest that the merger arbitrage strategy is profitable in
the American market, a proposition that is also supported by the literature
(Ravi, 2011). The promising results throughout the GFC have fostered
renewed interest in the merger arbitrage strategy. To date, studies focus-
ing on this strategy have been confined to the US, Canadian and UK
markets, with the study by Maheswaran and Yeoh (2005) the only one
to consider the Australian market.
Due to the limited studies outside the USA it is unclear whether the
profitability and risk–return characteristics for the US sample are a univer-
sal feature of the merger arbitrage strategy, or just an anomaly particular
to the US market. Hence further research is needed to examine profit-
ability outside the USA. In addition, takeover regulations vary consider-
ably across countries, affecting various factors of the deal such as: timing,
disclosure of information, revision of offer terms, ability of the bidder to
withdraw the offer, and timetable for the merger process to be completed
(Sudarsanam and Nguyen, 2008). The variation in political risks interna-
tionally also has the potential to inhibit the success of corporate control
contests. The results indicate that there is a considerable diverging variation
in the rates of success between Australia and the G7 nations.
For additional robustness, this study also calculates the historical prob-
abilities of a takeover eventuating. These results serve as an indicator of
the probability of investing in potential takeover targets before the offer
is forthcoming. This probability analysis can be utilized by hedge fund
managers and hedge fund replication investors to focus their investment
decisions and allocate their capital with a better chance of success.
Little is presently known about the profitability of the merger arbi-
trage trading strategy, despite it being employed by hedge fund managers
for decades.4 This study aims to address this issue by examining the
profitability of the merger arbitrage strategy in the context of the
Australian market.
The driving force behind the profitability of the merger arbitrage
strategy is the eventual success of the offer. The disparity in takeover
regulations between Australia and the USA affects timing, disclosure of
information, revision of offer terms, the ability of the bidder to withdraw
the offer, and the time frame in which the offer must be completed. All
these factors are believed to affect the overall success of a takeover, and
hence they are expected to reflect on the profitability of the merger arbi-
trage trading strategy. The rates of success in the Australian market have
been diverging in recent years, and are generally lower compared to G7
countries.
In addition, the Australian market is interesting due to its unique,
mining-orientated market structure. Also, offers and companies on the
Australian market tend to be smaller in comparison to international
markets—the Australian market represents approximately 2 percent of
the world’s stock market. This may have an adverse impact on the merger
arbitrage strategy, especially in the case of stock offers, since liquidity is
necessarily lower in the Australian market.
Studies by Mitchell and Pulvino (2001) scrutinize the methodologies
used in these prior studies, suggesting that simply calculating the annual-
ized returns overstates the profitability of the strategy. Mitchell and Pulvino
(2001) posit the use of a calendar-time portfolio, whereby the daily returns
for each deal are calculated, with active deals in each month going
to form the portfolio. The average daily returns are then compounded to
form a monthly portfolio. In addition to adopting the new methodology,
both the later studies employ substantially larger samples, including cash
and scrip offers of 4,075 and 1,901 respectively. The sample periods are
also longer, with the two studies covering the period from 1963 to 1998
between them.
The results of these studies are, however, still consistent with earlier
studies, although the returns are substantially lower, being 9.9 percent
and 8.88 percent per annum respectively. However, Mitchell and Pulvino
(2001) suggest that the strategy is not profitable after transaction costs.
The authors also suggest that the returns to the strategy are correlated
with market downturns, which is inconsistent with prior research that
indicates the strategy is market neutral (Maheswaran and Yeoh, 2005).
Prior research has indicated that the merger arbitrage strategy is highly
profitable in the USA, Canada, and the UK (Ravi, 2011). However, studies
on the Australian market are extremely limited, with only one study
conducted by Maheswaran and Yeoh (2005). As previously mentioned,

this study only considers cash deals. Existing studies also report large
cross-sectional variations in returns, ranging from 6.2 to 172 percent,
over different periods and across countries. This suggests that differing
regulations and market microstructures influence the profitability of
the strategy. Hence further investigation into the Australian market is
warranted. In addition, no study has investigated the short-term profit-
ability of the strategy. Given the semi-strong form of market efficiency
proposed by Fama (1970), it is expected that profits, on average, cannot
be generated in either the short or long term based on information con-
tained in the takeover announcement.
In addition, the increased volatility and risk present in the market during
recessionary periods are likely to adversely affect the market’s perception of
bid success, causing an increase in the profitability of the strategy during
these periods.
4.2 Data
The data sample used for this study comprises all public takeover offers
announced on the Australian Stock Exchange (ASX) for ten years from
1 January 2000 until 31 December 2009 (n 108 months). This sample
reflects the most up-to-date data available. In addition, several of the avail-
able databases are utilized to ensure maximum data integrity. The initial
list of offers was obtained from the Bureau van Dijk database, Zephyr. The
sample was then crosschecked with data available from Aspect Huntley.
Full details of the derivation of the model to be tested, the formulae, and
the literature base are provided in Ravi (2011).
To be included in the final sample the offers had to meet the following
criteria:
1. Both the bidder and target are public companies, and are listed on
the ASX.
2. The bidders seek to obtain a controlling stake by acquiring a majority
interest in the target firm (the acquirer must own more than 50 percent
of the target post transaction).
3. The bid is classified as completed, where completed means that the
offer is closed and the transaction end date is available.
4. A full set of information must be available for each offer.
Apart from these restrictions no other formal limitations were placed

on the data, since the objective of this study requires that all public
takeover bids be considered. The final sample consisted of 245 offers.

Of these, there were 180 unique targets, of which 43 received at least
one additional bid.
The full results of this study are detailed in Ravi (2011). For the purposes
of this chapter the generalized findings follow the following tabulated
and graphical representations of the full results.
The results presented in Table 4.1 and Figures 4.2–4.7 have several signifi-
cant implications for hedge fund managers and investors replicating
hedge funds and wishing to execute the plain vanilla merger arbitrage
strategy in the Australian market. In Figure 4.2 the graph illustrates
abnormal returns pertaining to target firm shareholders for all acquisi-
tions (n 245). The event window employed ranges from –42 to 42
days, where day 0 represents the announcement day. AAR represents the
Average Abnormal Return and CAAR represents the Cumulative Average
Table 4.1 Risk–return characteristics: merger arbitrage portfolio
Market index Plain vanilla Long-only

Mean 0.55% 1.60%* 1.53%*
Standard deviation 4.35% 7.84% 8.00%
Correlation with market – 0.143 0.210
index
Sharpe ratio 0.13 0.20 0.19
Value of $1 invested $1.66 $2.82 $2.75
24.00%
AAR (n = 245)
Daily average abnormal return (%)
CAAR (n = 245)
19.00%
14.00%
9.00%
4.00%
–1.00%
–42 –36 –30 –24 –18 –12 –6 0 6 12 18 24 30 36 42
Days
Figure 4.2 Short-term wealth effects: target shareholders

2.30%
AAR (n = 245)
1.80% CAAR (n = 245)
1.30%
0.80%
0.30%
–0.20% –42 –36 –30 –24 –18 –12 –6 0 6 12 18 24 30 36 42
–0.70%
–1.20%
Days
Figure 4.3 Short-term wealth effects: acquirer shareholders
4.00%
3.50% Daily average profit (n = 245)
3.00% Cumulative average profit
2.50%
2.00%
1.50%
1.00%
0.50%
0.00%
–0.50% 1 6 11 16 21 26 31 36 41
–1.00%
–1.50%
Days
Figure 4.4 The merger arbitrage spread
Abnormal Return. In Figure 4.3 the graph illustrates abnormal returns

pertaining to acquiring firm shareholders for all acquisitions (n 245).
The event window employed ranges from –42 to 42 days, where day 0
represents the announcement day. AAR and CAAR are as in Figure 4.2.
In Figure 4.4 the graph illustrates the movement in the arbitrage
spread (the difference between the traded price of the target firm’s
shares and the price offered by the acquirer under the takeover offer)
in the 42 days prior to offer completion at day 0. The figure shows this
spread for offers that were ultimately completed (n 140), offers that
were rejected (n 65) and offers that ultimately failed (n 40). A value
of 0 on the vertical axis indicates that the target’s share price is currently
trading at the offered price, while negative values indicate that the
Average spread (Rejected offers)

Average spread (Ultimately completed offers)
Average spread (Failed offers)
65.00%
55.00%
45.00%
Arbitrage Spread (%)
35.00%
25.00%
15.00%
5.00%
–5.00%
–42 –36 –30 –24 –18 –12 –6 0
–15.00%
Days (Until completion)
Figure 4.5 Short-term profitability: plain vanilla strategy
1.00%
Daily average profit (n = 245)
Cumulative average profit
0.50%
0.00%
1 7 13 19 25 31 37
–0.50%
–1.00%
–1.50%
–2.00%
Days
Figure 4.6 Short-term profitability: long-only strategy
target’s share price is currently trading above the offered price. Hence
a positive value indicates that the target’s shares are currently trading
below the offered price. In Figure 4.5 the graph illustrates the daily and
cumulative abnormal profits attributable to the plain vanilla merger
arbitrage strategy, over the arbitrary short term 42-day event window.
The graph incorporates all offers (n 245) over the entire ten year sample
period. Abnormal profits are calculated using the constrained (0, 1)
market model. Where an offer is made in cash the plain vanilla merger
arbitrage trading strategy involves taking a long position in the target
$3.50
Vanilla Strategy
$3.00
Investment value ($AUD)
Long Only
ASX All Ords Index
$2.50
$2.00
$1.50
$1.00
$0.50
0
9
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
/0
05
11
05
11
05
11
05
11
05
11
05
11
05
11
05
11
05
11
05
11
Months
Figure 4.7 Value of $1 invested: merger arbitrage strategy
company’s stock as soon as possible following a takeover announcement.

Where the offer is for scrip or a combination of cash and scrip an addi-
tional short position in the acquiring firm’s stock is required to hedge
the position.
In Figure 4.6 the graph illustrates the daily and cumulative abnormal
profits attributable to the long-only merger arbitrage strategy, over the
arbitrary short term 42-day event window. This strategy involves taking
a long position in the target firm only, and mitigates any short position
in the acquiring firm. The graph incorporates all offers (n 245) over
the entire ten year sample period. Abnormal profits are calculated using
the constrained (0, 1) market model.
In Figure 4.7 the graph illustrates the monthly return (calculated
using the calendar-time methodology), over the ten year sample period
(n 119 months), for the vanilla merger arbitrage strategy as well as the
long-only portfolio – a restricted variant of the plain vanilla strategy. The
long-only portfolio utilizes the same investment strategy as the vanilla
portfolio. However, a restriction is placed on short sales in the acquiring
firm. The cumulative monthly return on the ASX All Ordinaries Index is
also shown over the same period for comparative purposes. In accordance
with the strategy outlined in Section 4.3, where there are no active deals in
the month the investor is assumed to be out of the market (i.e. the investor
receives no returns).
Table 4.1 shows the mean, standard deviation, market correlation,
Sharpe ratio, and value of $1 invested in May 2000 for the plain vanilla
arbitrage portfolio as well as the long-only portfolio. The risk–return
characteristics are also shown for the market index (the ASX All Ordinaries
Index). In addition; t-tests were conducted for each of the means against
an H0 of 0 with one, two, and three asterisks indicating significance at the

1, 5, and 10 percent significance levels.
Primarily, analysis of the probabilities pertaining to takeovers reveals
that the chance of successfully picking a potential takeover target, without
exercising any discretion, is very slim (9.96 percent over ten years or
approximately 1 percent per annum on average). In addition, once
takeovers are announced they are successfully completed on average
71.1 percent of the time. If the initial bid is rejected the chance of a new
offer being successful increases if the bidding firm is different from the
firm that initially initiated the offer (45 percent compared to 52 percent
for an alternative bidder). While these results are not favorable to the
average speculator in the market, they provide evidence as to the profit-
ability of the merger arbitrage strategy for hedge funds.
The results from the investigation into the announcement period offer
spread indicate that the market is unable to distinguish between deals
that are ultimately successful and those that are unsuccessful. Therefore
hedge fund managers wishing to employ the merger arbitrage strategy
for part of their portfolio should invest in all deals, regardless of the offer
spread. If the strategy is followed religiously the plain vanilla strategy is
found to be highly profitable in the Australian market. Results suggest
that investors can expect to earn risk-adjusted returns of between 1.13
percent and 1.25 percent per month. This result, however, can only be
expected if the plain vanilla strategy is followed, as a long-only strategy
(short sales in the acquiring firm are restricted) is not found to yield
statistically significant returns.
Additionally, several factors, known at the time of the offer, have been
found to affect the overall profitability of the strategy. Hence it is possi-
ble for the investor to improve the performance of the strategy by only
investing in deals that meet certain criteria. For example, the investor
is advised to only execute the strategy on the initial bid. The results in
this study indicate that after this bid profitability declines significantly,
becoming negative after the third bid. However, returns after the first bid
are not significant, suggesting that no profits are available after this bid.
Finally, the profitability of the strategy was found to be correlated with
recessionary periods, suggesting that this is when the strategy should be
employed. However, the method of payment was also found to be cor-
related with different market environments. Overall, the results suggest
that in recessionary periods the strategy should only be executed on
deals offering scrip as the method of payment. However, if the strategy is
being employed in an expansionary market then only cash offers should
be considered.
4.4 Conclusion
The merger arbitrage trading strategy has historically been associated

with large hedge funds. The esoteric nature of the strategy has generally
kept retail investors from successfully implementing the strategy in their
private portfolios. The large profits purported to exist by hedge fund
managers as a result of implementing this strategy have attracted the
attention of researchers in recent years.
Numerous US studies have documented significant gains from the
merger arbitrage strategy. Significant differences between the USA
and international markets, such as takeover regulations and market
microstructure, suggest that the probability of success also differs inter-
nationally. Using a comprehensive sample of cash and scrip offers this
study investigates the profitability of the merger arbitrage trading strategy
in the Australian market. Taking this one step further, the influence
of various deal- and firm-specific factors were investigated in order to
determine whether it is possible to enhance the overall profitability of
the strategy.
The overall results of this study are consistent with prior studies in the
USA and with the only existing Australian study by Maheswaran and
Yeoh (2005). Using returns generated by the calendar-time approach, the
returns from the plain vanilla merger arbitrage trading strategy are found
to be between 1.13 and 1.25 percent per month or 13.56 and 15 percent
per annum. Therefore the strategy has been highly profitable in the
Australian market over the ten year sample employed in this study.
The strategy was also found to be profitable in the short term. By
employing standard event study methodology, over a 42-day window the
plain vanilla merger arbitrage strategy was found to generate statistically
significant returns of approximately 3 percent. Conversely, a portfolio
that was restricted to long positions only was not able to generate statisti-
cally significant returns. This result holds for both short- and long-term
investment horizons.
In addition to these findings, this study has established that the profit-
ability of the strategy diminishes as the target receives additional bids.
In fact, the profits attributed to additional bids were not statistically dif-
ferent from zero. Therefore investors looking to engage in this strategy
should only invest in the first bid in a series; investing in additional bids
is wealth destroying.
Finally, it is found, consistent with the prior US study by Mitchell and
Pulvino (2001), that the merger arbitrage strategy is correlated with reces-
sionary periods in the Australian market—the merger arbitrage strategy
is thus considered a contrarian investment strategy when applied to the

Australian market.
Overall, however, this study provides comprehensive and conclusive
evidence to suggest that the plain vanilla merger arbitrage strategy has
been profitable in the Australian market for hedge funds and investors
engaged in hedge fund replication over the ten years until 31 December
2009.
Notes
1. Also known as risk arbitrage; the two terms are used interchangeably through-
out this chapter.
2. The classical definition of arbitrage suggests the consummation of a riskless
profit from the simultaneous buying and selling of a security on different
markets—the risk arbitrage strategy does not earn a riskless profit.
3. Hedge Fund Research, Inc., is an American research company specialising in the
indexation and analysis of global hedge funds and alternative investments.
4. Ivan Boesky is credited with elucidating the strategy in the late 1970s. However,
it was not until the release of his best-selling book Merger Mania: Arbitrage: Wall
Street’s Best Kept Money-Making Secret in 1985 that the strategy was given any
credibility by the wider hedge fund community.
References
Bishop, S., Dodd, P., and Officer, R. R. (1987). Austrtalian Takeovers: The Evidence
1972–1985: St. Leonards, New South Wales: The Centre for Independent Studies.
Boesky, I. F. (1985). Merger Mania: Arbitrage: Wall Street’s Best Kept Money-Making
Secret. 1st edn. Upper Saddle River, NJ: Holt Rinehart & Winston.
Bradley, M., Desai, A., and Kim, E. H. (1983). “The Rationale Behind Interfirm
Tender Offers: Information or Synergy?” Journal of Financial Economics, 11(1–4):
183–206.
Branch, B. and Yang, T. (2003). “Predicting Successful Takeovers and Risk
Arbitrage.” University of Nebraska Working Paper, Lincoln, NE.
Brown, K. C. and Raymond, M. V. (1986). “Risk Arbitrage and the Prediction
of Successful Corporate Takeovers.” Journal of Financial Management, 15(3):
54–63.
Dodd, P. (1976). “Company Takeovers and the Australian Equity Market.”
Australian Journal of Management, 1(22): 15–35.
Fama, E. F. (1970). “Efficient Capital Markets: A Review of Theory and Empirical
Work.” Journal of Finance, 25(2): 383–417.
Hutson, E. (2000). “Takeover Targets and the Probability of Bid Success: Evidence
from the Australian Market.” International Review of Financial Analysis, 9(1):
45–65.
Hutson, E. and Partington, G. (1994). “Takeover Bids, Share Prices, and the
Expected Value Hypothesis.” UTS Business School Working Paper, Sydney,
New South Wales.
Larcker, D. F. and Lys, T. (1987). “An Empirical Analysis of the Incentives to

Engage in Costly Information Acquisition.” Journal of Financial Economics,
18(1): 111–26.
Maheswaran, K. and Yeoh, S. C. (2005). “The Profitability of Merger Arbitrage:
Some Australian Evidence.” Australian Journal of Management, 30(1): 111–26.
Mitchell, M. and Pulvino, T. (2001). “Characteristics of Risk and Return in Risk
Ravi, A., (2011). “The Merger Arbitrage Trading Strategy: Factors Influencing
the Profitability of Trading Takeovers in the Australian Market.” Unpublished
Honours Dissertation, Curtin University, Perth, Australia.
Samuelson, W. and Rosenthal, L. (1986). “Price Movements as Indicators of
Tender Offer Success.” Journal of Finance, 61(2): 481–99.
Sudarsanam, S. and D. Nguyen. (2008). “UK Evidence on the Profitability and
the Risk–Return Characteristics of Merger Arbitrage.” Centre for Research in
Economics and Finance, Working Paper, Cranfield University, Bedfordshire, UK.
5
Benchmarking of Replicated
Hedge Funds
5.1 Introduction
In the 1990s, the hedge fund universe consisted of approximately 500

hedge funds with an estimated total of $50 billion assets. At its peak in
late 2007, the hedge fund industry had grown to more than $2.8 trillion
in assets. As it grew, so did the complexity of hedge fund management.
Ongoing innovations, entry into new markets, and creative trading strate-
gies have made it more and more difficult for investors to understand the
market behavior of hedge funds.
In the traditional universe of mutual funds, passive approaches provide
a basis to evaluate a fund manager’s performance. Usually this performance
is measured by comparing fund performances with corresponding indices.
Since hedge funds vary their choices of strategy, it is rather difficult to com-
pare funds with above-average performance (Anson, 2004). The problem
of finding funds with superior performance emerges from the impossi-
bility of identifying the share of beta- and alpha-driven returns as part of
the absolute return. To identify the specific beta of hedge funds, financial
instruments such as derivatives, leverage, and short positions must be
taken into account. The resulting beta is then referred to as the alterna-
tive beta. Two issues are vital when dealing with a hedge fund beta: firstly,
alternative beta is often sold as hedge fund alpha; and secondly, the actual
problem regarding performance in the measurement of hedge funds lies in
the correct definition and isolation of beta (Jaeger and Wagner, 2005).
One possible approach to defining beta is to construct hedge fund
indices, which should provide a solid basis for benchmarking returns and
risk exposure. In the following sections we will therefore discuss several
existing benchmark instruments for hedge funds. Since all of them are
vulnerable to biases which generate distorting impacts on the suitability
61
62 Benchmarking of Replicated Hedge Funds
of benchmarks, we will show why factor-based benchmarking (which

also acts as the basis for most hedge fund replication strategies) will help
investors to manage their hedge fund investments. Empirical results in
terms of exact replication of hedge funds fail, however, to deliver satis-
factory clone results.
5.2 Hedge fund returns and benchmarking models
To understand hedge fund returns, we begin by defining them using

returns above risk-free rates. The three ways that an investor could earn
returns above risk-free rates are:
• Market beta: describes exposure to the systematic risks in publicly

traded markets such as stocks, bonds, and commodities.
• Carry returns: through exposures to alternative systematic risk that
do not exist in traditional buy-and-hold portfolios which require
skill to access.
• Alpha-driven returns: resulting from competitive advantages over other
financial market participants.1
While it is easy to access beta and carry returns, it is difficult for alpha
returns. Alpha is very fund-specific, is idiosyncratic to each fund, has
low correlation with the markets, and typically offers returns with low
volatility. However, good alpha is difficult to access because it is concen-
trated only in the highest quality hedge funds.
The critical requirements of an appropriate benchmark are representativity
and investability (Géhin and Vaissié, 2004). Thus several benchmark con-
cepts have been evaluated in order to access alpha, but many of them are
vulnerable to biases, which have a distorting influence on the suitability
of indices as benchmarks. In general two different benchmark concepts
can be distinguished: the first technique is based on the idea of creating
benchmarks using the underlying funds as elements of construction.2
The second technique refers to the factors which must be defined to
replicate a certain sample or index.
5.3 Benchmarking based on underlying funds
5.3.1 Non-investable hedge fund indices

Non-investable hedge fund indices try to provide an undistorted picture
of the hedge fund universe. Usually indices like the Royal Bank of Canada
Non-Investable Index have several sub-strategy indices which attempt
Martin D. Wiethuechter and Lajos Németh 63
to represent a particular hedge fund strategy within the hedge fund

universe. Despite its claim to deliver a true and fair image of the hedge
fund industry, this structure suffers from various biases.
5.3.1.1 Survivorship bias

Survivorship bias is caused by failed funds, or funds that go out of business.
The impact of survivorship bias depends mainly on two elements: firstly,
the number of funds that stop providing data to database vendors every
year; and secondly, the average difference in returns of dead and live funds.
Analyzing the approach that data vendors use to add new hedge funds to
existing records is crucial. Tremont Advisory Shareholders Services (TASS)
usually contacts new funds, whereas Hedge Fund Research, Inc. (HFR)
includes new funds upon request of those funds. Obviously, the active
approach of TASS results in a higher survivorship bias than HFR because
every new emerging fund is included in the TASS database. This bias is
due to the high number of hedge funds failing during the incubation
phase (Liang, 2000). The effect on performance caused by this bias varies
between 1.32 and 6.67 percent per annum (Géhin and Vaissié, 2004).
5.3.1.2 Selection bias

Databases rely on inputs from hedge fund managers. Due to non-
obligatory and inconsistent reporting regulations among database
providers, it is up to hedge fund managers to forward the necessary data
to database providers. It is to be expected that badly performing hedge
funds will be unwilling to report, which causes significant biases in
indices. Moreover, there are heterogeneous guidelines for inclusion into
databases. These rules embrace criteria such as duration of record track-
ing, minimum investment volume, and accessibility of new investors
(Géhin and Vaissié, 2004).
5.3.1.3 Backfilling bias

Adding a new hedge fund to indices implies that the past performance
of the fund is usually added as well. However, managers can decide
whether or not to report performance data from the fund’s inception to
databases. A manager who can present a successful track record might
choose to backfill the historical return data. This causes an upward bias
because new hedge funds have to surmount difficulties which are some-
times not logged in the records of the databases. Consequently, fund
performance appears to be superior during the incubation period, and
Fung and Hsieh (2000) have estimated an impact on performance of 1.4
percent per annum on average.
5.3.1.4 Short history bias

Short data history can lead to the phenomenon that market influences
are not taken into account when it comes to performance evaluation
(Fung and Hsieh, 2004). Despite the basic idea of hedge funds engen-
dering constant returns, the general market situation still influences
the performance of hedge funds. This effect can be observed in bear and
bull markets. There is evidence that the evaluation of short-term hedge
fund returns provides an unrealistic picture. The long-term performance
of hedge funds is approximately 34 percent lower than their short-term
performance (Agarwal and Naik, 2004).
5.3.1.5 Unambiguousness of strategies

Every database vendor applies its own strategy classification. Therefore the
implications of each set of strategy criteria differ, which is why individual
hedge funds are listed in different strategy indices.3 However, tracking
changes in a strategy is a difficult task because hedge funds usually do not
notify databases when the fund strategy is about to be modified (Géhin
and Vaissié, 2004). Consequently, non-investable indices cannot be con-
sidered as representative. As the term non-investable index suggests this
concept is inappropriate for investable benchmarks because non-investable
indices try to cover as many single hedge funds as possible.
5.4 Funds of hedge funds
According to Fung and Hsieh (2004), funds of hedge funds (FoHFs)

could be of more use than fund indices when it comes to the question
of benchmarking, as FoHFs reduce the impact of three different biases:
firstly, FoHFs are able to bypass selection bias because they are not
limited to funds that report to databases; secondly, backfilling bias can
be mitigated because only return data from the very first day on which
a fund is added to a fund of funds are considered, instead of ex-ante
returns. In order to use FoHFs as a benchmark, activities must be reported
to databases. FoHFs also mitigate survivorship bias. Past performance of
failed funds or funds that were targets of merging transactions remain
in the historical performance of a FoHF. However, potential distortions,
such as selection and backfilling bias, remain, as a few FoHFs stop
supplying databases with information.
Even though FoHFs do not eliminate all types of biases, they can be
considered as representative indices. However, the double fee effect and
the use of leverage can distort a fair evaluation of systematic risks of hedge
funds (Géhin and Vaissié, 2004). Moreover, the performance of FoHFs is

misleading, since often only funds with satisfying structural integrity pass
the due diligence process of FoHFs (Duc, 2004).
5.5 Investable hedge fund indices
Because the FoHF managers charge investors a fee for the construction
and administration of the vehicle, the overall performance of these FoHFs
should account for additional fees. However, the majority of FoHF man-
agers lack superior selection skills. Since the selection process shrinks the
engendered excess profit, passively managed FoHFs have been launched
by MSCI, S&P, HFR, Credit Suisse First Boston (CSFB)/Tremont, and others
(Lhabitant, 2007). The idea behind investable hedge fund indices is to
cut down selection costs by replacing an active picking process with an
index weighting scheme, and by shifting the attention from best per-
formance orientation to a more conservative approach. This method,
which is based on diversification of hedge fund strategies, provides a
high level of transparency and does not require expert knowledge in the
field of alternative investments (Géhin and Vaissié, 2004).
But investable hedge funds also face problems. For example, the effects
of backfilling bias as well as survivorship bias are restrained by the invest-
able attribute of this vehicle. Neither the liquidation of an underlying
fund nor the inclusion of new funds has a significant impact on the track
records (Géhin and Vaissié, 2004).
However, selection bias implies that the entire universe of hedge funds
is not represented, and is still an issue. In the case of investable hedge
fund indices, the design of these rules is determined by index providers.
Issues such as track record history, liquidity, and fund size are possible
criteria when constructing an investable hedge fund index. In order to
attract more investors, some index providers have extended their own set
of rules with criteria that appear to push performance to entice investors.
Moreover, the investable index method overcomes neither the issues of
non-uniform weighting and rebalancing portfolios nor of unambiguous
strategies of classification.
5.5.1 Closing hedge funds

The idea of an investable index is based on an assumption that underlying
financial instruments are open to new inflows of capital. However, this is
difficult to fulfill when the underlying assets are hedge funds. Many hedge
funds are usually closed to new capital once a certain volume is reached.
Moreover, some hedge funds with superior performance usually have strict
capacity limits and intensive liquidity restrictions. Thus the total expected
remaining capacity of capital is an important indicator of whether it is
legitimate to include a hedge fund in an investable hedge fund index.
5.5.2 Heterogeneity of indices

Heterogeneity can be subdivided into three different areas: overlapping
management, proportion of strategies, and performance difference among
indices. An analysis of the three most popular investable hedge
fund indices reveals that only a small part of the entire pool of hedge
fund managers is included in all of the three indices. For the purpose of
adequate samples, Duc (2004) used data from 2003 on the investable
hedge fund indices of S&P, Tremont, and MSCI. With a total number of
159 listed managers in the three indices, only 14 managers were covered
by two indices. Only three managers were commonly listed in all three
indices. Non-investable hedge fund indices are known to be more hetero-
geneous compared with traditional indices. Investable hedge fund indices,
however, are even more heterogeneous than non-investable hedge fund
indices. This occurs because the average investable hedge fund index
includes a smaller number of funds than their non-investable counterparts.
Duc (2004) states that the coverage of investable indices is 20 times
smaller compared with non-investable indices.
The second cause of heterogeneity is the varying exposure to strategies.
Even though index weighting methods are similar, strategy exposures
are different. The HFR Investable Hedge Fund Indices and MSCI apply
the value weighted approach, but the exposure to arbitrage and relative
value strategies at HFR accounts for 48 percent of total assets under
management, whereas in the case of MSCI, it is only 38 percent of total
asset value. Another example is the long/short equity strategy exposure of
S&P and MSCI. While MSCI invests 45 percent into the long/short strategy,
S&P allocates only 13 percent to this strategy (Duc, 2004).
The third type of heterogeneity is related to the different performances
of the indices. A comparison of the three most popular indices reveals
that the average correlation among the analyzed indices is only 0.77
(Géhin and Vaissié, 2004). The reliability of benchmarks based on invest-
able hedge fund indices is questionable since the three mentioned indices
provide rather diverging performance characteristics.
5.6 Indexing hedge fund indices
Indexing indices could be used to increase representativity of a

benchmark. One method to construct an index of indices is the equally
weighted scheme, which includes all listed indices of a particular strategy.

This approach will give an opportunity to cover as many funds of a certain
style as possible and to mitigate bias and heterogeneity (Duc, 2004).
As suggested by Amenc and Martellini (2002), the approach of analyz-
ing broad economic factors, such as equity risk, fixed income risks, and
slope risk, must be applied to the construction of index indices. This
method is selected to calculate portfolios which explain existing variance
at the highest level. In order to mitigate the issue of low representativity
and to come up with the best possible one-dimensional summary of
the analyzed competing indices, Amenc and Martellini used a Principal
Components Analysis. This technique results in the lowest loss of infor-
mation concerning a specific strategy available in the market. On aver-
age, every included index explains 79.12 percent of the entire computed
portfolio variance. The higher the correlation among hedge fund indices
of similar style, the higher is the fraction of variance explained by the
included index. After analyzing the loss information of a computed port-
folio and calculating the correlation coefficient of generated portfolios
measured over a period of three years, it was stated that the calculated
indices of indices bear a higher level of representativity than other exist-
ing indices (Amenc and Martellini, 2002).
However, the question of bias sensitivity remains. Since the bulk of
variance is explained by this construction method, bias is reduced to the
lowest possible level as well. In conclusion, indices of indices are appro-
priate as benchmarks regarding representativity. Despite the theoretical
qualification of indices of indices, these constructions are not applicable
as market benchmarks because neither the underlying indices nor the
indices of indices are investable, as claimed by investors.
5.7 Factor-based benchmarking methods
The analysis of the various benchmarking tools reveals a rather unsatisfy-

ing result. In order to tackle non-investability, the factor model approach
offers a practical solution. Instead of using indices as underlying factors,
the factor model relies on observable and investable assets, which also
represent return opportunities and risk exposures. The factor model
method combines representativity and investability. The underlying
assumption of a factor model is that a hedge fund’s returns can be
described as a combination of exposures to common factors, such as the
S&P 500 or interest rates.
Jensen (1968) established the foundation of performance replication
by using a single factor method to run a regression on returns of equity
mutual funds and the overall market performance. On the basis of Jensen’s
model, Sharpe (1992) developed an approach called Style Analysis to
describe an asset mix which replicates the style of a traditional equity
fund. Building on William Sharpe’s work, William Fung and David Hsieh
began in 1997 to apply factor analysis to hedge fund replication strategies
(Fung and Hsieh, 1997).
The mathematical formula for factor analysis is:
n
RiF ∑b R
k1
k k
i a ei
RiF return of the hedge fund in month i

b k the beta loading of the fund to factor k
Rik the consistent performance of hedge funds that cannot be explained
by the factors; does not change from month to month
ei the residual return in month i; this is the variation in returns that
cannot be explained by the model
The model requires factor returns to be specified and loadings to

these factors to be estimated using historical data. A risk factor can be
characterized as appropriate when contributing to reductions of the
tracking error on the corresponding index. One may easily suggest that
many economic factors affect hedge fund returns; however, from
both business and technical perspectives it is important to limit the
number of factors. Additionally, the model should be concise enough
for a portfolio manager to follow, and the number of factors should be
reduced to increase the explanatory power of the model. The explanatory
power of a model will diminish when too many factors are compared
to a limited number of monthly fund returns, which leads to model
over-fitting. According to J. P. Morgan (2010) the following major
factors affecting hedge fund returns should be included in a replication
model:4
5.7.1 Beta factors from various market indices

The beta factors are: equity risk, emerging market risk, credit risk, interest
rates, commodity prices, currency rates, and volatility.
5.7.2 Carry factors

Equity spreads: these can be represented by offsetting positions in dif-
ferent types of equities. Thus the relative performance of stocks is best
represented by their market capitalization and their value or growth style.
Other important characteristics are their price momentum, geographic

region, and industry sector.
Interest rate calendar spreads: exposure to yield curve changes can best be
mimicked by offsetting positions in bonds with different durations.
Volatility arbitrage:5 this can best be replicated by a factor that tracks
the difference between implied and realized volatility through variance
swaps.
Merger arbitrage: this can be replicated by an index that holds all stocks
that are currently involved in a pending merger.
FX carry returns: this can be mimicked by the use of an index that buys
currencies with high interest rates and sells currencies with low interest
rates.
Illiquidity: a useful factor to represent the amount of illiquid investments
within hedge funds is given by the fund’s lagged returns respectively the
autocorrelation of the fund’s returns.
5.7.3 Beta timing factor

Macro trend: as many global macro funds use the long-term trend of major
markets as a trading signal, a simple trend-following rule can be used as
replication factor.
Once a set of factors has been selected, linear regression is used to esti-
mate the factor loadings. The most common method is to apply a rolling
window on monthly data. Even though this method is relative simple
there are some drawbacks: the simple stepwise linear regression of the
factors cannot make up for these dynamic risks, because linear regression
represents the average exposure to a diverse set of risks of the past in-
sample period. However, the real past risk factors do not behave in a linear
and static way. Hence hedge fund risks must be mimicked by a dynamic
and nonlinear set of risk factors. Fung and Hsieh (2002b) try to compen-
sate for the dynamic elements by adding option-like elements as a look-
back straddle to the regression model. Despite the attempt to account for
dynamic factors, the authors admit that this method is not able to provide
sufficient additional explanatory power for real dynamic risk exposure.
Thus other modeling techniques, such as weighted least squares regression,
Kalman filters and neural networks should be used to improve the accuracy
of factor loading estimates with the same amount of historical data.
5.8 Evaluation of replicated hedge fund results
In the following we present empirical studies which apply the suggested

factor-based benchmarking model. The appraisal of these studies reveals
that the accuracy of this approach is not totally satisfactory. The results
for in-sample R² are insufficient, which means that the unexplained
variance of replicated returns is high. Moreover, out-of-sample results
indicate that volatility moves in an inappropriate range. An analysis of the
actual out-of-sample returns shows that they are usually outperformed by
their corresponding index.
5.8.1 Fung and Hsieh (2002b)

Fung and Hsieh (2002b) analyze five different fixed-income strategies by
combining a return-based style factor approach with an asset-based style
factor approach. They also use the instrument of look-back straddles to
reduce the effect of dynamic trading. This method helps to increase the
R² by 1 to 6 points. The study period from 1990 to 2000 reveals an R²
of 59 percent to 79 percent (see Table 5.1). In particular, the in-sample
results for the fixed-income convertible bond strategy and the fixed-
income high-yield strategy are well explained by the corresponding factor
models. However, there are no out-of-sample results available for the five
fixed-income strategies available.
Table 5.1 In-sample R² of five fixed-income strategies
Strategies In-sample R2 1990–2000

Fixed-income diversified 64%
Fixed-income convertible bond 70%–75%
Fixed-income mortgage-backed 59%–66%
Fixed-income arbitrage 66%
Fixed-income high-yield 78%–79%
Adapted from Fung and Hsieh (2002b).
5.8.2 Agarwal and Naik (2004)

Agarwal and Naik (2004) analyzed eight hedge fund strategies, focusing
on the in-sample period from 1990 to 2000 (see Table 5.2). They applied a
stepwise regression approach for the eight hedge funds. The factor model
uses buy-and-hold and option-based risk factors. The buy-and-hold factors
are based on elements such as currency, commodity, bond, and equity
indices, as well as book to-market-factors. The option-based factors refer
to European put and call options. In the first step, the authors defined
the factors using a significance level of 5 per cent for the inclusion of
factors in the stepwise regression. In the second step, the obtained factor
models are applied to the out-of-sample period from 2000 to 2001 and
assess the quality of the replicated portfolios. The analysis of the in-sample
data reveals a nonlinear risk–return relationship between most strategy
type indices. The payoff of five of the eight indices is explained to a large
extent by writing put options on the market index. According to Agarwal
and Naik (2004) this is either due to the fact that these five strategies relate
to economic activity and the money decline when equity markets show
negative development, or because managers want to create payoffs that are
similar to the payoffs from put option writing. This tactic is often applied
by managers due to their incentive structure and the objective of improv-
ing the Sharpe ratio. The adjusted R² ranged from 40 to 92 percent for the
in-sample period. Furthermore, the t-test and the Wilcoxon sign-test of the
out-of-sample returns reveal that performance differences between selected
HFR indices and the corresponding clone portfolios are significant.
Table 5.2 In-sample R² of eight equally weighted HFR

indexes

Event arbitrage 44.04%
Restructuring 65.57%
Event-driven 73.38%
Relative value arbitrage 52.17%
Convertible arbitrage 40.51%
Equity hedge 72.53%
Equity non-hedge 91.63%
Short selling 82.02%
Adapted from Agarwal and Naik (2004).
5.8.3 Fung and Hsieh (2004)

Fung and Hsieh (2004) conducted another study by combining the
diverse hedge fund strategies in a single factor model. Seven style factors
are derived from three individual hedge fund strategies. These factors are
than combined in a diversity-representing factor model. After defining
the factor model a regression of the HFR index on the seven selected
factors is performed. The analysis is designed to present R² for three time
periods. The in-sample analysis of the first period, from January 1994
to September 1998, reveals an R² of 69 Percent. During the second period,
from April 2000 to December 2002, the in-sample R² is 80 percent. The
third period included data from the first and second periods as well as
September 1998 to April 2000. During this period a coefficient of deter-
mination of 55 percent was calculated. No out-of-sample analyses were
conducted in this paper. The authors assume that the model will help
to spot the alternative betas, and thus alternative alphas as well. Despite
the statement that the seven asset-based style factors are able to explain
up to 80 percent of the return variation on a monthly base, the proposi-
tion that an undistorted comparison of hedge fund indices is possible
must be treated with caution, since the overall explanatory power of 55
percent is not considered to provide a sufficient in-sample fit.
5.8.4 Jaeger and Wagner (2005)

Jaeger and Wagner (2005) used a multilinear factor model to replicate
various strategies (see Table 5.3). The multifactor regression was applied to
a range of strategy indices provided by HFR. Subsequent to the factor
definition, a Replicating Factor Strategy (RFS) was calculated. The RFS
calculated monthly returns with underlying factors, which were derived
from a data regression on the corresponding indices for the in-sample
period of five years. The authors state that their model is able to explain
60 percent of the return variation on average. In particular, equity and
short selling strategy are explained well by their multifactor model.
However, a vital share of return variation remains unexplained for equity
neutral market and managed future strategies. Seven replicated port-
folios exhibit an R² which is below the 70 percent mark.
In order to evaluate the RFS, Jaeger and Wagner calculated the returns
of the model to compare the obtained performance with the real out-of-
sample performance of the corresponding investable and non-investable
indices (see Table 5.4). For the out-of-sample period, index data from
March 2003 to August 2005 are used. Only the replicated portfolio of
Table 5.3 In-sample R² of 11 HFR indexes

Equity hedge 88.5%
Equity market neutral 35.3%
Short selling 81.2%
Event-driven 79.3%
Distressed 68.4%
Merger arbitrage 52.9%
Fixed income arbitrage 40.5%
Convertible arbitrage 54.0%
Macro 49.7%
Managed futures 34.3%
Managed futures trend followers 35.4%
Adapted from Jaeger and Wagner (2005).

Table 5.4 Out-of-sample evaluation of the replicated clones of Jaeger and

Wagner
Strategy RFS HFRX HFRI

Equity hedge 27.8% 16.0% 32.8%
Equity market neutral 6.2% –3.9% 10.9%
Short selling –28.2% N/A –23.0%
Event-driven 29.8% 24.1% 40.0%
Distressed 20.1% 23.3% 44.8%
Merger arbitrage 13.0% 10.9% 15.3%
Fixed income arbitrage 7.8% N/A 16.3%
Convertible arbitrage 7.6% –5.3% 2.4%
Macro 16.7% 10.1% 24.6%
Managed futures 9.2% N/A N/A
Adapted from Jaeger and Wagner (2005).
the convertible arbitrage strategy outperformed the non-investable HFRI

index. Regarding the comparison with the investable HFRX index, the
replicated clone outperformed the benchmark in six of seven cases.
5.9 Conclusion
This chapter evaluates benchmark methodologies in the hedge fund

industry. We conclude that none of the analyzed fund-based benchmark
concepts fulfill the benchmark requirement of being representative
and investable. Hence none of the fund-based methods are eligible as
benchmarking tools. The factor-based approach which acts as the basis
for hedge fund replication strategies was introduced to offer a possible
solution based on the index of indices idea, but replacing the under-
lying indices by independent risk factors. These risk factors mimic a
representative index and provide an attribute of investability, as every
risk factor refers to observable and liquid assets. Empirical analyses
of in-sample and out-of-sample results show, however, that the factor
models presented fail to construct replication portfolios which exhibit
adequate accuracy.
Notes
1. The most common way for a hedge fund to gain a competitive edge is to
invest in research. Typically, long/short equity funds employ a team of experi-
enced analysts who conduct in-depth research on stock valuations in an effort
to uncover pricing anomalies and thus exploit advantages over other financial
market participants.
2. Since Géhin and Vaissié have published a comprehensive analysis of this

benchmarking approach, the following section on fund-based concepts refers
to the arguments of Géhin and Vaissié (2004).
3. Zurich provided five strategy indices, whereas Morgan Stanley Capital
International (MSCI) used over 190 in 2003 (Géhin and Vaissié, 2004).
4. For a more detailed overview about the “ABCs” of hedge fund investing see
Alpha, Beta and Carry: The ABCs of hedge fund investing (J.P. Morgan, 2010).
5. Many hedge funds are interested in exploiting the overpricing of options.
References
Agarwal, V. and Naik, N. Y. (2004). “Risks and Portfolio Decisions Involving
Hedge Funds.” Review of Financial Studies, 17(1): 63–98.
Amenc, N. and Martellini, L. (2002) “The Brave New World of Hedge Fund
Indices.” Working Paper, EDHEC Graduate School of Business, EDHEC-
Risk Institute, Nice, France. Available at http://www.edhec-risk.com/indexes/
edhec_position/index_html/attachments/brave_new_world.pdf
Anson, M. (2004). “Hedge Fund Indexes: Benchmarking the Hedge Fund
Marketplace.” Working Paper, EDHEC Graduate School of Business, EDHEC-Risk
Institute, Nice, France. Available at http://www.edhec-risk.com/research_news/
choice/RISKReview108089352936435435/attachments/HF%20Indexes%20-
%20Anson.pdf
Duc, F. (2004). “Hedge Fund Indices: Status Review and User Guide.” Working
Paper, 3A S.A.—Alternative Asset Advisors, Zurich, Switzerland. Available
at http://www.3-a.ch/filedownload.lbl?uid099929F9-6BD6-43B3-BB50-
AE2BBC53B146.
Fung, W. and Hsieh, D. A. (1997). “Empirical Characteristics of Dynamic Trading
Strategies: the Case of Hedge Funds.” Review of Financial Studies, 10(2): 275–302.
Fung, W. and Hsieh, D. A. (2000). “Performance Characteristics of Hedge Funds
and Commodity Funds: Natural vs. Spurious Biases.” The Journal of Financial
and Quantitative Analysis, 35(3): 291–307.
Fung, W. and Hsieh, D. A. (2002). “Benchmarks of Hedge Fund Performance:
Information Content and Biases.” Financial Analysts Journal, 58(1): 22–34.
Fung, W. and Hsieh, D. A. (2002b). “The Risk in Fixed-Income Hedge Fund
Styles.” Journal of Fixed Income, 12(2): 6–27.
Fung, W. and Hsieh, D. A. (2004). “Hedge Fund Benchmarks: A Risk-Based
Approach.” Financial Analysts Journal, 60(5): 65–80.
Géhin, W. and Vaissié, M. (2004). “Hedge Fund Indices: Investable, Non-
Investable and Strategy Benchmarks.” Working Paper, EDHEC Graduate
School of Business, EDHEC-Risk Institute, Nice, France. Available at http://
www.edhec-risk.com/edhec_publications/RISKReview1099410456098418642/
attachments/HF%20Indices%20Investable%20Non%20Investable%20and%20
Strategy%20Benchmarks.pdf.
J. P. Morgan. (2010). “Alpha, Beta and Carry: The ABCs of Hedge Fund Investing.”
Working Paper. J. P. Morgan, New York City, New York.
Jaeger, L., and Wagner, C. (2005). “Factor Modeling and Benchmarking of Hedge
Jensen, M. C. (1968). “The Performance of Mutual Funds in the Period 1945–1964.”

Journal of Finance, 23(2): 389–416.
Lhabitant, F.-S. (2007). “Hedge Fund Indices for Retail Investors: UCITS Eligible
or not Eligible?” Derivatives Use, Trading & Regulation, 12(4): 275–89.
Liang, B. (2000). “Hedge Funds: The Living and the Dead.” The Journal of
Financial & Quantitative Analysis, 35(3): 309–26.
Sharpe, W. F. (1992). “Asset Allocation Management Style and Performance
Measurement.” Journal of Portfolio Management, 18(2): 7–19.
6
Insight—Distributional Hedge
Fund Replication via State
Contingent Stochastic Dominance
Clemens H. Glaffig
6.1 Introduction
The topic of hedge fund replication is well established, both in the

academic world, with a growing list of literature, as well as in applications
with an active market for replication—and alternative beta products. Its
theoretical concepts were translated into market applications early on,
generating wide interest among investors, even though the resulting
products did not always perform to expectations. In Section 6.2 we com-
ment on the evolution of hedge fund replication, including a view as to
the role that replication techniques can play in applications in the future:
not constructing simple clones, but improving them on specific perform-
ance characteristics. Section 6.3 describes the role of hedge fund replica-
tion, and in Section 6.4 we give brief descriptions of previous replication
techniques. In Section 6.5 we highlight some of the pitfalls of common
replication techniques. Section 6.6 introduces a hybrid replication tech-
nique, combining the intuitive portfolio construction of factor-based
replication with the less ambitious goal of replicating certain preferred
and state-contingent characteristics of the return distribution rather than
replicating the performance path. This approach is specifically geared to
accommodate the role of replicating the good—and improving on the
bad—performance aspects of a given hedge fund target. In Section 6.7
we give an empirical application of the proposed replication technique
and compare it to factor-based regression, the most common replication
technique thus far. Section 6.8 concludes the chapter.
6.2 The evolution of hedge fund replication
The end of the bull market in equities at the turn of the century presented
investors with unaccustomed volatility and excessive losses. It resulted in
76
Clemens H. Glaffig 77
rethinking some aspects of the classical approach to asset management of

benchmark driven investments, diversifying into benchmark-free absolute
return strategies.
Hedge funds, thus far a rather mystical and secretive class of investment
strategies granting little insight and reporting infrequently, rose sharply in
popularity with mainstream investors. As it opened the segment to more
traditional investors, the need to understand, or at least shed some light
on, the inner workings of hedge funds arose. The first efforts in hedge
fund replication developed in that context, motivated by the desire to
detect and understand the main risk factors driving hedge fund perform-
ance by regressing historic hedge fund returns against the return of a set
of style factors.
While the initial attempt was to reproduce the historic performance
path of a given target fund, it was soon realized that these techniques
could also be used to replicate hedge fund returns on a forward-looking
basis. The vision was to construct a recipe for allocating funds within
a small universe of liquidly tradable instruments exhibiting identical
performance behavior than the target fund but being fully transparent,
highly liquid and with substantially lower fees than what the target fund
would charge.
The original hype and hope quickly disappeared, as it became apparent
that the ability to track a historic performance path successfully does not
necessarily translate into producing enough insight into the true nature
of the targeted strategy to construct a clone. The dynamic features,
specific trader’s talents, the granularity achieved by successful multi-
strategy funds, or other idiosyncrasies are too often just too dominant
a factor in hedge fund performance characteristics and are not readily
adapted to simple, semi-static modeling. Consequently, the results were
mixed at best.
To address these deficiencies hedge fund replication went on to concen-
trate on tracking indices. The hope was that idiosyncrasies would average
out for indices, and that more systematic and easily cloned features of
performance characteristics would dominate. Here, the results were more
promising, although they very much relied on combining the right instru-
ments and trading rules within the replicator to capture the dynamic and
nonlinear aspects that are still prevalent in style indices; see, for example,
Amenc et al. (2008) and Gupta, Szado, and Spurgin (2008).
To pursue the ambitious objective of path replication, new replica-
tion techniques, such as rule-based trading or distributional replication,
aimed to capture and replicate the general, essential features of specific
styles, referring to them as alternative beta strategies, with no or little
78 Insight—Distributional Hedge Fund Replication
HFR Index JP Alternative Beta Ref.

ML Factor Model Index GS Absolute Return Tracker
110.00
105.00
100.00
95.00
90.00
85.00
80.00
75.00
02.01.2008 02.01.2009 02.01.2010 02.01.2011
Figure 6.1 Performance graph for the daily performance from January 2, 2008
to March 31, 2011
ambition to clone the performance path any more; see, for example, Kat
and Palaro (2005) for distributional replication. For a recent comparative
survey of replication products in the market, see Tuchschmid, Wallerstein,
and Zaker (2009).
Competing with replication products are funds of hedge funds and,
more recently, portfolios of real hedge funds in the form of ETFs, promoted
by larger providers of managed account platforms with the advantage of
cost-efficient access. Figure 6.1 illustrates the daily performance graphs of
the HFR General Hedge Fund Index against three replication examples:
The GS Absolute Return Tracker, the JP Alternative Beta Reference and the
ML Factor Model Index.
6.3 The role of hedge fund replication
Replicating the essential performance features of a style is the objective

nowadays, followed by most applications of hedge fund replications in
the market. As products that capture and reflect the essential perform-
ance characteristics of general styles, they provide a liquid, transparent,
and cheap basis on which to construct a portfolio of hedge funds within a
coresatellite approach. They simplify and speed up the allocation
to different style sectors and even provide a means to shorten essential
aspects of specific style performance attributes, serving as an efficient
risk management tool for portfolios of hedge funds.
The scope could be extended to provide a basis for what investors really
want and expect from style-related alternative beta strategies: a product
that reflects the sum of all the good attributes of a hedge fund style,
with some of the undesired elements (tail correlation, etc.) removed: not
style clones but products that dominate style indices based on bespoke
criteria, while still exhibiting major performance characteristics of
the chosen style. The reflected bespoke preferences could go well beyond
correlation to existing portfolios or preferences with respect to moments
of the return distribution.
6.4 Replication techniques
Most replication techniques belong to one of two main classes: factor-

based and distributional replication. In addition there are a number of
hybrid as well as individual approaches.
6.4.1 Factor-based replication

Factor-based replication, the most common replication technique, is
based on Sharpe’s concept of style factors to capture the essentials of
an active management of classical strategies. Fung and Hsieh (1997)
extended this approach to hedge funds and hedge fund styles. Factor-
based replication is in general a parameterized, rule-based strategy, in
which the parameters are adjusted to best replicate or track a given
performance path over a specific history. It not only tries to match returns,
but also the sequential order in which returns are realized.
In the most common case, factor replication is a multilinear regression,
the parameters just the factor loadings or the betas with respect to stand-
ard investable market risk factors; the rules produced are consequently
just the static allocations according to these betas. More elaborate models
include option-based strategies and alternative indices reflecting certain
rule-based trading strategies as factors; see, for example, Fung and Hsieh
(1997, 2002), Gehin, Martellini, and MeyFredi (2007), Hasanhoddzic and
Lo (2007), Jaeger and Wagner (2005), and Spurgin (1999), and references
therein.
The advantage of factor-based replication lies in its simplicity. The
art and value added of this approach rests with the specification of the
factors: the quality of the replication depends heavily on capturing
the dynamic aspects, including carry aspects and tail event behavior,
which can be improved by including alternative indices reflecting non-
standard trading strategies (e.g. the BMX index) or individually defined
rule sets.
6.4.2 Distributional replication

Distributional hedge fund replication is the approach advocated by Kat and
Palaro. It is based on the dynamic replication developed in the framework
of contingent claim valuation. For a given portfolio, it results in matching
the joint distribution of the replicator and the portfolio to that of the
target and the portfolio on a percentile by percentile basis; see for example,
Gehin, Martellini, and MeyFredi (2007) and Kat and Palaro (2005, 2007).
In contrast to factor-based replication, its goal is not to approximate the
performance path but the distributional aspects, specifically the frequency
counts of returns, regardless of the sequential order in which these returns are
realized. It is thus less ambitious than factor-based replication. Distributional
replication is potentially less prone to over-fitting, while it may result in
completely different realizations of performance paths. The implementa-
tion, however, is clearly more complicated than factor-based replication.
6.4.3 Hybrid theory and other alternatives

Further approaches combine factor-based models with aspects of distribu-
tional replication; see, for example, Kazemi (2007) and Section 6.5 below.
Other examples use, for example, tracking techniques well established in
various non-financial applications based on Markov Chain Monte Carlo
methods that may be suited to deal with non-stationarities; see Roncalli
and Weisang (2008).
6.5 Some pitfalls in replication
Using hedge fund replication techniques to produce cheap and trans-

parent clones of hedge funds has thus far only been moderately suc-
cessful. Plenty of explanations for this lack of success have been given.
The most obvious deficiency of common replication techniques is the
lack of dynamics and the difficulty of modeling idiosyncrasies (“trader’s
talent”). More generally, any non-stationary trading behavior will be
difficult to model. While these problems weaken somewhat by follow-
ing less ambitious replication objectives or averaging out idiosyncrasies
by targeting indices instead of single funds, a number of challenges
remain, that can, even if the objective is only to gain insight into and
understanding of past performance, result in misleading conclusions.
We will highlight and repeat some—but certainly not all—of the
pitfalls of hedge fund replication.
• Over-fitting of factor-based replication: as with all parametric

approaches, adding ever more factors to a replicator will clearly improve
in-sample tracking performance, but is very prone to over-fitting.

Capturing the dynamics of the target necessitates the inclusion of
additional non-standard factors, deceiving us into overloading with
different factors that only marginally contribute to the explanation.
Prudent pruning is required. Closely related to this is:
• Factor misspecification: potential factor misspecifications are
manifold. The desire to capture the detail, dynamics, and nonlinear
performance behavior of hedge funds often leads to the inclusion
of a large number of highly inter-correlated factors. If those factors
do not play a role in the target fund’s strategy, the individual factor
loadings will often not optimize to zero but will produce combina-
tions of long and short positions that have no relevance to the target,
resulting not only in over-fitting, but also in a misconception of
the target’s risk and strategy. As another specific example, successful
market timing within the observed data frequency does create alpha.
Sometimes, for very pronounced timing talents, this may look like
being long a look-back option. Often though, this alpha looks very
much like option selling or some other positive carry strategy on the
larger observation scale. Typically, factor-based replication will model
this with an option selling factor, exhibiting an asymmetric, nega-
tively skewed performance behavior, which the target may not have.
• Misspecification of the distribution: distributional replication
depends very heavily on getting the distribution right. Real-life distribu-
tions are typically approximated by well-behaved and easy to handle
distribution classes. Given sparse data like monthly returns, these
approximations are very crude and poorly separate different strategies.
Fitting multivariate copulas to model multivariate distributions can-
not be properly done based on the available data history of most
funds or indices.
• Incomplete data: parameters of replication models are optimized
throughout some historical period. If those are incomplete (e.g., real
tail events are not included) behavior in extreme market phases will
not be reflected. Consequently, risk management and stop loss
behavior are not adequately modeled. In general, the data used may
be insufficient to properly separate different strategies, leading to
similar behavior in the given period, but potentially big perform-
ance differences in market phases not covered by the data used to
optimize parameters. This leads to misspecifications specifically for
rule-based trading replications.
• Data frequency: an increasing number of funds trade at high-
frequency intervals. Modern statistical arbitrage has evolved to a good
degree into high-frequency trading. Others, like short-term CTAs,

close out or minimize their risk at the end of each trading day.
Working with monthly return data to understand the aspects of such
trading strategies is meaningless. Even if more funds, specifically under
the format of UCIT III in Europe, report daily performance numbers,
high-frequency strategies or indices with a large component thereof
will be very difficult to replicate without very high misspecification
risk. Data frequency will also influence the factor choice: If trading
frequency differs systematically from observation frequency—for
example daily trading, monthly observations—the trading range and
relative closing level of factors will become more important than
absolute closing levels.
6.6 A new hybrid approach—dominance replication
In the following we highlight a hybrid approach of factor-based and

distributional replication. Distributional replication, as developed by
Kat and Palaro in a series of papers (see Kat and Palaro, 2005) is in gen-
eral a two-step process, based on the replication of contingent payoffs
as initially developed by Merton in the context of Black–Scholes option
pricing theory: for a given fund, H, a so-called reserve asset, R, which
drives the replication and an initial portfolio or benchmark B, a payoff
function g is replicated, such that for rB, rR, and rH denoting the respec-
tive period returns for B, R, and H
P( g ( rR , rB ) u1 , rB u2 ) P( rH u1 , rB u2 ) ∀u1 , u2 ∈ [ 0, 1] (6.1)
(see also Gehin, Martellini, and MeyFredi, 2007).

The replication of g follows the general theory of replicating contingent
claims.
If we identify the payoff function g with its replication X, i.e. X
g(R, B) and rX g(rR, rB), Equation (6.1) is equivalent to
P( g ( rR , rB ) < u1 | rB u2 ) P( rH u1 | rB < u2 ) ∀u1 , u2 ⇔ FX|B ( u ) FH|B ( u )∀u (6.2)
where FX|B , FH|B denote the conditional distribution function of X (respec-

tively H) given B.
Equation (6.2) states that the conditional distribution functions of
the replication and the fund are matched. Another way to express this
is to say that they have identical state-contingent first-order stochastic
dominance properties.
We will take this view to propose an alternative way for replication,

which we will refer to as dominance replication. Our motivation for a
new approach is threefold:
• We would like to replicate the preference-based state-contingent

distributional performance characteristics of a specified fund or index
to arbitrary fine detail.
• We would like an intuitive construction of a state-dependent repli-
cating strategy as given by factor replication, i.e., the allocation to
various risk factors in a state-contingent way.
• We would like to limit the potential for over-fitting, which is often
the price to be paid by adding too many parameters and incorporat-
ing dynamic strategies.
We will express state-contingent performance characteristics by con-

sidering the dominance of the state conditional distribution of returns
over target distributions, which reflect and represent these characteristics
in a state-contingent way: Let B denote some benchmark, which could
be an index, a given portfolio or a general state indicator of the market.
A realization of B will be denoted a state; state contingency will mean
conditioning on B b for some specific state b. Let (B) be a family
of one-dimensional conditional target distributions, conditioned on B,
reflecting the specific conditional performance preferences with | | N.
Each T ∈ will describe a state-contingent distributional target char-
acteristic, represented by its respective conditional target distribution
FT|Bb. A fund X will be measured against these target characteristics by
evaluating a version of integral stochastic dominance of the conditional
distribution function FX|Bb over FT|Bb. If two funds X and H have the
same dominance values for a given set of target characteristics, they will
be viewed to have the same state-contingent distributional performance
characteristics (relative to the chosen target set).
Example: let T describe a skew target relative to the benchmark B. This
could be represented by a distribution that, for each realization b of B,
is scattered around b with conditional expectation exceeding b by some
e > 0, that is, E(T|B b) b e. Two funds X and H with the same domi-
nance value over T will have the same conditional skew characteristic.
6.6.1 Definition
For a singlet T, we call
∞ ∞

T,B ( X) =
∫ ∫
−∞ −∞
( FT|B ( x | b) − FX|B ( x | b))dFT|B ( x | b)dFB ( b)
the state contingent dominance of X over the target T, contingent on the

benchmark B.
We call
⎧ ∞ ∞ ⎫
,B ( X ) ⎨
∫ ∫
⎩ ∞ ∞
( FTn |B ( x | b) FX|B ( x | b))dFTn |B ( x | b)dFB ( b), n 1, 2,… , N ⎬ ∈R N
⎭
{
Tn |B ( X ), Tn ∈ , n 1, 2,… , N }
the state contingent dominance vector of X over the performance target class
contingent on the benchmark B.
The vector
,B ( X) reflects to what extent X has better distributional
performance characteristics than the target characteristics we have
imposed by defining .
With this, we will call X a dominance replication of H if
,B ( X)
,B (H). Given a set Ξ of available strategies, we call X0 the best domi-

nance replication under if:
X 0 arg min{||
,B ( X )
,B (H) ||}
X ∈Ξ
Matching contingent dominance vectors seems to be less ambitious

than matching conditional distribution functions, as this is equivalent
to matching first-order stochastic dominance properties. However, as we
control the set , the distributional performance characteristics of arbitrary
fine detail can be reflected. With this flexibility, distributional replication
as defined by Equation (6.1) is a special case of dominance replication.
Let : {T , T|Bb0 x0 , b0 , x0 ∈ R } that is, comprises the set of all
“double delta” distributions dFT|Bb δx0 ( x)δb0 ( b) . Matching the domi-
nance set for X and H is equivalent to FX|Bb0 ( x0 ) FH|Bb0 ( x0 ), ∀b0 , x0 ,
which is Equation (6.2).
6.6.2 The replicator

We will consider replicators from a set of parameterized strategies. As
in factor-based replication, we choose a set of asset based style factors
(“ABSFs”) Ri, which can include standard market indices as well as
predefined strategy indices. The replicator set is then defined by
R: {X, X Σβi Ri , i 1, 2,… , M }
where the bi satisfy some normalization condition and are allowed to

be state- and path-dependent to allow for, for example, stop losses and
CPPI-like features.
The dominance replication strategy will then be the strategy in the

replicator set satisfying:
X 0 arg min{||
,B ( X )
,B (H) ||} arg min
X ∈R X ∈R
⎧⎪ ⎫⎪
⎨
⎩⎪
∑|
k
Tk ,B ( X )
Tk ,B ( H) |, k 1, 2,… , | |⎬
⎭⎪
While dominance replication is multilinear in the ABSFs, it is different

from factor-based multilinear regression: The optimization of the param-
eters is not done by minimizing the squared distance of the respective
returns, but by minimizing the distance of conditional dominance vectors
reflecting bespoke characteristics of the return distribution.
6.6.3 Motivation and value added for dominance replication

One of the goals to be achieved by a novel technique in hedge fund
replication is to improve on the deficiencies of previous approaches.
Factor-based replication tries to match the performance path and hence
the sequential returns, i.e. size and order of the realized returns. This
stringent objective easily leads to over-fitting and factor misspecification.
Dominance replication tries to match some chosen performance char-
acteristics in a weak distributional sense instead, reducing the risk of
over-fitting. Given that it matches integrated conditional distributions,
it is less sensitive to misspecification of the distribution than in the case
of distributional replication, which matches cumulative distribution
functions point by point. Dominance replication can retain some aspects
of path dependence by matching dominance over a whole set of state-
conditional distribution targets, each single target adding an anchor
with respect to sequential ordering. In addition, dominance replication
extends easily to an approach for constructing superior strategies in the
sense that they dominate given funds, styles, or indices on a bespoke
performance characteristics basis. As a drawback, implementing domi-
nance replication is clearly more elaborate than factor-based replication.
6.7 Empirical applications
We apply dominance replication to track the HFR Equity Hedge Index

and compare the resulting in-sample and out-of-sample results on a daily
basis to the standard factor-based replication, using the same ABSFs for
both cases. The HFR Equity Hedge Index provides delayed daily data back
to 2003. We optimize parameters for both approaches over a three-year
period, that is, over roughly 750 data points. We will then let the resulting
replicator run with the optimized parameters for the next month, that is,
roughly 20 trading days, at which point we will repeat the optimization
procedure to readjust the parameters. The initial starting date for the
out-of-sample period is January 1, 2010. The period for determining the
parameters ranges from January 1, 2006 to December 31, 2009. We will sub-
sequently move both the training period as well as the out-of-sample start-
ing point forward by one month each for the following 12 months, such
that the procedure produces 12 months of daily out-of-sample returns.
The ABSFs we use are: S&P 500, a CPPI strategy on the S&P 500, VIX
Index, BMX index of covered call writing, MSCI EM & EAFE, Russell 2000,
Russell 1000 Growth, and Russell 1000 Value. We use constant factor
loadings.
For dominance replication the benchmark we use is a market state
indicator, based on the S&P 500, distinguishing seven states: {very nega-
tive, negative, slightly negative, neutral, slightly positive, positive, very
positive}, defined via quantiles of the five-day S&P 500 returns. To estimate
the conditional distribution of replicating strategies, we use kernel-based
approximation of the empirical conditional distribution—there are enough
data points for all individual realizations of the state indicator to justify this
approach. The state-conditional target distributions against which we match
the respective dominance of replicator and HFR index are chosen as:
T0 |Bb ( x ) ~ N( 0, σ )∀ b
for the degree of dominance over state-independent pure random scatter-

ing around zero return, where N(0, σ ) denotes the Gaussian distribution
with zero mean and variance s.
Tk1 |Bb ( x) ~ N( 0, σ )1( b{ k }) , k 1, 2,… ,7
for the degree of dominance over state-contingent pure random scattering

around zero return, separately for each market state.
Tk2 |Bb ( x) ~ 0.5[{“negative tail event”}

{“positive mean event”}]1( b{ k }) , k 1, 2,… ,7
for the degree of dominance over state-contingent tail skews, separately

for each market state.
6.7.1 Results
The results for the out-of-sample replications are summarized in Tables
6.1 and 6.2. The out-of-sample r-squared of the standard factor-based
Table 6.1 Goodness of fit measures, using daily and monthly out-of-sample
return data for the period January 1, 2010 to December 31, 2010
Correlation r-squared
Daily Monthly Daily Monthly
Dominance 1% 61% 65% 66%
Regression 1.4% 84% 12% 23%
Table 6.2 Moments for the out-of-sample daily return series for the period
January 1, 2010 to December 31, 2010
Mean Std. Dev. Skew Kurtosis

Dominance 0.026% 0.588% 7.98 3.242
Regression 0.017% 0.257% –42.71 1.956
HFR Index 0.024% 0.401% –13.02 1.007
regression with daily return data is only 12 percent, while the out–of-sample
r-squared of the dominance replication is 65 percent. Improvements
specifically for the regression can be obtained by de-correlating the factor
set. However, as can be seen from Figure 6.2, the tracking result for the
regression is optically not nearly as bad as the r-squared would suggest:
A lot of investors can live with such a replication. Nevertheless, if the
objective is to track the index in some sense, the tables and the exhibit
indicate that dominance replication has captured the essentials and
gained insight into the inner workings of the index better than regres-
sion replication.
6.8 Conclusion
Dominance replication is a parameterized replication strategy in which

parameters are optimized not by trying to replicate the performance
path but by replicating preferred state contingent performance charac-
teristics. It concentrates on aspects of the return distribution that are
important for an individual application and can be tuned to arbitrary
fine detail, which makes it more general than classical distributional
replication. Dominance replication can easily be extended to go one
step beyond pure replication: replicate desired characteristics and
improve and dominate undesired characteristics.
HFR Equity Hedge Index Dominance Replicator
HFR Equity Hedge Index Regression Replicator
Figure 6.2 Out-of-sample performance graphs for the period January 1, 2010 to
December 31, 2010
References
Amenc, V., Gehin, W., Martellini, L., and MeyFredi, J. (2008) “Passive Hedge Fund
Replication—A Critical Assessment of Existing Techniques.” The Journal of
Alternative Investments, 11(2): 6983.
Fung, W., and Hsieh, D. (1997) “Empirical Characteristics of Dynamic Trading
Strategies: The Case of Hedge Funds.” The Review of Financial Studies, 10(2):
275302.
Fung, W., and Hsieh, D. (2002) “The Risk in Fixed Income Hedge Fund Styles.”
Journal of Fixed Income, 12(2): 627.
Gehin, W., Martellini, L., and MeyFredi, J. (2007). “The Myths and Limits of
Passive Hedge Fund Replication.” Edhec Position Paper, Edhec Business School,
Nice, France.
Gupta, R., Szado, E., and Spurgin, W. (2008). “Performance Characteristics of
Hedge Fund Replication Programs.” Working Paper, CIDSM, Amherst, MA.
Hasanhoddzic, J., and Lo, A. (2007). “Can Hedge Fund Returns be Replicated?:
The Linear Case”. Journal of Investment Management, 5(2): 545.
Jaeger, L., and Wagner, C. (2005). “Factor Modeling and Benchmarking of Hedge
Funds: Can Passive Investment Strategies Deliver?” The Journal of Alternative
Investments, 8(3): 936.
Kat, H., and Palaro, H. (2005). “Who Needs Hedge Funds? A Copula-Based
Approach to Hedge Fund Return Replication.” Working Paper #027, Alternative
Investment Research Centre Working Paper Series, Cass Business School, City
University, London.
Kat, H., and Palaro, H. (2007). “Alternative Routes to Hedge Fund Return
Replication.” Working Paper #037, Alternative Investment Research Centre
Working Paper Series, Cass Business School, City University, London.
Kazemi, H. (2007). “A Note on the Replication of Properties of Hedge Fund Returns.”
Working Paper, Isenberg School of Management, University of Massachusetts,
Amherst, MA.
Roncalli, T., and Weisang, W. (2008). “Tracking Problems, Hedge Fund
Replication and Alternative Beta.” Working Paper, Available at: http://ssrn.
com/abstract1325190, date posted: January 12, 2009.
Spurgin, R. (1999). “A Benchmark for Commodity Trading Advisor Performance.”
The Journal of Alternative Investments, 3(4): 1121.
Tuchschmid, N., Wallerstein, E., and Zaker, S. (2009) “Hedge Fund Clones are Still
to Count on.” Working Paper, Available at: http://ssrn.com/abstract1729109,
date posted: August 26, 2009.
7
Nonparametric Hedge Funds and
Replication Indices Performance
Analysis: A Robust Directional
Application
7.1 Introduction
Since the pioneering work by Treynor, Sharpe, and Jensen, many

performance measures have been introduced and empirically applied
for evaluating the performance of hedge funds (HF) and HF replication
indices. Recently, production frontier methods have been used in this
field (e.g., Gregoriou, Sedzro, and Zhu, 2005), since they do not require
the specification of a benchmark (such as in standard multifactor models)
and they do not assume any statistical properties of fund returns (e.g.,
normality assumption). In addition, they also have the considerable
advantage of being multi-dimensional.
In such a framework, the performance of each fund is measured
through its distance to the frontier of the production set. In a classical Data
Envelopment Analysis (DEA) setting, it relies on multiplicative measures
of a radial distance and requires to deal with strictly positive inputs and/
or outputs. This feature imposes strong constraints on the choice of the
inputs and outputs in the analysis (e.g., Gregoriou, Sedzro, and Zhu, 2005)
since data from funds regularly display zero or negative values. Moreover,
the DEA estimator is very sensitive to outliers, which might be frequent
in HF databases.
Recently, directional distance functions have been introduced
(Chambers, Chung, and Färe, 1998; Färe and Grosskopf, 2000) that
generalize both input and output distance functions and can be defined
with negative inputs or outputs. Simar and Vanhems (2010) propose
a simple method to compute these directional distances using Free
Disposal Hull (FDH) techniques and also introduce some robust versions
that are less sensitive to extreme observations.
90
Laurent Germain, Nicolas Nalpas, and Anne Vanhems 91
The motivation for introducing these nonparametric measures comes

from recent research in which the choice of a particular parametric
performance indicator has no significant influence on the ranking of
an HF (Eling and Schuhmacher, 2007). The aim of this chapter is then
to compare the ranking of HFs provided by classic performance measures
(Sharpe, Omega, and Calmar ratios) with those provided by robust
directional distance estimators introduced by Simar and Vanhems (2010).
In particular, we study the relative performance of HF indices which seek
to replicate the strategies used by HFs. The empirical analysis is conducted
using the TASS HF database and IQ replication indices. Because of HF
heterogeneity in terms of both investment strategies and the types of
market in which they operate, we separate our analysis by examining
eight HF classes: convertible arbitrage, emerging markets, equity market
neutral, event-driven, fixed-income arbitrage, global macro, long/short
equity, and funds of HF.
Contrary to the standard literature, our results suggest that when
different indicators of risk can be addressed simultaneously in a multi-
criteria analysis, which is made possible by the use of directional
distance measures, the rankings can differ greatly with respect to those
obtained with the Sharpe ratio. Moreover, although the introduction of
tradable indices that replicate HF strategies is suitable for an investor, it
seems possible to achieve better risk–return combinations by selecting
individual funds accurately.
The rest of the chapter is organized as follows. Section 7.2 presents the
main directional efficiency estimators studied in Simar and Vanhems
(2010). Section 7.3 describes the database and defines the perform-
ance indicators used for analyzing HF and replication indices rankings.
Section 7.4 presents the main empirical results. Section 7.5 concludes
the chapter.
7.2 Efficiency measures based on directional distances
7.2.1 Notation
In the classical setting of productivity analysis and technical efficiency
study, we consider a set of p inputs and q outputs used in the produc-
tion process. The production set is the set of technically feasible com-
binations of inputs and outputs and the efficient frontier is the upper
{ ( ) }
boundary of defined by ∂ ( x, y ) ∈ g1x, g y ∉ for all g > 1 . The
efficiency of a production unit at the level ( x, y ) is measured by its
distance to the efficient frontier. In the input-oriented case, the classical
92 A Robust Directional Application
{
Farrell–Debreu radial distance is defined by q( x, y ) inf q > 0 (q x, y ) ∈ . }
This indicates how much all the input quantities can be proportionately
reduced so that the output levels y can still be produced (a similar defi-
nition can be derived for the output-oriented case). The standard non-
parametric methods to estimate q( x, y ) are the FDH method introduced
by Deprins, Simar, and Tulkens (1984) and the DEA method when
is convex (Charnes, Cooper, and Rhodes, 1978; Banker, Charnes, and
Cooper, 1984).
7.2.2 Directional distance estimators

As recalled in the introduction, all the above methods rely on multipli-
cative measures of the distance and so require strictly positive inputs
and/or outputs. This can be critical when the data contain zero or nega-
tive values, as in financial databases. A natural idea suggested by several
authors is to translate the data to avoid negative values, but as pointed
out by Lovell and Pastor (1995), some specific DEA estimators can satisfy
translation invariance for inputs or outputs, but not for both. Such a
restriction may strongly constrain the choice of inputs and outputs.
Therefore we propose here to use a directional distance measure
introduced by Chambers, Chung, and Färe (1998) and studied more
extensively in Simar and Vanhems (2010). Consider a direction vector
g ( g x , g y ) ∈ Rpq that may depends on each value ( x, y ) and is arbitrar-
ily chosen by the researcher. The directional distance function projects
the input–output vector onto the production frontier in the direction
{ }
g and is defined by D( x, y ; g x , g y ) sup b > 0 ( x b g x , y b g y ) ∈ . As
recalled in Färe, Grosskopf, and Margaritis (2008), this definition covers
input (or output) radial distances as special cases: if g ( x, 0 ) or g ( 0, y ),
it is translation invariant and independent of unit of measurement.
By construction, it can be defined for any real values of inputs and
outputs.
From a computational point of view, Simar and Vanhems (2010)
show that the directional distance can be expressed as a standard radial
distance up to a simple transformation of the inputs/outputs space and
propose a simple method of estimation of D( x, y ; g x , g y ) using the FDH
technique (we refer to their paper for more technical details).
Simar and Vanhems (2010) also introduce robust directional distances
that are less sensitive to outliers and extreme observations. These robust
nonparametric estimators also benefit from nice properties ( n consist-
ency and asymptotic normal distribution) and are then very useful for
practitioners.1 The two classical robust distances introduced for radial
measures are the order-m efficiency distance (Cazals, Florens, and Simar,
2002) and the order-a quantile distance (e.g., Daouia and Simar, 2007).
Simar and Vanhems (2010) adapt both formulations to the directional
distance case. In particular, in the input-oriented setting and a ∈ ]0,1[ , the
order-a efficiency score Da gives an indication of how efficient a fund is,
compared to (1 a ) 100 percent of funds attaining at least a level y of
outputs, in the chosen direction g. Equivalently, for a fixed integer m ≥ 1,
the order-m efficiency score Dm gives an indication of how efficient a
fund is compared with m potential competing funds attaining at least y
in the chosen direction g. By construction, when m goes to infinity and
a goes to 1, then Dm and Da converge to D.
The practical choice of the parameters m and a impacts on the robust-
ness of the efficiency scores and some empirical methodology can be
applied to determine a value of m (and a ) depending on the number
of outliers. In our setting, we fix the levels of m and a so that around
5 percent of funds are considered as outliers. Moreover, the direction
chosen as the benchmark value is fixed to the average absolute value of
the inputs and the outputs.2
7.3 Data and performance indicators
7.3.1 Data
We use the June 2009 version of the Lipper TASS HF Database. The
“Live” and “Graveyard” files include 5,983 and 3,296 funds, respectively.
Combining the two types of funds helps us to mitigate survivorship bias
and allows us to get a maximum number of HF to analyze. We conduct
our analysis on eight investment styles: convertible arbitrage, emerging
markets, equity market neutral, event-driven, fixed-income arbitrage,
global macro, long/short equity, and funds of hedge funds, excluding the
latter category and convertible arbitrage when looking at the perform-
ance of IQ HF replication indices (for which they don’t exist). Funds
typically report their new data to vendors with a delay of a few months or,
if faster, only on a provisional basis. We therefore do not include the first
five months of 2009,3 and only employ the data on and before December
2008 (e.g., Getmansky, Lo, and Makarov, 2004). For each fund, we extract
the return net of fees and the net asset value at a monthly frequency. We
do not consider the few funds that either report performance gross rather
than net of fees, or quarterly instead of monthly. We exclude all funds
with missing data for any of the characteristics on the period from January
2004 to December 2008 (from April 2007 to March 2009 with respect to
HF replication indices performance analysis). We end up with a total of
1,995 HF (2146 for the HF replication indices analysis) in the sample.
Table 7.1 presents descriptive statistics of the funds in the two sample
periods. For each category of HF, we show the number of funds, as well
as the minimum, maximum, mean, standard deviation, skewness, excess
kurtosis, and annualized Sharpe ratio of the monthly return of an equally-
weighted average of the funds belonging in that category. The Jarque–Bera
statistic indicates the percentage of funds for which the normality hypoth-
esis is rejected at the conventional 5 percent level.
This shows that there is considerable heterogeneity in the historical
risk and return characteristics of the various categories of HF investment
styles. For example, over the period from January 2004 to December
2008, the monthly mean return ranges from –0.01 percent for convertible
arbitrage to 0.65 percent for emerging markets, and the monthly volatility
ranges from 2.09 percent for equity market neutral to 5.80 percent for
emerging markets.
Despite their heterogeneity, all categories do share a common char-
acteristic of strong tail risk (negative skewness) exposure. A more direct
measure of tail risk or “fat tails” is excess kurtosis—the normal distribu-
tion has an excess kurtosis of 0, so values greater than this represent
fatter tails than the normal. Not surprisingly, the two categories with the
most negative skewness—convertible arbitrage (–2.81) and fixed income
arbitrage (–1.46)—also have the largest excess kurtosis: 12.87 and 11.75,
respectively. This tail risk has been largely magnified by the effects of the
financial crisis. Whatever the strategy under consideration, the normality
hypothesis is rejected for most of the funds advocating the use of
performance measures that account for tail risk. Over the second sample,
this conclusion has to be mitigated since the normality assumption can
be accepted for more than half of the funds in several categories. Note
that the Sharpe ratios are negative for most strategies due to very negative
returns in the year 2008 highlighting the increasing correlation between
HF and regular asset classes during the financial turmoil.
Table 7.2 shows descriptive statistics of the IQ replication indices4
for each available category that matches the TASS database over the
sample from April 2007 to March 2009. Contrary to traditional HF
indices, such as the Dow Jones Credit Suisse HF indices, those of IQ are
tradable. IQ indices use HF replication strategies that seek to capture
the risk and return performance characteristics of major HF investment
styles by employing a proprietary, rule-based investment process that selects
components from a wide array of ETFs (commodities, currencies,
stocks, bonds, and real estate). Globally and surprisingly, the IQ HF
replication indices do not seem to represent adequately the universe
of funds of the TASS database: they seem to smooth the distribution of
Table 7.1 Monthly statistics
Categories No. of Min (%) Max (%) Mean (%) Std. dev. (%) Skewness Excess Sharpe Jarque–
Funds kurtosis ratio Bera (%)
Jan. 2004–Dec. 2008
Convertible arbitrage 31 –14.74 4.83 –0.01 3.01 –2.81 12.87 –0.36 90
Equity market neutral 95 –7.11 5.15 0.36 2.09 –0.97 6.46 0.17 68
Fixed-income arbitrage 55 –13.34 9.04 0.27 3.04 –1.46 11.75 0.17 93
Global macro 63 –10.45 9.98 0.64 3.68 –0.26 2.92 0.34 63
Long/short equity 654 –11.50 8.63 0.41 3.60 –0.80 3.32 0.11 69
Emerging markets 98 –20.89 11.56 0.65 5.80 –1.41 5.99 0.25 80
Event-driven 134 –10.91 6.57 0.39 2.95 –1.45 6.15 0.28 81
Funds of HF 865 –7.74 5.28 0.38 2.50 –1.28 3.63 –0.16 81
Apr. 2007–Mar. 2009
Equity market neutral 176 –7.01 5.59 0.03 2.88 –0.43 2.14 –0.25 40
Fixed-income arbitrage 113 –13.05 11.77 –0.16 4.97 –1.00 4.21 –0.23 69
Global macro 158 –9.48 10.24 0.56 4.46 –0.14 1.34 0.27 34
Long/short equity 1,258 –11.70 8.95 –0.35 4.76 –0.43 1.33 –0.44 29
Emerging markets 213 –18.63 11.78 –0.84 6.93 –0.78 2.54 –0.50 47
Event-driven 228 –10.97 6.76 –0.58 3.94 –0.91 3.05 –0.74 52
95
96
Table 7.2 Monthly statistics
IQ replication indices Min (%) Max (%) Mean (%) Std. dev. (%) Skewness Excess Sharpe ratio p-value of a
(Apr. 2007–Mar. 2009) kurtosis Jarque–Bera Test
Equity market neutral –2.10 2.78 0.10 1.09 0.23 0.30 –0.14 0.86
Fixed-income arbitrage –10.08 6.60 –0.25 3.15 –0.89 3.72 –0.16 0.00
Global macro –8.77 7.07 –0.40 3.16 –0.77 2.22 –0.21 0.02
Long/short equity –15.37 7.66 –1.15 4.62 –1.18 2.94 –0.31 0.00
Emerging markets –10.80 8.40 0.60 4.00 –0.72 1.66 0.09 0.08
Event-driven –10.53 5.47 –0.11 3.17 –1.28 4.02 –0.12 0.00
returns of an equally weigthed fund representing each strategy of the

TASS database.
Particular attention should be paid to the performance scores of the
equity market neutral IQ index, since most of its descriptive statistics
(lower standard deviation, positive (vs negative) skewness, and lower
excess kurtosis) favor a much better performance with respect to the indi-
vidual HF in the TASS Database. In fact, this index dominates 95 percent
of TASS’ funds with respect to the Sharpe ratio, together with the criteria
based on the skewness and the excess kurtosis coefficients.
7.3.2 Parametric performance indicators

We consider three parametric performance indicators: the Sharpe,
Calmar, and Omega ratios. The Sharpe ratio is frequently chosen as the
performance measure in the literature. Moreover several articles show
that HF ranking is not modified by introducing performance measures
that explicitly deal with tail risk (e.g., Eling and Schuhmacher, 2007).
Using historical monthly returns r1, … , rT for a particular fund or index,
the Sharpe ratio is calculated as follows:
∑
T
1
T
rt rf
t 1
Sharpe ratio
s
where s denotes the standard deviation of the monthly returns and

rf is the risk-free monthly interest rate. The latter is calculated as the
monthly geometric average of a rolling 1-month investment (monthly
compounding) in the US Libor 1-month over a year (0.298 percent over
the period January 2004 to December 2008 and 0.262 percent over the
period April 2007 to March 2009).
Drawdown-based measures (DD) are particularly popular in the HF
industry (Lhabitant, 2004). The maximum drawdown (denoted MD) of
a fund is the maximum loss incurred over the investment period. The
Calmar ratio allows returns to be determined on a downside risk-adjusted
basis by comparing the opportunity of gain to the potential maximum
loss. It is expressed as follows:
1/ T
⎛
∏ ⎞
T
⎜⎝ (1 rt ) 1⎟ rf
t 1 ⎠
Calmar ratio
MD
Lower partial moments (LPMs) measure risk by negative deviations

of the realized returns with respect to a minimal acceptable return t
(the risk-free rate in our analysis). The LPM of order n for a particular
fund or index is calculated as LPMn (1/T )Σ Tt 1(max( t rt , 0 ))n . Because
LPMs consider only negative deviations of returns from a minimal
acceptable return, they seem to be a more appropriate measure of risk
than the standard deviation. The Omega ratio uses a first-order LPM:
∑
T
1
T
rt rf
t 1
Omega ratio 1
LPM1
7.3.3 Input–output choice for the nonparametric methods

Regarding the choice of the input–output combination in the applica-
tion of nonparametric distance measures, outputs should be selected in
a way that represents what investors seek to maximize, and conversely
inputs should stand for what investors seek to minimize. The main
concern consists of choosing the numbers of both inputs and outputs
and to ensure that they accurately represent the preference of a typical
investor. Using too many inputs and outputs will be less helpful because
when the number of inputs and outputs increases, more HF or replication
indices tend to be efficient. We then restrict the analysis to a combina-
tion of three inputs and two outputs. From an economic point of view,
the input and output selection should be based on the investors’ utility
function. Since its typical representation exhibits preferences for odd
moments of the distribution of returns and reluctance for even moments
(Jurczenko, Maillet, and Merlin, 2006), we select the excess return and
the skewness as outputs and the excess kurtosis as an input. To account
for tail risk aversion and risk management issues that are fundamental
to every institutional investor, the absolute value of both the maximum
drawdown and the normal value-at-risk at 99 percent are the two other
inputs selected. The excess kurtosis is thus the only input for which we
can observe negative values. Since the output space includes negative
values (negative excess return and negative skewness for most funds
over the two sample periods), the DEA methods cannot be employed
with such a selection, since DEA requires the semi-positivity of either the
output or the input space.
Note that our focus is to compare HF and replication indices rankings
with traditional parametric performance indicators, which can take only
account of tail risk. Therefore, our input–output selection is strictly con-
strained to that. But as proved by Getmansky, Lo, and Makarov (2004),
the existence of a large serial correlation in HF returns, which basically
implies that the “true” risk of HF is underestimated (due to the holding of
illiquid assets and/or return smoothing practices), a multi-criteria analysis
of HF performance should account for serial correlation risk as well. The

introduction of several orders of Ljung–Box statistics would end up
with another set of negative inputs. Once again, contrary to a DEA
approach, the use of our directional measures will make possible to
cope with this concern.
7.4.1 The classical mean-variance framework

Since our selection of inputs and outputs contain negative values in both
domains, we cannot employ a DEA method (Eling, 2006; Gregoriou,
Sedzro, and Zhu, 2005) to rank HF and replication indices and we will
need to resort to FDH directional measures. To show the similarity
between the two methods, we choose in this section a framework that
makes this comparison fully accurate. To do so, we restrict the analysis
to one input (standard deviation of monthly returns) and one output
(monthly excess returns). This restriction to a one-dimensional space is
made necessary by the nature of the DEA optimization program needed
to calculate efficiency scores. Contrary to FDH methods, DEA methods
weight each input and output to obtain the best possible efficiency score.
In the case of multi-dimensional inputs/outputs, it can lead to a virtual
weight close to zero for a particular input/output if its value ranks very
unfavorably for the fund under consideration with respect to its peers.5
This issue is then completely circumvented in the single-input–output
case. Since the sole output space contains negative value, we choose
an input orientation and allow for variable returns to scale. The former
choice comes from the translation invariance property of DEA estimator
for outputs in that context, whereas the latter is related to the relation-
ship between HF size and performance (e.g., Eling, 2006).
Table 7.3 displays the Spearman rank correlations of HF efficiency scores
between DEA estimator and the (both classical and robust) directional
FDH ones described in Section 7.2. The results clearly show the great simi-
larity of these different methods for ranking HF since all correlations
are close to unity. We obtain similar findings for HF replication indices.
7.4.2 A multi-dimensional approach

We now turn to the multi-dimensional analysis that was fully described
in Section 7.3. To conserve space, we display results for only two HF
strategies on the sample from January 2004 to December 2008 as they
are very similar for each category: convertible arbitrage (Table 7.4) and
funds of funds (Table 7.5).
Table 7.3 Spearman rank correlations with DEA measures
Categories Dir. dist. Order-␣ dir. dist. Order-m dir. dist.

2004–08
Convertible arbitrage 0.971 0.965 0.985
Equity market neutral 0.952 0.893 0.947
Fixed-income arbitrage 0.995 0.992 0.995
Global macro 0.982 0.941 0.968
Long/short equity 0.977 0.919 0.921
Emerging markets 0.987 0.974 0.985
Event-driven 0.965 0.934 0.947
Funds of HF 0.942 0.915 0.946
Table 7.4 Spearman rank correlations
Convertible Sharpe Omega Calmar Dir. dist. Order-␣ Order-m

arbitrage dir. dist. dir. dist.
Sharpe 1
Omega 0.961*
Calmar 0.817* 0.879* 1
Dir. dist. 0.528* 0.630* 0.860* 1
Order-a dir. dist. 0.222 0.313 0.610* 0.875* 1
Order-m dir. dist. 0.481* 0.579* 0.833* 0.989* 0.898* 1
* Signiﬁcant at the 1 percent level.
Table 7.5 Spearman rank correlations
Funds of HF Sharpe Omega Calmar Dir. dist. Order-␣ Order-m

dir. dist. dir. dist.
Sharpe 1
Omega 0.993* 1
Calmar 0.890* 0.897* 1
Dir. dist. 0.649* 0.687* 0.786* 1
Order-a dir. dist. 0.483* 0.516* 0.737* 0.911* 1
Order-m dir. dist. 0.573* 0.609* 0.772* 0.979* 0.969* 1
* Signiﬁcant at the 1 percent level.
Tables 7.4 and 7.5 compare the rankings of two HF categories using the
three FDH directional models against the three parametric indicators
by means of the Spearman rank correlation. Our results are close to
those of Eling and Schuhmacher (2007) and Gregoriou, Sedzro, and Zhu
(2005): the correlations between parametric measures are very strong,
although slightly lower compared with Eling and Schuhmacher (2007),
whereas they are far from perfect when nonparametric directionals are
considered, ranging from 52.8 percent for convertible arbitrage to 64.9
percent for funds of funds regarding the correlation between the direc-
tional FDH model and the Sharpe ratio.
The correlations are always larger between directional FDH scores and
parametric ratios including a measure of tail risk (Calmar and Omega
ratios) than when only standard deviation (Sharpe ratio) is contem-
plated. This clearly indicates that tail risk constitutes an important factor
for ranking HF. In this regard, Sharpe ratios do not seem to be as appro-
priate, as advocated by Eling and Schuhmacher (2007). Another important
insight can be drawn from these results: unidimensional measures of
performance, as with all parametric ratios, do not seem to be able to take
fully into account the risk of investing in HF, which is protean by nature.
When considering the first sample (January 2004 to December 2008),
the DD risk seems to be a rather important factor as the correlations
between directional FDH scores and Calmar ratios are always larger than
those with Omega ratios (where the LPM of order one is considered
for measuring the tail risk), while this result does not systematically hold
when the second sample is examined (globally correlations fall by 20
percent in the second sample).
We shall recall that the FDH directional measures may also consider
the risk of serial correlation, which is not the case with traditional
parametric performance ratios (see Getmansky, Lo, and Makarov, 2004).
Finally, another key advantage of robust FDH directional distances
(order-a or order-m FDH estimators) is to allow investors to take account
of outliers in the data (see Section 7.2.2). Correlations with traditional
measures are then reduced by a factor ranging from 10 to 20 percent.
7.4.3 Performance of HF replication indices

In this final section, we analyze the relative performance of IQ HF
replication indices with respect to individual HF of the TASS database
(Table 7.6) for each category of HF. To facilitate the interpretation of
results, all funds and replication indices are grouped into deciles, the
top decile representing the 10 percent of best performing funds according
to the various indicators presented in Sections 7.2 and 7.3.
With the notable exception of the equity market neutral IQ index,
the replication indices are always less well ranked by nonparametric
measures relative to parametric indicators. This shows again that the
FDH directional measures capture risk factors that unidimensional para-
metric ratios are unable to apprehend. The rankings of various IQ repli-
cation indices are generally beyond the fifth decile. This illustrates that
Table 7.6 IQ replication indices ranks within the TASS HF database (in deciles)
Categories Sharpe Omega Calmar Dir. dist. Order-␣ Order-m

dir. dist. dir. dist.
Equity market neutral 7 6 5 1 1 1
Fixed-income arbitrage 5 5 5 8 8 8
Global macro 9 9 9 9 8 9
Long/short equity 8 9 9 9 9 9
Emerging markets 1 1 1 6 7 6
Event-driven 3 3 3 8 9 8
an investor can obtain better performance by appropriately selecting

individual HF rather than investing in IQ replication indices. However,
this conclusion should be qualified, as the IQ replication indices
represent fully liquid tradable assets while individual HF generally exhibit
lock-up periods.
The excellent relative performance of the equity market neutral IQ
index with respect to nonparametric measures is not that surprising
(see Table 7.2). Indeed, compared to individual HF this index exhibits
a positive skewness and a low excess kurtosis. Since these two statistics
belong, respectively, to our set of outputs and inputs, the nonparametric
ranking is extremely positively impacted. Conversely, its excess mean
return is in the average of individual HF. This explains the average para-
metric rankings. In particular, because this index is better ranked with
the MD statistics than with LPM1, we observe as expected a superior
ranking with the Calmar ratio compared to Omega and to Sharpe.
Symmetrically, the IQ index for emerging markets has a relatively
high excess mean return (18th over 214 funds) but displays an excess
kurtosis and a negative skewness in the average of individual HF. Given
that parametric measures are by nature unidimensional, they put more
weight on the excess mean return than FDH distance estimators for
which a large (positive) skewness is regarded also as a nice feature. If
the investor is primarily interested in return, the emerging market class
IQ replication index could be considered as a rewarding investment
strategy.
In fact, unlike the traditional parametric performance measures, the
FDH distance estimator is very flexible to suit any form of investors’
preference. The choice of the combination of inputs and outputs is then
critical. By adjusting the direction to which the distance to the efficient
frontier (the performance estimator) is measured, the decision maker is
also able to control for the relative importance of each input/output.
Therefore, using this method allows one to build a tailor-made tool to

rank HF and replication indices.
As the focus of this section is to highlight the discrepancy in ranking
HF replication indices through parametric and nonparametric measures,
we have restricted our choice of inputs/outputs to risk–return indicators
contained in classical performance ratios. This excludes taking account
of liquidity and correlation risks, which have been discussed as the major
source of risk when investing in HF. Again, it will necessarily make use of
a multi-dimensional approach that is fully permitted by directional FDH
models. In fact, including such risk factors would give a clear advantage
to replication indices over those of IQ, since they are fully liquid and less
autocorrelated.
7.5 Conclusion
Although recent parametric performance measures make it possible to

deal with higher moments of the probability distribution of HF returns,
they do not allow consideration of the very protean nature of risks in
investing in HF, such as the autocorrelation of returns, their correlation
with traditional asset classes (diversification in a portfolio perspective),
and constraints such as lock-up periods or transaction costs. Moreover,
due to their undimensional nature, they all concentrate on different
tail risk measures in isolation (DD, LPM, and conditional value at risk).
Recently, nonparametric models (DEA) have been applied to account
fully for the multiple risk–return attributes characterizing HF invest-
ment in a unique performance score. However, such models suffer from
several pitfalls, such as sensitivity to noise (outliers in the data) and the
necessary semi-positivity of the input and output spaces. We show that
to be fully comparable with classic HF performance indicators in a
traditional risk–return context these two issues can be tackled by using
a robust directional FDH approach. These models allow fund selectors
to construct a personalized tool incorporating their own criteria. This
is done through the choice of the various inputs and outputs and the
direction in which efficiency scores are calculated. This chapter shows
that IQ replication indices are globally less well ranked by nonpara-
metric measures with respect to parametric indicators. In particular,
they belong to the second half of the sample when they are mixed with
individual HF of the TASS database when traditional risk measures are
considered. This indicates that if investors are primarily interested in
the first four moments of the returns distribution they can achieve
better performance by accurately selecting individual funds.
Notes
1. See Daraio and Simar (2006) for an illustration of a mutual funds database
with radial distances.
2. Other directions could have been suggested; for standard possibilities see Färe,
Grosskopf, and Margaritis (2008). However, comparing the efficiency of each
fund to an average performance seems quite reasonable in this context.
3. Since IQ indices start reporting data in April 2007, we include the three first
months of 2009 to get a two-year interval in the performance analysis of HF
indices.
4. IQ indices data are taken from Bloomberg in January 3rd 2011. A description
of the different indices is available at http://www.indexiq.com/indexes/iniqh.
html.
5. Conversely, rather (artificial) large weights can be attributed to inputs/outputs
of a particular fund that compare favorably to its peers. Several methods exist to
mitigate this issue by either adding some additional constraints to the optimiza-
tion program or considering a cross-efficiency DEA model (Eling, 2006).
References
Banker, R. D., Charnes, A., and Cooper, W. W. (1984). “Some Models for
Estimating Technical and Scale Inefficiencies in Data Envelopment Analysis.”
Management Science, 30(9): 1078–92.
Cazals, C., Florens, J. P., and Simar, L. (2002). “Nonparametric Frontier Estimation:
a Robust Approach.” Journal of Econometrics, 106(1): 1–25.
Chambers, R. G., Chung, Y. H., and Färe, R. (1998). “Profit Directional Distance
Functions and Nerlovian Efficiency.” Journal of Optimization Theory and
Applications, 98(2): 351–64.
Charnes, A., Cooper, W. W., and Rhodes, E. (1981). “Evaluating Program and
Managerial Efficiency: an Application of Data Envelopment Analysis to
Program Follow Through.” Management Science, 27(6): 668–97.
Daouia, A. and Simar, L. (2007). “Nonparametric Efficiency Analysis: a Multivariate
Conditional Quantile Approach.” Journal of Econometrics, 140(2): 375–400.
Daraio, C. and Simar, L. (2006). “A Robust Nonparametric Approach to Evaluate
and Explain the Performance of Mutual Funds.” European Journal of Operational
Research, 175(1): 516–42.
Deprins, D., Simar, L., and Tulkens, H. (1984). “Measuring Labor Inefficiency in
Post Offices.” In: M. Marchand, P. Pestieau, and H. Tulkens (eds.), The Performance
of Public Enterprises: Concepts and Measurements. North-Holland, Amsterdam.
Eling, M. (2006). “Performance Measurement of Hedge Funds Using Data
Envelopment Analysis.” Financial Markets and Portfolio Management, 20(4):
442–71.
Eling, M., and Schuhmacher, F. (2007). “Does the Choice of Performance Measure
Influence the Evaluation of Hedge Funds?” Journal of Banking & Finance, 31(9):
2632–47.
Färe, R., and Grosskopf, S. (2000). “Theory and Application of Directional Distance
Functions.” Journal of Productivity Analysis, 13(2): 93–103.
Färe, R. S., Grosskopf, D., and Margaritis, D. (2008). “Efficiency and Productivity:
Malmquist and More.” In: H. Fried, C. A. Knox Lovell, and S. Schmidt (eds.),
The Measurement of Productive Efficiency, 2nd edn. Oxford University Press,

Oxford.
Getmansky, M., Lo, A., and Makarov, I. (2004). “An Econometric Analysis of
Serial Correlation and Illiquidity in Hedge-Fund Returns.” Journal of Financial
Economics, 74(3): 529–609.
Gregoriou, G. N., Sedzro, K., and Zhu, J. (2005). “Hedge Fund Performance
Appraisal Using Data Envelopment Analysis.” European Journal of Operational
Research, 164(2): 555–71.
Jurczenko, E., Maillet, B., and Merlin, P. (2006). “Hedge Funds portfolio Selection
with Higher-order Moments: A Non-parametric Mean-Variance-Skewness-
Kurtosis Efficient Frontier.” In: E. Jurczenko and B. Maillet (eds.), Multi-Moment
Asset Allocation and Pricing Models. John Wiley & Sons, Chichester.
Lhabitant, F. S. (2004). Hedge Funds: Quantitative Insights. John Wiley & Sons,
London.
Lovell, K. C. A., and Pastor, J. (1995). “Units Invariant and Translation Invariant
DEA Models.” Operations Research Letters, 18(3): 147–51.
Simar, L., and Vanhems, A. (2010). “Probabilistic Characterization of Directional
Distances and their Robust Versions.” Working Paper 1040. Institut de
Statistique, UCL.
8
Hedge Fund Cloning through State
Space Models
Roberto Savona
8.1 Introduction
Replicating hedge fund returns requires a clear understanding of com-

plex strategies implemented by fund managers. One way to do this is
through regression analysis, by which an investment fund’s return can
be expressed as a function of (Fung and Hsieh, 1997): (1) where it trades
(asset class-driven returns), (2) how it trades (strategy-driven returns),
and (3) how much it trades (leverage-driven returns).
Using regression analysis, return sources and corresponding risk expo-
sures of hedge funds are cloned by common financial instruments which
act as “complex reducers” while maintaining the same risk–return profile
of hedge funds. Because of their adaptive nature to financial dynamics,
hedge funds are notoriously difficult to replicate due to the significant
and often sudden shifts in investment strategies they exhibit over time.
Modeling time-varying betas is then essential in constructing effective
hedge fund “clones.”
With the objective of handling this problem, in this chapter we intro-
duce an asset pricing model within a Bayesian framework in which asset
returns are modeled by imposing a pseudo-stochastic process on the path
of risk loading. Using a three-equation system we model the hedge fund
return process, which is assumed to be latently correlated with a fund-
specific benchmark and where the sensitivity between returns and bench-
mark is time-varying and dependent upon some imperfect predictors. The
novelty of our model is to impose a structure on the system innovations
and then to try to explain how unexplained asset returns co-move with
unexplained benchmark returns through a pseudo-stochastic beta.
Our time-varying beta modeling represents a possible remedy for the
problem of missing time-varying risk exposure presented in Fung and
106
Roberto Savona 107
Hsieh (2007b), and offers the real possibility of hedge fund cloning by
means of “passive replication strategies” based on liquid underlying
assets aiming to replicate hedge fund returns and their systematic risk.
The remainder of the chapter is as follows. Section 8.2 introduces
the model. Section 8.3 presents the estimation procedure. Section 8.4
presents the empirical analysis and Section 8.5 concludes.
8.2 The model
The econometric representation of the model is as follows (Amisano

and Savona, 2008; Savona, 2009):
rp ,t α p β p ,t rm ,t ε p ,t (8.1)
(1 ϕL )( β p ,t μ) Γ′zt η p ,t (8.2)
rm ,t Λ′ zt um ,t . (8.3)
Equation (8.1) is the hedge fund excess return over the risk-free rate at
time t, bp,t the systematic risk exposure assumed to be time varying, rm,t
the fund specific benchmark return and ep,t the unexpected fund return.
Equation (8.2) is the time-varying beta, where L denotes the lag opera-
tor, f the persistence beta parameter, m the unconditional mean revert-
ing beta term, Γ the transposed vector of sensitivities, zt the vector of
instruments at time t, and hp,t the beta stochastic component to accom-
modate imperfect predictors in beta evolution. Equation (8.3) is the
hedge fund benchmark excess return over the risk-free rate obtained as
the expectation of the 71 Fung-Hsieh (FH) risk factor model (Fung and
Hsieh, 2004; 2007a,b). Λzt denotes the expectation at time t modeled as
a linear function of the same instruments in (2), with Λ representing the
transposed vector of sensitivities, and um,t is the unexpected benchmark
return at time t, then accommodating imperfect predictors. Analytically,
the 71 FH risk factors are:
1. Bond Trend-Following Factor,

2. Currency Trend-Following Factor;
3. Commodity Trend-Following Factor;1
4. Standard & Poor’s 500 index monthly total return;
5. Size Spread Factor (Wilshire Small Cap 1750 minus Wilshire Large
Cap 750 monthly returns);
6. Month-end to month-end change in the 10-year Treasury constant
maturity yield;
108 Hedge Fund Cloning through State Space Models
7. Credit Spread Factor (the month-end to month-end change in

the Moody’s BAA yield less the 10-year Treasury constant maturity
yield);
8. Emerging Market Risk Factor (the IFC Emerging Market Index).
As discussed in Savona (2009), zt are observable instruments used to

describe fully the dynamics of hedge fund returns. In other terms, these
variables play the role of Primitive Risk Signals (PRS), to be considered
as those signals that hedge fund managers use in changing their primi-
tive trading strategies. PRSs are also the financial instruments through
which we replicate hedge fund returns, and as such they are assumed
to be those latent factors that affect the hedge fund dynamics, but for
which the inner mechanism of such a relationship is partly obscured
by the complex nature of the trading rules followed by the managers.
Analytically, these instruments are:
1. CBOE Volatility Index (VIX);

2. Month-end to month-end change in the 3-month T-bill (TBILL);
3. Term spread (TERM), computed as the monthly difference between
the yield on 10-year Treasuries and 3-month Treasuries;
4. Innovations in the S&P 500 monthly standard deviation (INN) to
proxy the liquidity shocks in the US equity market.2
In order to derive scale-independent coefficient estimates for PRS

sensitivities, we standardized each of the four instruments.
To combine Equations (8.1), (8.2), and (8.3), we also impose a structure
on the system innovations that we assume follows the distribution:
⎡ ε p ,t ⎤ ⎛⎡ 0 ⎤ ⎡ σ 2 σ εη σ εu ⎤⎞
⎢ ⎥ ⎜⎢ ⎥ ⎢ ε ⎥⎟
⎢ η p ,t ⎥∼ N⎜⎢ 0 ⎥,⎢ σ ηε σ 2η σ ηu ⎥⎟. (8.4)
⎢ ⎥ ⎜ ⎥⎢ ⎟
⎣ um ,t ⎦
⎜⎢
⎝⎣ 0 ⎦ ⎣ σ uε σ uη σ 2u ⎥
⎦⎟
⎠
Such a structure is used because a non-negative covariance matrix helps

control for unobservable factors in describing the dynamics of hedge
fund returns through the beta process. In a sense, where the predictors
fail to explain the beta dynamics, the innovations try to measure what
is generally unobservable, namely the measurement error of observ-
able predictors. Instead of searching for other predictors that may
increase the explanatory power of beta variations, we bypass the prob-
lem by searching for how idiosyncratic asset returns, the inaccessible
Roberto Savona 109
beta variation component, and benchmark residuals co-move together,

as implied by unobservable/omitted/imperfect predictors.
8.3 Estimation approach
To estimate our three-equation system we developed a Bayesian approach

within a state space technology, hence treating the parameters of the
model as random variables. Let us start by denoting with ␪ [Σ,ap,f,m,Γ,Λ]
the parameters of the system where Σ is the covariance matrix of the
system innovations (Equation 8.4). Let us further denote by p(␤p,T ,␪) the
joint prior distribution, in which the values for bp,t are modeled as in
Equation (8.4), and by L the likelihood function expressed as
L p( rpT , rmT ␪) ≡ ∫ p( rpT , rmT ␪, ␤ p ,T )p(␤ p ,T ␪)d␤ p ,T
␤ p ,T [ β p ,0 , β p ,1 ,..., β p ,T ]′ ,
(8.5)
rpT [ rp ,1 , rp ,2 ,..., rp ,T ]′ ,
rmT [ rm ,1 , rm ,2 ,..., rm ,T ]′ ,
where bp,t are obtained through a simulation procedure that uses the
Kalman filter. Theoretically, the joint posterior distribution of param-
eters and latent variables (the betas) is
p(␤ p ,T , ␪ rpT , rmT ) ∝ p(␤ p ,T , ␪) p( rpT , rmT ␪, β p ,t ) . (8.6)
To simulate this posterior distribution we used Gibbs sampling-data

augmentation procedure, a Markov Chain Monte Carlo (see Johannes
and Polson, 2009 for financial applications of MCMC methods) tech-
nique that generates random samples from a given target distribution,
namely the joint posterior distribution of the parameters ␪ and the state
variables given the observed returns p(␤p,T,␪|rpT,rmT).3
8.3.1 Priors and posteriors

Prior distributions for the parameters of the model are constructed as
follows:
1. Given the initial sample t 1,2,…,T0 we estimate equation the rp,t

ap bp(Zt )rm,t ep,t by OLS where bp(Zt ) b0p Bpzt, then obtaining esti-
mates to be plugged into bˆp,t(Zt ) k0p jp(bˆp,t1(Zt ) k0p) Kpzt xp,t
which, in turn, is estimated by OLS. The resulting OLS estimates and
the estimated covariance matrix are then used as first and second
moments for a multivariate Gaussian distribution for the parameters of

Equation (8.2).
2. Pre-sample estimates of the betas are also used to estimate the inter-
cept in Equation (8.1). Its point estimate and its variance are used as
moments for a univariate Gaussian prior for ap.
3. OLS on pre-sample period is also used to estimate Equation (8.3). An
OLS point estimate of Λ and its OLS covariance matrix are used as
moments for a multivariate Gaussian prior.
4. Pre-sample OLS estimates of Equations (8.1), (8.2), and (8.3) gener-
ate residuals ␩ˆ t,t 1,2,…,T0. The sample covariance matrix of these
residuals is then used to calibrate a Wishart prior on the inverse of the
covariance matrix in Equation (8.4), that is, the covariance matrix of
the errors in the system, as follows
T0 ⎛4 T0 ⎞−1
Σ−1 ~ Wishart( ␯, S ), ␯
4
, S ⎜
⎜T ∑ t t⎟ .
␩ ^ ⎟
^ ␩ (8.7)
⎝ 0 t1 ⎠
In this way we account for the likely scale of those errors without
imposing too tight a prior.
Posteriors of parameters and latent variables (beta) of our three-

equation system are obtained by means of a Gibbs sampling-data augmen-
tation procedure, distinguishing five blocks: (1) the latent variables bt; (2)
the parameter in equation ap; (3) the parameters in equation c, f, g; (4) the
parameters in equation Λ; and (5) the parameters in Σ following Carter
and Kohn (1994). Given the priors being implemented, each of these
blocks is simulated from its conditional posterior distribution. Details on
these conditional distributions and on how to simulate from them are
given in Amisano and Savona (2008).
8.4 Empirical view of the cloning model
To test our model in cloning hedge fund returns we used data from
CSFB/Tremont monthly return indices over the period January 1994
to September 2008, splitting the time period into three intervals, the
first from January 1994 to December 1997 (the “pre-sample” for priors’
estimation), the second from January 1998 to December 2006 (“estima-
tion sample”) and the third from January 2007 to September 2008 (the
“validation sample” for out-of-sample analysis).
Hedge fund clones were computed using the expected beta times
the style benchmark, which in turn is the 71 FH risk factor model
Roberto Savona 111
projection. This allows us to control for time variation in systematic risk

exposure, using the PRSs as the rule-based dynamic risk loadings to be
used in modulating the 71 FH risk factor exposure. Mathematically,
rtclone E( β p ,t ) ⋅ rmt ≡ [ μ ϕ ( β p ,t−1 − μ) Γ′zt ]⋅ rmt (8.8)
Computationally, we first estimated the 71 FH risk factor model for

each CSFB/Tremont index; then we used the expectations as proxies
for the long-term style benchmarks. With these proxies, we next ran
the three-equation system using the 4 PRSs as instruments for beta and
benchmark processes, and the 30 days’ Treasury Bill rate was used as
the risk-free rate. The results for the posteriors of beta hyperparameters
are in Table 8.1.
8.4.1 In-sample analysis

In Table 8.2 we report the performance summary of the hedge fund
indices and corresponding clones over the estimation sample January
1998 to December 2006.
Firstly, note that the correlation between the hedge fund indices and
corresponding clones is on average 0.6139, indicating a relatively good
job in cloning the risk–return characteristics of hedge funds. For certain
categories, the correlation is not particularly high, such as for Multi-
Strategy (0.3323), Global Macro (0.3605) and Convertible Arbitrage
(0.3864), while for others the value is well above the 0.5.
On average, our Bayesian time-varying hedge fund clones significantly
outperformed the hedge funds. The annualized overall mean return of
the clones is greater than that of the hedge funds, while the annual-
ized standard deviation is significantly less, and equal to 6.34 percent
versus 7.8 percent of hedge funds. Based on the Sharpe ratio results, the
Multi-Strategy clone appeared as the best performer with 1.0126, while
Dedicated Short Bias was the worst performer with –0.2761.
In Table 8.2 we also report the probability of obtaining positive
returns (the Prob 0 in the table), computed as the number of posi-
tive returns over the total monthly observations in the period January
1998 to December 2006. Except for Dedicated Short Bias, our clones
obtain higher probability in that on average they significantly out-
perform hedge fund indices by 6.02 percent. Except for Dedicated Short
Biases and Event-Driven Multi-Strategy together with Event-Driven Risk
Arbitrage, the clones are all successful, with performances that signifi-
cantly exceed their hedge fund counterparts on a risk-adjusted basis. The
Min–Max comparison indicates that our clones are less prone to extreme
112
Table 8.1 Time-varying beta parameter estimates from January 1998 to

December 2006
Panel A: Instrument sensitivities

VIX TBILL TERM INN
Hedge Fund Index 0.0413 0.0701 –0.2429 –0.2185***
(HF Index)
Convertible Arbitrage (CA) 0.041 –0.2378** –0.8161*** –0.0588
Dedicated Short Bias (DSA) –0.0985** –0.1036** 0.1982*** –0.0255
Emerging Markets (EmM) –0.0496 0.1009*** –0.0105 0.0687
Equity Market Neutral 0.1191*** –0.0987** –0.3736*** –0.1868***
(EqMN)
Event-Driven (ED) –0.0399 –0.0569*** –0.2323*** –0.079**
Distressed (D) 0.0402 –0.0429 –0.028 –0.2531***
Multi-Strategy (M-S) 0.1198*** –0.0433 –0.4188*** –0.0359
Risk Arbitrage (RA) 0.299*** –0.0103 –0.2697*** –0.0143
Fixed Income Arb (FIA) 0.0737 0.0325 –0.0595 –0.1468***
Global Macro (GM) 0.0478 –0.1548 –0.8424*** –0.1801
Long/Short Equity (LSE) 0.0774*** 0.0052 –0.0382 –0.2365***
Managed Futures (MF) 0.5064*** –0.0139 0.2803*** –0.0029
Multi-Strategy (M-S) 0.2223*** 0.017 –0.4426*** 0.0305
Panel B: Long-run beta and persistence beta parameter
Long-run beta ( ␮) Persistence beta
parameter (␾)
Hedge Fund Index (HF Index) 1.033 0.4021
Convertible Arbitrage (CA) 1.7696*** –0.0687
Dedicated Short Bias (DSA) 0.987*** –0.2444
Emerging Markets (EmM) 1.1321*** 0.1904
Equity Market Neutral 1.2896*** 0.0928
(EqMN)
Event-Driven (ED) 0.7415*** 0.1344
Distressed (D) 0.3383*** 0.6649***
Multi-Strategy (M-S) 0.7652*** –0.0461
Risk Arbitrage (RA) 0.5702*** –0.2719
Fixed Income Arb (FIA) 0.2822*** 0.8007***
Global Macro (GM) 2.4736*** –0.0712
Long/Short Equity (LSE) 0.3029 0.6983
Managed Futures (MF) 1.3121*** –0.4601
Multi-Strategy (M-S) 1.5918*** –0.0677
Note: Panel A and B report estimates of Equation (8.2) relative to the instruments parameter,
and long-run beta together with the persistence parameter, respectively. ***, **, * denote
significance at the 0.01, 0.05, and 0.1 levels, respectively.
Table 8.2 Clones vs. hedge fund indices from January 1998 to December 2006
Correlation Meanyr StdDevyr SR Prob 0

HF Clone C minus HF Clone C minus HF Clone C minus HF Clone C minus
HF HF HF HF
HF Index 0.7143 0.0727 0.1361 0.0634 0.0685 0.0713 0.0028 0.1636 0.4138 0.2503 0.6759 0.7500 0.0741
CA 0.3864 0.0705 0.1230 0.0526 0.0491 0.0366 –0.0124 0.2155 0.7027 0.4872 0.7685 0.8981 0.1296
DSB 0.6818 –0.0354 –0.0838 –0.0484 0.1802 0.1230 –0.0572 –0.1109 –0.2761 –0.1652 0.4444 0.4352 –0.0093
EmM 0.8537 0.0727 0.1341 0.0615 0.1478 0.1623 0.0144 0.0758 0.1784 0.1026 0.6389 0.6667 0.0278
EqMN 0.4311 0.0816 0.1076 0.0259 0.0224 0.0217 –0.0007 0.6151 0.9801 0.3650 0.8796 0.9907 0.1111
ED 0.7068 0.0851 0.0829 –0.0022 0.0600 0.0360 –0.0240 0.2463 0.3932 0.1469 0.7870 0.7870 0.0000
D 0.6378 0.0991 0.1023 0.0032 0.0645 0.0350 –0.0295 0.2918 0.5641 0.2723 0.7963 0.7963 0.0000
M-S 0.7923 0.0779 0.0584 –0.0195 0.0647 0.0463 –0.0184 0.1966 0.1532 –0.0433 0.7500 0.7685 0.0185
RA 0.7182 0.0510 0.0264 –0.0246 0.0451 0.0259 –0.0192 0.1098 –0.0830 –0.1927 0.7315 0.8056 0.0741
FIA 0.6343 0.0364 0.0509 0.0145 0.0390 0.0268 –0.0122 0.0189 0.1835 0.1646 0.7500 0.8148 0.0648
GM 0.3605 0.0889 0.2902 0.2013 0.0874 0.1228 0.0353 0.1819 0.6028 0.4209 0.7315 0.8241 0.0926
LSE 0.7576 0.1031 0.0837 –0.0194 0.1070 0.0579 –0.0491 0.1868 0.2483 0.0615 0.6481 0.7222 0.0741
MF 0.5877 0.0628 0.1156 0.0528 0.1199 0.0984 –0.0215 0.0697 0.2400 0.1703 0.5278 0.5833 0.0556
M-S 0.3323 0.0754 0.1150 0.0396 0.0366 0.0231 –0.0135 0.3278 1.0126 0.6848 0.8056 0.9352 0.1296
Mean 0.6139 0.0673 0.0959 0.0286 0.0780 0.0634 –0.0147 0.1849 0.3795 0.1947 0.7097 0.7698 0.0602
StdDev 0.1697 0.0611 0.0236 0.2452 0.0470
t-value 1.7516 –2.3241 2.9708 4.7874
p-value 0.1034 0.0370 0.0108 0.0004
(continued)
113
Table 8.2 Continued
114
Min Max Cumulative Returns

HF Clone C minus HF Clone C minus HF Clone C minus
HF HF HF
HF Index –0.0771 –0.0611 0.0161 0.0836 0.0606 –0.0231 0.6543 1.2251 0.5708
CA –0.0483 –0.0296 0.0187 0.0340 0.0413 0.0073 0.6343 1.1074 0.4730
DSB –0.0884 –0.0792 0.0092 0.2255 0.0962 –0.1293 –0.3183 –0.7539 –0.4356
EmM –0.2319 –0.2070 0.0249 0.1517 0.1084 –0.0434 0.6540 1.2071 0.5531
EqMN –0.0101 –0.0005 0.0096 0.0233 0.0386 0.0153 0.7348 0.9680 0.2332
ED –0.1193 –0.0348 0.0845 0.0317 0.0342 0.0026 0.7656 0.7461 –0.0196
D –0.1261 –0.0266 0.0995 0.0384 0.0380 –0.0004 0.8918 0.9204 0.0286
M-S –0.1168 –0.0781 0.0387 0.0449 0.0375 –0.0074 0.7014 0.5258 –0.1756
RA –0.0631 –0.0564 0.0067 0.0364 0.0176 –0.0189 0.4590 0.2379 –0.2211
FIA –0.0711 –0.0325 0.0385 0.0189 0.0317 0.0128 0.3277 0.4581 0.1304
GM –0.1170 –0.0959 0.0211 0.1000 0.1419 0.0419 0.8005 2.6118 1.8113
LSE –0.1159 –0.0308 0.0852 0.1284 0.0908 –0.0377 0.9282 0.7532 –0.1750
MF –0.0871 –0.0580 0.0291 0.0979 0.1312 0.0333 0.5653 1.0408 0.4755
M-S –0.0491 –0.0159 0.0332 0.0261 0.0281 0.0020 0.6790 1.0352 0.3562
Mean –0.0944 –0.0576 0.0368 0.0744 0.0640 –0.0104 0.6056 0.8631 0.2575
StdDev 0.0306 0.0419 0.5501
t-value 4.4973 –0.9256 1.7516
p-value 0.0006 0.3715 0.1034
Note: The table presents performance summary of hedge fund indexes compared with corresponding hedge fund clones computed through the beta
replication of the system (8.1)–(8.4) in-sample. “Correlation” is the correlation between hedge funds and corresponding clones. “Meanyr” and “StdDevyr”are
the annualized mean return and standard deviation. “SR” and “Prob 0” are the Sharpe ratio and the probability to get positive returns computed as the
number of positive returns over the total monthly observations. “Min,” “Max,” and “Cumulative Returns” are the minimum, the maximum and the sum
of monthly returns over the period January 1998 to December 2006. ***, **, * denote significance at the 0.01, 0.05, and 0.1 levels, respectively.
Table 8.3 Clones vs. hedge fund indices from January 2007 to September 2008
Correlation Meanyr StdDevyr SR Prob 0

HF Clone C minus HF Clone C minus HF Clone C minus HF Clone C minus
HF HF HF HF
HF 0.8703 0.0112 0.0791 0.0679 0.0776 0.0121 –0.0655 –0.0784 1.1136 1.1919 0.5714 0.9524 0.3810
Index
CA 0.3094 –0.0883 0.1488 0.2371 0.1104 0.0139 –0.0965 –0.3152 2.4166 2.7317 0.4762 1.0000 0.5238
DSB –0.0509 0.0656 –0.0653 –0.1308 0.1658 0.0232 –0.1425 0.0581 –1.2122 –1.2702 0.5238 0.2381 –0.2857
EmM 0.3432 –0.0021 0.0537 0.0558 0.1146 0.0165 –0.0981 –0.0865 0.3755 0.4620 0.6190 0.8571 0.2381
EqMN 0.6218 0.0606 0.1124 0.0519 0.0248 0.0087 –0.0161 0.3294 2.6510 2.3216 0.8095 1.0000 0.1905
ED 0.5221 0.0177 0.0809 0.0632 0.0745 0.0080 –0.0665 –0.0565 1.7576 1.8140 0.6190 1.0000 0.3810
D 0.8696 –0.0060 0.0360 0.0420 0.0609 0.0133 –0.0476 –0.1812 0.0815 0.2627 0.5714 0.9524 0.3810
M-S 0.1685 0.0344 0.0820 0.0476 0.0864 0.0085 –0.0779 0.0072 1.6925 1.6853 0.6667 1.0000 0.3333
RA –0.2572 0.0391 0.0368 –0.0022 0.0494 0.0028 –0.0466 0.0397 0.4691 0.4293 0.5238 1.0000 0.4762
FIA 0.7428 –0.0456 0.0095 0.0551 0.0792 0.0048 –0.0744 –0.2839 –1.3622 –1.0783 0.5714 0.8571 0.2857
GM 0.5644 0.0836 0.2772 0.1936 0.0885 0.0195 –0.0690 0.1673 3.6207 3.4534 0.6667 1.0000 0.3333
LSE 0.8106 –0.0034 0.0136 0.0170 0.1001 0.0152 –0.0848 –0.1028 –0.3522 –0.2494 0.5714 0.8571 0.2857
MF –0.0160 0.0785 0.0709 –0.0075 0.1287 0.0128 –0.1160 0.1036 0.8757 0.7721 0.5238 1.0000 0.4762
M-S 0.6691 –0.0192 0.1226 0.1417 0.0770 0.0048 –0.0722 –0.1928 5.3971 5.5899 0.5238 1.0000 0.4762
Mean 0.4406 0.0161 0.0756 0.0594 0.0884 0.0117 –0.0767 –0.0423 1.2517 1.2940 0.5884 0.9082 0.3197
StdDev 0.3660 0.0891 0.0311 1.8405 0.1998
t-value 2.4957 –9.2401 2.6307 5.9876
p-value 0.0268 0.0000 0.0208 0.0000
(continued)
115
Table 8.3 Continued 116
Min Max Cumulative Returns

HF Clone C minus HF Clone C minus HF Clone C minus
HF HF HF
HF –0.0655 –0.0014 0.0641 0.0316 0.0126 –0.0190 0.0196 0.1384 0.1189
Index
CA –0.1226 0.0050 0.1276 0.0218 0.0171 –0.0046 –0.1545 0.2605 0.4149
DSB –0.0730 –0.0214 0.0516 0.1031 0.0048 –0.0982 0.1148 –0.1142 –0.2290
EmM –0.0893 –0.0076 0.0818 0.0548 0.0104 –0.0444 –0.0037 0.0939 0.0976
EqMN –0.0141 0.0020 0.0160 0.0165 0.0123 –0.0043 0.1060 0.1967 0.0907
ED –0.0575 0.0014 0.0589 0.0324 0.0096 –0.0228 0.0309 0.1415 0.1106
D –0.0518 –0.0119 0.0399 0.0208 0.0064 –0.0144 –0.0105 0.0630 0.0735
M-S –0.0617 0.0024 0.0642 0.0431 0.0105 –0.0326 0.0602 0.1434 0.0832
RA –0.0349 0.0015 0.0364 0.0322 0.0045 –0.0276 0.0683 0.0644 –0.0039
FIA –0.0680 –0.0037 0.0643 0.0207 0.0028 –0.0179 –0.0799 0.0166 0.0964
GM –0.0663 0.0076 0.0739 0.0444 0.0318 –0.0126 0.1462 0.4851 0.3388
LSE –0.0781 –0.0163 0.0619 0.0373 0.0052 –0.0321 –0.0059 0.0239 0.0298
MF –0.0479 0.0007 0.0486 0.0661 0.0130 –0.0531 0.1373 0.1241 –0.0132
M-S –0.0735 0.0065 0.0800 0.0302 0.0120 –0.0182 –0.0336 0.2145 0.2481
Mean –0.0646 –0.0025 0.0621 0.0396 0.0109 –0.0287 0.0282 0.1323 0.1040
StdDev 0.0259 0.0243 0.1560
t-value 8.9662 –4.4127 2.4957
p-value 0.0000 0.0007 0.0268
Note: The table presents performance summary of hedge fund indexes compared with corresponding hedge fund clones computed through the beta
replication of the system (8.1)–(8.4) out-of-sample. “Correlation” is the correlation between hedge funds and corresponding clones. “Meanyr” and
“StdDevyr”are the annualized mean return and standard deviation. “SR” and “Prob 0” are the Sharpe ratio and the probability to get positive returns
computed as the number of positive returns over the total monthly observations. “Min,” “Max,” and “Cumulative Returns” are the minimum, the
maximum, and the sum of monthly returns over the period January 2007 to September 2008. ***, **, * denote significance at the 0.01, 0.05, and 0.1
levels, respectively.
Roberto Savona 117
jumps in returns relative to hedge funds, and cumulative returns over

the entire period prove that the clones were more profitable in a “buy
and hold” perspective.
8.4.2 Out-of-sample analysis

The comparative analysis over the out-of-sample period January 2007 to
September 2008 corroborates the in-sample analysis offering convincing
evidence that our clones significantly outperformed hedge fund indices.
The results are in Table 8.3, where we observe correlations less than
those exhibited over the estimation sample (0.4406), with Dedicated
Short Bias, Event-Driven Risk Arbitrage, and Managed Futures showing
negative correlations. Such reduced correlations were actually “a good
news” in terms of performance, since the Sharpe ratios of our clones
were extremely high due to the moderate risk profile (on average the
annualized standard deviation was 1.17 percent versus 8.84 percent of
hedge fund indices) and higher excess returns (on average the annual-
ized mean return was 7.56 percent versus 1.61 percent of hedge fund
indices). We computed an average Sharpe ratio for hedge fund clones
of 1.2517 versus –0.0423 for hedge fund indices. And again, in terms of
probability of obtaining positive returns, return ranges (Min–Max) and
cumulative returns, our clones shown a risk–return pattern less risky and
more profitable than directly investing in hedge fund indices.
8.5 Conclusion
Using a Bayesian State Space Model we proposed a cloning procedure

for hedge fund indices based on liquid underlying assets which is easy
to replicate through common financial instruments. Our empirical
findings using CSFB/Tremont indices over the period January 1994 to
September 2008 proved that hedge fund returns can be both cloned
and outperformed in- and out-of-sample. Moreover, since our cloning
procedure is essentially based on an instrument-based beta variation
rule, we also reduce complexity in replicating the dynamic investment
strategies of hedge funds. It is noteworthy that, in doing so, we take into
account time variation in the systematic risk exposure and nonlinearity
in hedge fund returns.
Notes
1. The first three risk factors are the Trend-Following Risk Factors, namely the
primitive trend-following strategies proxied as pairs of standard straddles
constructed from exchange-traded put and call options, as described in Fung

and Hsieh (2001).
2. In Savona (2009), the market volatility is assumed to follow a mean-reverting
process with constant unconditional mean-reverting fundamental volatility
( )
that evolves according to vt vt 1 = cv vt −1 v f st , where vt and vt – 1 are the
market volatility at time t and t – 1, respectively; cv is the persistence volatility
parameter that shrinks the volatility process toward the long-run fundamental
volatility vf, assumed to be constant; s is the volatility surprise at time t assumed
to be Gaussian, namely our proxy for illiquidity shock.
3. See Amisano and Savona (2008) for more technical details of the estimation
procedure.
References
Amisano, G., and R. Savona, R. (2008). “Imperfect Predictability and Mutual
Fund Dynamics: How Managers Use Predictors in Changing the Systematic
Risk.” European Central Bank, Working Paper No. 881, Frankfurt.
Carter, C., and Kohn, R. (1994). “On Gibbs Sampling for State Space Models.”
Biometrika, 81(3): 541–53.
Fung, W., and Hsieh, D. A. (1997). “Empirical Characteristics of Dynamic
Trading Strategies: The Case of Hedge Funds.” Review of Financial Studies, 10(2):
275–302.
Fung, W., and Hsieh, D. A. (2001). “The Risk in Hedge Fund Strategies: Theory and
Evidence from Trend Followers.” Review of Financial Studies, 14(2): 313–41.
Fung, W., and Hsieh, D. A. (2004), “Hedge Fund Benchmarks: A Risk Based
Approach.” Financial Analyst Journal, 60(5): 65–80.
Fung, W., and Hsieh, D. A. (2007a), “Will Hedge Funds Regress towards Index-
like Products?” Journal of Investment Management, 2(4): 56–80.
Fung, W., and Hsieh, D. A. (2007b). “Hedge Fund Replication Strategies:
Implications for Investors and Regulators.” Banque de France Financial Stability
Review, 10(1): 55–66.
Johannes, M., and Polson, N. (2009) MCMC Methods for Continuous-Time
Financial Econometrics. In: Y. Ait-Sahalia and L. P., Hansen (eds.), Handbook of
Financial Econometrics, Vol. 2. Elsevier, Burlington, MA.
Savona, R. (2009). “Risk and Beta Anatomy in the Hedge Fund Industry.” EMFI
(Economics and Management of Financial Intermediation) Working Paper No. 1,
ADEIMF, Parma, Italy.
9
Hedge Fund Return Replication
via Learning Models
R. McFall-Lamm Jr.
9.1 Introduction
The desire to replicate hedge fund returns is a natural response to the

industry’s rapid growth from the early 2000s and its emergence as a $2
trillion “asset class.” After all, if the same return stream produced by
hedge funds is available with greater liquidity, more transparency, and
the avoidance of excessive fees, then a synthetic substitute offers tremen-
dous appeal. While an impartial observer might question why hedge
fund demand remains strong after the 2008 experience—when the
industry posted losses of 20 percent—the fact is that inflows continue at
a robust pace driven by ongoing commitments from institutions. Hedge
fund investing is now the status quo.
To date, three major issues are restraining a warm embrace of hedge
fund clones. First, many supporters of hedge fund return replication
have a vested financial interest in the commercial success of synthetics.
These include investment banks, which began to seriously market clones
as early as 2007, and academics, who promote their own versions of rep-
lication products. Their objectivity is open to question. Second, the jury
is still out on whether cloning actually works. Critics chastise the use of
opaque algorithms, a lack of intuitiveness, and clone failure in out-of-
sample testing. Third, distinctly different cloning approaches have been
proposed. One can attempt to produce a collinear stream of returns that
matches those of the hedge fund industry; one can endeavor to produce
returns that have the same distributions as hedge funds; or one can
design a set of rules that immolates the trading approaches of hedge
funds. Cloning remains controversial.
My own view is that, from an asset allocation perspective, a neces-
sary feature of hedge fund clones is that returns should be collinear
119
120 Hedge Fund Return Replication via Learning Models
with actual industry performance. Distribution replication fails in this

regard because the clone’s cross-asset correlations differ from those of
the hedge fund industry. This makes the use of standard portfolio opti-
mization approaches infeasible unless one defines clones as a distinct
asset class—something which is altogether different from investing in
hedge funds.
In this chapter I therefore focus on the efficacy of plain-vanilla
factor-based clones, asking whether they can be sufficiently enhanced to
substitute for real hedge fund exposure. My approach is different in that
I allow clones to learn over time and include a more palatable list of factors
than is typically considered. The conclusion is that cloning works if the
underlying models are properly designed. Furthermore, replication offers
the opportunity to calibrate results to deliver better performance than pro-
vided by the average hedge fund portfolio, thereby shifting the efficient
frontier outward. That said, from a practical perspective the use of clones
will likely continue to face at least one major challenge – many investors
believe they can select top quartile managers and may not be content with
lower returns than they believe they should have.
9.2 The factor-based approach to replication
Much of the work linking hedge fund returns to factors follows an essen-
tially standard methodology. For example, Hasanhodzic and Lo (2007)
regress hedge fund returns on six factors: the US dollar index, the return on
the Lehman Corporate AA intermediate bond index, the spread between
the Lehman BAA corporate bond and Treasury indices, commodities as
measured by GSCI total returns, and the first difference in the end-of-
month values of VIX. They conclude that rolling-window cloning offers
significant potential for replicating hedge fund returns and note that
results might be improved by including nonlinear factors and refining
the process.
Papageorgiou, Remillard, and Hocquard (2008) criticize Hasanhodzic
and Lo because they disregard the nonlinear factors identified by Mitchell
and Pulvino (2001), Fung and Hseih (2001), Agarwal and Naik (2004),
Chen and Liang (2006), and others, but admit that most of these factors
are not tradable and consequently of little use in constructing a replicat-
ing portfolio. Instead, they advocate the replication of hedge fund return
distributions as more appropriate, as detailed by Amin and Kat (2003)
and Kat and Palaro (2005). Kazemi, Tu, and Li (2008) also reject factor
replication and champion the cloning of hedge fund return distribu-
tion properties. Unfortunately, as already noted, distribution replication
R. McFall-Lamm Jr. 121
is unsatisfactory to asset allocators because if correlation with other

assets is zero or unknown, then the clone is not a real substitute for
hedge funds.
Amenc et al. (2008) also reject the factor-based approach because it fails
in out-of-sample tests. Moreover, they disparage distribution replication
for “practical reasons” and its lack of focus on the first moment, where
returns may be higher or lower than those produced by hedge funds.
Amenc et al. (2009) extend Hasanhodzic and Lo’s work to include non-
linear factors, but find that clones do not match actual hedge fund per-
formance in out-of-sample testing.
Based on this evidence, one might be inclined to conclude that
factor-based cloning is a lost cause. However, Roncalli and Weissang
(2008) suggest that replication can be improved via the Kalman filter
and proper model specification. The appeal of the Kalman filter is that
it allows the model to learn through time, thereby mimicking the well-
known ability of hedge fund managers to adapt quickly to new market
environments. In addition, clone rejectionists have failed to reach a
consensus on what factors make the most sense, and there is poten-
tial for improving replication robustness by expanding the candidate
universe.
9.3 Requirements for a successful replicating algorithm
Hedge fund trading is generally a very dynamic process where methods

are constantly revised and refined—today’s opportunities are different
from those of the past. For example, George Soros’ currency bet against
the Bank of England or John Paulson’s shorting of subprime debt represent
unique situations that required quick situational adjustment. Similarly,
managed futures funds, hedge funds engaged in high-frequency arbitrage,
and other quantitatively based managers are engaged in a perpetual “war
of models” where modifications to trading algorithms are constantly
made. In short, hedge funds exhibit rapid learning by swiftly responding
to market developments. Surely a successful clone should incorporate this
capability.
Proponents of cloning have attempted to take learning into account
by using rolling window regression. For example, Hasanhodzic and Lo
use trailing 24-month regressions. Unfortunately, rolling windows are a
crude and imperfect method of embedding learning in models, since past
data are weighted equally over an arbitrary time frame. In the case of a
24-month rolling window, each month’s information has the same value,
even experiences from two years ago, while information from 25 months
ago or more has no importance whatsoever. There is no memory fade

for two years and then everything is 100 percent forgotten. The human
thought process is not so disjoint, and learning models with continuous
memory decay are much more plausible.
As for the relevant factors to include in replicating algorithms,
legitimate candidates need to proxy the net exposures actually carried
by hedge funds. Clearly, this includes equities and equity cap spreads
due to the preponderance of equity long/short managers in the hedge
fund world. It also includes fixed income, commodities, and currencies,
which are the domain of global macro managers. Also, fixed income
managers have exposure to sovereign debt, credits, yield curves, and
duration spreads; convertible arbitrage funds own convertibles or their
underlying components with delta-short equity hedges; and distressed
debt managers hold assets in nearly bankrupt or insolvent companies.
Past research is often based on a haphazard list of possibilities that is
insufficiently inclusive, leading to model specification bias and poor
forecasting accuracy.
A last requirement for a successful clone is that it should accurately
account for costs. Most prior work has not allowed for the cost of short-
selling, which can be very significant for hedge funds since margin must
be posted and dividends or interest paid to the owners of shorted secu-
rities. In addition, some hedge fund strategies employ leverage, which
requires interest payments that subtract from performance. A replicating
algorithm that mimics hedge fund returns well in a hypothetical test may
underperform in the real world unless such costs are taken into account.
While this deficiency can be addressed by aiming for a sufficiently high
target return to offset costs, such calibration needs to be explicit to pro-
vide for an accurate assessment of results.
9.4 A simple cloning model
We attempt to take these issues into account and follow the basic
approach espoused by Branch and Evans (2006) in applying the Kalman
filter recursion. The model takes the form:
(1) yt ␤t xtet
(9.1)
(2) ␤t ␤t1 ␪t
where yt is the hedge fund return to be replicated, ␤t is the K 1 vector

of factor weights, and xt is the K 1 vector of factor returns with et and
␪t having zero means.
The realized return is yt bt′1xt . Model parameters are estimated via
recursive least squares (RLS), a special case of the Kalman filter:
(3) bt bt1 gtR1

t xt(ytbt 1xt)
(4) Rt Rt1 gt(xtxt Rt1)
where Rt is the covariance matrix of xt. If gt 1/t, the specification

is equivalent to ordinary least squares (OLS) where all information is
equally weighted. Because hedge fund managers are likely to pay more
attention to recent market behavior, a better method is constant gain
learning where gt g. This allows past information to be discounted geo-
metrically at the rate 1 – g, which means that forgetting occurs smoothly.
Figure 9.1 illustrates the differences in learning as embodied in the
weights on past information for OLS, a moving window of 24 months,
and constant gain learning with g set at 5 per cent and 10 percent, respec-
tively, over a 100 month sample. Undoubtedly, the pattern of memory
fade in the rolling window is quite strange with an abrupt tail-off, while
there is no memory wane at all in OLS.
While learning models have received much notoriety recently,
constant gain learning is really nothing new. It is equivalent to classic
weighted least squares with bt (Xt WXt)1Xt Wyt where W consists of
all zeros except the diagonal, which contains the progression [(1 g)t…
(1 g)2 (1 g)1 (1 g)0]1/2. However, the Kalman filter recursion is more
computationally efficient since it avoids the repeated inversion of large
matrices required by running weighted least squares through time.
10% Rolling window (24 months)

Constant gain g = .05
8% Constant gain g = .10
Full sample regression (100 months)
6%
Weight
4%
2%
0%
55 50 45 40 35 30 25 20 15 10 5 0
Months before present
Figure 9.1 The value of past information: rolling windows vs. constant gain
9.5 Implementation
9.5.1 Which hedge fund index to replicate?

An important issue not discussed so far is the question of what hedge
fund index to replicate. This is a crucial consideration because hedge
fund indices are subject to severe survivor bias and overstate returns.
For example, Xu, Liu, and Loviscek (2010) find that survivor bias in
composite hedge fund indices averaged 1.7 to 3.2 percent over the
1994–2009 period and might be even higher if one took into account
the time lapse between when a fund stopped reporting and when it
closed. Xu, Liu, and Loviscek’s results are consistent with earlier work
by Liang (2000, 2001) and Amin and Kat (2003), which found similar
levels of upward bias. Even worse, Bollen and Pool (2009) find that
hedge fund managers misrepresent the returns that are reported. As a
result, hedge fund indices are not credible as measures of returns and
building a clone to replicate sham composites is fallacious. Replicating
hedge fund returns by aggregating strategy performance is equally fool-
hardy, since the category indices are also subject to substantial bias.
This dilemma can be largely avoided by using fund-of-funds (FOF)
indices where survivor bias is minimal, since FOFs are long-lived and
rarely stop reporting or close if one fund in the portfolio experiences a
sharp drawdown or blows up. For example, Xu, Liu, and Loviscek report
survivor bias in FOFs of only 0.2–0.6 percent annually from 1994 to
2009. Additionally, “investable” hedge fund indices also generally avoid
survivor bias since the underlying funds are liquid and obliged to report
even if mishaps occur. Consequently, FOFs and investable indices are
reasonable proxies of the performance one might expect to obtain from
hedge fund investing.
The disadvantage of using FOF returns is that performance understates
what would be obtained if one invested in their own portfolio of hedge
funds—FOFs charge another layer of fees. Even so, this can be easily
addressed by calibrating the clone to target a return one or two percent
above that of FOFs. We choose the well-known Hedge Fund Research
FOF index to replicate, noting that the CISDM FOF index produces virtu-
ally identical results and that FOF returns are generally higher than those
of investable indices.
9.5.2 Factors for inclusion

Past studies are sometimes deficient in that they utilize factors designed
to rationalize the behavior of equity portfolio managers or employ factors
found to explain the behavior of single strategies. In both cases factors
could erroneously be deemed important because they are based on series

distorted by survivor bias. Here we use factors that intuitively correspond
with the net exposures carried by the dominant hedge fund strategies.
Our list includes equity returns (the S&P 500); time spreads (the Merrill
Treasury index less 3-month Treasury returns); equity cap spreads (the
Russell 2000 less the S&P 500 return); convertible debt returns (the Merrill
convertible index); credit spreads (Merrill high yield less Treasury index
returns); commodities (the return on the CRB index); currencies (the
return on the ICE dollar index); and volatility (as measured by the VIX
index). The result is a very simple nine variable, eight factor model that
is linear in the parameters.
This design scheme hopefully captures the essence of broad-based
hedge fund trading approaches. For example, it allows for the well-
known tendency of hedge funds to short Treasuries to support long posi-
tions in riskier assets; owning convertibles while delta-shorting equity
risk; taking outright long or short positions in fixed income, equities,
currencies, and commodities according to macro views; carrying net long
equity exposure whether the result of long/short positions, anticipated
events, or activist holdings; hedging by selling volatility; or investing
in distressed securities (where returns have nearly exactly paralleled
high-yield debt performance in recent years). Furthermore, the approach
allows for the extraction of net beta in market neutral trading approaches
such as merger, statistical, convertible, or fixed income arbitrage to the
extent that it exists. While such crude aggregation certainly blurs com-
plex nuances, all we need is for the clone to replicate the fundamental
nature of hedge fund exposure, and this may be enough.
9.5.3 Estimation process

The first five years of sample data (beginning in January 1990) are used
to make initial covariance matrix and parameter estimates. We then
move through time sequentially from February 1995 onwards, updat-
ing covariances and parameters as well as forecasting one period ahead.
The process is concluded as of December 2007 to provide a cut-off well
in advance of the 2008 financial crisis and the optimum value of the
learning gain parameter g is determined by minimizing the root mean
square error (RMSE) for the model over the 1995 to 2007 period. A pure
out-of-sample test is then conducted from January 2008 to December
2010 with g fixed at its predetermined optimum.
Short-selling and leverage costs are accounted for by charging such
positions at LIBOR plus 50 basis points and reducing the clone return by
the corresponding amount in each period. We charge for the full short
position even though in reality costs would likely be somewhat lower

via margin posting and offset allowances. We charge for leverage at the
same LIBOR plus 50 basis point rate. However, accounting for leverage
makes little difference in the analysis as the clone is more than 100
percent invested for only a few months of the sample.
The clone is initially set to target the actual performance of the HFR
fund-of-funds index over the 1995 to 2007 period on an adjusted cost
basis. The process is then repeated with the clone return target set 2 per-
cent higher. This second sequence is denoted as “clone-plus.” It provides
a more reasonable goal since it approximates what one might receive
from investing in a portfolio of hedge funds directly. Table 9.1 and
Figure 9.2 display the results.
The clone obviously replicates FOF performance very well within
sample, attaining exactly the same annualized return. There are small
Table 9.1 Clone vs. actual fund-of-funds returns
Metric Fund-of-funds Clone Clone-plus Clone without

learning
In-sample (Feb.
1995–Dec. 2007)
Months 155 155 155 155
Annualized return 9.0% 9.0% 11.0% 8.2%
Annualized volatility 5.7% 5.9% 5.9% 5.7%
Skew –0.35 –0.36 –0.34 –0.35
Kurtosis 4.55 6.14 6.03 5.08
Sharpe ratio 0.84 0.81 1.15 0.71
JB statistic 137.1 246.6 237.9 170.1
RMSE – 1.09 1.13 1.16
g – 0.046 0.048 –
Out-of-sample ( Jan.
2008–Dec. 2010)
Months 36 36 36 36
Annualized return –2.6% –2.2% 0.1% 0.7%
Annualized volatility 7.5% 8.4% 8.4% 9.2%
Skew –1.30 –1.60 –1.53 –1.61
Kurtosis 1.93 3.98 3.83 3.94
Sharpe ratio –0.44 –0.36 –0.09 –0.01
JB statistic 15.7 39.2 36.1 38.8
RMSE – 1.24 1.26 1.39
Fund of funds returns are from Hedge Fund Research.

differences in replicating higher-order distribution properties, however,

although the clone’s volatility and skew are very close to those of the
FOF index. Optimizing the learning parameter yields a value of 0.046 for
g, which produces a half life of approximately 15 months. This indicates
that the value of information from 15 months ago has half the weight of
current information and that the clone therefore learns very rapidly.
For the out-of-sample test from January 2008 through December 2010
(with g held at its optimized value of 0.046), there is some deterioration
in clone performance but not much. The clone delivers a slightly higher
return (–2.2 percent vs. –2.6 percent) but at the cost of greater volatility
(8.4 percent vs. 7.5 percent) and somewhat more negative asymmetry.
Nonetheless, the results are laudable given the extraordinary market
movements experienced over the period, which include a massive stock
market crash followed by an extraordinary rebound.
The clone-plus version of the model also delivers impressive results,
posting the targeted 2 percent outperformance with similar distribution
statistics both within and out-of-sample. Of particular interest is that
higher clone-plus returns are produced with only a very modest increase
in volatility versus the FOF index and the primary clone. This is accom-
plished with slightly faster learning with the optimum value of γ found
to be 0.048.
Despite the clone’s near-perfect replication of FOF returns, there is
noticeable performance drift at times due to tracking error. For example,
from 2001 to 2003 the clone outruns FOF returns while in 2006 and 2007
the clone substantially trails FOF performance as illustrated in Figure 9.2.
The tracking error issue shows up more conspicuously when one exam-
ines yearly clone returns, which are displayed in Table 9.2. While clone
performance is close to FOF returns for most years, there are occasions
when severe discrepancies occur, such as in 1998 and 2000 when clone
returns are directionally different from those of FOFs. As a result, cloning
clearly should be viewed as adequate only from a multi-year and long-run
perspective.
The results confirm that incorporating learning is of critical impor-
tance for improving replication precision. A comparison of clone returns
with and without learning makes this evident. The last column of Table
9.1 shows clone results when g is set to 1/t, which is equivalent to OLS.
Without learning, tracking error increases considerably and the clone
underperforms the FOF index by 0.7 percent annually with no gain in
higher-order performance characteristics. Out-of-sample, the no-learning
clone underperforms in higher moments and RMSE. Strangely, the mean
return of the no-learning clone is higher out-of-sample than actual
$450
Fund-of-funds
$400
Clone
$350 Clone plus
$300
$250
$200
$150
$100
95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11
Figure 9.2 Clone returns vs. fund-of-funds
Table 9.2 Clone vs. actual fund-of-funds returns by year
Year Fund-of-funds Clone Clone-plus

1995 12.5% 12.1% 13.4%
1996 14.4% 11.6% 14.4%
1997 16.2% 21.9% 25.3%
1998 –5.1% 5.9% 5.9%
1999 26.5% 24.9% 24.4%
2000 4.1% –1.4% –0.9%
2001 2.8% 4.9% 7.6%
2002 1.0% 5.6% 8.2%
2003 11.6% 7.8% 10.1%
2004 6.9% 5.8% 7.8%
2005 7.5% 2.5% 5.3%
2006 10.4% 7.0% 10.5%
2007 10.3% 9.8% 12.6%
2008 –21.4% –16.7% –15.3%
2009 11.5% 11.6% 13.4%
2010 5.6% 2.7% 4.3%
Fund of funds returns are from Hedge Fund Research.
FOF returns and the clones that learn. While this is no doubt a chance
occurrence, perhaps one does not want to learn too fast during market sell-
offs amid financial crises.
As for the factor positions utilized by the clone to replicate FOF per-
formance, Figure 9.3 shows net exposure month by month over the entire
150%
100%
Long
50%
0%
–50% Short
Convertibles Currencies
–100% High yield Treasuries
Small cap equities Large cap equities
Commodities
–150%
95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 10
Figure 9.3 Clone net positions
sample. Of note is the fairly sizeable long position consistently held by the
clone in convertible securities, which is in large part attributable to the fact
that convertible returns are most highly correlated with FOF performance
over time. The clone exploits this and thus intriguingly captures a sort of
nonlinear exposure, since convertibles decompose into corporate debt plus
an equity call option. Otherwise, the only major surprise is the generally
low net level of positions required for effective FOF return replication.
Finally, with respect to shorting and leverage costs, we find that they
often round to zero and at most subtract 0.2 percent of the clone’s return
in any month. The reason is primarily that the clone does not employ
significant short exposure—except in 2008 and 2009 when borrowing
costs are very low—nor use significant leverage.
9.7 An out-of-sample test using ETFs
One shortcoming of cloning is that factors are represented by indices

that cannot be reproduced without purchasing a large number of securi-
ties. Of course, this is not an intractable problem given the sophistication
of the market today and the ease of trading indices. Nonetheless, there
is a cost involved. To provide a quick and dirty judgment about whether
these costs make much of a difference, we build a clone portfolio consisting
of an appropriate basket of exchange-traded funds (ETFs) that are based
on indices similar to those included in our analysis. We start the exercise
$120
ETF clone portfolio
$115 Fund-of-funds index
$110
$105
$100
$95
M -09
Ju -09
Ju -09
Au -09
Se 09
O -09
N -09
D -09
Ja -09
Fe -10
M -10
Ap -10
M -10
Ju -10
Ju -10
Au -10
Se 10
O -10
N -10
D -10
Ja -10
11
g-
g-
n-
r
ay
n
l
p
ct
ov
ec
n
b
ar
r
ay
n
l
p
ct
ov
ec
Ap
Figure 9.4 Out-of-sample replication via ETFs
in April 2009, which is the first date that enough ETFs are available to
approximate the factors in the clone model.
For the S&P 500, the Russell 2000, and Treasury bonds we use SPY, IWM,
and IEI, respectively, which should provide nearly identical returns to our
factor indices less very small ETF fees. For convertible securities, the CRB
commodity index, high yield securities, and VIX, we employ CWB, GCC,
HYG, and VXX which are based on different but reasonably similar indices.
For the dollar index, we weight FXE, FXY, FXB, FXC, FXS, and FXF propor-
tionately to the dollar index’s weights for the euro, the yen, the pound, the
Canadian dollar, the Swedish krona, and the Swiss franc. The one liberty
taken is the use of BIL in lieu of SHY to represent cash returns.
Figure 9.4 displays results for the ETF clone using the out-of-sample
weights determined in the initial analysis. Surprisingly, we find the ETF
clone’s returns are nearly equal to those reported by HFR for the FOF
index over the April 2009 to January 2011 period—despite the fact that
one would expect slightly lower returns due to embedded ETF fees. This
provides yet further confirmation that cloning can be successful even
in a proximate case.
9.8 Conclusion
The key finding is that it appears that FOF returns or FOF returns plus 2
percent can be replicated via clones with reasonable accuracy, provided
that the appropriate factors are taken into account and more sophisticated
learning models are employed. The conclusion holds up through an
out-of-sample test that covers one of the most disruptive investment
environments ever—the 2008 and 2009 global financial crisis. The major
identifiable problem with synthetic cloning is that tracking error is not
zero, even within sample, and year-to-year clone performance can on
occasion depart substantially from that delivered by FOFs.
The immediate response from many investment professionals is
likely to be that achieving FOF returns—or 2 percent better—is not
enough. Many believe that they can do better by selecting top-quartile
hedge fund managers without adding incremental higher moment
risk. My rejoinder is that while better performance is certainly possible,
many investors held exactly the same view prior to 2008 only to dis-
cover that their hedge fund icons performed as badly as or worse than
the industry. Despite this, building your own hedge fund portfolio via
manager selection is paradigmatic and likely to remain so until a clone-
plus product goes live and demonstrates that it represents a truly viable
alternative.
A more pertinent question is whether FOF returns plus 2 percent is
enough from an asset allocation perspective. Normally, adding time-
varying exposure to assets such as commodities, currencies, convertibles,
and distressed debt might be expected to shift the efficient frontier
outward by increasing portfolio diversification. It was certainly the case
in the 1990s when hedge fund performance was exemplary and substan-
tial allocations were appropriate, as noted in Lamm (1999). However,
performance has deteriorated significantly for the past half decade and
more, so much so that a 60/40 percent stock and bond portfolio has
outperformed the FOF industry on both a total return and Sharpe ratio
basis since the mid-2000s. To justify a substantial hedge fund allocation
therefore requires expectations of better relative hedge fund performance
in the future, clone or no clone. While FOF returns plus 2 percent may
still justify at least a small allocation to hedge funds, the case may be
that an ultimately successful clone may need to target FOF returns plus
3 percent or more.
References
Agarwal, V., and Naik, N. Y. (2004). “Risks and Portfolio Decisions Involving
Hedge Funds.” The Review of Financial Studies, 17(1): 63–98.
Amenc, N., Gehin, W., Martellini, L., and Meyfredi, J.-C. (2008). “Passive Hedge
Fund Replication: A Critical Assessment of Existing Techniques.” The Journal of
Amenc, N., Martellini, L., Meyfredi, J.-C., and Ziemann, V. (2009). “Performance
of Passive Hedge Fund Replication Strategies.” Research report, EDHEC Risk
and Asset Management Research Center, Nice, France.
Amin, G., and Kat, H. (2003). “Welcome to the Dark Side: Hedge Fund Attrition
and Survivorship Bias over the Period 1994–2001.” The Journal of Alternative
Investments, 6(1): 57–73.
Bollen, N. P. B., and Pool, V. K. (2009). “Do Hedge Fund Managers Misreport
Returns? Evidence from the Pooled Distribution.” Journal of Finance, 64(5):
2257–88.
Branch, W. A., and Evans, G. W. (2006). “A Simple Recursive Forecasting Model.”
Economic Letters, 91(2): 158–66.
Chen, Y. and Liang, B. (2007). “Do Market Timing Hedge Funds Time the
Market?” Journal of Financial and Quantitative Analysis, 42(4): 827–56.
Fung, W., and Hsieh, D. A. (2001). “The Risk in Hedge Fund Strategies: Theory and
Evidence from Trend Followers.” Review of Financial Studies, 14(2): 313–41.
Hasanhodzic, J., and Lo, A. W. (2007). “Can Hedge Fund Returns be Replicated?
Kat, H. M., and Palaro, H. P. (2005). “Hedge Fund Returns: You Can Make Them
Yourself!” The Journal of Wealth Management, 8(2): 62–8.
Kazemi, H., Feng, T., and Ying Li (2008). “Replication and Benchmarking of
Hedge Fund Replication Programs.” The Journal of Alternative Investments, 11(2):
40–60.
Lamm, R. M. (1999). “Why Not 100% Hedge Funds?” The Journal of Investing,
8(4): 87–97.
Liang, Bing (2000). “Hedge Funds: The Living and the Dead.” Journal of Financial
and Quantitative Analysis, 35(3): 309–26.
Liang, Bing (2001). “Hedge Fund Performance 1990–1999.” Financial Analysts
Journal, 57(1): 11–18.
Mitchell, M., and Pulvino, T. (2001). “Characteristics of Risk and Return in Risk
Papageorgiou, N., Remillard, B., and Hocquard, A. (2008). “Replicating the
Properties of Hedge Funds.” The Journal of Alternative Investments, 11(2): 8–39.
Roncalli, T., and Weisang, G. (2008). “Tracking Problems, Hedge Fund Replication
and Alternative Beta,” Working Paper, University of Evry, Evry, France, South
Orange, NJ.
Xu, X., Liu, J., and Loviscek, A. (2010) “Hedge Fund Attrition, Survivorship Bias,
and Performance: Perspectives from the Global Financial Crisis,” Working
paper, Seton Hall University, February.
10
Linear Model for Passive Hedge
Fund Replication
10.1 Introduction and Literature
Considerable academic research has been conducted on the effectiveness

of hedge fund replication methods. In the last five years hedge fund
replication products have become part of investment banking. Funds or
structured products whose aim is to replicate the returns of hedge funds1
are offered by banks such as Goldman Sachs and Merrill Lynch. Their
success is due to two main reasons (1) the low interest rate experienced
in developed countries in the last decade and (2) the growing difficulty
for hedge funds to produce higher returns when risk premia are driven
down by huge flows of private and institutional capital.
Hedge fund replication products are cheap, transparent, and liquid
instruments that share most of the beneficial properties of hedge
fund investing. Replication of the average hedge fund payoff provides
insight into the investment process of the asset manager, but not the
portfolio manager’s trading strategy. The first paper to investigate hedge
fund risk exposure was by Fung and Hsieh (1997), who built on the work
of Sharpe (1992). Many papers have looked for the presence of systematic
risk factors in hedge fund returns by analyzing the exposure of CTAs
and hedge funds through style analysis. Following Fung and Hsieh (1997),
numerous papers have investigated linear and nonlinear risk factors with
the aim of improving the understanding of the dynamic trading strategies
employed by hedge fund managers (Fung and Hsieh, 1997, 2002, 2004,
2007; Agarwal and Naik 2000a, 2004). Other related research has tried to
reverse engineer hedge fund trading strategies, proposing simple and fea-
sible trading algorithms which are implemented to examine hedge fund
tactical asset allocation (e.g., Mitchell and Pulvino (2001) for merger arbi-
trage and Duarte, Longstaff, and Yu (2007) for fixed income arbitrage).
133
134 Linear Model for Passive Hedge Fund Replication
Using the TASS database, Hasanhodzic and Lo (2007) analyzed the out-
of-sample results of linear clones and compared them with original hedge
funds. Their results show that linear clones are inferior to their hedge
fund counterparts and argue that the transparency and lower cost could
make linear clones a valid and inexpensive substitute. This appears to be
true, especially after a credit crisis. The Markov-switching models and the
Kalman Filter applied respectively by Amenc et al. (2008) and Roncalli
and Teïletche (2008) show superior results to the standard rolling-window
OLS approach, at the cost of adding a level of complexity.
As observed by Rudolf et al. (1999), passive investment strategies could
be implemented in various ways, starting from different definitions
of tracking error depending on the investor’s objectives. Roll (1992)
defined tracking error as the variance of the error between a fund and
a benchmark by solving the minimization problem through quadratic
programming. In addition, Clarke, Krase, and Statman (1994) defined
the tracking error as the absolute difference between a portfolio and his
benchmark.2 This second definition has also the practical advantage
that managers rewarded by the difference between the portfolio and
the benchmark have a better perception of linear performance fees.
Goldstein and Taleb (2007, p. 84) studied the behavior of finance profes-
sionals and students when asked for estimates of the standard deviation.
They discover that:
Finance professionals, who are regularly exposed to notions of volatil-

ity, seem to confuse mean absolute deviation with standard deviation,
causing an underestimation of 25% with theoretical Gaussian variables.
In some fat tailed markets the underestimation can be up to 90%. The
mental substitution of the two measures is consequential for decision
making and the perception of market variability.3
Speranza (1993) showed that after imposing normal distributed returns

the weights under mean variance optimization and Mean Absolute
Deviation (MAD) coincide. Konno and Yamazaki (1991) presented and
analyzed the complete portfolio optimization using MAD technique.
Rudolf et al. (1999) proposed several mean absolute deviation optimiza-
tion models for equity portfolios and demonstrated that linear tracking
error optimization is equivalent to expected utility maximization and
lower partial moments minimization. We replicate the time series of
hedge fund payoffs and the distribution of returns. For this reason we do
not compare our strategy with the payoff distribution pricing model by
Kat and Palaro (2006a,b).
Giovanni Barone-Adesi and Simone Siragusa 135
10.2 Data
We use the Dow Jones Credit Suisse Hedge Fund Tremont Indexes in
order to replicate the average fund. We select Tremont because of its long
history and because it is an investable index.4 In Table 10.1 we report
the univariate properties of hedge funds strategies5 and the stock market
represented by the S&P 500 Index. As we can see the properties of hedge
funds generally have higher risk-adjusted return using the Sharpe ratio.
When we look at higher moments, a mixed picture is revealed. Only dedi-
cated short and managed futures report positive skewness. Furthermore,
kurtosis is also positive for all strategies except for managed futures,6 with
few exceptions where autocorrelations are substantially high.
This is also a well-known stylized fact and it is usually traced back to
stale prices and investment in illiquid assets. Only the market-neutral
and managed futures strategies are less autocorrelated than the S&P 500
Index. In Table 10.2 we regress the returns of each strategy against the
returns of the S&P 500 Index. We find that all hedge fund indexes, apart
from dedicated short and market-neutral strategies, have positive and
Table 10.1 Univariate statistics of S&P 500 and CSFB Tremont Indexes; data
from July 1996 to December 2010
Index Mean Std dev Skew Exc. Sharpe Autocorrelation (%)

(ann.%) (ann. %) kurt ratio
SP500 5.90 16.62 –0.63 0.64 0.35 10.35
Hedge fund 9.46 7.52 –0.32 3.06 1.26 22.70
index
Convertible 8.40 7.45 –2.78 15.49 1.13 55.53
arbitrage
Dedicated –2.86 17.66 0.75 1.52 –0.16 8.35
short
Emerging 9.78 14.24 –1.33 7.16 0.69 28.53
markets
Market 5.72 11.44 –11.15 138.10 0.50 6.59
neutral
Event-driven 10.36 6.35 –2.54 13.16 1.63 36.17
Fixed income 5.02 6.28 –4.20 26.62 0.80 53.11
arbitrage
Global macro 12.70 9.39 –0.27 4.34 1.35 12.64
Long short 11.06 10.36 0 3.40 1.07 18.14
Managed 8.30 11.79 0.05 –0.32 0.70 4.98
futures
Multi-strategy 8.55 5.26 –2.08 8.46 1.63 47.18
Table 10.2 Regression of the Dow Jones Tremont Hedge Fund Indexes on S&P
500 index
Index Alpha (%) Beta (%) Adj. R2 (%) F-stat

(t-stat) (t-stat)
Hedge fund index 0.63** 26.67** 34.33 91.44**
(4.69) (9.59)
Convertible arbitrage 0.60** 16.30** 12.71 26.20**
(3.89) (5.12)
Dedicated short 0.15 –81.25** 58.23 242**
(0.58) (15.56)
Emerging markets 0.53* 51.82** 36.19 99.13**
(2.13) (9.96)
Market neutral 0.37 19.59** 7.56 15.16**
(1.53) (3.89)
Event-driven 0.71** 24.04** 39.17 112.39**
(6.51) (10.6)
Fixed income arbitrage 0.35* 12.46** 10.34 20.95**
(2.67) (4.57)
Global macro 0.93** 14.44** 5.99 12.03**
(4.65) (3.47)
Long short 0.68** 40.42** 41.69 124.67**
(3.93) (11.16)
Managed futures 0.69** –5.65 0.06 1.10
(2.68) (–1.05)
Multi-strategy 0.62** 12.75** 15.77 33.40**
(5.88) (5.78)
Note:
* : indicates significance at 95 percent.
** : indicates significance at 99 percent.
significant alphas. The systematic exposure (beta) to market risk also

proved to be significant and positive for all strategies except for man-
aged futures. The F-statistic shows a strong linear relationship between
the stock market and hedge fund strategies. In fact, linearity is accepted
at 99 percent confidence level for all Tremont Indexes except for man-
aged futures.
10.3 Risk factors
The existing literature suggests different risk factors. In this chapter we

want to keep them as simple as possible, avoiding the use of spreads
or strategies. Similarly to Hasanhodzic and Lo (2007) we use a stock
market risk factor, a bond risk factor, a credit risk factor, a currency risk
factor, and a commodity factor. In addition, we also add volatility and

an emerging market risk factor. Our complete list7 includes
1. Stock risk factor: the S&P 500 total return represents the equity market
risk factor.
2. Bond risk factor: Fidelity Government Income Fund (FGOVX). The fund
invests in instruments related to US government securities and allocates
assets across different market sectors and maturities. The credit quality
is high (AAA) and the modified duration of the fund is medium-low.
3. Credit risk factor: Vanguard High Yield Corporate (VWEHX). The fund
invests in a diversified group of high-yielding, higher-risk corporate
bonds with medium- and lower-range credit quality ratings. At least
80 percent of assets are corporate bonds that are rated below BAA by
Moody’s or equivalent. The fund may not invest more than 20 percent
of assets in bonds with credit ratings lower than B or the equivalent,
convertible securities, and preferred stocks. The interest rate sensitivity
is classified as medium.
4. Currency risk factor: represented by the US Dollar Index return.
5. Commodity risk factor: represented by Gold futures.
6. Volatility risk factor: the first difference of the end-of-month value of
the CBOE Volatility Index.
7. Emerging markets risk factor: Fidelity Emerging Markets Fund
(FEMKX). The fund invests normally at least 80 percent of assets in
securities (common stocks) of issuers in emerging markets and other
investments that are economically tied to emerging markets.
10.4 Optimization models
Return-based style analysis uses a constrained quadratic optimization

to determine the combination of assets/strategies that best tracks the
performance of hedge fund managers. This suggests that the aim of
numerous investors is to track a certain benchmark return as close as
possible by minimizing the sum of the squared deviations between the
investment itself and a benchmark, that is, the tracking error variance
(Roll, 1992).
Let Y be the vector of returns of the hedge fund index that we want
to replicate, X the matrix of returns of the factors reported above, ␤ the
vector of the weights and ␧ the vector of the difference between the
hedge fund and the replica. The problem could be represented as
Y X␤ ␧, Y ∈ ℜT , X ∈ ℜT n , ␤ ∈ ℜn , ␧ ∈ ℜT
where n is the number of assets and T the number of observations. In

order to minimize the squared error we write:
min ␧␧ min( Y X␤)′( Y X␤)

␤ ␤
If we do not impose any restriction on ␤, then the estimation boils down

to the usual OLS estimation and has the properties to be BLUE.8 In order
to have meaningful weights we use the standard quadratic program-
ming and apply a constraint on the sum of weights (␤) of the replica
that must add up to 1.9
Using the same notation as above and starting from the definition of
the tracking error as absolute deviation of the portfolio from the bench-
mark (1(|Xβ Y|)), we can write a mean absolute deviation model as:
min 1′(| X␤ Y |), where 1′≡ (1,… , 1) ∈ ℜT

␤
In the second case, note that portfolio weights (␤) are determined by mini-
mizing the sum of the absolute deviations between the benchmark returns
and portfolio returns. The same constraint on the sum of the weights (␤)
as in the tracking error variance model is applied. As in Hasanhodzic and
Lo (2007), we adjust the weights of the replica for a gearing factor given
by the ratio of the volatility of the hedge fund index and the volatility of
the replication.
We run the Tracking Error Variance (TEV) minimization and Mean

Absolute Deviation (MAD) minimization models with a rolling window
period. In Table 10.3 we show results for 36 months in-sample estimation,
keeping the out-of-sample period at six months. We employ a six-month
rolling period because we wish to avoid high transaction costs and look-
ahead bias. During the out-of-sample period we do not change weights
and leverage. Table 10.3 reports statistics for every hedge fund index, the
style analysis optimization model and the mean absolute deviation opti-
mization model respectively. Looking at Table 10.3, the Mean Absolute
Deviation model (see MAD Indexes in Table 10.3) has, in general, higher
mean and lower volatility than the Style Analysis model (see TEV Indexes
in Table 10.3). We further report skewness and kurtosis for the indexes as
well as for both models.
Furthermore, the Sharpe ratio suggests that apart from convertible
arbitrage and managed futures strategies, higher risk-adjusted performance
Table 10.3 Performance comparison of 36 rolling window TEV and MAD clones with six months rebalancing
Index Mean St. dev Skew Exc. kurt Correlation Sharpe ratio Leverage
Hedge fund index 0.66 1.91 –0.20 4.32 – 1.24 –
Hedge fund TEV 0.31 1.93 –0.67 3.08 52.90 0.57 1.00
Hedge fund MAD 0.39 2.03 –0.28 2.26 54.11 0.68 0.95
Convertible arbitrage 0.59 2.44 –2.98 10.99 – 0.87 –
Convertible arbitrage TEV 0.31 2.15 3.96 28.6 8.49 0.51 1.22
Convertible arbitrage MAD 0.24 2.27 3.42 22.83 8.56 0.37 1.20
Emerging 0.93 3.39 –0.35 4.19 – 1.00 –
Emerging TEV 0.42 3.89 –0.27 2.52 77.64 0.38 1.02
Emerging MAD 0.53 3.66 0.01 1.86 79.69 0.52 1.00
Market neutral 0.27 3.66 –10.30 114.18 – 0.26 –
Market neutral TEV 0.64 3.27 4.60 31.01 4.85 0.70 1.10
Market neutral MAD 0.64 2.69 3.24 18.19 4.24 0.85 1.20
Event-driven 0.77 1.59 –1.22 2.89 – 1.75 –
Event-driven TEV 0.21 1.98 –0.15 1.31 51.16 0.37 1.01
Event-driven MAD 0.22 1.86 –0.98 3.44 58.42 0.41 0.95
Fixed income arbitrage 0.39 1.84 –4.49 30.06 – 0.75 –
Fixed income arbitrage TEV 0.31 2.42 0.27 9.07 54.85 0.45 1.05
Fixed income arbitrage MAD 0.34 2.28 0.41 13.68 56.13 0.53 1.10
Global macro 0.97 1.93 –0.35 4.31 – 1.84 –
Global macro TEV 0.37 4.71 –3.9 42.41 29.53 0.28 1.52
Global macro MAD 0.38 2.73 –1.24 11.21 33.24 0.49 1.08
Long short 0.69 2.80 0.55 4.15 – 0.89 –
Long short TEV 0.26 2.96 –0.52 1.73 56.92 0.31 1.04
(continued)
139
Table 10.3 Continued
140
Long short MAD 0.32 2.91 1.73 3.17 52.79 0.39 0.97
Managed futures 0.60 3.44 –0.00 –0.57 – 0.62 –
Managed futures TEV 0.76 3.67 –0.42 0.78 32.15 0.75 1.28
Managed futures MAD 0.63 3.76 –0.57 1.23 31.78 0.60 1.20
Multi-strategy 0.63 1.57 –2.00 8.31 – 1.44 –
Multi-strategy TEV 0.33 1.68 0.27 3.78 54.72 0.69 1.09
Multi-strategy MAD 0.35 1.56 –0.09 2.88 54.76 0.79 0.95
Table 10.4 Performance comparison of 48 rolling window TEV and MAD clones with six months rebalancing
Hedge fund index 0.59 1.62 –1.23 2.74 – 1.30 –
Hedge fund TEV 0.25 1.99 –0.42 1.70 58.63 0.44 1.08
Hedge fund MAD 0.27 2.04 –0.84 3.26 60.53 0.47 0.96
Convertible arbitrage 0.59 2.44 –2.18 10.99 – 0.87 –
Convertible arbitrage TEV 0.36 3.34 3.63 30.08 10.22 0.38 1.70
Convertible arbitrage MAD 0.35 3.07 2.71 19.41 –0.05 0.40 1.84
Emerging 0.82 2.98 –1.37 4.17 – 1.00 –
Emerging TEV 0.39 3.43 –0.66 1.59 79.60 0.40 0.99
Emerging MAD 0.54 3.40 –0.64 1.29 79.38 0.57 0.95
Market neutral 0.27 3.66 –10.30 114.81 – 0.26 –
Market neutral TEV 0.31 2.17 1.86 12.98 3.84 0.50 1.12
Market neutral MAD 0.52 2.15 2.78 10.68 6.26 0.86 1.56
Event-driven 0.74 1.62 –1.23 2.74 – 1.65 –
Event-driven TEV 0.12 1.86 –0.33 1.36 54.11 0.22 1.01
Event-driven MAD 0.17 1.90 –0.90 3.42 58.91 0.31 0.97
Fixed income arbitrage 0.39 1.84 –4.49 30.06 – 0.75 –
Fixed income arbitrage TEV 0.36 1.96 0.31 4.87 50.74 0.65 0.79
Fixed income arbitrage MAD 0.39 2.06 –1.07 12.35 64.63 0.67 0.92
Global macro 0.97 1.93 –0.35 4.31 – 1.84 –
Global macro TEV 0.26 2.31 0.08 2.40 13.42 0.40 1.37
Global macro MAD 0.31 3.38 –0.37 3.25 18.94 0.32 1.11
(continued)
141
142
Long short 0.52 2.20 –0.75 1.63 – 0.84 –
Long short TEV 0.16 3.05 –0.75 2.52 56.92 0.16 1.04
Long short MAD 0.18 3.37 –1.10 3.85 68.10 0.18 0.99
Managed futures 0.60 3.44 0.00 –0.65 – 0.62 –
Managed futures TEV 0.55 3.87 –0.43 0.79 29.04 0.51 1.43
Managed futures MAD 0.40 4.41 –0.91 2.51 23.30 0.32 1.41
Multi-strategy 0.55 1.60 –1.98 8.06 – 1.23 –
Multi-strategy TEV 0.23 1.53 0.06 3.20 53.88 0.53 1.11
Multi-strategy MAD 0.31 1.50 –0.45 2.22 48.52 0.73 1.02
is obtained by the MAD model when compared to the TEV model.

We also see that linear correlation with the original index is similar
or higher for the MAD model in almost all cases, except for Global
Macro and the Managed Futures index. We also observe that the aver-
age leverage, in almost all cases, is lower for the MAD model than the
Style Analysis model. In Table 10.4 we estimate the model with a longer
in-sample estimation period of 48 months. For the mean, standard devia-
tion, correlation, and Sharpe ratio the overall picture does not change
much, demonstrating the robustness of the previous results. Finally we
see that for both in-sample periods, only in the case of a market neutral
strategy do both the MAD and TEV models show better performance
than the original index. This can be traced back to the fact that this
strategy performed very badly during the credit crisis of 2008, losing
more than 40 percent in November.
10.6 Conclusion
Replication of hedge fund strategies is a challenging task given that we

only possess monthly data and we do not know exactly where the man-
ager invests. This paper investigates the behavior of the Mean Absolute
Deviation Model for the replication of hedge fund returns. The definition
of tracking error as the absolute deviation of the portfolio from a bench-
mark provides us not only with the advantage of a better perception of
the risk incurred in the replication process, but also demonstrates that it
works better than style analysis when applied to the mean absolute devia-
tion linear model. In almost all cases, the mean absolute deviation model
displays higher risk-adjusted return and higher correlation than the style
analysis model. Despite that, clones obtained from the mean absolute
deviation and style analysis models both show, in general, lower perform-
ance and higher volatility than the original indexes. In addition, we find
that the mean absolute deviation model represents an improvement over
usual quadratic programming.
Notes
1. A paper following the returns of hedge fund clones is Wallerstein, Tuchschmid,
and Zaker (2010), that analyze the first five year period of real returns
achieved by investment banks and asset manager firms.
2. Sharpe (1971) proposes a linear programming approximation.
3. This is proved by Jensen’s inequality.
4. Fung and Hsieh (2002, p. 25) noted that “TASS was constructed with the
purpose of being investable while HFRI is designed to be a proxy of the hedge
fund industry’s performance.”
5. In order to understand the categorization of CSFB Tremont, go to http://www.

hedgeindex.com/.
6. The Jarque–Bera normality test (not reported) is rejected in all cases except
Managed Futures.
7. Notice that contrarily to Lo et al. (2007) we do not use credit spread, but simply
a credit factor. VIX, which was excluded three years ago, is today much more
liquid.
8. Notice that Best Linear Unbiased Estimator (BLUE) properties are lost as soon
as we impose some restrictions.
9. No constraint on negativity of the weights should be applied as hedge funds
are free to short sell.
References
Amenc, N., Martellini, L., Meyfredi, J.-C., and Ziemann, V. (2008). “Passive
Hedge Fund Replication – Beyond the Linear Case.” EDHEC Risk and Asset
Management Research Centre, Working Paper. EDHEC Business School, Nice,
France.
Agarwal, V., and Naik, N. Y. (2000a). “Generalised Style Analysis of Hedge
Funds,” Journal of Asset Management, 1(1): 93–109.
Agarwal, V., and Naik, N. Y. (2004). “Risk and Portfolio Decisions Involving
Hedge Funds.” Review of Financial Studies, 17(1): 63–98.
Clarke, R. C., Krase, S., and Statman, M. (1994). “Tracking Errors, regret and tacti-
cal asset allocation.” The Journal of Portfolio Management, 20(3): 16–24.
Duarte, J., Longstaff, F. A., and Yu, F. (2007). “Risk and Return in Fixed-Income
Arbitrage: Nickels in Front of a Steamroller.” The Review of Financial Studies,
20(3): 769–811.
Fung, W., and Hsieh, D. A. (1997). “Empirical Characteristics of Dynamic
Trading Strategies: the Case of Hedge Funds,” Review of Financial Studies, 10(2):
275–302.
Fung, W., and Hsieh, D. A. (2002). “Benchmarks of Hedge Fund Performance:
Information Content and Measurement Biases.” Financial Analyst Journal,
58(1): 22–34.
Fung, W., and Hsieh, D. A. (2002). “The Risk in Fixed-Income Hedge Fund
Styles.” Journal of Fixed Income, 12(2): 16–27.
Fung, W., and Hsieh, D. A. (2004). “Hedge Fund Benchmarks: A Risk Based
Approach,” Financial Analysts Journal, 60(5): 65–80.
Fung, W., and Hsieh, D. A. (2007). “Will Hedge Funds Regress towards Index-like
Products?” The Journal of Investment Management, 5(2): 46–65.
Goldstein, D. G., and Taleb, N. N. (2007). “We Don’t Quite Know What We Are
Talking About When We Talk About Volatility.” Journal of Portfolio Management,
33(4): 84–6.
Hasanhodzic, J., and Lo, A. W. (2007). “Can Hedge-Fund Returns Be Replicated?:
Kat, H. M., and Palaro, H. P. (2006a). “Replication and Evaluation of Fund of
Hedge Funds Returns.” Alternative Investment Research Centre, Working
Paper, University of Reading, CASS Business School, Reading, UK.
Kat, H. M., and Palaro, H. P. (2006b). “Superstars or Avarage Joes? A Replication-

Based Performance Evaluation of 1917 Individual Hedge Funds.” Alternative
Investment Research Centre Working Paper, University of Reading, CASS
Business School, Reading, UK.
Konno, H., and Yamazaki, H. (1991). “Mean Absolute Deviation Portfolio
Optimization Model and its Application to the Tokyo Stock Market.” Management
Science, 37(5): 519–31.
Mitchell, M. L., and Pulvino, T. C. (2001). “Characteristics of Risk and Return in
Risk Arbitrage.” Journal of Finance, 56(6): 2135–75.
Roll, R. (1992). “A Mean/Variance Analysis of Tracking Error.” The Journal of
Portfolio Management, 18(4): 13–22.
Roncalli, T., and Teïletche, J. (2008). “An Alternative Approach to Alternative
Beta.” Journal of Financial Transformation, 24(2008): 43–52.
Rudolf, M., Wolter, H. J., and Zimmermann, H. (1999). “A Linear Model for
Tracking Error Minimization.” Journal of Banking and Finance, 23(1): 85–103.
Sharpe, W. F. (1971). “A Linear Programming Approximation for the General
Portfolio Analysis Problem.” Journal of Financial and Quantitative Analysis, 6(5):
1263–75.
Sharpe, W. F. (1992). “Asset Allocation: Management Style and Performance
Measurement.” Journal of Portfolio Management, 18(2): 7–19.
Speranza, M. G. (1993). “Linear Programming Models for Portfolio Optimization.”
Finance 14(1993): 107–23.
Wallerstein, E., Tuchschmid, N. S., and Zaker, S. (2010). “How Do Hedge Fund
Clones Manage the Real World?” The Journal of Alternative Investments, 12(3):
37–50.
11
Can Hedge Fund-Like Returns
be Replicated in a Regulated
Environment?
Iliya Markov and Nils S. Tuchschmid
11.1 Introduction
The European Union’s (EU) directive for Undertakings for Collective

Investment in Transferable Securities (UCITS) is a regulatory frame-
work that permits the replication of certain hedge fund-like investment
strategies. In this chapter we present a summary of the main aspects of
the framework, offer an analysis of their advantages and disadvantages,
and examine whether the regulations impose significant restrictions on
the investment opportunities of alternative UCITS funds, that is, those
UCITS funds that actually employ hedge fund-like investment strategies.
Additionally, we discuss swap arrangements, which allow alternative
UCITS funds to replicate hedge fund strategies which are not feasible in
the current regulatory environment. In those cases we examine the new
risks that appear as a result of these replication strategies.
As Tuchschmid, Wallerstein, and Zanolin (2010) point out, different
aspects of the UCITS framework affect the applicability of hedge fund-
like strategies and hence return in different ways. Firstly, the restriction
on the level of leverage of alternative UCITS funds leads to a different
risk–return profile compared to hedge funds. Secondly, the limita-
tions on eligible investment classes results in a different risk exposure.
Thirdly, the enhanced liquidity requirements confine the investment
opportunity set of alternative UCITS to the most liquid instruments,
which should limit losses during market downturns.
11.2 The UCITS regulatory framework and its advantages

and disadvantages
UCITS is a directive implemented by the EU in 1985 with a goal of creating

a single European financial services market by allowing the free EU-wide
146
Iliya Markov and Nils S. Tuchschmid 147
operation of investment funds registered in a particular EU member

state. Its aim is to maintain a high level of investor protection by
imposing constraints concerning the levels of diversification, liquidity,
and the use of leverage (Tuchschmid, Wallerstein, and Zanolin, 2010).
The severe limitations of the initial UCITS framework, however, ham-
pered investment interest. In response, several amendments, informally
known as UCITS III, were adopted in the 2000s (Gruenewald and Weber,
2009). They provided for the smoother cross-border operation of UCITS
funds and broadened the spectrum of eligible investment instruments
(Gruenewald and Weber, 2009).
11.2.1 Fundamentals of the UCITS framework

One central concept of hedge funds is that they are lightly regulated
investment vehicles, which can invest in a broad range of assets and
employ a wide variety of investment strategies. The name UCITS, on
the other hand, entails transferable securities, that is, “liquid, negotiable
instruments, whose risks are adequately captured by risk management,
and whose valuation is regular, accurate and comprehensive” (Beaudoin
and Olivier, 2010). According to the Committee of European Securities
Regulators’ (replaced by the European Securities and Markets Authority
on January 1, 2011) guidelines (2008), shares in companies, govern-
ment and corporate bonds, and most forms of derivatives on bonds
and shares are eligible instruments for UCITS funds. In addition, they
should be liquid and easily tradable.
Tuchschmid, Wallerstein, and Zanolin (2010) note that most jurisdic-
tions do not allow investments in physical commodities or certificates that
represent them. The main exception is the German regulation, which does
allow holdings in commodities certificates. Hedge fund, private equity,
and real estate holdings are also prohibited, although the Luxembourg
regulation does allow UCITS to invest in closed-ended real estate invest-
ment trust (REIT) funds and closed-ended hedge funds. An exemption in
the UCITS directive allows UCITS to hold up to 10 percent, often called
the “trash ratio,” in non-eligible assets. This in practice allows investments
in assets like hedge funds and private equity. Many jurisdictions also allow
investments in commodity and hedge fund indices.
Most jurisdictions have ruled out the use of short selling. In general,
UCITS funds are allowed to achieve short positions synthetically through
derivatives. The benefits of the restrictions on short selling, however,
are questionable. Long/short equity strategies are permissible under the
UCITS framework, but unlike in the traditional hedge fund model, UCITS
funds can only establish short positions via equity swaps or contracts for
148 Hedge Fund Replication in a Regulatory Environment
difference (CFD). According to Beaudoin and Olivier (2010), even though

this setup does not add significant modifications to the basic strategy,
it involves additional operational and counterparty risk consequences.
The limitation of counterparty risk to 10 percent of net asset value (NAV,
explained below) underlines the need for active collateral management.
The collateral posted with the equity swap needs to be regularly read-
justed and client money protection needs to be used for the remaining
collateral (interbank deposit spreads). Furthermore, the authors explain
that under the short selling restrictions strategies such as fixed income
arbitrage are difficult to employ because of the complications involved
in establishing a synthetic short position in non-equity instruments.
Since this strategy also requires a high degree of leverage, it is further
restricted by the leverage requirements outlined below.
France and Ireland do allow limited amounts of short selling
(Tuchschmid, Wallerstein, and Zanolin, 2010). There are, however, addi-
tional rules which require that the short position should be adequately
covered, either by the underlying asset or by an asset which is highly
correlated to the underlying. First of all, an asset’s correlation to the under-
lying may turn out not to be as good as expected. Secondly, the distinction
between naked and conventional short selling may not be economi-
cally justified. Gruenewald, Wagner, and Weber (2010) point out that in
reality clearing agencies have mechanisms of resolving delivery failures by
sourcing the stocks and debiting the short seller’s account. Thus the short
seller is exposed to the same risk of not delivering the stock on time. We
should also mention here the array of negative market-wide consequences
of the short selling restrictions, such as reduction of liquidity and ineffi-
cient price discovery, outlined by Fragnière and Markov (2011).
11.2.2 Risk management provisions

Tuchschmid, Wallerstein, and Zanolin (2010) explain that the most signifi-
cant risk management requirement for UCITS from an organizational
point of view is that the risk management team should be independent
of the portfolio management team. UCITS are also required to monitor
and measure the risk of their positions and how they contribute to the
overall risk profile of the portfolio. Many implementations of the UCITS
directive center the risk management regulations on the value at risk
(VaR) measure. Most jurisdictions require that the 99 percent monthly
VaR should not exceed 20 percent of NAV.
UCITS regulations give particular emphasis on the use of leverage,
liquidity, concentration risk, and counterparty risk. In terms of leverage,
UCITS funds are only allowed to access leverage by use of derivatives
(Tuchschmid, Wallerstein, and Zanolin, 2010). Leverage through

borrowing is prohibited. Since leverage through borrowing can amplify
any positive or negative return from an investment instrument, its
prohibition makes alternative UCITS funds’ returns much less volatile.
Research on cross-sectional data for a sample of alternative UCITS funds
and hedge funds shows that for some strategy categories the former are
indeed three to four times less volatile (Tuchschmid, Wallerstein, and
Zanolin, 2010). On the other hand, however, the reduction of leverage
entails a cost for “high octane strategies” (Beaudoin and Olivier, 2010).
As Tuchschmid, Wallerstein, and Zanolin (2010) state, the UCITS
directive also stipulates an array of rules concerning concentration and
counterparty risk. These rules require that exposure to any security
or money market instruments by the same issuer should not exceed 10
percent of NAV, and in combination with derivatives it should not exceed
20 percent of NAV. UCITS funds are also required to limit any individual
OTC derivative transaction to 10 percent of NAV if the counterparty is
a credit institution. The total exposure on all transactions towards one
issuer should not exceed 20 percent of NAV. There are exceptions, how-
ever, where UCITS funds can net their positions in OTC derivatives.
Limiting the amount of exposure to any given security or counterparty
effectively amounts to requiring alternative UCITS funds to be more
diversified than traditional hedge funds. Increased diversification reduces
volatility, which should make alternative UCITS funds safer investments.
On the other hand, similar to the restrictions on leverage, stricter diver-
sification requirements prevent alternative UCITS from taking full advan-
tage of booming segments of the market. Beaudoin and Olivier (2010)
point out that event-driven strategies, for example, are seriously affected
by the diversification requirements as they presuppose significantly more
concentrated portfolios. Other strategies, which also rely on concentrated
portfolios, such as global macro and fixed income, may also experience
difficulties under the UCITS framework. Fixed income arbitrage involves
highly leveraged positions, which also make it difficult to comply with
the diversification requirements (Beaudoin and Olivier, 2010). This is in
addition to the restrictions on short selling pertaining to this strategy.
Equity market neutral strategies will also need to be diversified enough to
comply with the UCITS regulation (Beaudoin and Olivier, 2010).
Tuchschmid, Wallerstein, and Zanolin (2010) point out that one of
the most important provisions of the UCITS directive is that it requires
UCITS funds to calculate their liquidity risk. They have to consider factors
such as bid–ask spread, breadth and depth of the secondary market, etc.
Specifically, they must be able to accommodate redemption requests of
20 percent of NAV at any time. UCITS funds have to offer redemption

facilities at least twice a month. Their maximum exposure to illiquid
assets is limited to 10 percent of NAV as long as they are able to meet
foreseeable redemption requests.
Enhanced liquidity requirements give alternative UCITS funds an edge
during market downturns. It has recently become clear that during finan-
cial crises even markets that were deemed liquid could suddenly freeze up.
In this environment, as Morris and Shin (2003) point out, short horizon
traders such as hedge funds start selling risky assets, thus pushing their
prices further down and creating a liquidity spiral. Sales become mutu-
ally reinforcing and lead to a “liquidity black hole.” This is confirmed by
Hameed, Kang, and Viswanathan’s (2007) research, which finds that neg-
ative market returns are associated with reduced liquidity. In this context,
during times of financial turmoil hedge funds try to shed illiquid assets
in a move labeled by Naes, Skjeltorp, and Ødegaard (2010) as “flight to
liquidity.” In this respect, alternative UCITS funds are likely to experience
lower portfolio restructuring costs as they are required to continuously
monitor their portfolio liquidity in order to meet redemption requests
with minimum delay.
Beaudoin and Olivier (2010) note that the 2008 financial crisis has
highlighted the liquidity problems faced by some hedge funds and their
low level of investor protection. Investors were frustrated with hedge
funds’ handling of the panicked fire sale of illiquid assets. Alternative
UCITS were prevented from investing in attractive thinly traded stocks
or other less liquid instruments such as mortgage-backed securities and
CDOs, which nevertheless offered competitive returns prior to the crisis.
Beaudoin and Olivier (2010) add that, in general, strategies focused on
distressed securities are not eligible due to the illiquid nature of under-
lying securities. Liquidity requirements also bar alternative UCITS from
deep value investing because it needs more time to realize. There are no
formal restrictions concerning convertible arbitrage strategies under the
UCITS framework (Beaudoin and Olivier, 2010). The convertible, how-
ever, may not be priced efficiently due to market illiquidity. Alternative
UCITS therefore have to focus on the most liquid part of the market,
which, despite ensuring a safer investment profile, offers fewer opportu-
nities (Beaudoin and Olivier, 2010).
11.2.3 Circumventing the rules

In quest for higher returns, alternative UCITS funds have started using
strategies to gain access to investment products, which are formally
prohibited under the UCITS framework. Laurent (2010) explains that an
alternative UCITS fund may replicate an offshore hedge fund’s returns

through the use of contracts for difference (CFD), equity swaps, or total
return swaps (TRS). After defining an appropriate index, an alternative
UCITS fund enters into a TRS agreement with an investment bank. The TRS
allows the fund to exchange the cash flows from assets held with the invest-
ment bank on its behalf (depositary bank) with the cash flows generated
from the index. Cua (2010) cites the case of Man AHL’s Singapore UCITS
fund. The latter has entered into a swap agreement with Deutsche Bank,
which in turn manages the fund’s volatility by controlling its exposure
to the index—in this case the AHL Trend Index, which replicates the risk
and return of a highly diversified portfolio of futures and forwards (Peters,
2010). In times of low volatility the exposure to the index can reach 100
percent, while during periods of increased volatility exposure is reduced.
Peters (2010) provides another example of an alternative UCITS fund
to take advantage of this replication strategy—the UK-based Man AHL
Diversity fund. According to Man AHL’s head of UK Distributions John
Bennett, UCITS regulations imposed significant diversification restric-
tions in terms of commodities exposure. Therefore instead of restricting
themselves to the eligible assets, they decided to access a broader plat-
form by tracking the AHL Trend Index. Their strategy complies with the
UCITS regulation of 10 percent maximum exposure to any counterparty
with Deutsche Bank posting eligible assets back to the product to miti-
gate counterparty risk. Man AHL aims for a 100 percent collateralization
of the swap. In the event that Deutsche Bank is financially distressed or
becomes insolvent, Man AHL is insured by the assets posted as collateral.
Another alternative UCITS fund with a similar replication strategy is run
by London-based Winton Capital Management. Their strategy is mostly
trend-following and they also track the performance of their index
through an OTC swap with Deutsche Bank (Deutsche Bank, 2010).
According to Danaher’s (2010) analysis, while technically satisfying the
UCITS III requirements, these replication strategies are outside the realm
of liquidity and transparency endorsed by the UCITS directive. Since
these trading techniques are very complex, many of the funds’ clients
will not be able to understand the risks inherent to them and whether
their returns are commensurate therewith.
Gruenewald and Weber (2009) accentuate the lack transparency of
such replication strategies by giving the example of a Luxembourg-based
alternative UCITS fund, Luxalpha, which became a feeder into Bernard
Madoff’s Ponzi scheme. This ignited an investigation after French nationals
lost their deposits in Luxembourg (Euractiv, 2010). The Madoff scandal,
as Gruenewald and Weber (2009) say, has not only started lawsuits against
a number of depositary banks but also revealed the uncertainties

surrounding EU regulated UCITS markets. According to French Finance
Minister Christine Lagarde, UCITS funds based in Luxembourg take advan-
tage of the slacker legislation, which, unlike in the case of France, does
not require them to return the full amount of funds (Euractiv, 2010). She
stressed the need for a more harmonized UCITS regulation across the EU
member states, especially with regard to the level of liability borne by
depositary banks (Gruenewald and Weber, 2009).
Danaher (2010) outlines two main categories of risk as far as swaps are
concerned. The first one is that as TRS strategies become more prolific,
regulators may start to step in and prohibit them. Secondly, when assets
in the UCITS portfolio perform better or worse than those held with the
bank, this creates a temporary asset or liability with the investment bank.
Even though it does not change the UCITS NAV, the UCITS fund will end
up owing money to the bank or vice versa. In a potential failure of the
bank, as in the example of Lehman Brothers, this creates an outstanding
liability at the bank. We should mention once again here the need for
active collateral management due to the 10 percent ceiling on counter-
party risk, which further complicates the swap arrangement.
From an alternative UCITS fund’s perspective there is another important
operational risk—the so-called tracking error. Fieldhouse and McIntosh
(2010) present the case of BlueTrend, an alternative UCITS fund operated by
BlueCrest—a Guernsey-based alternative investments company. BlueTrend
was discontinued due to tracking error that was higher than expected.
Tracking error in this regard is the difference between the expected per-
formance of an index and the realized returns from the replication strategy
compliant with the UCITS framework. While tracking error that leads to
improved performance is welcome, tracking error that leads to worse than
expected performance leaves investors disgruntled. Leda Braga, BlueTrend’s
president and head of systematic trading, pinpointed three main causes for
the higher than expected tracking error: the frequent liquidity provision
and the portfolio constraints stipulated by the UCITS framework and the
fact that offshore hedge funds have more latitude in terms of investment
classes. Fieldhouse and McIntosh (2010) suspect that BlueTrend will not
remain the only UCITS fund to close due to tracking error.
11.3 Comparison of the risk–return profiles of alternative

UCITS funds and hedge funds
Similar to hedge funds, most alternative UCITS funds specialize in

specific investment strategies. Table 11.1 lists some statistics on alternative
Table 11.1 Strategy breakdown of alternative UCITS funds at the end of March
2011
No. of Share of total AUM Share of total

funds no. of funds (%) (bn EUR) AUM (%)
CTA 35 5.62 4.04 3.69
Macro 148 23.76 23.92 21.84
Long/short equity 164 26.32 19.54 17.85
Equity market neutral 46 7.38 5.30 4.84
Event-driven 15 2.41 1.70 1.55
Fixed income 77 12.36 33.94 31.00
Emerging markets 49 7.87 6.41 5.86
Commodities 19 3.05 1.30 1.19
FX 42 6.74 3.02 2.76
Multi-strategy 28 4.49 10.32 9.43
Fund of funds 63 – 2.80 –
Source: UCITS Alternative Index, 2011, Alix Capital.
UCITS funds by strategy as of the end of March 2011. The table shows
that macro, long/short equity and fixed income are the predominant
strategies for alternative UCITS funds and constitute as much as
71 percent in terms of AUM. On the other hand, the table shows that
due to reasons explained above, strategies such as event-driven, equity
market neutral and commodities are not very popular. CTA strategies
are also relatively unpopular even though CTAs generally invest in
liquid derivatives and have few additional restrictions such as commodity
derivatives with physical delivery (Beaudoin and Olivier, 2010). On the
other hand, the number of long/short equity funds is large, even though
most jurisdictions restrict non-derivative short positions. As shall be
seen, alternative UCITS funds employing this strategy do not seem to
perform better than hedge funds.
Table 11.2 provides data on the annualized returns of alternative UCITS
funds indices and closely corresponding HFRX indices (Hedge Fund
Research investable indices constructed using UCITS III methodology)
for a number of strategies for the last 12 months and the last 36 months.
The hedge fund indices are expressed in USD while the alternative UCITS
indices are expressed in EUR. Their returns have therefore been computed
using monthly returns in excess of 1-month Libor rates for USD and EUR.
Thus, thanks to the covered interest parity both series are expressed as
hedged against currency variations and are hence comparable.
We can see that the performance of hedge funds is more varied than
that of alternative UCITS funds, especially during the last 12 months.
In terms of relative performance, for the last 12 months hedge funds
Table 11.2 Performance of alternative UCITS funds and hedge funds at the end
of March 2011
Alternative UCITS % Hedge Fund HFRX %

returns returns
Last 12 Last 36 months Last 12 Last 36 months
months (annualized) months (annualized)
Global index 0.03% –0.07% 3.63% –2.67%
CTA –1.14% 0.11% 1.27% 2.57%
Macro –0.66% 0.12% –1.16% –4.81%
Long/short equity 0.38% –0.29% 4.94% –3.00%
Equity market –1.38% –1.98% 4.18% –1.37%
neutral
Event driven 4.49% 0.04% 1.82% –1.54%
Fixed income 2.13% 1.40% 6.96% 0.96%
Emerging markets –0.06% 4.93% 17.11% 4.33%
Commodities 1.94% –4.58% –0.50% –2.17%
FX –1.24% –1.21% –1.39% –0.24%
Multi-strategy –1.29% –0.86% 6.55% 5.26%
Source: UCITS Alternative Index, 2011, Alix Capital; HFRX, 2011, Hedge Fund Research.
outperform alternative UCITS funds in six out of the ten strategy classes
and the global index by a large margin. For the last 36 months, alterna-
tive UCITS perform better than hedge funds in five strategy classes and
worse in the remaining five classes. Macro, long/short equity, and fixed
income alternative UCITS funds outperform hedge funds of the same
strategies. Since they represent more than two thirds of the total AUM
of the alternative UCITS industry, the performance of alternative UCITS
funds as expressed by the global index is significantly higher during the
last 36 months.
It was mentioned above that the stricter risk management process
and the enhanced liquidity requirements are likely to have a positive
effect on the returns of alternative UCITS funds during times of financial
turmoil. This is also evidenced by Table 11.2. Nevertheless, since periods
of massive financial distress have a relatively low probability of occur-
rence, we should view the results with care. The results for the last 12
months show that once the worst of the crisis was over, hedge funds
started to perform much better on average.
11.3.1 Risk return profile

Tuchschmid, Wallerstein, and Zanolin (2010) conduct a more detailed
analysis on the cross-sectional performance of alternative UCITS funds
and hedge funds. They study the mean performance and dispersion of
a sample of alternative UCITS funds belonging to three strategy groups,
namely long/short equity, global macro, and fixed income, as those are
the strategies employed by the majority of alternative UCITS funds. The
results are compared to those of hedge funds with a similar strategy and
statistical tests for the significance of the differences are performed. The
study period covers December 2006–July 2009 and only funds that cover
the whole study period are considered. Only euro-denominated share
classes net of fees are considered. The empirical results that the authors
obtain are important in the light of the preceding discussion.
In line with Table 11.2, Tuchschmid, Wallerstein, and Zanolin’s (2010)
results show that significant differences exist with regard to strategy class.
Long/short equity alternative UCITS have an annualized mean return of
1.01 percent, while hedge funds with the same strategy have an annual-
ized mean return of 0.98 percent. The equal return hypothesis cannot
be rejected in this case because the difference of 0.03 percent annually
is very small. In the global macro case the results are more ambiguous
because Kolmogorov’s two-sided test rejects the hypothesis of equal
distribution at the 95 percent confidence level while Wilcoxon’s test does
not. Nevertheless, global macro hedge funds have a higher annualized
mean return (3.87%) compared that of alternative UCITS funds (1.50%).
For the fixed income strategy, Kolmogorov’s two-sided test rejects
the hypothesis of equal distribution at the 95 percent confidence level.
Alternative UCITS funds have an annualized mean return of 2.47 percent,
while hedge funds have a negative return as low as –11.97 percent. The
sample of fixed income hedge funds, however, contains four negative
return outliers, whose removal boosts the annualized mean return to
–0.14 percent.
As Tuchschmid, Wallerstein, and Zanolin (2010) point out, a possible
explanation of the large difference in average performance is the liquidity
requirements and the restrictions on eligible assets for alternative UCITS
funds. The study period contains particularly large negative returns on
illiquid assets such as mortgage-backed securities and other structured
products during the financial crisis. These were held mostly by hedge
funds due to the strict liquidity requirements for UCITS funds. Table 11.2
above shows that in the more recent sample period, the performance of
fixed income hedge funds has improved whereas that of macro funds has
worsened compared to alternative UCITS funds. The empirical results of
Tuchschmid, Wallerstein, and Zanolin (2010) confirm the previous dis-
cussion that the restrictions imposed by the UCITS framework do affect
return. They have a serious impact on the replicability of certain hedge
fund strategies, which is evidenced by the lower return of global macro

alternative UCITS funds.
11.3.2 Dispersion of risk and return

The comparison of the cross-sectional means, however, can distort the
perception of relative performance as the mean is rarely achieved by
investors. Tuchschmid, Wallerstein, and Zanolin (2010) present the cross-
sectional dispersion in terms of the 10th and 90th percentiles of the annu-
alized mean return and standard deviation. In all cases, hedge funds show
higher levels of dispersion. For the group of long/short equity funds, for
both measures, the hypothesis that alternative UCITS and hedge funds
have equal dispersion is rejected at the 95 percent confidence level. For the
global macro strategy the results for the mean are more comparable. Yet,
the difference in the dispersions of the standard deviations is sizable.
The dispersions of the annualized mean return and standard deviation
for fixed income differ by a factor of 4 to 5, with alternative UCITS funds
exhibiting much lower levels of dispersion. This result is highly signifi-
cant as the hypothesis of equal dispersion is rejected at the 99.9 percent
confidence level. The results corroborate the previous discussion that
under the UCITS framework, alternative UCITS funds have lower volatility
and form a more homogeneous group compared to hedge funds.
11.4 Conclusion
This chapter presents an overview of the advantages and disadvantages

of the UCITS framework and its suitability for the replication of hedge
fund returns. We find that alternative UCITS funds exhibit lower volatility
and dispersion of return and provide better liquidity terms. Yet, as the
empirical results have shown, none of these characteristics can guarantee
high performance. The provisions of the EU directive on UCITS restrict
various hedge fund-like strategies such as investment in commodities,
distressed securities, and fixed income arbitrage and hinder the efficient
execution of various others, such as convertible arbitrage, event-driven,
and global macro.
In fact, liquidity regulations create a tradeoff between alternative UCITS
funds and hedge funds. They can certainly explain the differences in expo-
sure between the two categories of investment vehicles, with hedge funds
for instance being more exposed to credit and small cap risks (Tuchschmid,
Wallerstein, and Zanolin, 2010). Nevertheless, while frequent liquidity
redemptions may be beneficial for clients, they also prohibit exposure to
the risk and return of holding illiquid assets (Tuchschmid, Wallerstein,
and Zanolin, 2010). Moreover, the requirements on liquidity provision

and the increased risk management requirements for alternative UCITS
funds impose many additional operational costs.
Many alternative UCITS managers also find the restrictions on com-
modities exposure too confining. Contracts for difference and total return
swaps may offer a solution to this end by allowing alternative UCITS
funds to replicate certain hedge fund strategies. Their lack of transparency,
however, has sparked a row and the future of their appropriateness under
the UCITS framework remains to be seen. Moreover, swap arrangements
expose alternative UCITS funds to an array of operational and counter-
party risks that have not been rigorously analyzed. In conclusion, as
Fieldhouse and McIntosh (2010) point out, alternative UCITS funds may
not be suited to replicating all hedge fund-like strategies. Even though
there are some who believe that 95 percent of hedge fund strategies are
replicable under the UCITS framework, the number of skeptics is large.
References
1-Month EUR and USD Libor Rates. (2011). British Bankers Association. Available
at: http://www.global-rates.com/. Accessed: 18 April 2011.
Beaudoin, D., and Olivier, C. (2010). “UCITS III Funds: One Size Does Not Fit
All.” Swiss Hedge Magazine 2nd Half 2010. Available at: http://www.swisshed
gemagazine.ch/cms/zeigeBereich/58/zeigeBereich/59/2nd-half.html. Accessed:
11 April 2011.
The Committee of European Securities Regulators. (2008). “CESR/07-044b:
CESR’s Guidelines Concerning Eligible Assets For Investment By UCITS.”
Available at: http://www.esma.europa.eu/popup2.php?id=5280. Accessed: 18
March 2011.
Cua, G. (2010). “Managed Futures Fund For Retail Mkt.” Available at: http://
www.cpf.gov.sg/imsavvy/infohub_article.asp?readid%7B61734315-6929-
1521490216%7D. Accessed: 14 April 2011.
Danaher, S. (2010). “Total Return Swaps Expose UCITS Investors To Unknown
Risks.” Portfolio Adviser. Available at: http://www.portfolio-adviser.com/article/
total-return-swaps-expose-UCITS-investors-to-unknown-risks. Accessed: 14 April
2011.
Deutsche Bank. (2010). “DBX Systematic Alpha Index Fund.” Available at: http://
globaldocuments.morningstar.com/documentlibrary/Document/e036e4d523f
5677a9bca2b47cbd68140.msdoc/original. Accessed: 14 April 2011.
Euractiv. (2010). “Madoff Scandal Sparks EU Row.” Available at: http://www.
euractiv.com/en/financial-services/madoff-scandal-sparks-eu-row/article-178659.
Accessed: 16 April 2011.
Fieldhouse, S., and McIntosh, B. (2010). “Life After BlueTrend: Can CTA
Strategies Really Make The Grade In The UCITS World?” UCITS Hedge. Available
at: http://www.UCITShedgefunds.com/issue/201011/life-after-bluetrend.php.
Fragnière, E. and Markov, I. (2011). “Short Selling In France During The Crisis, The
Bans And What Has Changed Since The Euro Correction.” In: Greg N. Gregoriou
(ed.). Handbook of Short Selling. Elsevier, Burlington, MA, forthcoming.
Gruenewald, S., Wagner, A., and Weber, R. (2010). “Short Selling Regulation
After The Financial Crisis – First Principles Revisited.” International Journal of
Disclosure and Regulation, 7(2): 108–35; Swiss Finance Institute Research Paper,
No. 09–28. Available at: http://www.ssrn.com/abstract=1439652. Accessed: 8
April 2011.
Gruenewald, S., and Weber, R. (2009) “UCITS And The Madoff Scandal:
Liability Of Depositary Banks?” Butterworths Journal of International Banking
and Financial Law. Available at: http://www.ufsp.uzh.ch/finance/documents/
WeberGruenewald_UCITSdepositaries.pdf. Accessed: 16 April 2011.
Hameed, A., Kang, W., and Viswanathan, S. (2007). “Stock Market Declines
And Liquidity.” Journal of Finance, Forthcoming; AFA 2007 Chicago Meetings
Paper; EFA 2007 Ljubljana Meetings Paper. Available at: http://www.ssrn.com/
abstract=889241. Accessed: 9 April 2011.
HFRX. (2011). Hedge Fund Research. Available at: http://www.hedgefundre
search.com. Accessed: 18 April 2011.
Laurent, O. (2010). “Operational Risk In Alternative Investment Funds:
Understanding The UCITS Framework Takes Time.” UCITS Hedge. Available at:
http://www.UCITShedgefunds.com/issue/201007/operational-risk-in-alternative-
investment-funds.php. Accessed: 14 April 2011.
Morris, S., and Shin, H. S. (2003). “Liquidity Black Holes.” Cowles Foundation
Discussion Paper, No. 1434. Available at: http://www.ssrn.com/abstract=446600.
Naes, R., Skjeltorp, J. A., and Ødegaard, B. A. (2010). “Stock Market Liquidity And
The Business Cycle.” Journal of Finance, forthcoming. Available at: http://www.
ssrn.com/abstract=1158408. Accessed: 8 April 2011.
Peters, N. (2010). “Man AHL Tweaks UCITS Formula: The Man AHL Diversity
Offering.” UCITS Hedge. Available at: http://www.ucitshedgefunds.com/
issue/201003/man-ahl-tweaks-ucits-formula.php. Accessed: 13 April 2011.
Tuchschmid, N., Wallerstein, E., and Zanolin, L. (2010). “Will Alternative UCITS
Ever Be Loved Enough To Replace Hedge Funds?” Working Paper, Haute école
de gestion de Genève, Carouge, Switzerland. Available at: http://www.ssrn.
com/abstract=1686055. Accessed: 4 April 2011.
UCITS Alternative Index. (2011) Alix Capital. Available at: http://www.UCITS-
alternative.com/. Accessed: 18 April 2011.
12
A Factor-Based Application to
12.1 Introduction
Hedge fund returns are generally considered to be little correlated with

market returns. Skills and dynamic strategies are claimed to generate
more complex risk exposures that yield superior performance (alpha) or
complementary sources of risk premium (alternative beta) through bear
and bull markets by using a broad range of instruments, such as deriva-
tives, leverage, short selling, and arbitrage across markets. This market
neutrality feature of hedge funds would suggest that investing in hedge
funds, either directly or through funds of hedge funds, could be an effec-
tive tool of portfolio diversification, hence making it appealing for a large
range of institutional investors and high-wealth individuals.1 However,
hedge funds (1) provide limited liquidity, as resources are usually “locked
up” for 1–3 years; (2) impose high management fees (up to 5 percent a
year); and (3) offer poor transparency.
Against this backdrop, hedge fund replication has been proposed as
an alternative to investing in hedge funds while potentially obtain-
ing hedge fund-like returns. While these replication strategies may not
generate the same alpha (and/or alternative beta) as the original hedge
fund strategy, they may achieve pay-offs that are close enough in net
terms, that is, when liquidity, costs, and transparency considerations are
taken into account.
Three main approaches to hedge fund replication have been followed
so far.2 One is to duplicate mechanically hedge fund strategies in a system-
atic and quantitative fashion. Another is postulated on the assumption
that investors are interested in hedge funds because of their risk–return
characteristics (including volatility and correlation) rather than their
month-on-month returns; futures trading rules are then designed to
159
160 A Factor-Based Application to Hedge Fund Replication
generate returns with properties similar to those of hedge funds or hedge

fund indices. Because of their intent to replicate complicated hedge
fund strategies, these two approaches have been criticized for exposing
investors to operational risks and high operating costs. A third approach,
also referred to as the factor model approach, consists in estimating the
sensitivity of hedge fund returns to a series of risk factors. A portfolio of
stocks, bonds, and other securities can then be constructed with the same
month-on-month returns as the hedge fund or hedge fund index to be
replicated. Despite its straightforward appeal, the factor model approach
presents a number of drawbacks, such as missing variables (or factors),
linearity and normality assumptions, and lack of dynamic trading, which
is at the core of hedge fund strategies.
This chapter estimates a factor-based model using data from the Hedge
Fund Research (HFR) database. The focus is on replicating hedge fund
performance across strategies and time, specifically, around times of height-
ened market volatility during recent episodes of global financial distress.
12.2 Methodology
The relationship between the returns on hedge fund strategies (Rt) and
the returns on individual factors (Fit) at time t can be expressed as:
m
Rt ∑w F
i1
i i
t t i 1,… , m and t 1,… , T (12.1)
where wti is the weight of the ith individual strategy, which identifies a
time-varying allocation across strategies. These weights correspond to
the local beta of a portfolio of assets that proxy for the individual factors
and can be recovered by regressing individual hedge fund returns on the
risk factors.
The specification is similar to that presented in Ennis and Sebastian
(2003), Capocci and Hübner (2004), and Hasanhodzic and Lo (2006):
Rti αi + βi1 F1t … βim Fmt εit (12.2)
where ai and bi are the intercept and slope of the regression and eit is the
error term.
Table 12.1 reports the list of factors. These are assumed to cover the
various sources of market risk: stock market, bond market, currencies,
commodities, credit, and volatility. These factor returns can be realized
through relatively liquid instruments so that the returns on hedge fund
Marco Rossi and Sergio L. Rodríguez 161
Table 12.1 Variable definitions
Risk factors
S&P 500 Index
It is a capitalization-weighted index of 500 stocks maintained by the S&P Index
Committee. It includes 500 leading companies in leading industries of the
US economy, capturing 75 percent coverage of US equities. See http://www.
standardandpoors.com/home/en/us for further details.
USD Exchange Rate Spot Index
The index is calculated as a geometric average of exchange rates in a basket of
six major currencies. The currencies and their weights are determined by the
New York Board of Trade. Source: Bloomberg.
Moody’s Corporate AAA Bond Yields
Yields derived from pricing data on corporate bonds in the US market with
current outstanding amounts over US$100 million. The bonds have maturities
as close as possible to 30 years. Source: Bloomberg.
Credit Spread
The difference, in percentage points, between relatively high-risk corporate
yields (Moody’s Corporate BAA Bond Yields) and the US Generic Government
30 year yield computed by Bloomberg.
S&P Commodity Index
This index, maintained by the S&P Index Committee, is calculated primarily on
a world production-weighted basis and is comprised of the principal physical
commodities that are the subject of active, liquid futures markets. See http://
www.standardandpoors.com/indices/sp-gsci/en/us/.
JPM EMBI Plus
Constructed by JP Morgan, EMBI Plus tracks returns for actively traded external
debt instruments in emerging markets. It includes US dollar-denominated
Brady bonds, Eurobonds, and traded loans issued by sovereign entities. See
http://www.jpmorgan.com/pages/jpmorgan/investbk/solutions/research/EMBI.
VIX
The CBOE Volatility Index (VIX) measures expectations of near-term volatility
conveyed by the S&P 500 stock index option prices, with the price of each
option reflecting the market’s expectation of future volatility. See http://www.
cboe.com/micro/VIX/vixintro.aspx.
Hedge funds’ investment strategies3
Equity Hedge
Equity Hedge strategies maintain positions both long and short in primarily
equity and equity derivative securities; managers would typically maintain at
least 50 percent exposure to, and may in some cases be entirely invested in,
equities, both long and short.
Event-Driven
Investment managers maintain positions in companies currently or
prospectively involved in corporate transactions of a wide variety including
(continued)

but not limited to mergers, restructurings, financial distress, tender offers,
shareholder buybacks, debt exchanges, security issuance, or other capital
structure adjustments. Security types can range from most senior in the capital
structure to most junior or subordinated, and frequently involve additional
derivative securities.
Macro
Investment managers trade a broad range of strategies in which the investment
process is predicated on movements in underlying economic variables and the
impact these have on equity, fixed income, hard currency, and commodity
markets.
Relative Value
Investment positions are maintained based on an investment strategy
predicated on realization of a valuation discrepancy in the relationship
between multiple securities; security types range broadly across equity, fixed
income, derivative, or other security types.
Fund of Funds
Fund of Funds invest with multiple managers through funds or managed
accounts. The strategy designs a diversified portfolio of managers with the
objective of significantly lowering the risk (volatility) of investing with an
individual manager. A manager may allocate funds to numerous managers
within a single strategy, or with numerous managers in multiple strategies.
replication strategies can be achieved in practice. In particular, there are

forward contracts for each of the component currencies of the US dollar
index and future contract for the stock and bond indexes and for the
component of the commodity index.
12.3 Data and Some stylized facts
The HFR Database reports monthly returns on approximately 6,700 hedge

funds from 1995 to 2010. Hedge funds are categorized in five different
strategies and 35 sub-strategies, reflecting a wide range of risk–return
combinations; data availability, however, varies across hedge funds. To
limit survivorship bias, the study focuses only on hedge funds with data
available from January 1995 to December 2010. This reduces the number
of hedge funds to 332; instant return history bias remains.4 Table 12.1
provides also for a description of the investment strategies included in
this study.
To assess hedge fund performance during periods of financial distress,
returns are estimated during three periods in which the level of financial
assistance provided by the International Monetary Fund (Figure 12.1)
rapidly expanded. These are: Period I (April 1997 to May 1999), Period II
80
Period I Period II Period III
60
40
20
0
1995 2000 2005 2010
Time
Figure 12.1 IMF credit outstanding (in SDR billions)

Source: International Financial Statistics, IMF.
(November 2000 to September 2003), and Period III (from March 2008
to December 2010).
Basic hedge fund performance is summarized in Table 12.2. Panel A
reports the number of funds included in each strategy as well as summary
statistics of annualized monthly returns for the full sample and crisis
periods. For the full sample, the highest median return was produced by
event-driven strategies (12.1 percent), while the lowest median return was
generated by macro strategies (8.2 percent). Mean-adjusted volatility—
measured by the standard deviation to mean ratio—differs significantly
across strategies, from almost 8 percent for macro strategies to 4 percent
for fund of funds strategies. Across strategies, median returns fall across
periods of financial distress, except for the macro strategy, which reports
the lowest returns during Period III. In terms of mean-adjusted volatility,
Period III appears clearly the most volatile period, with ratios ranging
from –103 for fund of funds to 42 for equity hedge strategies. Panel B
shows that the best median return occurred in 1995, while the worst was
in 2008; the most volatile years are 1998, 2002, and 2010 with standard
deviation to mean ratios of 23, 16, and –6, respectively. In general, based
on this metric, the late 1990s to early 2000s period looks more volatile
than other years in the sample.
Performance of risk factors is reported in Table 12.3 both across crisis
periods (Panel A) and over time (Panel B). Panel A lists summary statistics
Table 12.2 Hedge funds’ annualized monthly returns, percent 164
Panel A
Returns
Hedge Fund Number of Std. Dev./
Strategy Funds Mean Std. Dev. Median 75th Percentile 25th Percentile Mean
Full Sample: 1995–2010
Equity Hedge 106 12.8 94.5 12.1 52.9 –16.4 7.4
Event-Driven 31 11.6 48.3 11.5 26.7 0.2 4.2
Fund of Funds 106 9.0 38.9 9.8 23.7 –3.4 4.3
Macro 62 13.9 103.0 8.2 52.3 –18.7 7.4
Relative Value 27 10.0 39.7 9.5 21.6 1.9 4.0
Period I: April 1997–May 1999
Equity Hedge 106 15.6 124.0 16.4 76.3 –20.3 7.9
Event-Driven 31 16.8 78.0 16.2 34.8 1.4 4.6
Fund of Funds 106 9.5 48.4 10.7 32.5 –5.3 5.1
Macro 62 11.6 131.0 7.8 47.6 –24.5 11.3
Relative Value 27 9.7 67.2 11.7 23.3 4.7 7.0
Period II: Nov. 2000–Sep. 2003
Equity Hedge 106 8.1 89.5 7.2 47.0 –21.3 11.1
Event-Driven 31 9.4 46.4 7.6 20.8 –2.4 4.9
Fund of Funds 106 7.1 31.2 6.4 15.7 –2.1 4.4
Macro 62 17.1 102.4 9.3 65.2 –17.9 6.0
Relative Value 27 10.2 30.5 9.1 20.3 1.7 3.0
Period III: March 2008–Dec. 2010
Equity Hedge 106 2.7 111.2 6.2 52.9 –27.3 41.5
Event-Driven 31 3.7 63.0 8.9 27.7 –9.0 17.1
Fund of Funds 106 –0.4 43.3 4.9 19.3 –12.0 –103.0
Macro 62 7.3 93.8 3.9 37.7 –18.5 12.8
Relative Value 27 4.6 62.6 6.5 24.6 –2.4 13.6
Panel B
All Hedge Funds
1995 21.4 68.8 17.1 40.8 0.2 3.2
1996 21.2 75.5 17.4 42.6 1.2 3.6
1997 20.3 72.3 15.9 47.8 –3.4 3.6
1998 4.9 114.1 7.7 38.3 –17.0 23.3
1999 24.6 92.8 14.4 50.5 –5.0 3.8
2000 10.7 88.9 9.1 35.1 –14.6 8.3
2001 7.4 71.5 7.4 27.6 –10.0 9.7
2002 3.9 63.4 3.7 19.7 –12.4 16.4
2003 20.5 52.9 12.7 35.9 1.1 2.6
2004 9.5 46.8 7.3 24.9 –5.4 4.9
2005 9.7 45.4 9.6 27.6 –8.6 4.7
2006 11.7 45.3 11.2 28.3 –3.2 3.9
2007 13.0 49.2 12.1 32.9 –4.7 3.8
2008 –16.3 103.5 –9.5 18.4 –35.9 –6.3
2009 18.2 68.9 12.8 37.9 –4.6 3.8
2010 9.6 57.7 8.1 31.4 –8.9 6.0
Source: Authors’ calculations using the HFR database.

165
Table 12.3 Risk factors: annualized monthly returns, percent
166
Panel A
Returns
Risk Factors Mean Std. Dev. Median 75th Percentile 25th Percentile Std. Dev./Mean
S&P 500 7.9 71.5 15.0 53.8 –21.0 9.1
USD Exchange Rate –0.4 32.9 –0.8 16.4 –17.8 –76.5
Moody’s Corp AAA 6.3 0.9 6.2 7.2 5.5 0.2
Credit Spread 2.0 0.8 1.8 2.4 1.4 0.4
S&P Commodity 11.4 113.5 17.2 85.8 –34.5 10.0
JPM EMBI Plus 14.2 60.8 20.7 44.5 –5.7 4.3
VIX 23.8 677.5 –4.9 224.6 –72.7 28.4
S&P 500 30.4 80.1 60.5 97.6 –13.1 2.6
USD Exchange Rate 3.9 26.6 10.3 25.7 –16.7 6.9
Moody’s Corp AAA 6.8 0.4 6.7 7.0 6.5 0.1
Credit Spread 1.6 0.3 1.4 1.9 1.4 0.2
S&P Commodity –8.3 103.6 –13.8 30.7 –47.7 –12.5
JPM EMBI Plus 4.6 134.6 32.9 62.9 –26.2 29.4
VIX 37.5 1,090.9 –47.5 145.8 –73.9 29.1
S&P 500 –10.2 84.5 –10.4 24.0 –53.3 –8.3
USD Exchange Rate –7.2 32.7 –2.9 15.7 –20.7 –4.5
Moody’s Corp AAA 6.5 0.6 6.6 7.1 6.2 0.1
Credit Spread 2.4 0.3 2.4 2.6 2.3 0.1
S&P Commodity 1.9 110.1 –2.9 94.8 –39.1 58.3
JPM EMBI Plus 14.3 53.4 20.2 52.4 –15.2 3.7
VIX 16.0 528.2 –1.8 449.9 –75.8 33.1
S&P 500 0.5 110.2 14.6 86.0 –44.2 205.1
USD Exchange Rate 3.2 49.8 8.7 27.7 –21.0 15.5
Moody’s Corp AAA 5.3 0.4 5.3 5.6 5.1 0.1
Credit Spread 2.8 1.2 2.4 3.2 1.8 0.4
S&P Commodity 2.5 179.4 27.0 111.6 –45.2 70.9
JPM EMBI Plus 9.6 53.6 15.9 28.5 0.7 5.6
VIX 18.0 1,337.5 –42.4 251.3 –84.2 74.1
Panel B
S&P 500 USD Exchange Rate Moody’s Corporate AAA 1/ Credit Spread 2/
Year Mean Std. Dev. Median Mean Std. Dev. Median Mean Std. Dev. Median Mean Std. Dev. Median
1995 34.3 19.3 39.2 –3.7 31.8 –6.8 7.6 0.5 7.5 1.4 0.1 1.4
1996 20.9 42.5 21.2 4.1 17.5 2.9 7.4 0.3 7.4 1.3 0.1 1.4
1997 32.5 67.7 67.7 13.5 32.2 18.1 7.3 0.3 7.3 1.3 0.1 1.3
1998 29.5 99.8 69.0 –5.3 24.9 –5.3 6.5 0.1 6.5 1.7 0.3 1.5
1999 20.5 53.4 40.1 8.3 21.1 9.1 7.0 0.4 7.2 2.0 0.2 1.9
2000 –8.9 74.2 –19.8 8.0 35.8 8.3 7.6 0.2 7.6 2.5 0.3 2.6
2001 –11.4 89.8 –3.4 6.8 27.2 8.8 7.1 0.1 7.1 2.5 0.2 2.5
2002 –21.8 94.4 –19.8 –12.5 29.8 –5.0 6.5 0.2 6.5 2.5 0.1 2.5
2003 27.1 45.0 17.8 –14.4 31.0 –14.2 5.7 0.3 5.7 1.9 0.4 1.8
2004 9.3 26.9 15.6 –6.8 23.8 –7.2 5.6 0.2 5.5 1.4 0.1 1.4
2005 3.3 29.2 –0.7 13.0 22.3 13.6 5.2 0.1 5.3 1.5 0.2 1.5
2006 13.8 20.6 15.9 –8.1 21.3 –4.2 5.6 0.2 5.5 1.6 0.1 1.6
2007 4.0 37.3 14.6 –8.2 17.3 –7.1 5.6 0.2 5.5 1.7 0.2 1.7
2008 –37.1 96.8 –23.7 6.8 51.9 9.3 5.6 0.3 5.6 3.2 1.3 2.6
2009 26.3 105.0 50.5 –3.7 44.2 –8.2 5.3 0.2 5.3 3.2 1.1 3.0
2010 14.7 86.6 29.3 2.2 49.1 11.8 4.9 0.3 4.9 1.9 0.2 1.8
167
(continued)
168
S&P Commodity Index JPM EMBI Plus VIX

Year Mean Std. Dev. Median Mean Std. Dev. Median Mean Std. Dev. Median
1995 17.7 39.0 14.3 38.3 69.3 38.3 14.9 294.8 155.1
1996 6.5 49.0 10.9 40.2 46.9 41.0 97.8 646.8 10.0
1997 –17.2 75.3 13.5 14.4 67.5 40.3 35.4 639.8 –4.2
1998 –23.0 85.3 –27.9 –8.9 189.9 5.7 46.7 1804.8 –60.2
1999 49.2 102.4 61.2 27.2 60.1 45.3 7.7 221.9 41.2
2000 30.3 118.2 29.5 16.3 45.6 16.1 22.0 378.8 30.0
2001 –30.8 61.1 –31.7 –0.2 45.5 –6.8 6.8 599.1 5.3
2002 41.2 85.4 41.4 15.5 67.6 20.4 42.9 589.6 28.9
2003 14.5 132.8 39.2 29.3 33.1 32.6 –27.0 379.5 –10.1
2004 21.9 106.9 25.8 12.2 35.1 21.5 –21.3 273.9 –37.5
2005 42.8 121.0 76.0 12.0 21.1 21.8 0.3 335.4 –36.6
2006 2.6 98.9 15.4 10.7 23.1 12.4 9.3 468.9 –23.4
2007 42.4 70.3 63.0 6.6 18.1 9.0 143.5 924.1 –12.5
2008 –37.4 270.0 –40.4 –8.2 82.3 4.9 183.2 2912.8 –37.7
2009 53.1 100.8 27.0 26.2 26.4 20.9 –40.6 287.3 –32.3
2010 23.3 106.3 40.6 12.1 26.3 22.5 7.5 932.6 –7.6
Source: Authors’ calculations using data from Bloomberg.

Note: Moody’s Corporate AAA is expressed as annual rate in percent. Credit spread is expressed as the difference in percentage points of annual interest
rates.
1/ Annual interest rate.
2/ Percentage points.
for the whole sample period, indicating that extreme performance is

dominated by movements in S&P 500, S&P Commodity Index, and
JPM Emerging Market Bond Index. For the full sample, the highest
median return comes from investing in emerging market bonds (20.7
percent), while the lowest comes from holding US dollars (–0.8 percent).
The median change in the volatility index VIX is negative. During sample
Period I, the S&P 500 has the highest median return (60.5 percent), while
the lowest belongs to the S&P Commodity Index (–13.8 percent). During
sample Period II, the Emerging Market Bond Index posts the best per-
formance (20.2 percent), while the S&P 500 the worst (–10.4 percent).
During sample Period III, top returns are achieved by commodities
(27 percent), the lowest by the US dollar (8.7 percent). The return on
Moody’s Corporate AAA Bonds falls from 6.7 percent per year during
sample Period I to 5.3 percent per year during sample Period III; the
average for the whole sample is 6.3 percent. By contrast, credit spreads
rise from 1.4 percentage points during sample Period I to 2.4 percentage
points for sample Period III. Panel B displays summary statistics for risk
factors over time. Based on median returns, the worst years for the S&P
500 are 2000, 2002, and 2008; for the US dollar, 2003, 2007, and 2009;
for commodities and emerging markets, 1998, 2001, and 2008.
Table 12.4 presents simple correlations between hedge fund returns
and risk factors for the complete sample and sub-periods of financial
distress; correlation coefficients statistically significantly different from
zero are marked with an asterisk. For the complete sample and for all the
strategies considered, the S&P 500, S&P Commodity Index, and the JPM
EMBI Plus are positively and statistically significantly correlated to hedge
fund returns, while the US dollar exchange rate and VIX are negatively
and statistically significantly correlated to hedge fund returns. Returns
on Moody’s Corporate AAA bonds and credit spreads are positively and
negatively statistically significantly correlated, respectively, to all returns
but those achieved by macro strategies.
Estimated correlations reflect the fact that potential independent
variables share the same association with the dependent variable. For
instance, S&P 500, Moody’s Corporate AAA bonds, S&P Commodity
Index, JPM EMBI Plus, US dollar exchange rate, credit spreads, and VIX
would all be contributing, though with the opposite expected sign, to
explain hedge fund returns.
Table 12.4 also indicates that the associations between hedge fund
returns and risk factors differ across periods of financial turmoil and
strategies. The US dollar, for instance, is positively related to hedge
fund returns during sample Period I, but negatively related during
170
Table 12.4 Simple correlations between hedge funds’ returns and risk factors
Hedge Fund S&P 500 USD Exchange Moody’s Corp Credit Spread S&P Commodity JPM EMBI VIX
Strategy Rate AAA Index Plus
All Funds 0.28* –0.10* 0.03* –0.10* 0.18* 0.21* –0.21*
Equity Hedge 0.39* –0.11* 0.03* –0.11* 0.18* 0.27* –0.28*
Event-Driven 0.42* –0.09* 0.04* –0.17* 0.19* 0.32* –0.32*
Fund of Funds 0.36* –0.07* 0.06* –0.17* 0.25* 0.31* –0.28*
Macro 0.04* –0.12* 0.02* –0.02 0.15* 0.04* –0.03*
Relative Value 0.29* –0.09* 0.05* –0.11* 0.18* 0.30* –0.27*
All Funds 0.26* 0.08* 0.05* –0.01 0.07* 0.23* –0.15*
Equity Hedge 0.37* 0.01 0.08* 0.00 0.10* 0.34* –0.24*
Event-Driven 0.40* 0.09* 0.07* –0.01 0.07 0.39* –0.30*
Fund of Funds 0.41* 0.21* 0.13* –0.08* 0.11* 0.38* –0.22*
Macro –0.04* 0.05* –0.06* 0.06* 0.04 –0.07* 0.10*
Relative Value 0.22* 0.16* 0.07 –0.05 –0.03 0.25* –0.21*
All Funds 0.16* –0.05* –0.07* –0.07* 0.05* 0.14* –0.15*
Equity Hedge 0.38* 0.01 –0.12* –0.17* 0.01 0.22* –0.34*
Event-Driven 0.35* 0.03 –0.15* –0.17* –0.02 0.26* –0.32*
Fund of Funds 0.13* –0.07* –0.06* –0.07* 0.04* 0.19* –0.15*
Macro –0.18 –0.17* 0.00* 0.09* 0.18 –0.02* 0.14*
Relative Value 0.22* –0.08* –0.08 –0.04 0.04 0.25* –0.22*
All Funds 0.42* –0.31* –0.22* –0.12* 0.40* 0.33* –0.36*
Equity Hedge 0.55* –0.37* –0.22* –0.13* 0.45* 0.41* –0.44*
Event-Driven 0.57* –0.35* –0.31* –0.21* 0.53* 0.48* –0.50*
Fund of Funds 0.50* –0.34* –0.32* –0.19* 0.56* 0.42* –0.48*
Macro 0.12* –0.22* –0.08* –0.03 0.18* 0.02 –0.07*
Relative Value 0.46* –0.30* –0.36* –0.14* 0.49* 0.54* –0.45*
Source: Authors’ calculations using the HFR database and data from Bloomberg.
* Indicates that correlation is statistically significant at the 5 percent significance level.
171
sample Period III; the negative statistical significance reflected for the
full sample appears to stem mainly from the strength of the association
during sample Period III. The correlation between hedge fund returns
and Moody’s Corporate AAA Bonds switches from positive during
sample Period I to negative during sample Periods II and III. Correlation
for credit spread across strategies appears weaker during sample Period I
than during the other two sample sub-periods. S&P Commodity Index is
more strongly correlated with hedge fund returns during sample Period III
than during sample Period II.
12.4 Results
This section discusses the results of estimating the relationship

between hedge fund returns and the risk factors—described in Table
12.1—according to the formulation shown in Equation (12.2). These
are estimated using a panel data approach for the whole sample and
sub-periods of financial distress, as well as controlling for investment
strategy.5 The objective is to estimate the portfolio’s betas that could help
replicate average hedge fund returns, while an estimation of the inter-
cept (alpha) would provide information on the hedge fund manager’s
ability to outperform the market. To capture the potentially significant
heterogeneity among different strategies, fixed effects are introduced
in the estimation. As managerial skills may vary across hedge funds
and strategies, the model as specified in Equation (12.2) is estimated
for each of the 332 hedge funds in the sample. Also, to capture the
possibility that hedge fund heterogeneity is random and uncorrelated
with the risk factors, random effects are incorporated in the panel data
estimation.
Table 12.5 lists estimated coefficients and standard errors for the four
different samples—whole sample and three sub-periods—without con-
trolling for differences in investment strategies. For the full sample all
risk factors are statistically significant, with signs in line with the cor-
relation discussed earlier, suggesting some degree of robustness in the
estimated coefficients; the intercept is negative, but not statistically
significant. For sample Period I, all risk factors but Moody’s Corporate
AAA Bonds, credit spread, and S&P Commodity Index are significant; the
intercept is negative and non-statistically significant. For sample Period II
all risk factors but Moody’s Corporate AAA Bonds are significant, with
the significant coefficients following the sign pattern described by the
simple correlations. For sample Period III only the JPM EMBI Plus
is not statistically significant; the intercept is positive and statistically
Table 12.5 Panel regression: fixed effects and random effects 1
Dependent variable: Hedge fund’s returns
Independent Variables Full Sample: Period I: April Period II: Nov. Period III: March
1995–2010 1997–May 1999 2000–Sep. 2003 2008–Dec. 2010
Estimated Standard Estimated Standard Estimated Standard Estimated Standard
Coefficient Error Coefficient Error Coefficient Error Coefficient Error
Fixed Effects
S&P 500 0.1796 0.0056* 0.2832 0.0211* 0.0402 0.0152* 0.1918 0.0124*
USD Exchange Rate –0.0306 0.0079* 0.1760 0.0314* –0.1006 0.0191* 0.0698 0.0194*
Moody’s Corp AAA 0.2077 0.0190* 0.2392 0.2221 0.1347 0.0901 –1.1202 0.1434*
Credit Spread –0.2157 0.0248* 0.0736 0.2671 –0.7758 0.1630* 0.2970 0.0512*
S&P Commodity Index 0.0796 0.0029* 0.0122 0.0110 0.0459 0.0066* 0.1308 0.0069*
JPM EMBI Plus 0.0908 0.0054* 0.1035 0.0164* 0.0882 0.0141* –0.0077 0.0212
VIX –0.0072 0.0013 0.0352 0.0045* –0.0243 0.0051* –0.0156 0.0028*
Intercept –0.2332 0.1347 –1.5651 1.8324 1.6409 0.4540* 5.2986 0.7026*
R-Squared 0.10 0.08 0.04 0.22
F-Statistic 2/ 997 109 72 446
Random Effects
S&P 500 0.1796 0.0056* 0.2832 0.0211* 0.0402 0.0152* 0.1918 0.0124*
USD Exchange Rate –0.0306 0.0079* 0.1760 0.0313* –0.1006 0.0191* 0.0698 0.0193*
Moody’s Corp AAA 0.2076 0.0190* 0.2392 0.2216 0.1347 0.0901 –1.1209 0.1427*
Credit Spread –0.2157 0.0248* 0.0736 0.2666 –0.7758 0.1630* 0.2970 0.0510*
(continued)
173
174
Independent Variables Full Sample: Period I: April Period II: Nov. Period III: March
1995–2010 1997–May 1999 2000–Sep. 2003 2008–Dec. 2010
Estimated Standard Estimated Standard Estimated Standard Estimated Standard
Coefficient Error Coefficient Error Coefficient Error Coefficient Error
S&P Commodity 0.0796 0.0029* 0.0122 0.0110 0.0459 0.0066* 0.1308 0.0069*
JPM EMBI Plus 0.0908 0.0054* 0.1035 0.0164* 0.0882 0.0141* –0.0077 0.0211
VIX –0.0072 0.0013* 0.0352 0.0044* –0.0243 0.0051* –0.0156 0.0028*
Intercept –0.2323 0.1351 –1.5651 1.8286 1.6409 0.4548* 5.3023 0.6991*
R-Squared 0.10 0.08 0.04 0.22
Chi-Squared Statistic 3/ 6,982 763 504 3,154
Note 1: Moody’s Corporate AAA is expressed as annual rate. Credit spread is expressed as the percentage difference in annual rates.
2: In all cases the null hypothesis (all estimated coefficients are equal to zero) is rejected.
3: In all cases the null hypothesis (all estimated coefficients are equal to zero) is rejected.
significant. The R-squared is relatively low, although the hypothesis

that all coefficients are equal to zero is easily rejected. As expected for
relatively large sample sizes, regressions with fixed and random effects
produce similar results as estimated coefficients and standard errors (i.e.
t-statistics) are the same in almost all cases.
Table 12.6 reports the estimation results when fixed effects across
investment strategies are incorporated. For the full sample, credit
spreads and volatility (VIX) have in general a negative effect on fund
returns across strategies. With the exception of hedge funds pursuing
macro strategies, in most cases the impact is statistically significant.
Only hedge funds pursuing event-driven strategies report a positive and
statistically significant intercept, indicating that for most strategies it
may be possible to replicate hedge fund returns by investing in liquid
instruments. For most strategies the goodness of fit is relatively higher
in comparison with the R-squared estimated without controlling for
the investment strategy (11 percent). As before, a model with random
effects does not generate any significantly different results.
Estimation results, however, differ considerably across periods of
financial distress, particularly with regard to the ability of the hedge
fund manager to overperform the market. For sample Period I, the
intercept is statistically significant in only one of the five strategies con-
sidered; specifically, it is negative for equity hedge, and not statistically
significant for the rest. For sample Period II, the intercept is significant
in three out of the five strategies and is positive in all cases. Finally, for
sample Period III, the intercept is positive and statistically significant for
all strategies; the variation explained by the model is above 40 percent
for most strategies. These findings suggest that, aside from the possibility
that other risk factors, including the size of assets under manage-
ment, could help explain hedge fund returns, risk factors cannot fully
replicate hedge fund returns during times of financial distress. To be
sure, other elements are likely to play an important role in generating
alpha, especially when specific expertise can be brought to bear on port-
folio allocation decisions dynamically in a heightened volatile market
environment.
Table 12.7 shows estimation results for each of the hedge funds in the
sample in each of the four sample periods. For each strategy, summary
statistics—average, standard deviation, maximum, and minimum—for
the estimated coefficients are reported. The number of estimated coeffi-
cients, statistically significant and non-significant, positive and negative,
is also reported. For instance, for the full sample period, using the 106
hedge funds with equity hedge strategy, the average intercept is –0.459
176
Table 12.6 Panel regression: fixed effects by hedge fund investment strategy
Dependent variable: Hedge funds’ returns
Hedge Fund S&P 500 USD Moody’s Credit S&P JPM VIX Intercept R-Squared F-Statistic 1/
Strategy Exchange Corp Spread Commodity EMBI
Rate AAA Index Plus

Equity Hedge
Coefficient 0.3840* –0.0266 0.2115* –0.1681* 0.0782* 0.1048* –0.0080* –0.3898
Standard Error 0.0118 0.0166 0.0397 0.0518 0.0061 0.0114 0.0028 0.2819 0.17 598
Event-Driven
Coefficient 0.1930* 0.0169 0.1093* –0.3891* 0.0471* 0.1058* –0.0115* 0.7381*
Standard Error 0.0125 0.0176 0.0421 0.0550 0.0065 0.0121 0.0030 0.2989 0.20 219
Fund of Funds
Coefficient 0.1124* 0.0521* 0.1808* –0.3320* 0.0801* 0.1135* –0.0057* –0.0021
Standard Error 0.0057 0.0080 0.0191 0.0249 0.0029 0.0055 0.0013 0.1353 0.19 696
Macro
Coefficient –0.0093 –0.2101* 0.3253* 0.0075 0.1118* 0.0047 –0.0033 –1.0463*
Standard Error 0.0178 0.0249 0.0598 0.0780 0.0092 0.0171 0.0042 0.4241 0.03 51
Relative Value
Coefficient 0.0591* –0.0128 0.1421* –0.2596* 0.0458* 0.1277* –0.0134* 0.2269
Standard Error 0.0118 0.0165 0.0396 0.0516 0.0061 0.0113 0.0028 0.2807 0.14 120
Equity Hedge
Coefficient 0.4254* –0.0705 1.4235* 1.1560* –0.0213 0.2021* 0.0456* –11.4357*
Standard Error 0.0435 0.0646 0.4575 0.5504 0.0227 0.0338 0.0092 3.7754 0.16 73
Event-Driven
Coefficient 0.2574* 0.1391 0.5207 0.2737 –0.0214 0.1539* 0.0150 –3.3979
Standard Error 0.0568 0.0844 0.5975 0.7188 0.0297 0.0441 0.0120 4.9308 0.18 25
Fund of Funds
Coefficient 0.2830* 0.3383* –0.1771 –1.0218* 0.0332* 0.0860* 0.0359* 2.7691
Standard Error 0.0197 0.0293 0.2072 0.2493 0.0103 0.0153 0.0042 1.7098 0.25 129
Macro
Coefficient 0.1554* 0.2945* –1.2743 0.0643 0.0745* –0.0615 0.0446* 8.9594
Standard Error 0.0640 0.0950 0.6726 0.8091 0.0334 0.0497 0.0135 5.5506 0.03 6
Relative Value
Coefficient 0.0489 0.2772* 0.3763 –0.0835 –0.0437 0.1053* –0.0064 –1.8927
Standard Error 0.0571 0.0848 0.6005 0.7224 0.0298 0.0443 0.0121 4.9552 0.09 10
Equity Hedge
Coefficient 0.2983* –0.0052 0.4933* –1.9018* 0.0518* 0.1176* –0.0182 2.1052*
Standard Error 0.0311 0.0391 0.1848 0.3344 0.0135 0.0289 0.0104 0.9313 0.16 100
Event-Driven
Coefficient 0.1367* 0.0395 0.0363 –1.0379* 0.0119 0.1325* –0.0074 2.9932*
Standard Error 0.0342 0.0430 0.2030 0.3675 0.0148 0.0318 0.0115 1.0233 0.15 27
Fund of Funds
Coefficient –0.0120 –0.0669* 0.2039* –0.7216* 0.0161* 0.0973* –0.0140* 0.8146*
Standard Error 0.0137 0.0172 0.0814 0.1473 0.0060 0.0127 0.0046 0.4102 0.05 30
Macro
Coefficient –0.3509* –0.4044* –0.4975 0.9060 0.1168* –0.0046 –0.0640* 1.9204
Standard Error 0.0473 0.0596 0.2812 0.5089 0.0206 0.0440 0.0159 1.4172 0.09 29
Relative Value
Coefficient 0.0189 –0.0712* 0.0206 –0.1296 0.0161 0.0990* –0.0170 0.8674
Standard Error 0.0263 0.0330 0.1560 0.2823 0.0114 0.0244 0.0088 0.7863 0.09 13
177
(continued)
178
Hedge Fund S&P 500 USD Moody’s Credit S&P JPM VIX Intercept R-Squared F-Statistic 1/
Strategy Exchange Corp Spread Commodity EMBI
Rate AAA Index Plus

Equity Hedge
Coefficient 0.4431* 0.1462* –0.8363* 0.4681* 0.1361* 0.0482 –0.0106 3.2311*
Standard Error 0.0261 0.0408 0.3013 0.1075 0.0145 0.0445 0.0059 1.4762 0.33 249
Event-Driven
Coefficient 0.1811* 0.2675* –0.8262* 0.0330 0.1450* 0.2134* –0.0192* 4.3322*
Standard Error 0.0292 0.0456 0.3365 0.1205 0.0163 0.0497 0.0066 1.6478 0.42 105
Fund of Funds
Coefficient 0.0653* 0.1647* –1.1923* 0.1774* 0.1472* 0.0264* –0.0263 5.7181*
Standard Error 0.0116 0.0181 0.1340 0.0478 0.0065 0.0198 0.0026 0.6567 0.40 347
Macro
Coefficient 0.0475 –0.4630* –1.5132* 0.3777* 0.0648* –0.4990* –0.0084 8.0387*
Standard Error 0.0353 0.0552 0.4073 0.1457 0.0197 0.0602 0.0079 1.9953 0.09 28
Relative Value
Coefficient 0.0449 0.3937* –1.3859* 0.2124 0.1819* 0.5122* –0.0056 6.5815*
Standard Error 0.0307 0.0481 0.3544 0.1269 0.0172 0.0524 0.0069 1.7355 0.43 96
Note: In all cases the null hypothesis (all estimated coefficients are equal to zero) is rejected.
Table 12.7 Individual OLS regressions: summary indicators for estimated coefficients and adjusted R-Sq
Dependent variable: Hedge Funds’ returns
Hedge Fund S&P USD Moody’s Credit S&P JPM VIX Intercept Adj.
Strategy 500 Exchange Corp Spread Commodity EMBI R-Squared
Rate AAA Index Plus
Equity Hedge
Average 0.384 –0.026 0.220 –0.163 0.078 0.106 –0.008 –0.459 0.345
Standard Error 0.454 0.202 0.448 0.551 0.089 0.192 0.025 2.649 0.218
Minimum –1.454 –0.853 –1.384 –1.693 –0.191 –0.341 –0.079 –6.648 –0.015
Maximum 1.230 0.671 1.717 1.469 0.325 1.050 0.082 8.812 0.955
Significant (Number of Coefficients)
Positive 80 6 21 5 43 26 3 1
Negative 12 11 1 12 3 4 6 7
Non-Significant
Positive 8 36 64 36 44 56 28 44
Negative 6 53 20 53 16 20 69 54
Event-Driven
Average 0.194 0.018 0.115 –0.385 0.047 0.106 –0.011 0.692 0.343
Standard Error 0.181 0.080 0.244 0.451 0.043 0.134 0.015 2.182 0.146
Minimum –0.014 –0.080 –0.545 –1.582 –0.019 –0.008 –0.047 –4.879 0.071
Maximum 0.764 0.322 0.888 0.285 0.177 0.599 0.024 6.871 0.636
Positive 25 0 8 0 14 14 0 5
Negative 0 0 0 12 0 0 8 3
(continued)
179
180
Rate AAA Index Plus
Non-Significant
Positive 5 19 16 8 15 15 4 15
Negative 1 12 7 11 2 2 19 8
Fund of Funds
Average Coefficient 0.113 0.053 0.188 –0.326 0.080 0.114 –0.005 –0.069 0.334
Standard Error 0.126 0.113 0.233 0.353 0.045 0.094 0.012 1.507 0.142
Minimum –0.602 –0.249 –0.439 –2.056 0.022 –0.127 –0.062 –7.613 0.029
Maximum 0.378 0.554 1.170 1.205 0.284 0.487 0.033 5.121 0.596
Positive 67 14 43 0 98 72 0 5
Negative 2 3 1 47 0 0 10 4
Non-Significant
Positive 28 59 49 9 8 26 25 45
Negative 9 30 13 50 0 8 71 52
Macro
Average Coefficient –0.006 –0.208 0.340 0.023 0.111 0.004 –0.003 –1.185 0.084
Standard Error 0.234 0.251 0.484 0.428 0.181 0.261 0.023 3.106 0.136
Minimum –0.458 –0.904 –1.158 –1.430 –0.198 –0.861 –0.045 –8.117 –0.019
Maximum 0.802 0.218 1.348 0.833 1.115 0.815 0.061 8.367 0.507
Positive 8 1 14 0 18 12 1 1
Negative 1 22 2 0 1 3 0 3
Non-Significant
Positive 16 13 35 31 33 17 25 17
Negative 37 26 11 31 10 30 36 41
Relative Value
Average Coefficient 0.059 –0.012 0.145 –0.257 0.046 0.128 –0.013 0.203 0.272
Standard Error 0.138 0.071 0.233 0.389 0.048 0.210 0.025 1.242 0.178
Minimum –0.410 –0.217 –0.279 –1.157 –0.053 –0.004 –0.125 –2.274 –0.022
Maximum 0.376 0.119 0.656 0.556 0.143 1.075 0.002 2.722 0.783
Positive 8 0 9 0 15 16 0 2
Negative 1 3 0 10 1 0 4 3
Non-Significant
Positive 13 13 13 6 7 9 5 12
Negative 5 11 5 11 4 2 18 10
Equity Hedge
Average 0.425 –0.071 1.424 1.156 –0.021 0.202 0.046 –11.436 0.361
Standard Error 0.543 0.693 3.908 4.487 0.188 0.411 0.061 31.879 0.259
Minimum –1.479 –1.455 –11.912 –8.989 –0.701 –0.678 –0.121 –137.206 –0.106
Maximum 1.626 3.403 15.777 14.961 0.524 2.145 0.272 83.581 0.934
Positive 52 2 8 4 1 11 4 5
Negative 4 8 4 4 4 4 0 7
Non-Significant
Positive 36 41 60 57 46 65 84 32
Negative 14 55 34 41 55 26 18 62
(continued)
181
182
Rate AAA Index Plus
Event-Driven
Average 0.257 0.139 0.521 0.274 –0.021 0.154 0.015 –3.398 0.369
Standard Error 0.335 0.457 2.254 3.220 0.143 0.306 0.047 18.144 0.232
Minimum –0.389 –0.973 –8.661 –9.306 –0.401 –0.379 –0.052 –47.457 –0.083
Maximum 1.591 1.818 4.727 9.991 0.469 1.185 0.225 71.909 0.770
Positive 10 1 2 1 0 5 0 1
Negative 0 0 0 1 1 0 1 0
Non-Significant
Positive 18 22 21 11 8 23 21 9
Negative 3 8 8 18 22 3 9 21
Fund of Funds
Average Coefficient 0.283 0.338 –0.177 –1.022 0.033 0.086 0.036 2.769 0.363
Standard Error 0.211 0.312 1.978 1.997 0.069 0.151 0.034 16.048 0.229
Minimum –0.593 –0.154 –8.356 –5.308 –0.155 –0.405 –0.080 –82.242 –0.202
Maximum 0.853 1.246 10.665 6.643 0.384 0.651 0.130 62.021 0.807
Positive 55 32 1 0 1 8 14 1
Negative 0 0 0 3 0 0 0 0
Non-Significant
Positive 50 65 41 28 74 75 82 69
Negative 1 9 64 75 31 23 10 36
Macro
Average Coefficient 0.155 0.294 –1.274 0.064 0.075 –0.061 0.045 8.959 0.080
Standard Error 0.304 0.714 2.499 3.756 0.257 0.496 0.057 19.926 0.244
Minimum –0.459 –2.737 –6.930 –10.754 –0.656 –1.566 –0.140 –23.993 –0.240
Maximum 0.733 1.695 3.612 8.155 1.367 1.204 0.217 54.073 0.622
Positive 4 5 0 0 4 6 0 1
Negative 1 2 1 1 0 3 0 0
Non-Significant
Positive 41 43 16 30 37 18 53 43
Negative 16 12 45 31 21 35 9 18
Relative Value
Average Coefficient 0.049 0.277 0.376 –0.083 –0.044 0.105 –0.006 –1.893 0.218
Standard Error 0.229 0.508 1.227 2.056 0.100 0.302 0.095 9.804 0.235
Minimum –0.847 –0.265 –3.555 –7.536 –0.469 –0.314 –0.468 –24.665 –0.093
Maximum 0.506 1.700 3.232 3.360 0.055 1.489 0.066 25.456 0.896
Positive 4 9 0 1 0 3 0 1
Negative 0 0 0 1 1 0 1 0
Non-Significant
Positive 17 8 22 10 8 14 20 10
Negative 6 10 5 15 18 10 6 16
Equity Hedge
Average 0.298 –0.005 0.493 –1.902 0.052 0.118 –0.018 2.105 0.388
Standard Error 0.543 0.319 1.836 2.347 0.114 0.232 0.074 8.951 0.267
(continued)
183
184
Rate AAA Index Plus
Minimum –1.805 –1.095 –6.478 –7.847 –0.253 –0.520 –0.211 –19.586 –0.156
Maximum 1.516 1.645 6.326 3.075 0.554 0.731 0.172 36.525 0.958
Positive 40 1 19 0 6 10 1 5
Negative 11 5 3 10 1 2 3 4
Non-Significant
Positive 33 53 50 23 69 68 48 58
Negative 22 47 34 73 30 26 54 39
Event-Driven
Average 0.137 0.040 0.036 –1.038 0.012 0.132 –0.007 2.993 0.247
Standard Error 0.293 0.163 0.550 1.386 0.046 0.155 0.047 5.272 0.189
Minimum –0.654 –0.171 –1.383 –5.675 –0.165 –0.056 –0.234 –3.674 –0.164
Maximum 1.077 0.780 1.284 0.918 0.130 0.864 0.040 24.697 0.673
Positive 7 0 1 0 0 9 0 4
Negative 1 0 0 3 0 0 1 0
Non-Significant
Positive 17 17 17 5 21 20 17 16
Negative 6 14 13 23 10 2 13 11
Fund of Funds
Average Coefficient –0.012 –0.067 0.204 –0.722 0.016 0.097 –0.014 0.815 0.271
Standard Error 0.231 0.170 0.625 1.341 0.063 0.084 0.035 4.064 0.182
Minimum –1.854 –0.864 –0.976 –4.281 –0.094 –0.182 –0.261 –18.160 –0.125
Maximum 0.375 0.351 4.654 8.445 0.530 0.609 0.041 14.008 0.698
Positive 10 3 10 0 2 35 0 10
Negative 11 12 0 24 0 0 2 2
Non-Significant
Positive 46 27 58 21 60 67 22 57
Negative 39 64 38 61 44 4 82 37
Macro
Average Coefficient –0.351 –0.404 –0.498 0.906 0.117 –0.005 –0.064 1.920 0.194
Standard Error 0.526 0.597 1.246 2.437 0.163 0.296 0.085 7.911 0.184
Minimum –1.269 –2.152 –3.014 –3.872 –0.157 –1.081 –0.343 –19.522 –0.140
Maximum 1.095 0.714 1.621 6.801 0.747 0.836 0.062 25.932 0.732
Positive 4 1 0 0 4 3 0 3
Negative 18 8 0 0 2 0 1 0
Non-Significant
Positive 10 12 20 39 46 31 16 27
Negative 30 41 42 23 10 28 45 32
Relative Value
Average Coefficient 0.019 –0.071 0.021 –0.130 0.016 0.099 –0.017 0.867 0.176
Standard Error 0.186 0.122 0.618 1.203 0.065 0.132 0.051 2.706 0.263
Minimum –0.631 –0.464 –1.169 –4.383 –0.061 –0.034 –0.245 –2.539 –0.126
Maximum 0.494 0.082 2.340 1.896 0.298 0.669 0.026 6.708 0.777
Positive 2 0 1 1 0 2 0 1
Negative 2 2 0 0 0 1 0 1
185
(continued)
186
Rate AAA Index Plus
Non-Significant
Positive 10 10 15 12 14 24 11 11
Negative 13 15 11 14 13 0 16 14
Equity Hedge
Average 0.444 0.145 –0.840 0.468 0.136 0.046 –0.011 3.253 0.587
Standard Error 0.567 0.414 1.874 0.953 0.184 0.453 0.054 9.344 0.289
Minimum –1.771 –1.405 –7.734 –1.993 –0.238 –1.156 –0.130 –34.127 –0.113
Maximum 1.582 1.298 2.853 4.373 1.127 2.032 0.226 37.240 0.985
Positive 65 25 1 17 45 10 4 7
Negative 9 7 10 6 2 10 18 2
Non-Significant
Positive 22 45 34 60 41 47 40 58
Negative 10 29 61 23 18 39 44 39
Event-Driven
Average 0.181 0.268 –0.826 0.033 0.145 0.213 –0.019 4.332 0.619
Standard Error 0.183 0.298 0.925 0.546 0.162 0.429 0.029 4.845 0.194
Minimum –0.112 –0.192 –4.269 –1.280 –0.053 –0.454 –0.093 –5.638 0.229
Maximum 0.702 1.460 1.071 1.104 0.706 1.921 0.031 22.913 0.868
Positive 14 13 1 3 17 12 0 2
Negative 0 0 0 3 0 2 7 1
Non-Significant
Positive 15 14 1 17 8 10 10 28
Negative 2 4 29 8 6 7 14 0
Fund of Funds
Average 0.066 0.164 –1.222 0.179 0.147 0.024 –0.026 5.881 0.610
Standard Error 0.112 0.208 0.809 0.523 0.106 0.215 0.022 3.568 0.208
Minimum –0.279 –0.961 –6.200 –0.799 –0.228 –1.111 –0.105 –1.807 –0.090
Maximum 0.419 0.870 0.343 4.572 0.869 0.874 0.099 25.340 0.882
Positive 25 41 0 5 92 7 0 16
Negative 3 0 20 0 0 2 43 0
Non-Significant
Positive 51 55 2 70 13 55 4 86
Negative 27 10 84 31 1 42 59 4
Macro
Average 0.047 –0.462 –1.501 0.377 0.065 –0.497 –0.008 7.972 0.211
Standard Error 0.323 0.419 2.281 0.842 0.328 0.571 0.054 11.055 0.246
Minimum –0.286 –1.862 –9.022 –1.283 –0.400 –1.668 –0.134 –17.960 –0.172
Maximum 1.210 0.415 3.777 2.412 1.849 0.758 0.119 44.820 0.892
Positive 5 0 2 7 6 1 2 4
Negative 0 24 5 0 4 24 0 2
(continued)
187
188
Rate AAA Index Plus
Non-Significant
Positive 21 8 10 34 27 15 21 47
Negative 36 30 45 21 25 22 39 9
Relative Value
Average 0.049 0.277 0.376 –0.083 –0.044 0.105 –0.006 –1.893 0.218
Standard Error 0.229 0.508 1.227 2.056 0.100 0.302 0.095 9.804 0.235
Minimum –0.847 –0.265 –3.555 –7.536 –0.469 –0.314 –0.468 –24.665 –0.093
Maximum 0.506 1.700 3.232 3.360 0.055 1.489 0.066 25.456 0.896
Positive 4 19 3 5 21 17 0 7
Negative 1 0 8 1 0 0 5 3
Non-Significant
Positive 11 6 3 11 3 7 13 14
Negative 11 2 13 10 3 3 9 3
percent, with standard deviation of 2.649; in one case the intercept is

positive and statistically significant, whereas in seven cases it is negative
and statistically significant. Only hedge funds with event-driven and
relative value strategies report positive average intercepts. Event-driven
strategies report the highest proportion of positive and statistically
significant intercepts (17 percent). Considering all the funds and all
the strategies in the sample, the proportion of positive and statistically
significant coefficients increases from 3 percent in sample Period I to 11
percent in sample Period III. The proportion of intercepts not statisti-
cally different from zero for the full sample period is 90 percent, which
compares with proportions of 95 percent, 91 percent, and 87 percent,
in sample Periods I, II, and III, respectively. These findings underscore
the role of specific factors, such as manager’s skills, in generating
returns that cannot be replicated using available market instruments.
12.5 Conclusion
This chapter estimates a factor-based model using data from the

Hedge Fund Research (HFR) database. The focus is on the replication
hedge fund performance across strategies and time. The relationship
between hedge fund returns and the risk factors is estimated using a
panel data approach with both fixed and random effects on the entire
sample as well as on individual strategies. In addition, estimations are
carried out also for three sub-periods of heightened financial distress,
as evidenced by an increase in borrowing from the International
Monetary Fund.
The results show that for the full sample all risk factors are statisti-
cally significant, with signs in line with observed correlation. There
appears to be no statistically significant alpha. When the relationship is
estimated in the context of each strategy, results show that only hedge
funds pursuing event-driven strategies report a positive and statistically
significant intercept, indicating that for most strategies it is possible to
replicate hedge fund returns by investing in liquid instruments. Finally,
estimation results differ considerably across periods of financial distress,
particularly with regard to the ability of the hedge fund manager to
overperform the market. In this regard, results focusing on the perform-
ance of each individual hedge fund across strategies and time periods
provide further evidence that managers’ skills have a positive impact
on hedge fund returns particular at times of heightened volatility in the
context of global financial distress.
Acknowledgments
The authors would like to thank Oksana Khadarina for help with the
database. The views expressed herein are those of the authors and should
not be attributed to the International Monetary Fund, its Executive
Board, or its management.
Notes
1. Ennis and Sebastian (2003) found no evidence of market neutrality.
2. See Kat (2007) and Roncalli and Teïletche (2007) for a review of these methods.
3. For a complete definition of the strategies see HFR (2011) Strategy and
Regional Classifications, Hedge Fund Research, Inc. Information also available
at: http://www.hedgefundresearch.com/index.php?fuseindices-str#2561.
4. For a detailed discussion of the various types of biases in hedge funds data see
Capocci and Hübner (2004).
5. For detailed discussion on panel data estimation methods, see Wooldridge
(2008), Chapters 13 and 14, and Green (2011), Chapter 14.
References
Capocci, D., and Hubner, G. (2004). “Analysis of Hedge Fund Performance.”
Journal of Empirical Finance, 11(1): 55–89.
Ennis, R., and Sebastian, M. (2003). “A Critical Look at the Case for Hedge Funds:
Lessons from the Bubble.” Journal of Portfolio Management, 29(4): 103–12.
Green, W. H. (2011). Econometric Analysis. Prentice Hall, Upper Saddle River, NJ.
Hasanhodzic, J., and Lo, A. W. (2006). “Can Hedge-Fund Returns Be Replicated?
The Linear Case.” MIT Laboratory for Financial Engineering, Working Paper,
Cambridge, MA.
Kat, H. M. (2007). “Alternative Routes to Hedge Fund Return Replication:
Extended.” Cass Business School Research Paper No. 0037, available at SSRN:
http://ssrn.com/abstract=939395.
Roncalli, T., and Teïletche, J. (2007). “An Alternative Approach to Alternative
Beta” available at SSRN: http://ssrn.com/abstract=1035521.
Wooldridge, J. (2008). Introductory Econometrics: A Modern Approach. South-
Western College Publisher, Florence, KY.
Index
abnormal return 49, 53–5 delivery failure 148

algorithms 32, 119, 121–2, 133 distributional replication 77–82, 84–5,
allocation 2, 75, 78–9, 83, 105, 119, 87
131, 133, 144–5, 160, 175 distribution-matching 31–2
alpha 1–2, 6, 12–3, 16, 28, 61–2, 72, dominance replication 82–7
74, 81, 136, 151, 157, 159, 172, dynamic averaging 16, 27
175, 189 dynamic mix 21–3, 26
alternative alpha 72
alternative beta 1, 12–3, 29, 47, 61, efficiency measure 91
72, 76–9, 89, 132, 145, 159, 190 ETFs 3, 12, 35, 78, 94, 129–130
acquirer 49, 52, 54 event study 58
autocorrelation 30, 33–4, 69, 103, 135 expected utility maximization 134
backfilling bias 63–5 factor loadings 15, 17, 20, 69, 79, 81,
bayesian framework 106 86
bayesian model averaging 15, 21, 28 factor misspecification 81, 85
benchmarking 13–4, 29, 47, 61–2, 64, factor weights 5–6, 122
66–70, 72–4, 89, 132 factor-based models 15, 80
beta 1–2, 5–6, 12–3, 29, 39–43, 47, fees 1–2, 13, 31, 65, 77, 93, 119, 124,
61–2, 68–9, 72, 74, 76–9, 89, 106–112, 130, 134, 155, 159
114, 116–8, 125, 132, 136, 145, financial distress 154, 160, 162–3,
159–160, 172, 190 169, 172, 175, 189
beta factors 68 free disposal hull 90
beta timing factor 69
bidder 50–2, 57 half life 127
hedge fund indices 2, 4, 12, 23, 26,
CAAR 53–4 28, 32, 46, 61–2, 65–7, 72, 74–5, 111,
calibration 6, 18, 122 113, 115, 117, 124, 147, 153, 160
clone 1–3, 5–17, 27, 29–30, 32, 47, hedge fund replication 2–3, 6, 13,
62, 71, 73, 76–80, 89, 106, 110–1, 15–6, 18, 20–2, 25–8, 30–2, 36, 38,
113–7, 119–122, 124–131, 134, 139, 44, 46–8, 50, 59, 62, 68, 73, 76–8,
141, 143, 145 80, 82, 84–6, 88–9, 118, 131–4, 136,
clone returns 127–8, 130 138, 144, 148, 150, 152, 154, 156,
clone-plus 126–8 158–160, 162, 172, 190
composite models 39, 44 heterogeneity 66–7, 91, 94, 172
concentration risk 148 HFRI 3–4, 7–8, 10–11, 22–4, 34, 41,
convergence 4, 48 43, 73, 143
counterparty risk 148–9, 151–2, 157 hurdle rate 2
covariance 108–110, 123, 125 hybrid replication 76
covariance matrix 108–110, 123, 125 hybrid theory 80
DEA 90, 92, 98–100, 103–5 inputs 19, 63, 90–3, 98–9, 102–4
deal failure 48–9 investable index 62, 64–5, 135
191
192 Index
Kalman filter 2, 6, 8–10, 12, 17, 69, primary clone 127

109, 121–3, 134 primitive risk signals 108
kurtosis 4, 17, 24, 27, 30–1, 39, 41, principal component analysis 5
43–4, 87, 94–8, 102, 105, 126, 135, production frontier 90, 92
138 productivity analysis 91, 104
learning models 119–120, 122–4, 126, quantile regression 16, 19

128, 130–2
likelihood function 109 regulated environment 146
linear models 15 replicating algorithm 121–2
liquidity 2, 30–1, 51, 65–6, 69, 103, replication 1–3, 6, 10, 12–3, 15–8,
105, 108, 118–9, 146–152, 154–9 20–32, 36, 38, 44, 46–8, 50, 59, 62,
liquidity risk 149 67–9, 73, 76–91, 93–4, 96, 98–9,
lock-up 2, 102–3 101–3, 107, 114, 116, 118–122, 124,
126–134, 136, 138, 143–6, 148,
market-timing 16 150–2, 154, 156, 158–160, 162, 172,
mean absolute deviation 134, 138, 143, 189–190
145 replicators 15, 17, 22–3, 26, 84
mean variance optimization 134 ridge regression 16, 19
mimic 17, 31–2, 46, 69, 73, 121–2 risk constraints 30, 32, 36, 38, 44, 46
multifactor models 10, 13, 90 risk factors 1, 3, 7, 9, 31, 69–70, 73,
77, 79, 83, 101, 103, 107, 117, 133,
neural networks 69 136, 160–1, 163, 166, 169–170, 172,
noise 18, 103 175, 189
no-learning clone 127 rolling samples 16, 18, 23
non parametric 105 rule based 13, 31–2, 77, 79, 81, 94, 111
nonlinear optimization 31
selection bias 23, 63–5
OLS 5–6, 15–20, 24–7, 109–110, 123, sensitivity 19, 67, 103, 106, 137, 160
127, 134, 138, 179 short history bias 64
omega 37, 91, 97–8, 100–2 skewness 4, 17, 24, 27, 30–1, 33, 39,
optimization 5, 20, 31, 39, 46, 85–6, 41, 43–4, 94–8, 102, 105, 135, 138
99, 104, 120, 134, 137–8, 145 smoothing 33, 47, 98
optimization function 39 space models 106, 108, 110, 118
outputs 90–3, 98–9, 102–4 stepwise regression 16, 19, 70
stochastic dominance 76, 82–4
passive 1–3, 6, 12–4, 28–9, 46–7, 61, style factors 71–2, 77, 79, 84
65, 74, 88–9, 107, 131–4, 136, 138, survivorship bias 63–5, 93, 132, 162
144 systematic risk 1, 13, 31, 62, 64, 107,
passive hedge fund replication 2, 6, 111, 117–8, 133
13, 28, 46, 88, 131–4, 136, 138, 144 systematic exposure 1, 136
passively managed 2, 65
performance 2–3, 7, 12, 17, 22–3, 28, tail event 79, 81, 86
31–3, 36–9, 44, 46–7, 57, 61, 63–8, tail risk 94, 97–8, 101, 103
71–2, 74–85, 87–91, 93–4, 97–9, takeover 48, 50–4, 56–60
101–5, 111, 114, 116–7, 120–2, 124–9, target 17, 22, 25, 48–50, 52–9, 64,
131–2, 134, 137–9, 141, 143–5, 151–6, 76–7, 80–1, 83–6, 109, 122, 124,
159–160, 162–3, 169, 189–190 126–7, 131
ponzi 151 technical efficiency 91
Index 193
thick modeling 16, 21–2, 27 tracking error variance 137–8

thin modeling 16, 21–2, 27 trading algorithms 32, 121, 133
time-varying betas 106 trash ratio 147
tracking error 8, 12, 17, 20, 22–7,
31, 68, 127, 131, 134, 137–8, 143–5, UCITS framework 146–7, 149–150,
152 152, 155–8

Hedge Fund Replication PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hedge Fund Replication PDF

Uploaded by

Copyright:

Available Formats

Hedge Fund Replication

This page intentionally left blank

List of Tables vii

1 Can We Really “Clone” Hedge Fund Returns? Further

10 Linear Model for Passive Hedge Fund Replication 133

1.1 Selected hedge fund indices and statistics (January

7.5 Spearman rank correlations 100

1.1 Comparison of average Sharpe ratios of clones vs. indices

factor replication with an innovative version of distributional replication.

Greg N. Gregoriou has published 42 books, 60 refereed publications in

Maher Kooli is Professor of Finance at the School of Management,

Giovanni Barone-Adesi is Director of the Swiss Finance Institute at the

Laurent Germain is Professor of Finance and the Head of the Finance

Clemens H. Glaffig is principal partner of Panathea Capital Partners

Mathematics at the University of California and research fellow at the

Richard D. F. Harris is a Professor of Finance in the XFI Centre for

Peter Mayall is a lecturer in finance at the Curtin University of

Iliya Markov has an MSc in Operational Research with Finance from

at the American University in Bulgaria and a full scholarship at the

R. McFall-Lamm Jr. is Chief Investment Officer at Stelac Advisory Services,

Nicolas Nalpas is Professor of Finance at Toulouse Business School

Lajos Németh holds an MSc in Management from EBS Business School

Anthony Ravi is a past honors student at the School of Economics and

Sergio L. Rodríguez received his PhD in Economics from UCLA, and

Marco Rossi received his PhD in Quantitative Economics jointly from

Roberto Savona is Associate Professor of Financial Markets and

Sameer Sharma is the Head of International Funds at Investment

fund-related research and consulting while maintaining research links

John Simpson is a Professor at Curtin University in Western Australia in

Simone Siragusa obtained his Bachelor’s degree from the University

Jérôme Teïletche is the Head of Systematic Strategies and Funds of Hedge

to this, Nils was Senior Vice-President of Synchrony Asset Management, a

Anne Vanhems is Professor of Statistics and Econometrics at Toulouse

Martin D. Wiethuechter is a doctoral research assistant at EBS Business

We examine the possibility of creating hedge fund “clones” using liquid

Hedge fund replication is a growing interest in the financial industry.

We propose a method for hedge fund replication using a factor-based

To make a truly informed investment decision, an investor needs to

We propose a new hybrid hedge fund replication technique, which com-

The objective of this chapter is to evaluate the performance of hedge

The majority of hedge fund managers modify their trading techniques

Traditional models of factor replication are based on squared error mini-

UCITS are a series of directives issued by the European Commission

transparency. For some, the importance of UCITS is inevitable and the

Limited liquidity, high management fees, and poor transparency provide

While investors generally consider hedge fund investments as pure alpha

pre-specified hurdle rate. Passive hedge fund replication can, however,

The rest of this research proceeds as follows: Section 1.2 discusses

1.2 Data and methodology

The factor-based approach is applied to the first generation of hedge

b̂jk is the estimated exposure of the return of hedge fund j to factor k,

Mean Standard deviation Sharpe Skewness Kurtosis Jacques– Q-stat

The main challenge of this factor-based process is the selection of

In sample return α ∑ (βi ABS factori ) Error term (1.3)

The ABS factors are generally selected in three ways:

1. Arbitrary factor specification. This approach could increase the risk

Out-of-sample return = ∑ ( βi factori ) (1.4)

We consider two main methods to replicate hedge fund returns:

βit1 βi ,t bi ,t1 (1.6)

1.4 Empirical results

Table 1.2 provides a comparison between the performance of fixed

Hedge fund Fixed weight clone Extended Kalman filter clone

Factors used in the Factors (final period) in the

HF _ Clonet xt′ βt1 rft , t τ 1,..., T . (2.2)

Σ Tt t1 HFt HF _ Clonet . Next, we look at hit ratios HIT (T t )

the weight attributed to observation t i , i 0,..., t 1, is equal to

Cl , t 1 denotes the metrics used to implement the mix. We concentrate

min f ( x) var( rhf ,t rp ,t ) (3.1)

LPM p rb E ⎡⎣rp | rp rb ⎤⎦ (3.16)

UPMhf E ⎡⎣rhf | rhf rb ⎤⎦ rb (3.17)

LPM hf rb E ⎡⎣rhf | rhf rb ⎤⎦ (3.18)