Facts

about
Factors

Paula Cocoma
Megan Czasonis
Mark Kritzman
David Turkington
Why use factors?

Factors may be less correlated than assets and therefore provide better
diversification.
Factors may be more effective than assets at reducing noise.

Investors may be more skilled at predicting factors than assets.

Factors may have less estimation error than assets.

Diversification
Diversification a specious argument

Assets define the opportunity set. It is impossible to create a more efficient in-sample portfolio with the
same constraints and of the same periodicity by regrouping assets into factors.

Principal components (all factors uncorrelated) Asset classes (correlations range from -0.16 to 0.82)
7% 7%

6% 6%

5% 5%

4% 4%
Excess return

Excess return
3% 3%

2% 2%

1% 1%

0% 0%
0% 5% 10% 15% 20% 25% 0% 5% 10% 15% 20% 25%
Volatility Volatility

Frontier with leverage Frontier without leverage Frontier with leverage Frontier without leverage

Notes: This analysis incorporates the following asset classes: U.S. large cap, U.S. small cap, EAFE equities, emerging equities,
global sovereigns, U.S. government bonds, U.S. corporate bonds, commodities, and hedge funds. It is based on monthly returns over
the period Jan 1990 through Dec 2013. Excess returns represent the return over the risk-free rate. All data obtained from
DataStream.
Noise Reduction
Noise reduction: Means

Does consolidating toward factors reduce noise more effectively than consolidating
toward assets?

Portfolio returns are less noisy than the returns of the assets within them due to
diversification.

However, we have already shown that efficient frontiers comprised of factors are
identical to efficient frontiers comprised of assets.

It follows, therefore, that we cannot reduce dispersion around portfolio means more
effectively by using factor groupings as opposed to asset groupings .
Noise reduction: Covariances

If asset returns are normally distributed and serially independent, the average noise in
covariances across a group of assets is no less than the average noise in covariances
across the individual assets that form that group.

For assets with non-normal distributions or serial dependence, different methods of

grouping may result in different levels of noise.

Nonetheless, there is no reason to expect factor groupings to produce less noise than
asset class groupings.
Predictability
Superior predictability an untestable hypothesis

Some investors may be more skilled at predicting factors than assets.

However, some investors may be more skilled at predicting assets than factors.
It is impossible to test these hypotheses generically; they are investor specific.
Superior predictability an untestable hypothesis

Some investors may be more skilled at predicting factors than assets.

However, some investors may be more skilled at predicting assets than factors.
It is impossible to test these hypotheses generically; they are investor specific.

Nonetheless, those who prefer to predict factors are faced with the additional
challenge of predicting how these factors map onto factor mimicking portfolios.
Estimation Error
Sources of estimation error
Small-Sample Independent-Sample Mapping Interval
Error Error Error Error

Vary: Small samples Vary: Forecasting sample Vary: Factor mapping Vary: Measurement interval
Hold constant: Hold constant: Hold constant: Hold constant:
Forecasting sample Sample size Sample size Sample size
Factor mapping Factor mapping Forecasting sample Forecasting sample
Measurement interval Measurement interval Measurement interval Factor mapping
Small-sample error

The realization of parameters from a small sample will likely differ from the
parameter values of a large sample from which it is selected.

We call this small-sample error.

2
1 ,, ,, ,, , , ,
, =
, ,
=1

When A and B are the same asset, this formula will measure the error in the
standard deviation of that asset.
Small-sample error

The realization of parameters from a small sample will likely differ from the
parameter values of a large sample from which it is selected.

We call this small-sample error.

2
1 ,, ,, ,, , , ,
, =
, ,
=1

m, j indicates monthly estimates m alone indicates monthly estimates

from a 36-month testing subsample from the full sample

When A and B are the same asset, this formula will measure the error in the
standard deviation of that asset.
Independent-sample error

The realization of parameters from a future sample will likely differ from the
parameter values of an independent historical sample.

We call this independent-sample error.

2
1 ,, ,, ,, ,,
,,
,,

, = (, )2
, ,
=1

Notes: In rare instances, small sampler error exceeds the mean squared error between two independent samples. In these cases, we
record small sample error as the mean squared error between the two samples, and record independent sample error as zero.
Independent-sample error

The realization of parameters from a future sample will likely differ from the
parameter values of an independent historical sample.

We call this independent-sample error.

2
1 ,, ,, ,, ,,
,,
,,

, = (, )2
, ,
=1

m, j indicates monthly estimates ^

m, j indicates monthly estimates
from a 36-month testing subsample from a 36-month independent
subsample immediately preceding
the testing subsample

Notes: In rare instances, small sampler error exceeds the mean squared error between two independent samples. In these cases, we
record small sample error as the mean squared error between the two samples, and record independent sample error as zero.
Mapping error

Assets define the opportunity set for investing. A desired factor exposure must be
mapped onto a portfolio of assets to be investable.

The mapping that best tracks a factor in the future will likely differ from the mapping
of an independent historical sample.

We call this mapping error.

Mapping error applies only to factors.

The calculations for small sample error and independent sample error for factors do
not reflect mapping error.
Mapping error

Mapping error can be isolated by comparing the covariance of the best-fit factor-
mimicking portfolio to the covariance of a factor-mimicking portfolio estimated from
an independent sample, holding the evaluation period constant.

2
1 ,,
,,
,,
,, ,, ,,
, =
, ,
=1
Mapping error

Mapping error can be isolated by comparing the covariance of the best-fit factor-
mimicking portfolio to the covariance of a factor-mimicking portfolio estimated from
an independent sample, holding the evaluation period constant.

2
1 ,,
,,
,,
,, ,, ,,
, =
, ,
=1

^ ^
A, B indicate factor mappings A, B indicate factor mappings
derived from the independent derived from the testing subsample
subsample
Interval error

Parameters estimated from monthly or higher-frequency returns often differ from

those estimated from lower-frequency returns, within the same sample.

We call this interval error.

Interval error

January 1990 to December 2013

Emerging Markets Stock Return 9.30%

U.S. Stock Return 9.50%

Correlation of Monthy Returns 69%

January 2005 to December 2007

Emerging Markets - US Stocks 121%

January 2011 to December 2013

Emerging Markets - US Stocks -62%

Interval error

U.S. and emerging markets stocks: monthly returns

Correlation = 0.69
20%

10%
Emerging markets equity

0%

-10%

-20%

-30%

-40%
-20% -15% -10% -5% 0% 5% 10% 15%

U.S. equity
Interval error

U.S. and emerging markets stocks: annual returns

Correlation = 0.44
100%

80%

60%
Emerging markets equity

40%

20%

0%

-20%

-40%

-60%

-80%
-60% -40% -20% 0% 20% 40% 60%

U.S. equity
Interval error

U.S. and emerging markets stocks: triennial returns

Correlation = 0.04
250%

200%
Emerging markets equity

150%

100%

50%

0%

-50%

-100%
-50% 0% 50% 100% 150%

U.S. equity
Interval error: Long-horizon and short-horizon volatility

The volatility of the cumulative continuous returns of x over q periods is given by:

( xt xt q 1 ) x q 2k 1 (q k ) x , x
q 1
t t k
Interval error: Long-horizon and short-horizon volatility

The volatility of the cumulative continuous returns of x over q periods is given by:

( xt xt q 1 ) x q 2k 1 (q k ) x , x
q 1
t t k

This term reflects annualization in the

absence of lagged effects
Interval error: Long-horizon and short-horizon volatility

The volatility of the cumulative continuous returns of x over q periods is given by:

( xt xt q 1 ) x q 2k 1 (q k ) x , x
q 1
t t k

This term captures the impact of auto-

correlation
Interval error: Long-horizon and short-horizon correlation

The correlation between the cumulative returns of x and the cumulative returns of y over q
periods is given by:

( xt xt q 1 , yt yt q 1 )

q xt , yt k 1 ( q k )( xt k , yt xt , yt k )
q 1

q 2k 1 ( q k ) xt , xt k q 2k 1 ( q k ) yt , yt k
q 1 q 1
Interval error: Long-horizon and short-horizon correlation

The correlation between the cumulative returns of x and the cumulative returns of y over q
periods is given by:

( xt xt q 1 , yt yt q 1 )

q xt , yt k 1 ( q k )( xt k , yt xt , yt k )
q 1

q 2k 1 ( q k ) xt , xt k q 2k 1 ( q k ) yt , yt k
q 1 q 1

This term captures the lagged cross-correlation

between x and y
Interval error: Long-horizon and short-horizon correlation

The correlation between the cumulative returns of x and the cumulative returns of y over q
periods is given by:

( xt xt q 1 , yt yt q 1 )

q xt , yt k 1 ( q k )( xt k , yt xt , yt k )
q 1

q 2k 1 ( q k ) xt , xt k q 2k 1 ( q k ) yt , yt k
q 1 q 1

This term captures the auto- This term captures the auto-
correlation of x correlation of y
Interval error

Interval error can be isolated by comparing a covariance matrix estimated from low-
frequency returns to a covariance matrix estimated from high-frequency returns.

2
1 ,, ,, ,, /12 ,, ,, ,,
, =
, ,
=1
Interval error

Interval error can be isolated by comparing a covariance matrix estimated from low-
frequency returns to a covariance matrix estimated from high-frequency returns.

2
1 ,, ,, ,, /12 ,, ,, ,,
, =
, ,
=1

tri, j indicates implied 3-year m, j indicates monthly estimates

estimates from a 36-month testing from the same 36-month testing
subsample sample
Total covariance error

We summarize the degree of error associated with an entire covariance matrix by

averaging the squared errors across all of the elements of the matrix and then taking
the square root.

1
= 2
(, )2

=1 =1
Composite instability score

The four sources of estimation error are independent from one another, which
means we can sum the variances of each error and then take the square root of this
sum to compute a composite instability score.

() = 2 + 2 + 2 + 2
Classifications and Data
Asset classes, fundamental factors and principal components

Asset Classes
U.S. Large Cap S&P 500
Equity
U.S. Small Cap Russell 2000
U.S. Government Bonds Barclays U.S. Aggregate Government
Fixed Income
U.S. Corporate Bonds Barclays U.S. Aggregate Corporate
Commodities S&P GSCI Commodities
Alternatives
Hedge Funds HFRI Fund of Funds Composite
Fundmanetal Factors
Inflation U.S. Consumer Price Index, Seasonally Adjusted
Macro
Growth One year ahead U.S. GDP growth forecast
Term Premium 10-year minus 2-year U.S. treasury yield
Fixed Income
Credit Premium Baa U.S. corporate yield minus 10-year U.S. treasury yield
Small Cap Premium Fama-French Small-minus-Big factor
Equity
Value Premium Fama-French High-minus-Low factor

Notes: Our analysis uses monthly data over the period Jan 1990 through Jul 2014. We proxy U.S. large cap, U.S. small cap, U.S. government bonds, U.S.
corporate bonds, commodities, and hedge funds with the following indices: S&P 500, Russell 2000, Barclays U.S. Agg Government, Barclays U.S. Agg.
Corporate, S&P GSCI Commodity, and HFRI Fund of Funds Composite, respectively. We proxy inflation, growth, term premium, credit premium, small premium,
and value premium with the following: U.S. CPI, U.S. GDP growth forecast, 10-2 yr U.S. Treasury, Baa Corp-10 yr U.S. Treasury, Fama-French SMB factor, and
Fama-French HML factor, respectively.
To analyze fundamental factors, we build six factor-mimicking portfolios by regressing each factor on the six asset classes.
To analyze principal components, we build six portfolios using principal components analysis (each portfolio corresponds to an eigenvector).
Principal components
Principal components
Principal components
Principal components
Industries, attributes and principal components

10 portfolios formed
Industries
on GICS level I

10 portfolios formed
Size
on capitalization

10 portfolios formed
Value
on book-to-market

10 portfolios formed
Momentum
on trailing 1y return

Top 10 principal
Principal Components
components

Notes: Our analysis uses monthly returns over the period Jan 1989 through Jan 2015. We narrow the universe of 400 stocks (based on the constituents in the
MSCI U.S. Index as of Jan 2015) to the 194 stocks that have full price, market cap, and book-to-market history over this sample.
To form industry portfolios, we use GICS classifications.
To form attribute portfolios, we use average market cap, book-to-market ratios, and 1-year returns, respectively.
To form principal components portfolios, we run a PCA on the 194 stocks and form portfolios corresponding to the top 10 eigenvectors.
* Because we run the PCA on 194 stocks using 36-month sub-samples, we restrict our universe to only the top 36 principal components.
Results
Asset classes, fundamental factors, and principal components

Composite Instability Score: Sources of Error

0.62

0.43 0.44
0.41 0.39

0.31 0.31
0.28 0.27
0.25
0.23

0.00

Asset Classes Fundamental Factors Principal Components

Small-Sample Error Independent-Sample Error Mapping Error Interval Error

Asset classes, fundamental factors, and principal components

Composite Instability Score: Total Covariance Error

0.86

0.68

0.58

Asset Classes Fundamental Factors Principal Components

Industries, attributes, and principal components (10 groups)

Composite Instability Score: Sources of Error

0.60

0.41
0.38 0.39
0.37
0.34
0.30
0.27 0.27
0.26

0.19

0.00

Industries Attributes (average) Principal Components

Small-Sample Error Independent-Sample Error Mapping Error Interval Error

Industries, attributes, and principal components (10 groups)

Composite Instability Score: Total Covariance Error

0.80

0.67

0.57

Industries Attributes (average) Principal Components

Noise reduction in covariances
Summary

It has become fashionable to use factors instead of assets as the building blocks for
forming portfolios.
Some have argued that factors offer greater potential for diversification, but this
argument is specious.

Others argue that factor groupings reduce noise more effectively than asset groupings,
but this too is untrue.

It is also argued that it is easier to predict factors than assets, but this argument is
generically untestable.
Nonetheless, it may be the case that factor parameters are more stationary than asset
parameters.

We find no evidence that factors produce more stable results taking into account small-
sample error, independent-sample error, mapping error, and interval error. Instead, we
find the opposite to be true.
Conclusion

Why use factors?

Conclusion

Why use factors?

In our view, the case is yet to be made that investors should use factors
rather than assets as building blocks for forming portfolios.

However, investors may be able to gather useful insights by attributing

portfolio performance to factor exposures in addition to asset exposures.

And understanding a portfolios factor exposures may help investors to

identify concentration risk and to hedge unwanted exposures, which
otherwise might go undetected.