Professional Documents
Culture Documents
ard
Aw
zi/ A ng
evy
ob al
Jac nnu
oz th nci
sL
ab e 20 ou
n F h nn
tei T A
rns
Be
JPM.iprjournals.com
QUANTITATIVE STRATEGIES: FACTOR INVESTING |
CONTRIBUTORS: DEMIR BEKTIĆ | JENNIFER BENDER | THOMAS BLACKBURN | DAVID BLITZ | JACK DAVIES | KRISTIN FERGIS
KATELYN GALLAGHER | DAVE GIBBON | TARUN GUPTA | HAN HAO | PHILIP HODGES | KED HOGAN | JOOP HUIJ | BRYAN KELLY
GEORGI KYOSEV | ASHLEY LESTER | FEIFEI LI | NONGCHAO LI | LIONEL MARTELLINI | JOHN M. MULVEY | DIRK SCHIERECK
JOSEPH (YOSEOP) SHIM | SARA SHORES | JOSEPH SIMONIAN | JOSEPHINE SMITH | TIMO SPIELMANN | XIAOLE SUN
FOURTH EDITION
EDUARD VAN GELDEREN | MILAN VIDOJEVIC | MICHAEL WEGENER | JOSEF-STEFAN WENZLER | CHENWEI WU
SPONSORED BY
| MARCH 2019
INTRODUCTION
S
ince 2016, five quantitative special and then evaluate the merits of mean–variance
issues have been published by The factor portfolios against alternative weighting
Journal of Portfolio Management focusing schemes. In the opinion of the authors, alter-
on two themes: factor investing and native weighting schemes have arguably weak
multiasset strategies. This issue is the fourth theoretical foundations. They then empiri-
special issue on factor investing. There are 11 cally show that a large part of the outper-
articles in this issue. formance of alternative weighting schemes
The article by Tarun Gupta and is attributable to a few well-known factors.
Bryan Kelly presents the robust momentum Bender, Blackburn, and Sun argue that factor
behavior in 65 widely studied characteristic- portfolios that are sensibly constructed gen-
based equity factors around the world and erate a similar or higher information ratio by
explains how this behavior can be exploited explicitly harnessing the factors in an efficient
by a time-series momentum trading strategy risk- and transaction cost–aware manner.
that scales factor exposures based on recent Three possible approaches portfolio
behavior. Their time-series momentum managers can employ in response to time-
strategy involves aggregating individual factor varying factor returns are to (1) ignore the
timing strategies and dominates all individual short-term variation, (2) develop a short-term
timing strategies. Relative to untimed fac- prediction model for generating excess
tors, factor timing employing their proposed returns by tactically changing positions, and
strategy generates economically and statisti- (3) pursue defensive factor timing. In their
cally large excess performance. Their results article, Kristin Fergis, Katelyn Gallagher,
indicate that the momentum phenomenon is Philip Hodges, and Ked Hogan describe the
driven in large part by persistence in common third approach, defensive factor timing, as
return factors and not solely by persistence in proactively using market signals to reduce
idiosyncratic stock performance. The Gupta– exposures to individual factors or a portfolio
Kelly strategy complements stock momentum of factors when performance outcomes are
because both enter optimized multifactor unattractive. The signals they use are mea-
portfolios with significant positive weights. sures of aggregate risk tolerance, the effec-
Markowitz mean–variance portfolio tiveness of diversification, and valuation
construction can be used to create efficient, measures. In contrast to well-known market
cost-effective portfolios that capture timing applications, defensive factor timing
factor exposure. Jennifer Bender, Thomas is done infrequently, seeking to preserve cap-
Blackburn, and Xiaole Sun accomplish this ital during bad times. The authors provide
AMBASSADOR BOARD
Clifford S. Asness Mohamed A. El-Erian William N. Goetzmann Bruce I. Jacobs
AQR Capital Management Allianz and Chair of President’s Yale University Jacobs Levy Equity Management, Inc.
Global Development Council
Turan Bali Campbell R. Harvey Daniel Kahneman
McDonough School of Business Robert Engle Duke University Princeton University
Georgetown University New York University, and Man Group plc
Stern School of Business Andrew W. Lo
MIT
ADVISORY BOARD
Michele Aghassi Lev Dynkin Petter N. Kolm Karthik Ramanathan
AQR Capital Management Barclays Courant Institute Consultant
Sergio M. Focardi of Mathematical Sciences, Marc R. Reinganum
Carol Alexander New York University
University of Sussex Pôle Universitaire Léonard de Vinci Senior Investment
and The Intertek Group Mark Kritzman Management Executive
Noël Amenc
EDHEC-Risk Institute Russell J. Fuller Windham Capital Management Vincenzo Russo
Fuller & Thaler Asset Management Assicurazioni Generali S.p.A.
Mark J.P. Anson Wai Lee
Commonfund Gary L. Gastineau Wells Fargo Asset Management Laurence B. Siegel
ETF Consultants.com, Inc. Research Foundation
Stan Beckers Jim Kyung-Soo Liew of CFA Institute
London Business School Chris Gowlland Johns Hopkins
Delaware Investments Carey Business School António Baldaque da Silva
Jennifer Bender BlackRock
Steven P. Greiner Marcos López de Prado
State Street Global Advisors Charles Schwab & Co., Inc. Eric Sorensen
AQR Capital Management PanAgora Asset Management
Richard Bernstein Ronald Hua and Cornell University
Richard Bernstein Advisors LLC Goldman Sachs Asset Management Meir Statman
Lionel Martellini Santa Clara University
Michele Leonardo Bianchi Brian J. Jacobsen EDHEC-Risk Institute
Bank of Italy Wells Fargo Funds Management Stoyan V. Stoyanov
Dimitris Melas Stony Brook University
David C. Blitz Robert Jones
Robeco Asset Management MSCI Radu S. Tunaru
System Two Advisors,
Arwen Advisors John Mulvey University of Kent
Gerald W. Buetow, Jr.
BFRC Services, LLC Ronald N. Kahn Princeton University M. Barton Waring
BlackRock Anna Obizhaeva Barclays Global Investors (retired)
Peter Carr
New York University Jang Ho Kim New Economic School Andrew B. Weisman
Kyung Hee University Windham Capital Management
Joseph Cerniglia Wesley Phoa
University of Pennsylvania The Capital Group Companies Robert E. Whaley
Woo Chang Kim Vanderbilt University
Denis Chaves KAIST Edward E. Qian
The Vanguard Group and Princeton University PanAgora Asset Management Jarrod Wilcox
Wilcox Investment
Jim Clayton Will Kinlaw Xiao Qiao
Cornerstone Real Estate Advisers LLC State Street Global Exchange Paraconic Technologies US Inc. Frank Zhang
and University of Connecticut Yale School of Management
C.G. (Kees) Koedijk Svetlozar T. Rachev
Kevin Dowd Tilburg School of Economics Texas Tech University Guofu Zhou
Durham University Business School and Management and FinAnalytica Washington University in St. Louis
The Journal of Portfolio Management ® (ISSN# 0095-4918) (USPS# 060-830) is published six Subscriptions: For subscription enquiries please contact a member of the team at
times a year (quarterly, plus an extra issue in the Spring and Fall) by Pageant Media, Ltd., +1 646 931 9045/+44 20 7139 1600 or at iprjournal@pageantmedia.com. Printed by The
41 Madison Avenue, 20th Floor, New York, NY 10010. Phone (646) 891-2110. Copyright Manson Group, St Albans, UK.
© 2019 Pageant Media, Ltd. POSTMASTER: Send address changes for The Journal of Port- The opinions expressed are those of the individual authors. IPR Journals is not responsible for
folio Management to Pageant Media, Ltd., 41 Madison Avenue, 20th Floor, New York, NY the accuracy, completeness, or timeliness of the information contained in the articles herein. If you
10010 USA, Att: Journals Department. Reproduction in whole or in part without written require expert advice, you should seek the services of a competent professional. No statement in
permission is prohibited. this journal is to be construed as a recommendation to buy or sell securities. Product names men-
Manuscripts: The Journal of Portfolio Management considers but assumes no responsibility for tioned in the articles in this journal may be trademarks or service marks of their respective owners.
unsolicited manuscripts. Unsolicited manuscripts will not be returned. Receipt of all manuscripts
will be acknowledged within two weeks. Instructions for submission of manuscripts are printed
at the end of the journal.
BEST ARTICLE
Behavioral Efficient Markets
Meir Statman
Winter 2018
OUTSTANDING ARTICLES
Proverbial Baskets Are Uncorrelated Risk Factors! A Factor-Based Framework
for Measuring and Managing Diversification in Multi-Asset Investment Solutions
Lionel Martellini and Vincent Milhau
Multi-Asset Strategies Issue, December 2017
Buyback Derangement Syndrome
Clifford Asness, Todd Hazelkorn, and Scott Richardson
Spring 2018
The Impact of Volatility Targeting
Campbell R. Harvey, Edward Hoyle, Russell Korgaonkar, Sandy Rattray,
Matthew Sargaison, and Otto Van Hemert
Fall 2018
*Articles authored by Frank Fabozzi were not eligible for an award. Authors were not permitted to vote for their own articles.
The ballots were tallied by Institutional Portfolio Research Journals.
QUANTITATIVE SPECIAL ISSUE 2019 VOLUME 45 NO. 3
requires an understanding of how factor characteristics index-based exchange-traded funds to private funds or
drive portfolio returns. undertakings for collective investment in transferable
securities. In this article, the authors assess the impact
On the Theory and Practice of constraints common to these investment vehicles
of Multifactor Portfolios 87 through the lens of hypothetical equity momentum and
value factor strategies. As constraints—leverage, trading
Ashley Lester frequency, and risk levels—on the factor strategies are
relaxed, risk-adjusted potential returns may improve.
Investors interested in factor investing often seek expo- Conversely, moving from an unconstrained implemen-
sure to several factors, not just one or two. The decision tation to a low-turnover, long-only implementation may
on how to implement multiple exposures may have a decrease the Sharpe ratios of momentum and value strat-
considerable effect on performance. In this article, the egies by as much as 60%.
author considers the investment choice between a set of
single-factor subportfolios and a single integrated port-
folio. He defines factor exposures in a way that precisely
Trade-Off in Multifactor
relates relative factor exposures to the number and cor- Smart Beta Investing: Factor
relation of factors and predicts analytically consequent Premium and Implementation Cost 115
changes in risk and return. Theoretically, integrated
Feifei Li and Joseph (Yoseop) Shim
portfolios increasingly outperform segregated portfo-
lios when the number of factors is large and average The authors study the impact of the inclusion of the
correlation is low. He tests his predictions by generating momentum and size factors, and the selectiveness in
over 1,000 matched pairs of historical portfolios using a stock screening, on the performance and implementa-
wide range of factor definitions. This contrasts with the tion cost of a multifactor strategy. Whereas the cost of
existing literature, which typically focuses on a handful implementing a stand-alone momentum strategy is quite
of portfolios using particular versions of factors. The high because of its extraordinarily high turnover, adding
empirical results strongly support the theoretical predic- momentum to a mix of the value, low beta, profitability,
tions. In practical terms, an integrated portfolio of four investment, and size factors helps lower tracking error
orthogonal factors generates twice the factor exposure of and improves the information ratio without a signifi-
the corresponding set of single-factor portfolios, twice cant increase in implementation cost because offsetting
the outperformance, and a 40% higher information ratio. trades cancel each other out. Contrary to conventional
wisdom, the inclusion of the size factor does not nega-
Implementation M atters: Relaxing tively affect the multifactor strategy’s trading costs
Constraints Can Improve the Potential in light of its relatively low turnover. Additionally,
Returns of Factor Strategies 101 including the size factor improves the performance and
coverage of the multifactor strategy, which otherwise
Jack Davies, Dave Gibbon, Sara Shores, would be constructed with only large and mid-size
and Josephine Smith companies. As expected, by using factor portfolios with
more concentrated holdings, investors can improve the
Eligible investors seeking factor exposures may have a performance of a multifactor strategy, but these benefits
choice of different investment vehicles to implement come at the expense of high turnover and high trading
an investment strategy, ranging from fully transparent, costs. The authors specify portfolio design as a conscious
“For many years, Marcos has led the way towards the adoption of machine
learning techniques in finance,” said Frank J. Fabozzi, Editor of JPM. “His many
publications have introduced innovative ways of thinking about financial
problems and solving them in practice. Our ‘Quant of the Year Award’
recognizes the totality of work by a researcher, and I think Marcos’ name was
in everyone’s mind from the onset of the selection process.”
QUANTITATIVE SPECIAL ISSUE 2019 VOLUME 45 NO. 3
choice made on the trade-off between the effective har- markets. Yet, the relevance of these factors in global
vesting of the factor premium and low-cost implementa- credit markets is less explored, although equities and
tion. They highlight the importance of the thoughtful bonds should be related according to structural credit
portfolio construction work required to deliver a smart risk models. In this article, the authors investigate the
beta multifactor strategy that serves investors’ needs. impact of the four Fama–French factors in the US and
European credit space. Although all factors exhibit eco-
Factor Investing from nomically and statistically significant excess returns in
Concept to Implementation 125 the US high-yield market, the authors find mixed evi-
dence for US and European investment-grade markets.
Eduard van Gelderen, Joop Huij, Nevertheless, they show that investable multifactor
and Georgi Kyosev portfolios outperform the corresponding corporate
bond benchmarks on a risk-adjusted basis. Finally, their
Mutual funds following factor investing strategies based results highlight the impact of company-level charac-
on equity asset pricing anomalies, such as the small-cap, teristics on the joint return dynamics of equities and
value, and momentum effects, earn significantly higher corporate bonds.
alphas than traditional actively managed mutual funds.
The authors report that a buy-and-hold strategy for a
random factor fund yields 110 basis points per annum
Factors in Time: Fine-Tuning
in excess of the return earned by the average traditional Hedge Fund Replication 159
actively managed mutual fund. However, they find Joseph Simonian and Chenwei Wu
that the actual returns that investors earn by investing
in factor mutual funds are significantly lower because Hedge fund replication has become a cottage industry
investors dynamically reallocate their funds both across in investing. Among the most popular hedge fund rep-
factors and factor managers. Although factor funds have lication frameworks are factor models based on ordi-
attracted significant fund f lows over their sample period, nary least squares (OLS) regression, a development that
it appears that fund f lows have been driven by factor is no doubt due to its simplicity and familiarity among
funds earning high past returns and not by the funds investment practitioners. Despite their widespread use,
providing factor exposures. The authors argue that, the OLS regression-based factor models that form the
rather than timing factors and factor managers, inves- basis for many hedge fund replication programs are often
tors would be better off by using a buy-and-hold strategy overfitted to a single sample, severely undercutting their
and selecting a multifactor manager. predictive effectiveness. As a remedy to the latter short-
coming, in this article the authors apply the regularization
Extending Fama–French Factors method known as “ridge regression” to the replica-
tion of hedge fund strategies. Ridge regression works
to Corporate Bond M arkets 141
by formally imbuing a regression with additional bias
Demir Bektić, Josef-Stefan Wenzler, in exchange for a reduction in the variance between
Michael Wegener, Dirk Schiereck, training and test samples. Using a simple yet robust meth-
and Timo Spielmann odology, the authors show how to dynamically calibrate
the predictively optimal level of bias without significantly
The explanatory power of size, value, profitability, reducing the backward-looking explanatory power of a
and investment has been extensively studied for equity given model. In doing so, the authors demonstrate that
ACTIONABLE RESEARCH—
SMARTER INVESTING
THE IPR JOURNALS PLATFORM is the go-to resource for all
independent research for those within the investment
management community. Delve further than ever
before into our content on our new global platform.
UNRESTRICTED
ACCESS
11 JOURNALS
400 ARTICLES ANNUALLY
COVERING 60+ TOPICS
A RCHIVE OF OVER
10,000 ARTICLES
P
Tarun Gupta rice momentum is most com- factor momentum and stock momentum
is a managing director at monly understood as a phenom- are correlated, they are also complementary.
AQR Capital Management
enon in which assets that recently Factor momentum earns an economically
LLC in Greenwich, CT.
tarun.gupta@aqr.com enjoyed high (low) returns relative large and statistically significant alpha after
to others are more likely to experience high controlling for stock momentum. Nor does
Bryan K elly (low) returns in the future. It is customarily factor momentum displace stock momentum.
is a vice president at AQR implemented as a cross-sectional trading Because of stock momentum’s especially
Capital Management LLC strategy among individual stocks ( Jegadeesh strong hedging benefit with respect to value,
in Greenwich, CT, and
professor of finance at Yale
and Titman 1993; Asness 1994) or long-only we find a significant benefit to combining
School of Management in equity portfolios (Moskowitz and Grinblatt factor momentum, stock momentum, and
New Haven, CT. 1999; Lewellen 2002). It has an impressive value in the same portfolio. 1
bryan.kelly@aqr.com and robust track record of risk-adjusted per- In recent decades, academic literature
formance (Asness et al. 2014; Geczy and and industry practice have accumulated
Samonov 2016). dozens of factors that help explain the co-
Grouping stocks based on relative cross- movement and average returns among indi-
section performance has led many to inter- vidual stocks. We build and analyze a large
pret momentum as a strategy that isolates collection of 65 such characteristic-based fac-
predominantly idiosyncratic momentum tors that are widely studied in the academic
(e.g., Grundy and Martin 2001; Chaves literature. From this dataset, we establish
2016). In this article, we document robust factor momentum as a robust and pervasive
momentum behavior among the common phenomenon based on the following facts.
factors that are responsible for a large fraction Serial correlation in returns is the
of the covariation among stocks. A portfolio basic statistical phenomenon underlying
strategy that buys the recent top-performing momentum and is thus the launching point
factors and sells poor-performing factors (i.e., for our analysis. First, we show that indi-
that exploits factor momentum) achieves sig- vidual factors exhibit robust time-series
nificant investment performance above and momentum (Moskowitz, Ooi, and Pedersen
beyond traditional stock momentum. On 2012), a performance persistence phenom-
a standalone basis, our factor momentum enon by which an asset’s own recent return
strategy outperforms stock momentum,
industry momentum, value, and other com- 1
This is especially true when value is con-
monly studied investment factors in terms structed following the HML-Devil refinement of
of Sharpe ratio. Furthermore, although Asness and Frazzini (2013).
5(96853
'&$3
'3
63
&3
(3
;5'$7
,92/
6($621$/
2/
6+257,17
0$;5(7
(1708/7
'(/$<
$'09
$%1,19
;5'09
,6.(:
52$
12$
$'9(57&+*
30
(&216
/9*%0
(36020
&$6+
30&+*
*6
)6&25(
;5'6
$72&+*
52(
60%
=6&25(
$,0
92/0751
&(*7+
/9*
26&25(
&$372
$&&
1:&&+*
*5352)
:+
/12$&+*
33$&+*
$66(7*
%0
,668(
,668(
(17%0
$72
1&2$&+*
,9*
40-
)5((&)
;),1
%$%
,1'020
020
/75
785129(5
5(5
686*5
675
,668(
(17%0
±
(1708/7
(3
(36020
3DQHO%6KDUSH5DWLR
)5((&)
'3
(3
63
&3
,92/
;5'$7
6($621$/
2/
0$;5(7
6+257,17
(1708/7
'(/$<
$'09
,6.(:
$%1,19
;5'09
52$
12$
$'9(57&+*
30
(&216
/9*%0
(36020
;5'6
*6
30&+*
60%
/9*
92/0751
:+
=6&25(
&$6+
52(
$,0
)6&25(
$72&+*
&$372
&(*7+
26&25(
*5352)
(17%0
%0
,668(
$72
$&&
,668(
$66(7*
,1'020
1:&&+*
40-
33$&+*
020
)5((&)
%$%
;),1
/12$&+*
,9*
1&2$&+*
/75
785129(5
686*5
5(5
675
63
(3
5(96853
&3
'&$3
(1708/7
,92/
2/
6($621$/
0$;5(7
6+257,17
;5'$7
$'09
'(/$<
52$
$%1,19
12$
;5'09
,6.(:
/9*
$'9(57&+*
*6
(17%0
(&216
92/0751
%0
60%
,668(
30
&(*7+
$66(7*
$,0
33$&+*
&$6+
52(
)6&25(
30&+*
&$372
$72&+*
(36020
26&25(
1&2$&+*
,668(
/12$&+*
,9*
=6&25(
;),1
$&&
$72
)5((&)
*5352)
1:&&+*
:+
40-
/9*%0
%$%
;5'6
,1'020
020
/75
686*5
785129(5
5(5
675
Note: * and ** signify that a weight estimate is significantly different from zero at the 5% or 1% level, respectively.
j ,t +1 = si , j ,t × f i ,t +1 ,
f i ,TSFM
FACTOR MOMENTUM
1 j
Factor Persistence si , j ,t = min max
σ i , j ,t
∑f i ,t −τ+1 , −2 ,2 (1)
τ=1
We begin our analysis by investigating the pri-
mary statistical phenomenon underlying momentum: Unpacking Equation 1, we use the scaling term si,j,t
serial correlation in returns. In Exhibit 2, we report to time positions in factor i based on the factor’s return
monthly first-order autoregressive coefficients (denoted over the formation period (t − j to t). If formation returns
as AR(1)) for each factor portfolio along with 95% con- are positive, it buys the factor; if negative, it sells the
fidence intervals. When zero lies outside the confidence factor. We convert recent returns to z-scores by dividing
interval, it indicates that the estimate is statistically sig- by σi,j,t , which is the annualized factor volatility, over
nificant at the 5% level (or, equivalently, the t-statistic the previous three years (for short formation windows,
is greater than 1.96 in absolute value). j < 12) or over the previous 10 years (if j ≥ 12), and we
The strength and pervasiveness of one-month cap z-scores at ±2.4
own-factor serial correlation is stunning. Of our 65 fac- The benefits of factor timing can be assessed in
tors, 59 have a positive monthly AR(1) coefficient, and terms of alpha by regressing the scaled factor on the
the coefficient is statistically significant for 49 of these. raw factor
For comparison, the monthly AR(1) coefficient for the
excess market return is 0.07 during our sample, which f i ,TSFM
j ,t = α i , j + βi , j f i ,t + ei , j ,t
Moskowitz, Ooi, and Pedersen (2012) demonstrated
is powerful enough for implementing a time-series Exhibit 3, Panel A, reports the annualized per-
momentum strategy. The average AR(1) coefficient of centage alphas from the time-series strategy with a one-
our factors is 0.11, and 50 of them have an AR(1) coef- month formation period for each factor, as well as 95%
ficient larger than that of the market. 3 This is a first confidence intervals. The performance of time-series
indication that it may be possible to time factors based momentum in individual factors is extraordinarily per-
on their own past performance. vasive. It is positive for 61 out of 65 factors and statistically
significant for 47 of these. To provide a clearer inter-
pretation in terms of risk–return trade-off, Exhibit 3,
3
We believe that own-factor persistence may be even stronger
than these results portray because any illiquidity imbalance in a Our findings are robust to other estimation choices for σi,j,t ,
4
factor will tend to create some negative serial correlation, and we including shorter windows and exponentially weighted moving
are not directly accounting for that here. averages, and to other caps such as ±1.
±
±
,668(
*6
,668(
60%
;5'6
%$%
(&216
020
*5352)
)5((&)
$72
&$372
'&$3
&3
(3
5(96853
$&&
63
'3
(1708/7
6+257,17
,1'020
;5'$7
26&25(
:+
0$;5(7
)6&25(
$,0
&(*7+
(36020
=6&25(
52(
&$6+
,92/
6($621$/
30
%0
2/
/9*%0
(17%0
30&+*
;),1
$'9(57&+*
'(/$<
$72&+*
/12$&+*
33$&+*
;5'09
92/0751
$%1,19
,9*
/9*
$66(7*
1:&&+*
$'09
12$
1&2$&+*
52$
,6.(:
40-
/75
5(5
675
785129(5
686*5
Panel B, shows Sharpe ratios for each individual factor TSFM earns an annualized average return of 12.0%.
momentum strategy: It exceeds 0.20 for 56 factors and The last bar in Panel A reports the alpha from regressing
is statistically significant for 48 of them. the one-month TSFM return on the equal-weighted
Our overall TSFM strategy combines all individual average of raw factor returns. The equal-weighted
factor time-series momentum strategies into a single portfolio of raw factors is itself an impressive strategy,
portfolio. In particular, TSFM aggregates timed factors earning an annualized Sharpe ratio of 1.07. Nevertheless,
(with formation window j) according to the portfolio of individual factor momentum strategies
generates a highly significant 10.3% alpha (t-statistic of
TSFM j ,t = TSFM Long
j ,t − TSFM Short
j ,t 4.6) after controlling for the average of untimed factors.
The last bar in Panel B reports the annual Sharpe ratio
where of the combined factor momentum portfolio. It is 0.84,
exceeding the Sharpe ratio of every individual factor
TSFM Long =
∑1 f TSFM
i { si , j ,t >0} i , j ,t +1
and
momentum strategy.
Exhibit 4 explores how TSFM performance
j ,t
∑1 s
i { si , j ,t >0} i , j ,t changes with alternative implementations. We form
TSFM Short
=
∑1 f TSFM
i { si , j ,t ≤0} i , j ,t +1
the momentum signal using look-back windows of one
month (1-1) up to five years (1-60). We also split out
j ,t
∑1 s
i { si , j ,t ≤0} i , j ,t the 11 months excluding the most recent month (2-12)
for more direct comparability with UMD and the four
That is, the long and short legs are rescaled to form years excluding the most recent year (13-60) to compare
a unit leverage ($1 long and $1 short) TSFM portfolio. the role of long-term versus short-term return trends.
±
±
$,0
$'9(57&+*
$%1,19
'(/$<
,6.(:
6+257,17
&$6+
&(*7+
)6&25(
/12$&+*
12$
52$
92/0751
26&25(
30
(1708/7
*6
(36020
;),1
&3
*5352)
$66(7*
,668(
;5'6
&$372
)5((&)
(3
:+
$72
;5'$7
/9*%0
2/
,1'020
$72&+*
30&+*
1&2$&+*
$&&
5(96853
1:&&+*
,668(
;5'09
,9*
'&$3
6($621$/
33$&+*
(&216
40-
60%
$'09
020
52(
%0
/9*
(17%0
63
=6&25(
'3
0$;5(7
%$%
,92/
76)0
5(5
686*5
675
785129(5
/75
3DQHO%$QQXDOL]HG6KDUSH5DWLR
±
±
±
$,0
$'9(57&+*
$%1,19
'(/$<
,6.(:
6+257,17
;5'09
'&$3
92/0751
&(*7+
,1'020
)6&25(
52$
&$6+
:+
;5'6
6($621$/
5(96853
/12$&+*
;5'$7
12$
$72&+*
30&+*
,668(
26&25(
/9*%0
020
$&&
,668(
*5352)
33$&+*
30
$'09
&3
(3
(&216
1&2$&+*
60%
'3
*6
(1708/7
,92/
/9*
1:&&+*
)5((&)
0$;5(7
40-
$72
;),1
$66(7*
&$372
,9*
(36020
%0
52(
63
(17%0
=6&25(
2/
%$%
76)0
5(5
675
785129(5
/75
686*5
Panel A reports the raw TSFM average return for With a look-back as long as five years, TSFM earns
each formation period, and Panel B reports annualized 7.1% per annum with a Sharpe ratio of 0.72. Panel A
Sharpe ratios. The 12-month TSFM strategy achieves shows that although one-month factor momentum is the
an expected return of 9.5% and a Sharpe ratio of 0.70. overall performance driver, large positive and significant
3DQHO&αRQ(:)DFWRU 3DQHO'αRQ))
±
0RQWKVLQ)RUPDWLRQ3HULRG 0RQWKVLQ)RUPDWLRQ3HULRG
contributions from longer (nonoverlapping) 2-12 and section of all factors. CSFM buys (sells) factors that have
13-60 formation windows remain as well, with Sharpe recently outperformed (underperformed) peers, rather
ratios of 0.54 and 0.53, respectively. than sizing factor exposures based on their own recent
When we benchmark TSFM against the equal- performance. For example, if all factors recently appre-
weighted average factor (EW, shown in Panel C) or ciated, TSFM will take long positions in all of them.
against the Fama–French f ive-factor model (FF5, CSFM, on the other hand, will be long only the relative
Panel D), the excess performance of TSFM is little outperformers and will short those with below-median
affected. For a one-month formation, EW explains less recent returns (despite their recent positivity).5
than one-sixth of the TSFM average return, and at one Exhibit 5 explores the performance of CSFM with
year it explains less than one-third. The Fama–French various look-back windows for portfolio formation (in
model explains less than one-tenth of TSFM’s average analogy with Exhibit 4). The results show that CSFM
return for all formation windows. and TSFM have similar behavior. The Sharpe ratios of
CSFM are slightly inferior to TSFM, and it has slightly
Cross-Section Factor Momentum smaller alphas with respect to the equal-weighted
±
0RQWKVLQ)RUPDWLRQ3HULRG 0RQWKVLQ)RUPDWLRQ3HULRG
3DQHO&αRQ(:)DFWRU 3DQHO'αRQ))
0RQWKVLQ)RUPDWLRQ3HULRG 0RQWKVLQ)RUPDWLRQ3HULRG
portfolio of raw factors, but their performance patterns returns for each momentum variable, including 1-month
are otherwise closely aligned. and 12-month look-back windows for TSFM and
CSFM, along with the excess market portfolio. Two
Factor, Stock, and Industry Momentum features of this plot stand out. First is the comparatively
steep slope of TSFM. This is consistent throughout the
Next, we directly compare various incarnations of sample rather than being driven by a few good runs.
the momentum effect against each other, including factor (One-month CSFM shares a similarly steep slope, but
momentum (TSFM and CSFM), stock-level momentum the 12-month implementation drops off substantially).
(UMD), short-term stock reversal (STR), and industry Second is the sharp drawdown of UMD, when stock
momentum (INDMOM, following Moskowitz and momentum experienced a loss of 31% from March to
Grinblatt 1999 and Asness, Porter, and Stevens 2000, May 2009 (Daniel and Moskowitz 2016). INDMOM also
for which we use a 1-12 formation strategy). To make a experienced a drawdown of 24% over this time. In con-
clearer comparison among average returns, we rescale trast, factor momentum entirely avoided the momentum
all five series to have an ex post annualized volatility crash. Over the same months, 12-month TSFM and
of 10%. CSFM earned 16% and 15%, respectively (one-month
Exhibit 6 provides a preliminary visual comparison versions of TSFM and CSFM both earned 18%).
of momentum strategies. It shows the cumulative log
,1'020
5P5I
±
Exhibit 7
Momentum Strategy Return Correlations
It is well known that stock momentum is con- momentum from a 2-12 strategy, as well as STR, which
centrated in intermediate formation windows of 6 to captures short-term stock reversals that arise in a 1-1
12 months. With very short look-backs (one month) strategy. We compare each of these to TSFM and CSFM
or at long horizons, stocks experience reversals rather with a range of formation choices ranging from 1 month
than momentum. To gain a basic understanding of to 60 months and again splitting out 2-12 and 13-60.
co-movement in strategies, particularly with respect to Exhibit 7 highlights an interesting distinction in the
different formation periods, Exhibit 7 reports momentum time series dynamics of different momentum strategies.
correlations. We include UMD, which describes stock When factor momentum is based on an intermediate
$QQXDO
±
±
±
(>5@ αYV80' αYV,1'020 αYV675 αYV&6)0
3DQHO%&6)05HODWLYH3HUIRUPDQFH
$QQXDO
±
±
±
(>5@ αYV80' αYV,1'020 αYV675 αYV76)0
±
αYV76)0 αYV&6)0
3DQHO%,1'020
$QQXDO
±
±
αYV76)0 αYV&6)0
3DQHO&675
$QQXDO
αYV76)0 αYV&6)0
,QGLY 7DQJHQF\3RUWIROLR:HLJKWV
)DFWRU 6KDUSH
76)0
76)0
±
76)0
76)0
&6)0
&6)0
&6)0
&6)0 ±
80'
675
0.7
60%
+0/
50:
&0$
6KDUSH
Note: * and ** signify that a weight estimate is significantly different from zero at the 5% or 1% level, respectively.
1-1, 2-12, and 13-60 versions of TSFM and CSFM, Column 4 considers the optimal combination of
as well as the 1-12 versions of each. We also include TSFM with UMD and the Fama–French factors. In
UMD, INDMOM, and STR. Finally we investigate this case, 2-12 TSFM takes an exact zero weight and
combinations with the Fama–French five-factor model. is replaced by a significantly positive weight of 0.10 on
Superscripts of single and double asterisks signify that UMD. This combination earns a Sharpe ratio of 1.65
a weight estimate is significantly different from zero at (the five Fama–French factors on their own achieve a
the 5% or 1% level, respectively. tangency Sharpe ratio of 1.09). The same conclusion
The second column reports the standalone Sharpe emerges if we simultaneously include UMD and STR
ratios for each factor. The remaining columns, labeled 1 alongside TSFM (Column 6), where all three enter the
to 7, report tangency portfolio weights among various tangency portfolio with significantly positive weights.
sets of factors. Column 1 shows that the ex post efficient Among the Fama–French factors, MKT, conserva-
combination of 1-1, 2-12, and 13-60 TSFM puts the tive minus aggressive (CMA), and robust minus weak
heaviest weight (0.47) on 1-1 TSFM but also puts sig- (RMW) are significant contributors to tangency across
nificantly positive weight on 2-12 and 13-60 (0.22 and the board.
0.31, respectively). This tangency combination achieves The diversif ication benef its from combining
a Sharpe ratio of 1.07. For CSFM, the tangency portfolio momentum factors with value factors become more
is dominated by a weight of 0.67 on the 1-1 component. pronounced when using the HML-Devil refinement
Column 3 shows that the optimal combination of TSFM of Asness and Frazzini (2013), which incorporates more
and CSFM takes a highly levered position in TSFM timely price data in its value signal construction and
with a large negative offsetting position in CSFM. This significantly outperforms the traditional Fama–French
result restates the fact that TSFM and CSFM are highly HML. Exhibit 11 shows that the correlation of UMD
correlated but have alphas of opposite signs with respect and HML-Devil is −0.64, whereas UMD is only −0.18
to one another. correlated with Fama–French HML. Likewise, the
76)0 &6)0
80' ,1'020
80' ±
,1'020 ± ±
+0/ ± ± ± ± ± ± ± ±
+0/'HYLO ± ± ± ± ± ± ± ± ±
correlation of 1-12 TSFM drops from −0.02 with HML momentum with that of other price trend factors.7
to −0.37 with HML-Devil. In terms of formation window, STR is the natural stock-
Motivated by the potential for stronger hedging level benchmark for one-month factor momentum, and
benefits, Exhibit 12 investigates the impact of replacing UMD and INDMOM are most natural for comparison
HML with HML-Devil in our tangency portfolio anal- with 12-month factor momentum. In both cases we see
ysis. Three observations emerge from this table. First, that factor momentum turnover is comparable to, but
our central conclusions regarding factor momentum are slightly lower than, its stock-level counterpart. Panel A
unchanged—it remains a strong contributor to optimal also shows that, like other momentum varieties, factor
multifactor portfolios. Second, HML-Devil takes a momentum involves substantially more trading than
large and statistically significant portfolio weight in Fama–French factors.
all cases, in contrast with the general insignificance Panel B of Exhibit 13 compares the performance
of Fama–French HML in Exhibit 10. Third, UMD of strategies net of transaction costs. Our calculations
becomes one of the most important components of assume costs of 10 bps per unit of turnover (based on
the tangency portfolio thanks to the added diversifica- the estimates of Frazzini, Israel, and Moskowitz 2015).
tion benefits of coupling UMD and HML-Devil.6 In Red bars represent the net annualized Sharpe ratio for
summary, factor momentum and stock momentum are each strategy, along with the gross Sharpe ratio in blue
most effectively used in tandem when devising optimal for comparison. The main takeaway from the exhibit is
portfolios. that although trading costs indeed eat into the perfor-
mance of factor momentum, its net performance con-
IMPLEMENTABILITY tinues to exceed that of UMD, INDMOM, STR, and
the Fama–French factors. For example, the net Sharpe
Momentum strategies are high turnover by ratio of TSFM 1-12 is 0.63, versus 0.70 gross. The next
nature; thus, trading costs are a first-order consider- best net Sharpe ratio among stock-level price trend fac-
ation for understanding the practical usefulness of factor tors is 0.51 for UMD, and the best among Fama–French
momentum in portfolio decisions. Panel A of Exhibit 13 factors is 0.45 for RMW.
compares the average annualized turnover of factor Lastly, Panel B sheds new light on the findings in
Exhibit 9: It reveals that the strong performance of STR
after controlling for factor momentum is illusory. Even
on a standalone basis, the performance of short-term
6
Exhibit 12 highlights the benefits of combining value and reversal is entirely wiped out by transaction costs.
momentum strategies (a point previously emphasized by Asness,
Moskowitz, and Pedersen 2013). Our factor momentum findings
naturally call for an investigation into an analogous factor value
7
strategy that times factors based on factor-level value signals (as Average annualized turnover is defined as the sum of abso-
discussed, for example, by Asness 2016a, 2016b and Asness et al. lute changes in portfolio weights each month, averaged over all
2000). Although an exploration of factor value, and in particular months and multiplied by 12. This describes total two-sided trading
the benefits of combining it with factor momentum, is beyond the volume (both entering and exiting positions) as a fraction of gross
scope of this article, it is a fascinating direction for future research. asset value.
,QGLY 7DQJHQF\3RUWIROLR:HLJKWV
)DFWRU 6KDUSH
76)0
76)0
±
76)0
76)0
80'
,1'020
+0/'HYLO
0.7
60%
50:
&0$ ±
6KDUSH
Note: * and ** signify that a weight estimate is significantly different from zero at the 5% or 1% level, respectively.
FACTOR MOMENTUM AROUND THE WORLD portfolio that aggregates individual time-series factor
momentum strategies has a Sharpe ratio of 0.73 (versus
In this section, we show that each of our main 0.84 in the United States) and earns an alpha of 6.6%
factor momentum conclusions from the US sample is per year after controlling for the equal-weighted port-
strongly corroborated in international equity markets. folio of raw (untimed) factors.
We study three international samples. The Europe Second, international factor momentum demon-
sample includes Austria, Belgium, Switzerland, strates extraordinarily stable performance regardless
Germany, Denmark, Spain, Finland, France, the of formation window (shown in the left-most bars of
United Kingdom, Greece, Ireland, Italy, Netherlands, Exhibit 15). TSFM and CSFM earn essentially the same
Norway, Portugal, Sweden, and Israel. The Pacific average return whether they use a short look-back of
sample includes Australia, Hong Kong, Japan, New one month, all the way through a long look-back of five
Zealand, and Singapore. The broadest international years. As in the US sample, this is a remarkable diver-
sample we consider is global (without the United States) gence from stock-level continuation patterns, where a
and combines Europe, Pacific, and Canada. Because of one-month window gives rise to reversal but a one-year
data limitations, we study only 62 of the original 65 window captures momentum.
factors in the international sample. 8 Third, international factor momentum demon-
First, individual factor returns are highly per- strates large and significant excess performance after con-
sistent. The average AR(1) coefficient is 0.10 (versus trolling for other varieties of international momentum,
0.11 in the United States), is positive for 51 of 62 fac- including UMD, INDMOM, and STR (Exhibit 15).
tors, and is signif icant for 30 of these. Exhibit 14 The TSFM alpha versus UMD is significantly positive
shows that the success of individual factor time-series for all formation windows except 13-60.
momentum strategies (one-month formation) work as Fourth, TSFM and CSFM are more than 0.95 cor-
well for international factors as they do for the United related for all formation windows. Yet TSFM tends to
States. The alpha of momentum-timed factors versus possess positive alpha relative to CSFM, and CSFM earns
raw factors is positive for 55 of 61 factors and is sig- negative alpha versus TSFM; this finding indicates that,
nificant for 22 of these (versus 61 of 65 positive in as in the US sample, TSFM more efficiently captures the
the United States, 47 of those significant). The TSFM benefits of factor momentum.
8
Excluded are ADVERTCHG, AD2MV, and AIM.
$YHUDJH$QQXDO7XUQRYHU
$
/
5
'
%
7
0
:
0
&0
67
60
0
.
50
+
0
)0
)0
0
)0
)0
'
76
&6
,1
76
&6
3DQHO%6KDUSH5DWLR
$
/
5
'
%
7
0
:
0
&0
67
60
0
.
50
+
0
0
)0
0
)0
)0
)
'
76
&6
,1
76
&6
*URVV 1HW
Fifth, the performance of UMD and INDMOM from Exhibit 17 are qualitatively similar to the US
is explained by factor momentum. Exhibit 16 shows that analysis in Exhibit 10. US and international TSFM 1-1
UMD’s alpha is essentially zero, and INDMOM has a share a correlation of 0.62, and the US and international
negative alpha after controlling for either TSFM or CSFM. 1-12 versions are 0.64 correlated. The ex post tangency
Sixth, international tangency portfolio analysis in portfolio that combines US and international TSFM 1-1
Exhibit 17 highlights the additivity of factor momentum earns an annual Sharpe ratio of 0.83; individually, they
to the broader set of investment factors.9 The conclusions each earn 0.73.
In further (unreported) robustness analyses, we
9
All momentum variables are based on international equi- find that the majority of the performance of the factor
ties. However, because our data begin earlier than Ken French’s momentum strategy arises from dynamically adjusting
international five-factor data, we use the US Fama–French factors.
;5'6 $&&
675 1:&&+*
$72&+* &$6+
1:&&+* 675
$&& ;5'6
,9* 92/0751
'(/$< 5(96853
*5352) '(/$<
785129(5 $72&+*
30&+* &(*7+
&(*7+ 33$&+*
3DQHO%$QQXDOL]HG6KDUSH5DWLR
'&$3 30&+*
;5'$7 12$
60% '&$3
/12$&+* $72
&$372 2/
*6 /12$&+*
2/ (36020
52( 1&2$&+*
33$&+* *6
;),1 52(
12$ ,6.(:
;5'09 785129(5
$%1,19 60%
26&25( 686*5
30 $%1,19
6($621$/ 26&25(
686*5 30
(36020 $66(7*
1&2$&+* ;),1
(3 ;5'$7
$66(7* (&216
/9*%0 /9*%0
52$ ;5'09
,1'020 52$
(1708/7 )5((&)
0$;5(7 &3
,6.(: (3
(&216 (1708/7
:+ %0
,92/ ,668(
/75 =6&25(
=6&25( /9*
'3 )6&25(
/9* '3
%0 ,668(
)5((&) (17%0
)6&25( 63
63 %$%
(17%0 /75
%$% :+
Time Series Momentum for Individual Factors (global ex. US, one-month formation window)
&3 ,92/
020 40-
,668( ,1'020
40- 0$;5(7
,668( 020
76)0 76)0
12
10
6
Annual %
–2
–4
–6
E[R] α vs. UMD α vs. INDMOM α vs. STR α vs. CSFM
12
10
6
Annual %
–2
–4
–6
E[R] α vs. UMD α vs. INDMOM α vs. STR α vs. TSFM
6
Annual %
–2
–4
α vs. TSFM α vs. CSFM
Panel B: INDMOM
8
4
Annual %
–2
–4
–6
α vs. TSFM α vs. CSFM
Panel C: STR
8
4
Annual %
–2
–4
α vs. TSFM α vs. CSFM
7DQJHQF\3RUWIROLR:HLJKWV
,QGLY
)DFWRU 6KDUSH
76)0
76)0
76)0
76)0
&6)0
&6)0
&6)0 ±
&6)0 ±
80'
675 ±
0.7
60% ±
+0/ ± ± ±
50:
&0$
6KDUSH
Note: * and ** signify that a weight estimate is significantly different from zero at the 5% or 1% level, respectively.
factor weights over time, rather than from taking static complementary with stock momentum, as both enter
long (short) bets on factors that have higher (lower) optimized multifactor portfolios with signif icant
average returns unconditionally. We also find that the positive weights (particularly when combined with
performance of factor momentum is not dependent on HML-Devil).
using dozens of fine-grained factors. Instead, with a set An interesting aspect of factor momentum is its
of only six broad theme factors,10 we reproduce the same stability with respect to the definition of recent per-
basic factor momentum phenomenon found in the 65 formance. Whether the look-back window is as short
factor dataset. as one month or as long as five years, our strategy
identifies large positive momentum among factors.
CONCLUSION This contrasts sharply with stock momentum, which
exhibits reversal with respect to recent one-month
We document robust persistence in the returns performance, momentum at intermediate horizons of
of equity factor portfolios. This persistence is exploit- around one year and again reversal for windows beyond
able with a time-series momentum trading strategy two years.
that scales factor exposures up and down in propor- Factor momentum is a truly global phenom-
tion to their recent performance. Factor timing in this enon. It manifests equally strongly outside the United
manner produces economically and statistically large States, both in a large global (excluding the United
excess performance relative to untimed factors. We States) sample and f iner Europe and Pacif ic region
aggregate individual factor timing strategies into a subsamples.
combined time-series factor momentum strategy that Taken alongside the evidence of time series
dominates all individual timing strategies. TSFM is momentum in commodity, bond, and currency factors
(Moskowitz, Ooi, and Pedersen 2012), our findings of
10 momentum among equity factors—in the time series,
The six theme factors are valuation, momentum, earnings
quality, sustainable growth, management, and risk. Each theme in the cross section, and around the world—support the
aggregates a set of closely related subfactors—for example, valuation conclusion that factor momentum is a pervasive phe-
includes book-to-market, earnings-to-price, and dividend yield. nomenon in financial markets.
(continued)
AR = The Accounting Review; BAR = The British Accounting Review; CFR = Critical Finance Review; CMBA = The Chicago MBA: A Journal
of Selected Papers; JAE = Journal of Accounting and Economics; JAR = Journal of Accounting Research; JF = The Journal of Finance; JFE = Journal
of Financial Economics; JFM = Journal of Financial Markets; JFQA = Journal of Financial and Quantitative Analysis; JFR = Journal of Financial
Research; ROF = Review of Finance; RFS = The Review of Financial Studies; WP = unpublished paper.
COMPANY or DEPARTMENT
NEED ACCESS to
the JOURNALS?
We offer IP authenticated access to our
content which will allow your team to find
the research and insights they need.
FOR INFORMATION ON HOW TO CUSTOMIZE A PLAN TO MEET YOUR NEEDS AND BUDGET
CONTACT A REPRESENTATIVE at (646) 891-2154 or journals-sales@pageantmedia.com.
N
Jennifer Bender ew portfolio construction ideas Moore 2005), which weights securities by
is a senior managing have proliferated rapidly in fundamentals such as book value or earn-
director at State Street
the last decade. These include ings instead of by market cap. The intriguing
Global Advisors in Boston,
MA. numerous contestants, from part was the fact that fundamental indexing
jennifer_bender@ssga.com fundamental weighting to minimum vari- exhibited phenomenal performance in back-
ance to risk parity to maximum diversifi- tests against market-cap-weighting schemes.
T homas Blackburn cation. But too much of a good thing has From there, a cottage industry exploded,
is an assistant vice created confusion for investors now at risk including risk parity or equal risk contribution
president at State Street
Global Advisors in Boston,
of embracing methods without a full under- (weighting based on risk contribution), not to
MA. standing of their potential limitations. be confused with risk weighting (weighting
tj_blackburn@ssga.com Fifty years ago, times were much sim- based on risk), equal weighting (weighting
pler. Investors sought to maximize wealth/ based on the number of securities), diversity
X iaole Sun return while minimizing the volatility of weighting, and many others. Just to make
is a vice president at State
that return stream. The main question was matters more confusing, some methods such
Street Global Advisors in
Boston, MA. how to define this utility function. The as minimum variance and maximum Sharpe
xiaole_sun@ssga.com mean–variance framework (Markowitz ratio were many times incorrectly labeled as
1959) emerged as the most prevalent model “new” weighting schemes when in fact they
for capturing investor preferences. These were variants of the original mean–variance
return-maximizing investors sought sources paradigm.
of return that could improve return while The time is ripe for sanity to be restored.
reducing risk. Stock pickers looked at qualita- It is well known that the theoretical founda-
tive aspects of individual companies, whereas tions for alternative weighting schemes are
quantitative investors gravitated toward fac- weak. Supporters have pointed instead to the
tors—for example, security characteristics empirical evidence, mainly how well these
that were shown by the academics to be able weighting schemes perform in backtests.
to explain the cross section of stock returns Here, we isolate the return drivers behind the
across broad universes. most popular alternative weighting schemes
A curious thing happened in the and show that they come from a handful of
mid-2000s. Investors, possibly tired of well-known factors.
boom–bust equity market cycles, started to Beating the market is not easy and never
challenge the idea of market-cap-weighted has been. Identifying robust systematic and
benchmarks. One of the first challengers repeatable sources of return has always been
was fundamental indexing (Arnott, Hsu, and the objective of investors and should continue
38 Clash of the Titans: Factor Portfolios versus Alternative Weighting Schemes Quantitative Special Issue 2019
to be. Investors should refocus on isolating sources of the expense of factor exposure. As investors increasingly
return that are rewarded and building portfolios that recognize these limitations, this opens the door to opti-
capture these sources in the most risk-efficient and cost- mization-based methods that can balance competing
efficient way. Separating the source of return from how objectives.1
to build the portfolio is a critical distinction. Factor Our preferred algorithmic framework is based on
portfolios do exactly that. mean–variance optimization. The framework has been
an asset allocation workhorse for decades. Mean–vari-
AN OLDIE BUT GOODIE: EXTENDING ance assumes investors care only about the first two
THE MEAN–VARIANCE FRAMEWORK moments of portfolio return—return and risk. (Mean–
TO FACTOR PORTFOLIOS variance itself was a special case of Markowitz’s 1952
original expected utility maximization framework, it
Ross (1976) was among the first to note that one being consistent with the latter under the assumption
way to understand the returns to stocks was to model that either investors have quadratic utility or returns are
them as a function of exposures to various factors, which normally distributed).
could be thought of as attributes relating to a set of First-generation quant models, as by Grinold and
securities’ returns. Although a wide range of factors Kahn (1994), in fact used a mean–variance framework
has been proposed (macroeconomic factors, statistical with factors, much the same as in factor portfolios today.
factors, technical factors, and fundamental factors, to Thus, we are not claiming to invent anything new; we
name a few), the most widely cited today are those pro- are merely reviving a well-known framework in the
duced by the seminal research of Fama and French (1992, context of harnessing factor premiums.
1993)—value, size—and extended by Carhart (1997) to We define an objective function that maximizes
momentum. factor exposure scaled by risk aversion. That is, the
The term factor portfolios, or factor investing, is rela- portfolio objective is to maximize exposure to intended
tively newer, born of the growth of smart beta in the last factor(s) in a risk-aware way. As first proposed by Bender,
decade. Smart beta started with simple rules-based ways Sun, and Wang (2016), a factor portfolio should seek to
to capture factors. In fact, fundamental indexing is often deliver as much exposure to the targeted factor(s) as pos-
viewed as one of the first examples of smart beta. First- sible while minimizing risk and ensuring the portfolio is
generation smart beta used simple rules-based portfolio sufficiently liquid and not too expensive to trade.
construction because investors preferred their simplicity The traditional return–risk objective function can
and ease of understanding. (One of the benefits of the be summarized as
rules-based approach is a direct link between each secu-
rity’s factor profile and weight. Rules-based portfolios Max w ′r − λw ′ ∑ w (1)
overweight stocks that rank highly on the factor and
underweight stocks that rank low on the factor.) For our in which w is the vector of optimal weights, r is the
own research on smart beta rules-based strategies, see vector of expected returns for all securities, l is the risk
Bender, Brandhorst, and Wang (2014) and Bender and aversion, and S is the covariance matrix for all securities.
Wang (2015) for example. Applying this framework to factor models is a natural
We have long recognized the limitations to extension because factor investing rests on the belief
capturing factors in rules-based portfolios. First, one
can take risk into account by adjusting all factor scores 1
Comparing rules-based portfolios to optimization-based
by security volatilities, but that only takes into account portfolios is, in our view, comparing apples to oranges. Rules-
individual security volatility and not correlations based portfolios have the benefit of being more transparent in how
between securities. Second, one can control country a security’s factor characteristics translate into its weight in the
and sector active weights by establishing limits and redis- portfolio. For many investors, that transparency is desirable from a
tributing excess weights pro rata back to securities such governance and attribution perspective. Optimization-based port-
folios, however, are more efficient at balancing multiple, often-
that the limits are met; however, this involves increased competing objectives. To most efficiently maximize factor exposure
computational complexity. Third, turnover can only be while minimizing risk and controlling for other objectives such
mitigated using simple buffer rules, but almost always at as concentration and turnover, the latter is a preferable approach.
40 Clash of the Titans: Factor Portfolios versus Alternative Weighting Schemes Quantitative Special Issue 2019
Exhibit 1
Various Weighting Schemes Explored in Past Papers
Note: *EDHEC’s maximum Sharpe ratio strategy uses semi-deviation or downside risk as estimates for expected returns.
weighting schemes” versus “heuristic-based weighting of minimum variance); see Christoffersen et al.
schemes.” In that vein, we list popular rules-based (2012) and Amenc, Goltz, and Martellini (2013)
(heuristic) weighting schemes: • Maximum diversification: Weights of securities
are selected to maximize the ratio of weighted-
• Fundamental indexing: Weighting by fundamen- average asset volatilities to portfolio volatility; see
tals; see Arnott, Hsu, and Moore (2005) Choueifaty and Coignard (2008)
• Equal weighting: All securities are assigned the • Risk parity/equal risk contribution: Weights are
same weight; see DeMiguel, Garlappi, and Uppal based on contribution to portfolio risk such that
(2009) each asset has the same contribution; see Qian
• Diversity weighting: Weights of securities are based (2011); Maillard, Roncalli, and Telietche (2010);
on market diversity, measured as the level of con- and Carvalho, Lu, and Moulin (2012)
centration of capital in the market; see Fernholz, • Maximum deconcentration: A modified form
Garvy, and Hannon (1998) of equal weighting that maximizes the effective
number of stocks; see DeMiguel, Garlappi, and
Apart from equal weighting, the heuristic Uppal (2009) and Amenc, Goltz, and Martellini
weighting schemes require an expected return source (2013)
to be a special case of mean–variance optimality. • Maximum Sharpe ratio: Weights are selected to
Optimization-based portfolio construction maximize the Sharpe ratio of the portfolio where,
schemes include in practice, the forecast returns are often assumed
to be proportional to security volatilities; see Tobin
• Minimum variance: Weights are selected to min- (1958)
imize the variance of the portfolio; see Clarke,
de Silva, and Thorley (2011) Exhibit 1 summarizes the optimization-based
• Maximum decorrelation: Weights are selected to weighting schemes.
minimize the variance of the portfolio, assuming All of the weighting schemes shown in Exhibit 1
securities all have the same volatility (a special case can be equivalent to a mean–variance optimal portfolio
given a concrete set of assumptions. These are shown
42 Clash of the Titans: Factor Portfolios versus Alternative Weighting Schemes Quantitative Special Issue 2019
Exhibit 2
Performance Summary (USD gross returns, January 1995 to December 2016, results using three-factor principal
component analysis model, S&P 500 Universe)
Exhibit 3
Comparison of Factor Exposures Across Approaches (sample period January 1995 to December 2016; results
using three-factor principal component analysis model, Fama–French Factors plus BAB)
Notes: BAB = betting against beta; CMA = conservative minus aggressive (investment); HML = high minus low (value); RMW = robust minus weak
(profitability); SMB = small minus big; WML = momentum.
Statistically significant loadings adjusted for Newey-West are in bold. Results adjusted for both Newey-West and Bonferroni corrections are qualitatively
similar and available upon request.
Sources: SSGA; Kenneth French and AQR data libraries.
not expect all six factors to have equally high exposures Fama–French factor returns are sourced from Kenneth
individually; however, the sum/average of the exposures French’s website (http://mba.tuck.dartmouth.edu/pages/
should be high.) What is of interest is whether any of the faculty/ken.french/data_library.html). The BAB factor
other weighting schemes, none of which are designed returns are sourced from AQR’s website (https://www.
to have high factor loadings, do in fact. Exhibit 3 shows aqr.com/Insights/Datasets). For consistent comparison to
factor loadings using the traditional Fama–French defini- the universe used for the strategies, the market factor we
tions plus the betting against beta (BAB) (low volatility) use is the return to the cap-weighted universe composed
factor identified by Frazzini and Pedersen (2011). All of the S&P 500 Index constituents.
0HDQ±
0D[LPXP 5LVN 9DULDQFH %PN
0D[LPXP 6KDUSH 3DULW\ 0LQLPXP 0D[LPXP 0D[LPXP %DVHG)DFWRU &DS
'LYHUVLILFDWLRQ 5DWLR (5& 9DULDQFH 'HFRQFHQWUDWLRQ 'HFRUUHODWLRQ 3RUWIROLR :WG
$QQXDOL]HG5HWXUQ
±
± ±
± ± ±
±
$QQXDOL]HG([FHVV5HWXUQ
± ± ± ± ± ± ± ±
±
± ± ±
± ± ± ± ±
7UDFNLQJ(UURU
±
±
±
±
,QIRUPDWLRQ5DWLR
± ± ± ± ± ± ± ±
±
± ± ±
± ± ± ± ±
Source: SSGA.
The intercepts, or unexplained returns, are low, portfolios conceptually could be a special case of our
and none are statistically significant. All weighting framework, as long as the expected returns (numer-
schemes except for maximum decorrelation have sig- ator) are factor exposures. However, here we follow
nificant exposures (of the right sign) to at least two of the the route of Chow et al. (2011) and Qian, Alonso, and
nonmarket factors. This indicates that these weighting Barnes (2015) for the maximum Sharpe portfolios, so
schemes are indirectly deriving performance from some the expected returns to securities have nothing to do
of the factors. Risk parity and maximum deconcentra- with factor exposures but are instead proportional to
tion stand out as having statistically significant exposures their semi-volatilities.
to many (five or greater) factors. Maximum decorrela- To understand the robustness of these results, it is
tion is the only weighting scheme that does not have important to look at how sensitive they are to
any significant exposures, though the magnitude of the
exposures to SMB, CMA, and BAB is positive and not • Different time periods
trivial. • Different specifications for the covariance matrix
The results overall are striking because none of • Different sets of explanatory factors
the weighting schemes are designed to capture fac-
tors either implicitly or explicitly. Minimum variance Sensitivity of Results to Time Periods
portfolios are not factor portfolios. Their purpose is to
minimize risk. Maximum diversification is a spin on Exhibit 4 shows the performance by subperiod.
minimum variance, using again volatilities and corre- There is in fact a sizable amount of variation across
lations, not factor exposures. Maximum Sharpe ratio periods. The mean–variance portfolio does not have
44 Clash of the Titans: Factor Portfolios versus Alternative Weighting Schemes Quantitative Special Issue 2019
Exhibit 5
Probability of Outperforming the Benchmark in Any Three-Year Period (USD gross returns, January 1995 to
December 2016; results using three-factor principal component analysis model, S&P 500 Universe)
90.0%
79.0%
Probability of Outperforming
80.0% 74.7%
68.1%
in any 3-year Period
70.0%
60.0% 53.7% 56.3%
51.5%
50.0% 46.3%
40.0%
30.0%
20.0%
10.0%
0.0%
Risk Parity/ERC
Maximum Sharpe Ratio
Maximum Deconcetration
Maximum Decorrelation
Source: SSGA.
the highest information ratio in all subperiods, but than to build the most robust possible mean–variance
taken across all periods, it is among the most consistent. portfolio.
During the first period (1995–1999), the mean–variance Exhibit 5 plots the probability of outperforming
portfolio has the highest (least negative) information the market-cap-weighted portfolio in any three-year
ratio. During the second period (2000–2004), the equal period, another way to assess consistency in perfor-
risk contribution and maximum deconcentration port- mance. (We note that overlapping periods are used here,
folios exhibited higher information ratios; however, the so we do not assign any statistical significance to the
mean–variance portfolio is not far behind. The third estimates). They are, however, good relative indicators.
period (2005–2009) is the one period in which the The mean–variance-based framework does not in fact
mean–variance portfolio has the weakest information have the highest probability of outperformance; the risk
ratio, and the last period (2010–2016) is again dominated parity portfolio is the highest.
by the mean–variance portfolio. We acknowledge that
there is enough intertemporal variation to warrant some Sensitivity of Results to the Covariance
caution in extrapolating that the mean–variance port- Matrix Specification
folio will always be the best performer. The weakness
in 2005–2009 is in part driven by the mean–variance The covariance matrix used in the results so far is
portfolio having higher concentration (lower effec- constructed using principal component analysis with 60
tive number of names) than the alternative weighting months of returns. Prior papers use different approaches
schemes. That said, none of the alternative weighting in generating the covariance matrix. This is similar to
schemes consistently exhibit the highest information Amenc et al. (2012), though they estimated principal
ratio in all periods. Moreover, our aim in this article component analysis with two years of weekly returns.
is to compare the drivers of returns across representa- Qian, Alonso, and Barnes (2015) employed five years of
tive portfolios using different weighting schemes rather historical monthly returns with exponentially decayed
2
In their nomenclature, equal weighting, diversity weighting, ologies. Meanwhile, minimum variance, maximum Sharpe, and
and fundamental weighting are heuristic-based weighting method- maximum diversification are optimization-based methodologies.
46 Clash of the Titans: Factor Portfolios versus Alternative Weighting Schemes Quantitative Special Issue 2019
(largest 1,000 stocks) over the 1968–2012 period and asset classes, securities, or factors. Risk parity, as con-
used a single-factor risk model for the covariance matrix. ceptualized early on at Bridgewater and AQR, focused
They found that, although all three outperform market on asset classes, specifically bonds versus equities. In this
cap weighting, minimum variance and risk parity are article, we apply the notion of equal risk contribution
superior to maximum diversification based on risk- to securities—that is, each stock contributes the same
adjusted returns. This is consistent with our findings. amount of risk to the final portfolio. This is similar to
Lastly, Qian, Alonso, and Barnes (2015) provided Chow et al. (2011) and Badaoui and Lodh (2013). We
the most recent contribution to the literature. They ana- acknowledge that risk parity/ERC may be justified at
lyzed four different weighting schemes: equally weighted, the asset class level and not at the securities level. For
minimum variance, maximum diversification, and risk instance, risk parity for asset classes has been shown to
parity. They called these four weighting schemes “naïve be less sensitive to input parameters than other risk-based
beta” because all four naively assume that at least one allocation schemes; see Kaya and Lee (2012).
input (expected returns, volatilities, correlations) is the Separately, requiring equal risk from securities
same across all securities. For their empirical tests, they is different from requiring each factor to contribute
used the 10 Global Industry Classification Standard S&P the same amount of risk to the portfolio. We equally
500 sectors as the assets (in contrast to Chow et al. 2011 weight factor scores in the objective function that would
and Clarke, de Silva, and Thorley 2013 who used indi- follow if expected returns for each factor were equal. We
vidual securities as the assets). Qian, Alonso, and Barnes believe this is an appropriate assumption for an investor
(2015) found that, for all four weighting schemes, risk- without an explicit factor forecasting model. A port-
adjusted returns are higher than market cap weighting. folio constructed by weighting factors proportional to
However, they argued that the risk parity approach is risk contribution is only optimal if factors are uncor-
the best of the three because the minimum variance and related and if all factors have the same Sharpe ratio (i.e.,
maximum diversification approaches are highly sensi- each factor is rewarded in proportion to its own risk).
tive to risk inputs (volatilities and correlations) resulting Empirically, the historical evidence does not support
in concentrated portfolio holdings and high portfolio this assumption. For instance, for single-factor portfolios
turnover. built with similar levels of tracking error, momentum,
In contrast to Qian, Alonso, and Barnes (2015), and size generally have higher return than low volatility.
the framework we present here is not as sensitive to risk One point to note is that proponents of equal risk
inputs. The main difference is that in the Chow et al. weighting factors often cite the fact that factor returns
(2011) and Qian, Alonso, and Barnes (2015) papers, the are hard to forecast, so weighting by risk makes sense.
security risk estimates have a first-order impact on the We do not disagree, but the optimal solution in this
optimal portfolio. In the minimum variance optimiza- situation is to assume factor returns are equal, not that
tions, the volatility and correlation estimates are the factors are proportional to risk.
only source of information. In the maximum Sharpe
ratio optimizations, both papers assume a Sharpe ratio SUMMARY
at which the expected return is proportional to the risk
estimates; thus, again, the optimal portfolio is very sensi- Factors can be captured today in eff icient,
tive to the volatility estimates. In our framework, vola- conceptually clean, cost-effective portfolios. Modern
tilities and correlations are only one input. The other computing has led to cheaper and easier ways to access
critical input is the expected factor returns/exposures. data, execute robust portfolio construction techniques,
These balance out the risk estimates and overcome the and trade securities quickly and efficiently. We can now
main criticisms of the two prior papers. meet the practical requirements of making pure factors
investable. Factor portfolios are very intuitive: They are
A SPECIAL NOTE ON RISK PARITY based on the belief that factors drive returns and risk and
that we can harness them easily.
One final point we want to highlight concerns risk Here we resurrect Markowitz’s mean–variance
parity and equal risk contribution. Weighting schemes framework for modern-day factor portfolios, which not
that allocate to risk equally across assets can pertain to coincidentally are similar to the first-generation quant
48 Clash of the Titans: Factor Portfolios versus Alternative Weighting Schemes Quantitative Special Issue 2019
Fama, E. F., and K. R. French. 1992. “The Cross-Section ——. Portfolio Selection: Efficient Diversification of Investments.
of Expected Stock Returns.” The Journal of Finance 47 (2): New York: Jon Wiley & Sons, 1959.
427–465.
Merton, R. 1980. “On Estimating the Expected Return on
Fama, E. F., and K. R. French. 1993. “Common Risk Fac- the Market: An Exploratory Investigation.” Journal of Financial
tors in the Returns on Stock and Bonds.” Journal of Financial Economics 8: 323–361.
Economics 33 (1): 3–56.
Michaud, R., and R. Michaud. 2007. “Estimation Error and
Fernholz, R., R. Garvy, and J. Hannon. 1998. “Diversity Portfolio Optimization: A Resampling Solution.” Journal of
Weighted Indexing.” The Journal of Portfolio Management 24 Investment Management 6 (1): 8–28.
(2): 74–82.
Qian, E. 2011. “Risk Parity and Diversification.” The Journal
Frazzini, A., and L. Pedersen. 2011. “Betting against Beta.” of Investing 20 (1): 119–127.
Journal of Financial Economics 111 (1): 1–25.
Qian, E., N. Alonso, and M. Barnes. 2015. “The Triumph
Grinold, R. C., and R. N. Kahn. Active Portfolio Manage- of Mediocrity: A Case Study of Naïve Beta.” The Journal of
ment: A Quantitative Approach for Providing Superior Returns and Portfolio Management 41 (4): 19–34.
Controlling Risk. New York: McGraw-Hill, 1994.
Ross, S. 1976. “The Arbitrage Theory of Capital Asset
Kan, R., and G. Zhou. 2007. “Optimal Portfolio Choice with Pricing.” Journal of Economic Theory 13 (3): 341–360.
Parameter Uncertainty.” The Journal of Financial and Quantita-
tive Analysis 42 (3): 621–656. Tobin, J. 1958. “Liquidity Preference as Behavior towards
Risk.” The Review of Economic Studies 25 (2): 65–86.
Kaya, H., and W. Lee. 2012. “Demystifying Risk
Parity.” March 2012, https://papers.ssrn.com/sol3/papers. Zhou, G. 2008. “On the Fundamental Law of Active Port-
cfm?abstract_id=1987770. folio Management: What Happens If Our Estimates Are
Wrong?” The Journal of Portfolio Management 34 (4): 26–33.
Maillard, S., T. Roncalli, and J. Telietche. 2010. “The Prop-
erties of Equally-Weighted Risk Contribution Portfolios.”
The Journal of Portfolio Management 36 (4): 60–70. To order reprints of this article, please contact David Rowe at
d.rowe@pageantmedia.com or 646-891-2157.
Markowitz, H. 1952. “Portfolio Selection.” The Journal of
Finance 7: 77–91.
A
K ristin Fergis well-balanced strategic allocation Put another way, defensive timing focuses on
is an associate at to factors, or asset classes, is the risk mitigation, which manages second and
BlackRock, Inc., in
foundation of a well-constructed higher moments, rather than on incremen-
New York, NY.
kristin.fergis@blackrock.com long-term portfolio. However, tally increasing returns, which targets the
factor premiums vary over time. There first moment. The risk reductions are taken
K atelyn Gallagher are three possible approaches that inves- infrequently and only for those factors that
is a director at BlackRock, tors could take in response to time-varying are very expensive or offer few diversifica-
Inc., in New York, NY. factor returns. The first approach is to ignore tion benefits. During periods when aggre-
katelyn.gallagher@blackrock.com
the short-term variation—since Markowitz gate market risk tolerance is unusually low,
P hilip Hodges (1952), we have built strategic allocations or correlations unusually high, it may also
is a managing director to access sources of return that should be be appropriate to reduce risk for the total
at BlackRock, Inc., in rewarded in the long run. Second, investors portfolio.
San Francisco, CA. might develop a short-term prediction model To illustrate the concept of defensive
philip.hodges@blackrock.com
that aims to generate excess returns by tac- factor timing, we work with a portfolio of
K ed Hogan tically changing positions around strategic macro factor exposures in a multiasset context:
is a managing director factor weights (e.g., Hodges et al. 2017). A economic growth, real rates, inf lation, credit,
at BlackRock, Inc., in disadvantage of this approach is that it may emerging markets, and liquidity. These fac-
San Francisco, CA. lead to short-term underperformance rela- tors are rewarded macro risk premiums, and
ked.hogan@blackrock.com tive to the strategic allocation, and those risks we define these factors so that, to the prac-
may increase with the frequency and size of tical extent possible, they maximize exposure
the tactical decisions. The third approach is to the underlying macroeconomic source of
to practice defensive timing: to recognize risk. For example, nominal bonds are affected
that during certain periods, it may be appro- by changes in both real rates and inf lation.
priate to reduce risks to certain factors or An inf lation factor portfolio goes long nom-
across a portfolio with an aim to preserve inal bonds and short real bonds (linkers) to
capital during bad market regimes. concentrate on only inf lation-related com-
In this article, we explore this third ponents (or inflation breakevens). Both devel-
approach to defensive factor timing. Defen- oped market and emerging market equity has
sive factor timing focuses on mitigating the tended to perform strongly when the global
magnitude of losses by periodically reducing economy is growing, but certain risks present
risk, as opposed to aiming to outper- in emerging market equities are either absent
form a strategic benchmark by continually or minimal in developed market equities (see,
over- or underweighting different factors. among many others, Harvey 1995). Thus, an
&XPXODWLYH9DULDQFH([SODLQHG
E\3ULQFLSDO&RPSRQHQWV
3& 3& 3& 3& 3& 3& 3& 3&
global asset class returns data can be represented by a Economic Intuition for Hypothetical
small number of factors, it has little to say beyond the Macro Factor Portfolios
fact that we need only model a few underlying factors.
To move to a tradeable representation of a small number In Exhibit 3, we show that these hypothetical
of factors based on economic rationale, we follow Chen, macro factor portfolios—economic growth, real rates,
Roll, and Ross (1986) and choose to work with macro inf lation, credit, emerging markets, and liquidity—each
factors as shown in Exhibit 2. capture a fundamental risk premium and are constructed
These macro factors are intuitive and have been so as to provide relatively clean exposure to the under-
the focus of many in the literature. For example, eco- lying premium. Exhibit 3 compares the realized returns
nomic growth is the foundation of consumption-based of the economic growth, real rates, and inf lation trade-
asset pricing and real business cycle models (see, among able factor portfolios with the respective underlying
many others, Lucas 1978). Policymakers treat separately economic phenomena.
the effects of real rates and inf lation (Taylor 1993) and Panel A graphs gross domestic product (GDP) sur-
assign a role to credit in channeling monetary policy prises, calculated as the difference between realized GDP
(Bernanke and Gertler 1995). and GDP expectations, sourced from Consensus Economics
We form each of the macro factors to be a hypo- survey forecasts. The correlation between the series from
thetical long–short portfolio to maximize exposure to 1997 to 2017 is 0.6. The premiums investors expect from
the underlying macro factor. This is especially impor- economic growth reflects compensation for bearing the
tant for emerging markets, which have significant cor- risk that GDP will slow below expectations. For example,
relations with developed markets, but there are specific during the financial crisis of 2008, realized GDP was sig-
risks for emerging markets in excess of their exposure nificantly under expectations. During this time, we see
to developed equity (see Bekaert and Harvey 2003), the economic factor track the decline in GDP with losses
and the liquidity factor because illiquidity risk increases in developed equities, commodities, and listed real estate.
during market downturns (see Amihud, Mendelson, and In Panel B of Exhibit 3, we show returns of the real
Pedersen 2005). rates factor (which is the difference in returns between real
government bonds and cash rates) and surprises in real yields.
Notes: For illustrative purposes only using hypothetical factors and not representative of an actual investment or account. As of June 30, 2018.
The surprises in real yields on the left-hand side of the chart a portfolio exposed to these macro factors may be able
are sorted in reverse order to illustrate that when yields to defensively time the factor exposures using various
surprise on the downside, we expect both real bonds and indicators.
the real rates factor to deliver a positive return. The cor-
relation between these two series is 0.8. There is a notable INDICATORS FOR DEFENSIVE
negative surprise in real yields in 2009, which corresponds FACTOR TIMING
to rising real bonds during the recovery from the financial
crisis. There is a large positive real yield surprise in 2013, We start with a robust strategic multifactor port-
when Federal Reserve Chair Ben Bernanke surprised bond folio that is deliberately diversified across the underlying
markets by announcing a decrease in large-scale bond pur- drivers of asset class returns. The simplest approach is to
chases, effectively winding down the quantitative easing allocate an equal amount of risk to each factor portfolio.
program that had been in place since 2008. This coincides Investors with unique investment goals, such as higher
with one of the worst years for the real rates factor, which returns or reduced risk, can adjust the factor weights to
falls when real rates rise more than expected. meet those bespoke requirements, as discussed by Bass,
We construct the inf lation factor as the difference Gladstone, and Ang (2017).
between a portfolio of nominal government bonds and For the purposes of this article, we use a hypo-
inf lation-protected bonds. This factor has tended to thetical strategic factor allocation constructed such that
deliver positive returns when realized inf lation falls 30% of the risk is attributable to economic growth and
below expectations, as Panel C illustrates. For example, 30% to real rates, and the remaining 40% is driven by
inf lation expectations fall in the financial crisis of 2008, an equally weighted combination of inf lation, credit,
and traditional safe havens such as US Treasuries rise emerging markets, and liquidity. The total portfolio
faster than their inf lation-linked counterparts; con- is scaled to have an unconditional volatility of 10%.
sequently, the inf lation factor rises significantly, cap- By upweighting economic growth, the investor has a
turing the downside inf lation surprise. The correlation higher exposure to a factor that historically has had
between the inf lation factor and inf lation surprises is larger long-term Sharpe ratios than the other factors;
0.9 over 1997 to 2017. similarly, the investor benefits from higher exposure
Having demonstrated that these macro factor to real rates, which have tended to do well when
portfolios are good proxies for the underlying eco- policymakers seek to lower real rates below expecta-
nomic risks, we now illustrate how an investor holding tions to stimulate economic growth during bad times
Exhibit 3
Relating Macro Factors to Economic Shocks
3DQHO$(FRQRPLF*URZWK)DFWRUDQG6XUSULVHVLQ(FRQRPLF*URZWK
(FRQRPLF)DFWRU5HWXUQ
*'36XUSULVH
±
±
±
±
±
±
±
± ±
± ±
*'36XUSULVH/+6 (FRQRPLF)DFWRU5HWXUQ5+6
3DQHO%5HDO5DWHV)DFWRUDQG6XUSULVHVLQ5HDO5DWHV
±
±
5HDO5DWHV)DFWRU5HWXUQ
5HDO<LHOG6XUSULVH
±
±
±
(continued)
±
±
,QIODWLRQ)DFWRU5HWXUQ
,QIODWLRQ6XUSULVH
±
±
±
±
,QIODWLRQ6XUSULVH/+6 ,QIODWLRQ)DFWRU5HWXUQ5+6
Notes: Past performance is not indicative of future results. Returns do not account for any fees or transaction costs. If fees were included, returns would be
lower. For illustrative purposes only using hypothetical factors and not representative of an actual investment or account. The modeled performance is cal-
culated with the benefit of hindsight and knowledge of factors that may have positively affected its performance and has inherent limitations. This analysis
cannot account for risk factors that may affect actual portfolio performance.
However, during periods of low risk tolerance, the denote the ranking by risk. The RTI is then calculated
positive correlation between risk and return can turn as the correlation between the two rank vectors:4
negative: When investors panic, they tend to f lee from
risk assets, driving down the price while bidding up the RTI = corr [ q( Rt ), q(σ t )]
price of safer assets. During these periods of extreme
negative sentiment, even well-diversified portfolios Typically, positive RTIs accompany periods of
can suffer sharp negative returns as market dynamics increasingly positive investor sentiment, or risk-on
break down and investors are penalized for holding environments, whereas periods of declining investor
risky assets.3 sentiment, often coinciding with a f light to quality, will
The risk tolerance indicator (RTI) aims to identify see a negative RTI. Investor sentiment tends to per-
such low-frequency, high-loss periods ex ante. Let q( Rti ) sist, so rising (falling) investor sentiment usually leads
denote the ranking of asset i by return, and let q(σ it )
4
Note that we could also have defined the RTI equal to the
correlations between the quantities themselves. Taking ranks allows
the indicator to be applicable to asset pricing models that have a non-
linear dependence between risk and expected return and prevents
3
There is a substantial literature following, French, Schwert, the dominance of extreme values. Rank statistics are commonly
and Stambaugh (1987), documenting a pronounced contempora- used in robust statistical inference. For well-behaved, especially
neous negative correlation between realized returns and volatility, normally distributed, data, there is little difference between using
which we capture in the RTI. The relation between expected Pearson correlations on raw series versus Spearman rank correla-
returns and volatility is much harder to pin down empirically but tions—see, for example, Huber (1981). A similar rank correlation
should theoretically be positive over most of the range of volatility measure is computed by Kumar and Persaud (2002) to assess senti-
(see Ang and Liu 2007). ment in the equity market.
5ROOLQJ<U5HWXUQ06&,:RUOG,QGH[KHGJHG
5LVN7ROHUDQFH,QGLFDWRU
± ±
± ±
± ±
± ±
Notes: Indexes are unmanaged, and it is not possible to invest directly in an index. Past performance is not a guarantee of future results.
5ROOLQJ<U5HWXUQ
'LYHUVLILFDWLRQ5DWLR
±
±
±
Notes: The 20/80 portfolio represented by MSCI World Index (hedged) and Bloomberg Barclays Global Aggregate Index (hedged), respectively. Indexes
are unmanaged, and it is not possible to invest directly in an index. Past performance is not a guarantee of future results.
is low tend to correspond with periods of significant that during 2013 when the diversification ratio decreases,
market stress, when tail risks are elevated. the RTI remains broadly positive (see Exhibit 4), illus-
By construction, our strategic multifactor port- trating the benefit of having multiple defensive timing
folio seeks to maximize diversification by balancing indicators.
risk across historically rewarded macroeconomic fac-
tors, so it is vulnerable to correlation spikes that reduce Valuation Indicators
the benefits of diversification. A short-term measure
of portfolio diversification could serve as an indicator Factors, like all investments, become richer or
for reducing portfolio risk against a background of cheaper compared to their historical long-term averages.
increasing correlations. We construct factor valuation indicators to measure their
Exhibit 5 illustrates historic diversification ratios prices relative to intrinsic value. Each macro premium is
against the rolling one-year return of a 20/80 stock/ aggregated and smoothed over time to produce a mea-
bond portfolio, consisting of 20% MSCI World Index sure of how the valuation of the factor compares to its
(hedged) and 80% Bloomberg Barclays Global Aggregate own history. A positive valuation score indicates cheap-
Index (hedged). We use a 20/80 portfolio to represent ness, and a negative score indicates richness. A factor is
a simple multiasset portfolio with balanced equity and fairly valued compared to its own historical long-term
interest rate risks, which normally benefits from diver- average when its valuation score is zero. The time-
sification. The Taper Tantrum of 2013 is an extreme series scoring also adjusts for long-term risk premiums
episode of deteriorating diversification benefits, when embedded in each factor. We describe our valuation
correlations between equities and bonds spike toward methodology for each factor.
one. In Exhibit 5 we see the diversification ratio and the Economic growth. We use a variant of Shiller’s
20/80 portfolio decline in tandem. It is worth noting cyclically adjusted price-to-earnings ratio (CAPE)
([WUHPHO\&KHDS
±
±
5LFKHU
±
([WUHPHO\5LFK
±
±
±
±
3DQHO%,QIODWLRQ)DFWRU
&KHDSHU !
([WUHPHO\&KHDS
±
±
5LFKHU
±
± ([WUHPHO\5LFK
±
±
±
3DQHO&&UHGLW)DFWRU
&KHDSHU !
([WUHPHO\&KHDS
±
±
5LFKHU
±
([WUHPHO\5LFK
±
±
±
±
(continued)
([WUHPHO\&KHDS
±
±
5LFKHU
±
([WUHPHO\5LFK
±
±
±
±
3DQHO((PHUJLQJ0DUNHWV)DFWRU
&KHDSHU !
([WUHPHO\&KHDS
±
±
5LFKHU
±
([WUHPHO\5LFK
±
±
±
±
3DQHO)/LTXLGLW\)DFWRU
&KHDSHU !
([WUHPHO\&KHDS
±
±
5LFKHU
±
([WUHPHO\5LFK
±
±
±
±
Note: For illustrative purposes only using hypothetical factors and not representative of an actual investment or account.
5LVN7ROHUDQFH,QGLFDWRU57,
'LYHUVLILFDWLRQ5DWLR'5
±
±
±
±
$SU
-XO
2FW
-DQ
$SU
-XO
2FW
-DQ
$SU
-XO
2FW
-DQ
$SU
-XO
2FW
-DQ
$SU
-XO
2FW
-DQ
$SU
-XO
2FW
3HULRGV+LJKOLJKWHGLQ&DVH6WXGLHV 57, '5
standard deviation thresholds.5 Although the definition consistently experiences a drawdown in each scenario,
of infrequent action is subjective, we estimate that defen- the metrics that serve as our forward-looking indicators
sive actions might take place, on average, once every exhibit dissimilar behavior across the scenarios. For the
12–18 months. first and second scenarios, the RTI f lags the imminent
Exhibit 7 plots the RTI and diversification ratio downturn, whereas in the third the diversification ratio
from 2006 through the second quarter of 2018. We does so. This underscores the importance of having a
overlay the four time periods on which the following mosaic of indicators to make defensive timing deci-
case studies are based: European Crisis 2010, European sions robust. Over the sample from January 2006 to
Crisis 2012, Taper Tantrum of 2013, and China and Oil June 2018, the correlation between the RTI and the
Concerns of 2015. In the case studies that follow we see diversification ratio is -0.2, suggesting our two indica-
that although our hypothetical macro factor portfolio tors are complementary.
Our analysis is hypothetical and represents actions
5
that could be taken on the diversified macro factor
In terms of hypothesis testing, we are focused on Type I portfolio in response to elevated risks highlighted by
errors. In practicing defensive timing, on some occasions we will
reject the null (no regime change of good times) in favor of the
one or more of the indicators. In the case studies that
alternative (regime change to bad times). Thus, some false posi- follow, we describe the market backdrop, comment on
tives are inevitable. The use of extreme thresholds means that we how it affects the defensive timing indicators, and show
will fail to reject with some small probability, corresponding to the how the indicators could be used to inform changes
p-value of a classical test. In practice, we can reduce the frequency to the portfolio in an effort to mitigate portfolio loss.
of Type I errors by contextualizing the information we receive
We compare the simulated drawdowns of our de-risked
from the signal. For example, in September 2016, the diversification
ratio sharply declines to its lowest level since the Taper Tantrum in hypothetical macro factor portfolio against the simu-
2013. This indicates a breakdown in cross-asset correlations and an lated performance of our baseline hypothetical macro
environment that would warrant reducing risk. However, a further factor portfolio from the second section to examine
assessment of market conditions indicates a benign environment, the efficacy of our risk-reducing portfolio changes.
and correlations increase as all asset classes move higher.
Exhibit 8
Valuation Indicators during Case Study Periods
3DQHO$(XUR&ULVLV
&KHDSHU !
±
5LFKHU
±
±
±
5HDO5DWHV ,QIODWLRQ &UHGLW (FRQRPLF (PHUJLQJ /LTXLGLW\
*URZWK 0DUNHWV
3DQHO%(XUR&ULVLV
&KHDSHU !
±
5LFKHU
±
±
±
5HDO5DWHV ,QIODWLRQ &UHGLW (FRQRPLF (PHUJLQJ /LTXLGLW\
*URZWK 0DUNHWV
(continued)
&KHDSHU !
±
5LFKHU
±
±
±
5HDO5DWHV ,QIODWLRQ &UHGLW (FRQRPLF (PHUJLQJ /LTXLGLW\
*URZWK 0DUNHWV
3DQHO'&KLQDDQG2LO&RQFHUQV
&KHDSHU !
±
5LFKHU
±
±
±
5HDO5DWHV ,QIODWLRQ &UHGLW (FRQRPLF (PHUJLQJ /LTXLGLW\
*URZWK 0DUNHWV
Note: For illustrative purposes only using hypothetical factors and not representative of an actual investment or account.
We plot the drawdowns of the theoretical de-risked with the scaling down of real rates exposure, mitigates
and the baseline portfolios during the period from May losses in our de-risked factor portfolio: The maximum
to September 2013 in Exhibit 9, Panel C. As both equi- drawdown is -7.3% versus -9.5% in the baseline port-
ties and bonds fall, our pro rata reduction in risk, coupled folio. This perfect-storm environment of heightened
±
±
±
±
±
0D\ -XQ -XO $XJ 6HS
%DVHOLQH)DFWRU3RUWIROLR 'H5LVNHG)DFWRU3RUWIROLR
3DQHO%(XUR&ULVLV
±
±
±
±
0D\ -XQ -XO $XJ
%DVHOLQH)DFWRU3RUWIROLR 'H5LVNHG)DFWRU3RUWIROLR
3DQHO&7DSHU7DQWUXP
±
±
±
±
±
-XQ -XO $XJ 6HS
%DVHOLQH)DFWRU3RUWIROLR 'H5LVNHG)DFWRU3RUWIROLR
(continued)
±
±
±
±
±
±
-XO $XJ 6HS 2FW
%DVHOLQH)DFWRU3RUWIROLR 'H5LVNHG)DFWRU3RUWIROLR
Notes: Past performance is not indicative of future results. Returns do not account for any fees or transaction costs. If fees were included, returns would be
lower. For illustrative purposes only using a hypothetical macro factor portfolio and not representative of an actual investment or account. The modeled per-
formance is calculated with the benefit of hindsight and knowledge of factors that may have positively affected its performance and has inherent limitations.
This analysis cannot account for risk factors that may affect actual portfolio performance.
cross-asset correlations and broadly unchanged investor economic slowdown in China at the center of the market
sentiment demonstrates the need for multiple, uncor- correction, commodity prices reach their lowest point
related indicators to effectively implement defensive since 2008, and the VIX Index spikes to more than
factor timing. 50, its highest level since the financial crisis. The RTI
continues to spiral downward and ends the month of
China and Oil Concerns 2015 August at -60%. In October 2015, markets begin to nor-
malize as investors start to accept the reality of a Chinese
Financial markets respond to three main sources slowdown; upon this normalization we feel comfortable
of uncertainty during the summer of 2015: China, increasing portfolio risk back to normal levels.
Greece, and commodities. Throughout the month of During the period between July and October
July, Chinese equity markets remained highly volatile 2015, the de-risked hypothetical portfolio has a subtly
despite measures taken by authorities to stabilize prices, improved drawdown profile compared to the baseline
and commodities suffered intense liquidation pressure, portfolio, -4.5% versus -5.5%, respectively, as illus-
led by oil. Greece dominated headlines after a negative trated in Exhibit 9, Panel D. Although the hypothetical
vote on the terms of the bailout referendum on July 5. portfolios have similar drawdown profiles at the start
We see in Exhibit 7 the deterioration of investor senti- of the period, the improvement becomes more pro-
ment throughout July, as the RTI falls to -43% by the nounced in late August and September at the peak of
end of the month. This rapid drop in the RTI again sig- the market correction, when the RTI further deterio-
nals a defensive timing opportunity, and we reduce risk rates to -80%.
in our macro factor portfolio by 20%. (Note that at this
time, there are no warning signs from the diversification CONCLUSION
ratio in Exhibit 7 or the factor valuations in Exhibit 8).
The RTI effectively signals in July what will Defensive factor timing periodically, but infre-
become the worst month in financial markets since the quently, takes down risk exposures to a given factor or
financial crisis. During the month of August, with the the full portfolio during adverse market conditions. This
Ang, A., and J. Liu. 2007. “Risk, Return, and Dividends.” Harvey, C. 1995. “The Risk Exposure of Emerging Equity
Journal of Financial Economics 85 (1): 1–38. Markets.” World Bank Economic Review 9: 19–50.
T
David Blitz raditional asset allocation models capitalizations (small-caps), higher book-to-
is the head of quantitative are centered around capturing market (value), higher profitability and lower
equity research at Robeco
risk premiums offered by distinct asset growth (quality), and higher recent
Asset Management
in Rotterdam, the asset classes such as public and returns (winners) have delivered high returns
Netherlands. private equities, corporate and government relative to their peers with lower scores on
d.c.blitz@robeco.nl bonds, and commodities. Decades of empir- these characteristics.
ical research on asset prices have uncovered Widely used asset pricing models take
M ilan Vidojevic a set of factors that drive expected returns the capital asset pricing model (CAPM) as
is a quantitative
researcher at Robeco
within and across asset classes. Explicit, top- a starting point and then augment it with
Asset Management in down allocation to factor premiums (e.g., other factors (e.g., the Fama and French
New York, NY, and size, value, momentum, quality, and low 1993, 2015 three- and five-factor models).
holds a position at Vrije volatility) is now becoming mainstream. Yet These models inspire a top-down approach
Universiteit Amsterdam many questions remain about how to effi- to factor investing, wherein investors allo-
in Amsterdam, the
ciently gain exposures to these premiums. cate a certain amount to each of these fac-
Netherlands.
m.vidojevic@robeco.com Factors are systematic drivers of stock tors. For instance, an investor who wishes
returns—they explain why certain stocks go to capture the value premium could allo-
up or down at the same time (co-move) and cate a part of the portfolio to the theoret-
why certain stocks command higher expected ical Fama–French high minus low (HML)
returns than others. In practice, investment value factor or, more likely, a long-only
portfolios that provide exposures to these value index in practice. A pitfall of the top-
factors are constructed by sorting stocks down approach, however, is that each factor
on certain characteristics and buying those portfolio is constructed in isolation, with a
that score favorably and, if implemented in complete disregard for all other factors. To
a long–short setting, short selling those that address this problem, we propose a bottom-
score unfavorably. Over the years, researchers up approach instead, wherein the expected
have identified a number of factor character- return of each stock is a function of all its
istics that are key determinants of expected factor characteristics.
stock returns. Some examples include market To illustrate, take two hypothetical
capitalization (size), book-to-market (value), value stocks. These stocks, by definition,
gross or operating profitability, asset growth have a high book-to-market characteristic,
(investment), and past return (momentum). which contributes positively to their long-
In particular, holding all other character- term expected returns. However, these stocks
istics constant, stocks with lower market can be completely different in terms of other
3RVLWLYH,PSOLHG5HWXUQV 1HJDWLYH,PSOLHG5HWXUQV
3DQHO%5HDOL]HG([FHVV5HWXUQVPRQWK
RI*HQHULFDQG(QKDQFHG6LQJOH)DFWRUV
6PDOO 9DOXH :LQQHUV 3URILWDEOH /RZ,QYHVWPHQW
*HQHULF (QKDQFHG
the integrated and enhanced single-factor strategies are results or why many factors appear to be stronger in
designed to provide the same level of factor exposure the small-cap segment of the market. We show that the
(i.e., have comparable overall factor characteristics), their differences in performance among all these portfolios
model-implied and historically realized returns are in can be fully explained by the differences in their factor
fact not statistically different. characteristics.
Lastly, through the lens of our model, we also Although the characteristics-based model that we
address some longstanding questions about why dif- apply in this article can be used to forecast out-of-sample
ferent weighting schemes have delivered different stock and portfolio-level expected returns, this is not
Exhibit 3
Portfolios Sorted on Implied Return
5HDOL]HG5HWXUQ
0RGHO,PSOLHG5HWXUQ
reward to a one standard deviation higher profitability to 8.76% a year, is in line with the realized excess return
is 14 bps, the reward to a one standard deviation higher across all stocks over this sample period.
momentum is 26 bps, and the reward to a one standard
deviation lower investment is 14 bps, all per month. MODEL VALIDATION
These results are fully in line with the existing asset
pricing literature, which has previously established the Using our multi-factor characteristics-based return
existence of these factor premiums. model, each month we calculate the implied return of
Contrary to the predictions of the CAPM and the each stock over the following month, based on its cur-
Fama–French asset pricing models, market beta is not rent factor characteristics. To validate the predictions
priced in the cross-section of stock returns, whereas the of our model, we first sort stocks every month into five
constant is positive and significant, instead of zero. This portfolios, based on their model-implied returns. We use
is a manifestation of the well-known low-beta anomaly, value weighting (i.e., the weight of a stock in a portfolio
which is also fully consistent with the existing litera- is proportional to its market capitalization). We further
ture; see, for instance, Blitz (2014), Clarke et al. (2014), calculate the implied and realized returns of these five
and Blitz and Vidojevic (2017). The positive constant portfolios over the subsequent month. In Exhibit 3, we
effectively ref lects the equity risk premium (i.e., the fact show that the average implied and realized returns line
that average stock returns are higher than the risk-free up closely over the full sample period.
return), which market beta fails to capture. Our esti- A more formal way to evaluate the fit of our model
mate of the constant, 0.73% a month, which translates is to rerun Fama–MacBeth regressions with realized
0.35%
0.30%
Excess Return Per Month
0.25%
0.20%
0.15%
0.10%
0.05%
0.00%
Small Value Winners Profitable Low Investment
–0.05%
Implied Realized
excess stock return on the left-hand side and the implied is close to zero; this is because better profitability
stock returns on the right-hand side. If the model fit is characteristics tend to go hand in hand with worse size
good, we expect to see a slope coefficient equal to 1. and value characteristics, which, on balance, implies
We estimate this model and find a coefficient of 1.05 market-relative returns close to zero according to our
(t-statistic of 7.11). We further test if this coefficient model. We elaborate on this issue in the next section.
is statistically different from 1 and find that it is not Realized returns of the profitability portfolio do show a
(t-statistic of 0.34). We conclude that our simple model positive excess return, but much lower than that for the
does a good job matching the first moment of the return other factors. Although our characteristics-based model
distribution. seems to have some trouble explaining the performance
We further examine the accuracy of our model’s of the profitability portfolio, the difference between the
predictions by considering various generic factor port- realized and predicted return is also statistically insig-
folios. Specifically, we select the most attractive 20% nificant for this factor portfolio.
of stocks every month based on their size, value, prof-
itability, investment, or momentum characteristics. THE CHARACTERISTICS OF GENERIC
Exhibit 4 shows how the average implied returns relate FACTOR STRATEGIES
to the average realized returns of these generic factor
portfolios. Again we observe that realized and implied Generic factor strategies select stocks with the
returns line up nicely, with the difference between them best scores on one particular factor, regardless of their
being statistically indistinguishable from zero in all five other factor characteristics. A generic value strategy, for
cases. The size, value, momentum, and investment port- instance, may buy a stock that has strong value charac-
folios show the best fit. Interestingly, the average implied teristics but, at the same time, poor size, momentum,
market-relative return for the profitability portfolio profitability, or investment characteristics. On balance,
60%
50%
40%
30%
20%
10%
0%
Q3
Q3
Q3
Q3
Q3
Q2
Q2
Q2
Q2
Q2
Q4
Q4
Losers
Q4
Winners
Q4
Q4
Small
Value
Big
Growth
HighInv
LowInv
Unprof
Prof
Negative Implied Return (Relative to Market)
Positive Implied Return (Relative to Market)
taking all this information into account, this may imply a by the size (ln(Mcap)) characteristics. However, this
negative, instead of a positive, return for this stock, rela- portfolio also gets a positive boost from having had a
tive to the market. With our multi-factor characteristics- positive exposure to the value factor and a negative con-
based return model we are able to assess, at each point in tribution from a negative exposure to the profitability
time, how much of a generic factor portfolio is invested factor. The value strategy has also experienced a negative
in stocks with negative implied returns relative to the contribution from the profitability factor, on average.
market. Panel A of Exhibit 1 shows the average results The low-investment portfolio has primarily benefited
for the top quintile factor portfolios, and Exhibit 5 from a high value of the (inverse of the) investment
presents results for all five quintiles of the five factor characteristic; however, a substantial part of its outper-
strategies. Returning to the example of a generic value formance comes from a high value exposure. This is not
portfolio, the graph shows that this strategy invests about surprising given that Fama and French (2015) showed
20% in stocks that have a negative predicted market- that their value (HML) and investment (conservative
relative return. This is clearly a non-negligible part of minus aggressive [CMA]) factors are closely related and
the portfolio. Similar weights are found for the generic that, in fact, the CMA factor subsumes HML in the
small-cap strategy (about 20%), the generic momentum spanning regressions, making it obsolete.
strategy (about 15%), and the generic investment strategy Exhibit 7 shows the percentage of months in our
(about 30%). We also see that going from the top to the sample when each of the six characteristics of these factor
bottom quintile for each factor, the weight in stocks portfolios were lower (unfavorable) than that of the
with implied outperformance declines monotonically. market (i.e., the targeted characteristic is going against
We next focus on the small, value, winners, profit- other characteristics). For instance, the small-cap port-
able, and low investment portfolios (i.e., the top port- folio never has a worse size (ln(Mcap)) characteristic than
folios) and in Exhibit 6 present the result of the factor the (capitalization-weighted) market; in 12% of months
contribution decomposition implied by our model. it has a lower value (ln(BtM)) characteristic, and in 51%
For instance, the small-cap portfolio has a full-sample of months it has a worse momentum score. However, the
implied outperformance over the market of 30 bps a small-cap portfolio systematically goes against the two
month (3.6% a year), the vast majority of which is driven quality factors—in every single month the profitability
0.4%
0.3%
Implied Return
0.2%
0.1%
0.0%
–0.1%
–0.2%
Small Value Winners Profitable Low Investment
5HDOL]HG5HWXUQ
6PDOO 9DOXH :LQQHUV 3URILWDEOH /RZ
,QYHVWPHQW
$OO6WRFNV ,5! $W/HDVW $W/HDVW $W/HDVW $OO
Exhibit 9
Realized Versus Implied Returns on Enhanced Single-Factor Strategies
0.9%
0.7%
Implied Return
0.5%
0.3%
0.1%
–0.1%
–0.1% 0.1% 0.3% 0.5% 0.7% 0.9%
Realized Return
the number of stocks in the portfolio decreases from, Exhibit 9 shows that the realized returns on these
on average, 302 with no constraints to 276, 273, 223, portfolios align with their model-implied values; that
96, and only 13. A similar pattern is observed for other is, the returns of these portfolios can be attributed to
factors, albeit not always monotonic, because very con- their factor characteristics. None of the differences
centrated portfolios can be subject to a fair amount of between the realized and implied returns are statistically
stock-specific risk that is in general diversified away in significant, although we do observe a couple of more
broader portfolios. notable deviations from the 45-degree line. The largest
3RUWIROLR:HLJKW
XJ
XJ
XJ
XJ
XJ
XJ
XJ
XJ
XJ
XJ
$
$
$
$
$
$
$
deviations are for the size and investment portfolios mid-1960s, late 1970s, and early 2000s. On average, 9.8%
that are constrained to have non-negative exposures of the market portfolio consists of stocks with implied
to all five factors; these are also the most concentrated returns lower than bonds. Systematically removing
portfolios, on average containing just over 10 stocks, so such stocks from the equity market portfolio increases
naturally the fit of our return model is somewhat lower. its excess return over 10-year US bonds from 0.36%
Nevertheless, even for these portfolios, the difference to 0.42% per month, which amounts to a 16% increase
between the average implied and realized returns is sta- in realized returns for investors, in relative terms. The
tistically indistinguishable from zero. difference in returns is statistically significant at the
A similar analysis can be conducted for the equity most conservative levels (t-statistic of 3.22). Thus, our
market portfolio, which many investors track passively multi-factor characteristics-based return model can also
to earn the equity risk premium. The existence of fac- be used to enhance the capitalization-weighted market
tors that are priced in the cross-section implies that the portfolio.
market portfolio is not mean–variance efficient; that is,
it is suboptimal to hold the market in the presence of MIXED-SLEEVE VERSUS INTEGRATED
factor portfolios. Using our multi-factor characteristics- MULTI-FACTOR PORTFOLIOS
based return model, we can calculate at each point in
time which stocks have an implied excess return that We next compare the performance of the inte-
is not merely below average but is even less than the grated and mixed-sleeve multi-factor portfolios. Clarke,
excess return on 10-year US Treasury bonds. Over de Silva, and Thorley (2016); Bender and Wang (2016);
our July 1963 to December 2017 sample period, the and Fitzgibbons et al. (2017) all argued that integrated
average excess return on a portfolio of 10-year US bonds portfolios, which are constructed by investing in
amounted to 0.18% per month. stocks with a high combined rank on multiple factors,
Exhibit 10 plots the weight of the market port- deliver higher absolute and risk-adjusted returns than
folio that is invested in stocks that have a lower model- multi-factor portfolios constructed by mixing single-
implied excess return than that of bonds at each point factor sleeves. On the other hand, Amenc et al. (2017)
in time. This weight varies over time between less than emphasized some of the shortcomings and risks of inte-
5% and more than 25%, with the peaks occurring in the grated portfolios, such as inefficiency, instability, and
control for these ancillary exposures.”). In our article, strategy). A fourth possibility is that the small-cap and
we emphasize the importance of thinking about port- large-cap portfolios have comparable factor characteris-
folio characteristics from a holistic standpoint because tics but that the factor premiums are larger in the small-
both the targeted, as well as the non-targeted, factor cap segment, in which case a linear expected return
exposures drive expected returns. model is not appropriate. Lastly, it could be the case that
Another difference in our approaches is in how factors that are not accounted for drive these differences.
the exposures are matched: Ghayur, Heaney, and Platt Another well-known phenomenon is that the
(2018) matched factor exposures over the full sample equally-weighted factor portfolios tend to generate
period and not at each point in time (footnote 8 in higher returns than their value-weighted counterparts.
their paper). This means that their approach could lead Because equally-weighted factor portfolios invest more
to substantial deviations in any given time period. In in smaller stocks, all of the aforementioned explanations
our article, we account for both the targeted and the may also apply here. In addition, equally-weighted factor
non-targeted factor characteristics at each point in time portfolios might profit from the periodic rebalancing
because the implied stock and portfolio returns are cal- that is required to maintain equal weights over time
culated monthly. Our results are consistent with those (i.e., a potential rebalancing premium).
of Leippold and Rueegg (2017), who concluded that the In this section, we decompose the implied returns
two approaches yield similar results if the portfolios are of these portfolio pairs using our characteristics-
constructed in comparable ways. based multi-factor model and pinpoint the reason for
the observed return differences between small-cap
EXPLAINING OTHER FACTOR PORTFOLIO and large-cap factor portfolios and between equally-
RETURN DIFFERENCES weighted and value-weighted factor portfolios.
Exhibit 13 shows the implied and realized perfor-
It is well-known that small-cap factor portfo- mance difference between top quintile value-weighted
lios tend to generate higher returns than large-cap factor portfolios in the small-cap space and top quintile
factor portfolios (Fama and French 2008), but it is not value-weighted factor portfolios in the large-cap
entirely clear what drives this spread. One possibility space, where stocks are split into two size groups at
is that the outperformance of the factors in the small- the median market capitalization value of our stock
cap universe is driven by a higher exposure of these universe. Exhibit 14 shows the implied and realized
portfolios to the size premium. Another possibility is performance difference between equally-weighted and
that the outperformance comes from a higher expo- value-weighted factor portfolios. Both graphs show
sure to the targeted factor—for instance, if the small- a close match between implied and realized returns,
cap value strategy has higher value characteristics than which indicates that our model is able to explain the
the large-cap value strategy. A third possibility is that observed return differences quite well. The tables below
the outperformance comes from different exposures to the graphs present these numbers in a tabular format
the non-targeted factor premiums (e.g., if the small- and confirm that the difference between realized and
cap value strategy has better momentum, investment, implied returns is not statistically significant. In other
or profitability characteristics than the large-cap value words, the superior performance of the small-cap factor
5HDOL]HG5HWXUQ
,PSOLHG5HWXUQ
portfolios compared to the large-cap factor portfolios, the small-cap winners portfolio not only benefits from
and the superior performance of equally-weighted factor a higher exposure to the small-cap factor but also from
portfolios compared value-weighted factor portfolios, a higher exposure to the momentum factor itself; that
appears to be fully in line with the implied returns that is, small stocks have higher momentum scores than
can be derived from the factor characteristics of these large stocks, on average.
portfolios. This means that there is no need to assume Exhibit 16 shows these breakdowns for the equal
that the same factor exposure is rewarded more in the versus value-weighted portfolios. We find that the dif-
small-cap space than in the large-cap space, nor is there ference in performance across factors constructed using
a need to assume that equally-weighted factor portfolios these two weighting schemes can to a large extent be
benefit from some kind of rebalancing premium. attributed to the small-cap and value tilts that equal-
Exhibit 15 shows the breakdown of the implied weighting and rebalancing back to the equal weight
return contributions across the factor characteristics introduce.
for the small-big factor portfolios. Clearly, the reason
why small-cap factor portfolios outperform their CONCLUSION
large-cap counterparts can be mostly attributed to a
higher exposure to the size factor (i.e., lower market Eff icient implementation of factor investing
capitalization). All small-cap factor legs tend to have requires an understanding of how factors perform inde-
better value characteristics but worse prof itability pendently and how they are related to one another. In
scores. In the case of the past winners, it appears that this article, we apply a characteristics-based multi-factor
5HDOL]HG5HWXUQ
,PSOLHG5HWXUQ
Exhibit 15
Implied Return Decomposition for Factor Portfolios in Small- versus Large-Cap Universes
,PSOLHG5HWXUQ
±
9DOXH :LQQHUV 3URILWDEOH /RZ,QYHVWPHQW
,PSOLHG5HWXUQ
±
±
9DOXH :LQQHUV 3URILWDEOH /RZ,QYHVWPHQW
return model to show that investing in generic single- that our model can explain performance differences
factor portfolios leads to suboptimal investment results between factor portfolios in the small-cap and in the
because stocks that score highly on one factor can go large-cap space, as well as equal- and value-weighted
strongly against other factors and, on net, have nega- portfolios. Therefore, an efficient implementation of
tive market-relative model-implied and realized returns. factor-based, or any kind of equity strategies requires
We show that removing such stocks from the portfolio an understanding of the underlying portfolio factor
yields substantially higher returns. We further show that characteristics.
if, in addition to excluding stocks with implied under- The discussion in this article is confined to the
performance relative to the market, we also exclude question of how investors should approach building
stocks that have negative exposures to one, two, three, an expected return model for equity portfolios. In the
or four non-targeted factors, we can enhance these real world, investors’ preferences and constraints cause
single-factor strategies even more. Our model can also portfolios to have a significant exposure to factors that
attribute performance differences between mixed-sleeve are not rewarded with higher returns but certainly do
and integrated multi-factor portfolios to differences contribute to the risk. An example is industry expo-
in their factor characteristics. As generic single-factor sures—an investor should not expect to be compen-
sleeves are an inefficient way to obtain factor exposure, sated for taking industry bets unless they are a result
mixing them into a portfolio also results in suboptimal of the rewarded factor characteristics of the underlying
multi-factor portfolios. However, a mixed-sleeve port- industry. Although an effective management of the
folio of enhanced single factors that do not invest in unrewarded sources of risk is of paramount importance,
stocks with implied underperformance is able to match one has to be cognizant of the fact that, in the long run,
the performance of a bottom-up integrated multi-factor expected portfolio returns depend on the priced port-
portfolio with similar factor exposures. Finally, we show folio characteristics.
REFERENCES Fama, E., and J. MacBeth. 1973. “Risk, Return, and Equi-
librium: Empirical Tests.” Journal of Political Economy 81:
Amenc, N., F. Ducoulombier, M. Esakia, F. Goltz, and 607–636.
S. Sivasubramanian. 2017. “Accounting for Cross-Factor
Interactions in Multi-Factor Portfolios without Sacrificing Fitzgibbons, S., J. Friedman, L. Pomorski, and L. Serban.
Diversification and Risk Control.” The Journal of Portfolio 2017. “Long-Only Style Investing: Don’t Just Mix, Inte-
Management 43 (5): 99–114. grate.” The Journal of Investing 26 (4): 153–164.
Asness, C., A. Frazzini, R. Israel, T. Moskowitz, and Ghayur, K., R. Heaney, and S. Platt. 2018. “Constructing
L. Pedersen. 2018. “Size Matters, If You Control Your Junk.” Long-Only Multi-Factor Strategies: Portfolio Blending vs.
Journal of Financial Economics 129 (3): 479–509. Signal Blending.” Financial Analysts Journal 74 (3): 70–85.
Bender, J., and T. Wang. 2016. “Can the Whole Be More Haugen, R., and N. Baker. 1996. “Commonality in the
Than the Sum of the Parts? Bottom-Up versus Top-Down Determinants of Expected Stock Returns.” Journal of Finan-
Multi-Factor Portfolio Construction.” The Journal of Portfolio cial Economics 41 (3): 401–439.
Management 42 (5): 39–50.
Hou, K., C. Xue, and L. Zhang. 2015. “Digesting Anomalies:
Blitz, D. 2014. “Agency-Based Asset Pricing and the Beta An Investment Approach.” The Review of Financial Studies 28
Anomaly.” European Financial Management 20: 770–801. (3): 650–705.
Blitz, D., and M. Vidojevic. 2017. “The Profitability of Low ——. “Replicating Anomalies.” Working paper, 2017.
Volatility.” Journal of Empirical Finance 43: 33–42.
Leippold, M., and R. Rueegg. 2017. “The Mixed vs the
Clarke, R., H. de Silva, and S. Thorley. 2014. “The Not-So- Integrated Approach to Style Investing: Much Ado about
Well-Known Three-And-A-Half-Factor Model.” Financial Nothing?” European Financial Management 24 (5): 1–27.
Analysts Journal 70: 13–23.
Lewellen, J. 2015. “The Cross-Section of Expected Stock
——. 2016. “Fundamentals of Efficient Factor Investing.” Returns.” Critical Finance Review 4: 1–44.
Financial Analysts Journal 72 (6): 9–26.
Novy-Marx, R. 2013. “The Other Side of Value: The Gross
——. 2017. “Pure Factor Portfolios and Multivariate Regres- Profitability Premium.” Journal of Financial Economics 108 (1):
sion Analysis.” The Journal of Portfolio Management 43 (3): 1–28.
16–31.
Fama, E., and K. French. 1993. “Common Risk Factors in To order reprints of this article, please contact David Rowe at
the Returns on Stocks and Bonds.” Journal of Financial Eco- d.rowe@pageantmedia.com or 646-891-2157.
nomics 33: 3–56.
A
Ashley L ester s factor investing strategies have approaches.2 Second, existing empirical work
is global head of moved from academia to main- is typically relatively small in scale, involving
multi-asset research
stream practical application over the comparison of a limited number of
and systematic investment
at Schroders in London, recent years, the question of portfolios, sometimes under quite special
UK. how best to build portfolios consisting of circumstances.3
ashley.lester@schroders.com more than one factor has become increas- This article makes two contributions
ingly pressing. Proponents of traditional to the debate. First, it makes an explicit and
smart-beta–style approaches argue in favor quantitative set of theoretical predictions
of building factor portfolios through a set of relating to simple bottom-up and top-down
single-factor subportfolios (e.g., a value port- multifactor portfolios. These predictions are
folio, a momentum portfolio, and so on). We highly tractable because they use the sim-
will follow the literature in referring to this plest possible building blocks. We derive
approach as top down. A growing literature analytic expressions for the point-in-time
argues instead that combining information factor exposure of portfolios formed from
about each factor at the stock level and then stocks with factor characteristics above some
building a single, combined portfolio—an percentile cut-off. As Amenc et al. (2017)
approach known as bottom up—is likely to
produce superior outcomes.1
2
The persistence of this discussion A notable exception that develops a theoretical
framework addressing a similar question was under-
ref lects two shortcomings in the existing
taken by Clarke, de Silva, and Thorley (2016). Their
literature. First, although there is a general framework focused on the relative information ratios
intuitive sense that the advantages of com- of various portfolio construction choices. Although
bination are larger when the factors are less this is a powerful framework, the machinery is less
correlated, there is (at this stage) no explicit well suited to highlighting the simple test of the linear
theory of why this is the case or of whether factor model we highlight here, in particular its appli-
cation over a very large number of empirical tests.
there are other empirically testable propo- 3
Fitzgibbons et al. (2017) reported the results
sitions that might differentiate the two of experiments on simulated data but did not make
entirely clear how their results were generated, nor did
1
The main author who explicitly argues in favor they explicitly link their results to a theory of expo-
of single-factor portfolios (albeit with some modifi- sure. Ghayur, Heaney, and Platt (2016) ran a larger
cations) is Amenc et al. (2017), although others also number of experiments than most but focused on
hedge their bets (e.g., Ghayur, Heaney, and Platt 2016). the question of relative performance for given factor
The combination side was taken by Bender and Wang exposure. This contrasts with most practically avail-
(2016) and Fitzgibbons et al. (2017). able approaches to factor investing.
88 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
because the variance of each single-factor score is unit factor but also of the extent to which the chosen factor is
by construction. That is, the distribution of total factor correlated with other factors. In particular, if the chosen
scores is inf luenced by the set of correlations among all factor is negatively correlated with some other factor,
of the factors. The concept of correlation used here (and attempts to increase the intensity of exposure to the
throughout) refers to the point-in-time correlation of chosen factor inevitably result in ever-greater negative
the factor scores across stocks, not a time-series measure incidental exposure to the secondary factor.
of factor return correlation. For a portfolio that selects the top x-percent of
Under these assumptions, we can use the inverse stocks, we can substitute in the inverse Mills ratio, and
Mills ratio to calculate the average factor score of a the resulting score is
portfolio that consists of the top x-percent of stocks,
equally weighted. The inverse Mills ratio is a useful φ( Φ −1( x ))
E(Fk ) = (1 + ( K − 1)ρk )
property of the normal distribution that allows us to 1− x
calculate the expected value of the truncated normal
distribution. Technically, where ρk is the average correlation between factor k and
the other K - 1 scored factors, and the expression on the
z − µ
φ
right is the transformed inverse Mills ratio that defines
σ the expected value of the top x-percent of observations
E[X X > z ] = µ + σ
z − µ of a standard normal.
1− Φ
σ The total factor score for a multifactor portfolio
formed on the top x-percent of stocks can be derived
in which f and F are, respectively, the probability dis- from Equation 1 and is
tribution function and cumulative distribution function
of the standard normal distribution. Finally, we need φ( Φ −1( x ))
E(F{K } ) = K + ∑ ∑ ρ jk
to know the expected value of another factor, Y, given j ∈K k ∈{K / j } 1− x
that we have formed a portfolio using a single factor, X.
φ( Φ −1( x ))
Under our assumptions, this is simply = K + ∑( K − 1)ρ j
j ∈K 1− x
E[Y X ] = βX
φ( Φ −1( x ))
Cov( X ,Y ) = K (1 + ( K − 1)ρ ) (3)
= X 1− x
Var ( X )
= ρX where ρ is the average correlation across all factors.
It will be helpful later in interpreting the empirical results
Therefore, the expected factor score over all fac- to note that the factor score of the underlying benchmark
tors (not just one), Fk, of a single-factor portfolio formed portfolio (with stock weights the same as those used in
using scores for factor k, is forming the factors, here equal weights) is zero.
Finally, we can use Equation 2, the total factor score
of a single-factor portfolio, to derive factor exposure for
E(Fk ) = E ∑s j the equally weighted set of single-factor portfolios that
j ∈K
forms a top-down portfolio. The total factor score of an
equally weighted set of K single-factor portfolios is
= E sk +
∑ ρ jk sk
j ∈{N /k }
1 φ( Φ −1( x ))
E K ( E(Fj )) =
K
∑(1 + ( K − 1)ρ j ) 1− x
= 1+
∑ ρ jk E(sk ) (2)
j
j ∈{N /k } φ( Φ −1( x ))
= (1 + ( K − 1)ρ ) (4)
1− x
This expression highlights that the combined
factor score of any single-factor portfolio is a func- Combining Equations 3 and 4 yields an ana-
tion not only of the intensity of exposure to the chosen lytic expression for the ratio of factor exposure in a
±
&RUUHODWLRQ
bottom-up portfolio to that in a similarly formed top- the factor exposure ratio is larger when the number of
down portfolio: factors is larger; equivalently, the effect of more factors
is larger when correlation is more negative. The effect
E(F{K } ) K (1 + ( K − 1)ρ ) K is illustrated in Exhibit 1.
= = (5)
E K ( E(Fj )) (1 + ( K − 1)ρ ) 1 + ( K − 1)ρ So far, we have only predicted factor exposures,
but the point is to predict performance. We, there-
which is strictly greater than one for all levels of average fore, make the simplest possible assumptions regarding
correlation less than one. returns and risk. Adopting the workhorse linear model
of expected factor returns (and ignoring time subscripts
throughout), we assume that returns to stock i are driven
Observation 1: Number of Factors
by the underlying factor returns, rj, with an intensity
Suppose that there are K factors and the factors are proportional to stock i’s exposure to the factor, and an
orthogonal. Then from Equation 5, it is apparent that idiosyncratic error term, ei:
the ratio of factor exposures is K .
ri = ∑sij r j + εi (6)
j ∈N
Observation 2: Correlation
The correspondingly simple model of stock risk is
From Equation 5, it is obvious that the ratio of factor
driven by idiosyncratic risk, which for stock i has stan-
exposures is a strictly decreasing function of correlation.
dard deviation hi, and factor risk, which for factor j has
standard deviation sj:
Observation 3: Correlation
and Number of Factors K
2
σ = ηi + ∑sij σ j (7)
2
i
It is possible to show that the cross-partial deriva- j =1
tive of Equation 5 is negative, but we omit the proof.
This means that the inf luence of reduced correlation on Equation 6 generates a trivially simple mapping
from factor exposures to expected returns: Any prediction
90 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
for exposures is simply replicated for expected returns. K , whereas risk in the bottom-up portfolio is constant.
The linear factor model, therefore, extends observations This result exactly offsets the previous result on factor
1–3 to expected returns, rather than purely exposures. exposure: If the investor cares about risk-adjusted
Under this assumption, the bottom-up approach gen- returns, so far they are indifferent between the two
erates greater returns than the top-down approach as a portfolio construction techniques.
result of the extra factor exposure generated. However, Now consider the effect of idiosyncratic risk. The
whereas observations 1–3 describe point-in-time factor result for the bottom-up portfolio is unchanged because
exposure by definition, the real world is not so simple. the number of stocks selected is not a function of the
In particular, the error term in returns means that only number of factors.
expected returns are linear in factor exposure: Realized The result for the top-down portfolio, however,
returns will not be perfectly linear in exposure even if depends on how the number of stocks varies with the
the model is true. number of factors. Recall that the top x-percent of stocks
Generating similarly neat predictions concerning are selected for each factor. If there were no overlap
risk—and by extension risk-adjusted returns—is more across factor portfolios, the number of stocks would
difficult, but assuming that all sources of risk are mutu- increase proportionally with the number of factors, and
ally orthogonal permits some progress. We then can the previous result would continue to hold exactly (NTD
derive expressions for bottom-up and top-down risk is equal to a constant times K). Inevitably, however,
(see proofs in the Appendix). some stocks will be chosen in more than one portfolio.
Portfolio risk for the top-down portfolio is For orthogonal factors, the probability that a particular
stock is chosen in no factor portfolio is (1 - x)K . There-
1
η2 fore, the fraction of stocks chosen in the top-down port-
IMR 2 σ 2j 2
σTD = i + folio is 1 - (1 - x)K . This fraction increases at a rate less
N TD K than proportional to K, and therefore risk in the top-
down portfolio decreases at a rate less than proportional
where ηi2 is the average idiosyncratic variance across all to 1/ K .
stocks, σ 2j is the average across all factors of the factor
risk associated with unit factor exposure, IMR is the Observation 4: Portfolio Risk
transformed inverse Mills ratio defined earlier, and NTD
For orthogonal factors, risk in bottom-up portfo-
is the total number of stocks in the top-down portfolio.
lios is constant with respect to the number of underlying
We have ignored single-stock concentration risk arising
factors, whereas risk in top-down portfolios falls at a rate
from overlapping holdings across the underlying single-
less than 1/ K .
factor portfolios.6 Under the same conditions, risk in the
bottom-up portfolio is
Observation 5: Risk-Adjusted Returns
1
η2 2
For orthogonal factors, bottom-up portfolios
σ BU = i + IMR 2 σ 2j
N BU accumulate expected returns proportional to K and
maintain constant risk. Therefore, risk-adjusted returns
It is easiest to gain intuition for this result if we increase with the number of factors at rate K .
temporarily ignore idiosyncratic risk. In that case, risk For orthogonal factors, top-down portfolios
in the top-down portfolio decreases in proportion to maintain constant expected returns as the number
of underlying factors increases, and risk falls at a rate
less than proportional to K . Therefore, risk-adjusted
6
In particular, any stock held by more than one underlying returns increase at a rate less than K . Thus, for
factor portfolio will generate excess concentration risk relative to orthogonal factors, risk-adjusted returns are greater
stocks held only once. For simplicity, we ignore this channel of
concentration risk from multiple holdings of the same stock. How-
in bottom-up portfolios than in top-down portfolios,
ever, see later in the main text for much more on the effect on risk and their relative advantage increases with the number
of overlapping holdings arising from the number of stocks held. of factors.
92 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
Exhibit 3
Performance and Risk Measures, February 1995 to December 2016
1XPEHURI)DFWRUV
6LQJOH)DFWRU 7RS'RZQ %RWWRP8S 7RS'RZQ %RWWRP8S 7RS'RZQ %RWWRP8S 7RS'RZQ %RWWRP8S
5HWXUQV
5LVN
± ± ± ± ± ± ± ± ±
,QIRUPDWLRQ5DWLR
$YHUDJH+ROGLQJV
$YHUDJH&RUUHODWLRQ ±
([SRVXUH
1XPEHURI3RUWIROLRV
Notes: Figures are based on daily total returns in US dollar terms. Returns, risk, and information ratios are relative to an equal-weighted benchmark
based on the stocks in MSCI All Country World Index. Standard errors are in parentheses.
Source: MSCI All Country World Index.
%RWWRP8S
%RWWRP8S
±
± ±
± ± ±
7RS'RZQ 7RS'RZQ
1XPEHURI)DFWRUV 1XPEHURI)DFWRUV
%RWWRP8S
%RWWRP8S
7RS'RZQ 7RS'RZQ
returns across portfolios with a given number of factors tight relationship between top-down and bottom-up
is a way to deal with measurement error in expected performance, and the relationship is almost exactly as
returns. However, we can also examine directly whether, predicted by observation 1. The theoretical relationship
for a given number of factors, relative excess returns are is shown as a solid line, and the estimated empirical rela-
aligned with those expected. (This is not possible across tionship (without intercept) is shown as a dashed line. In
all portfolios unconditionally because the predicted slope all cases, the relationship is tight and highly significant.
of the relationship differs with the number of factors.) A notable feature of Exhibit 5 is the scatter around
Exhibit 5 shows the results across portfolios the predicted relationship. Judged across the many
for a given number of factors. In all cases, there is a tested portfolios, the relationship appears very close,
94 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
particularly given the approximations needed to arrive as the number of factors increased, whereas risk in
at the theoretical prediction. However, as noted earlier, top-down portfolios would decline at a rate less than
standard practice is to judge bottom-up and top-down 1/ K . Exhibit 3 shows that this is indeed the case: Risk
techniques (or alternative portfolio construction tech- is almost identical on average across the bottom-up port-
niques more broadly) using just a handful of portfolios. folios as the number of factors changes, whereas average
The scatter highlights that such practices are likely to risk for top-down portfolios declines with the number
discover spurious results purely as a result of the random- of factors. Exhibit 8 shows these relationships graphi-
ness in realized returns. cally, along with the 1/ K line, all indexed to a value
of unity for the average of single-factor portfolios. As
Observations 2 and 3: Correlations predicted, average risk for top-down portfolios declines
and Number of Factors at a slower rate than 1/ K , and the rate of divergence
grows as the number of factors increases.
The results so far have focused on the role of the The prediction of observation 5 was based on the
number of factors in generating superior outperfor- declining rate at which new stocks are added to the
mance. We now consider the predictions relating to top-down portfolio as the number of factors increases
correlation. These predictions are more difficult to test because of overlapping holdings across the underlying
than the return predictions because correlation is a prop- single-factor portfolios. It is, therefore, valuable to check
erty of each factor at each point in time but may vary whether the number of stocks held in the portfolios cor-
significantly over time. Pragmatically, we deal with this responds with the predicted number. For bottom-up
problem by taking a simple time-series average of the portfolios, the number of stocks held should be invariant
exposure correlation for each factor targeted in a given to the number of factors targeted, and it is.
portfolio. For top-down portfolios, Exhibit 3 shows the
Exhibit 6 shows graphically the extent to which average number of stocks held in each set of portfolios.
our time-series average is an effective proxy for the Using 238 as the base, we can predict how many stocks
overall effect of correlation on portfolio factor expo- the top-down portfolios should hold, on average, if fac-
sure. The y-axis represents, for each portfolio, the ratio tors are orthogonal. The results of this exercise, along
of average factor exposure of the bottom-up portfolio with the corresponding actual number of stocks held,
to that of the corresponding top-down portfolio. The are shown in Exhibit 9. Again, there is a close cor-
results are significantly noisier than those for outper- respondence. That the portfolios actually hold slightly
formance, as might be expected, but qualitatively they fewer stocks than they should is consistent with the small
map well to the theoretical relationship. The advantage positive average correlation across the factors.
of bottom-up portfolios is greater for those portfolios
built using factors that are more negatively correlated, Observation 5: Risk-Adjusted Returns
on average. Moreover, this effect is larger when the port-
folio is built from more factors (compare with Exhibit 1). Given all the preceding, and in particular the
Exhibit 7 reports the same experiment but replaces results concerning returns and risk, observation 5 fol-
exposures with relative outperformance. This relation- lows immediately: Risk-adjusted returns are higher in
ship is somewhat weaker again because the effect of the bottom-up portfolios, and the difference grows with the
underlying randomness in portfolio returns is magnified number of targeted factors. The notable feature of this
when taking the ratio of outperformance. Nevertheless, result is the quantitative importance of the idiosyncratic
the broad shape of the function, and the relative slope risk mechanism that slows the rate of risk reduction in
as the number of factors changes, remains consistent the top-down portfolios. Risk adjusted, the bottom-up
with the theoretical prediction. portfolios outperform by less than 20% in two-factor
combinations but by 47% in five-factor combinations.
Observation 4: Portfolio Risk Finally, Exhibit 10 shows risk-adjusted returns of
bottom-up portfolios relative to top-down portfolios
Observation 4 predicted that, for orthogonal fac- grouped by number of factors (analogous to Exhibit 5).
tors, risk in bottom-up portfolios would remain constant In this case, the theoretical line is simply the 45-degree
3
2
2.5
1.5
2
Exposures
Exposures
1.5
1
1
0.5
0.5
0 0
–0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 –0.4 –0.3 –0.2 –0.1 0 0.1 0.2 0.3 0.4 0.5
Average Correlation Average Correlation
4
5
3.5
3 4
Exposures
Exposures
2.5
3
2
1.5 2
1
1
0.5
0 0
–0.3 –0.2 –0.1 0 0.1 0.2 0.3 –0.25 –0.2 –0.15 –0.1 –0.05 0 0.05 0.1 0.15 0.2
Average Correlation Average Correlation
96 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
Exhibit 7
Ratio of Excess Returns versus Correlation
1XPEHURI)DFWRUV 1XPEHURI)DFWRUV
5DWLRRI([FHVV5HWXUQV
5DWLRRI([FHVV5HWXUQV
± ± ± ± ± ± ±
$YHUDJH&RUUHODWLRQ $YHUDJH&RUUHODWLRQ
1XPEHURI)DFWRUV 1XPEHURI)DFWRUV
5DWLRRI([FHVV5HWXUQV
5DWLRRI([FHVV5HWXUQV
± ± ± ± ± ± ± ±
$YHUDJH&RUUHODWLRQ $YHUDJH&RUUHODWLRQ
line ref lecting the base case of equal risk-adjusted return used this expression to examine whether the ex post
(equivalently, no overlapping holdings). The growing performance of these portfolios matched their ex ante
gap between the empirical result and the base-case result factor exposure. Rather than focusing on just one or
as the number of factors grows confirms the key insight. two factor portfolios, we examined data on a total of
1,008 matched portfolios using various factor defini-
CONCLUSION tions to abstract from an otherwise inevitable focus on
particular factors, factor definitions, or portfolio con-
We have derived an analytically tractable formula struction techniques.
relating factor exposure in simple bottom-up portfolios The experimental data strongly support the
to that in corresponding top-down portfolios. We then contention that factor returns are approximately linear
&RQVWUXFWLRQ
%RWWRP8S 7RS'RZQ URRW.
Exhibit 9
Top-Down Stock Holdings: Empirical and Theoretical
1400
1200
1000
800
600
400
200
No overlaps Predicted Empirical
0
1 2 3 4 5
in factor exposures. In particular, rather than per- betas. Moreover, our investigative framework provides a
forming qualitative tests of the relative outperformance novel benchmark against which to judge whether there
of bottom-up portfolios, we were able to examine quan- may be nonlinearities in expected returns to particular
titatively how the outperformance of bottom-up factor factors or to particular combinations of factors. We hope
portfolios relative to their top-down counterparts was that this framework plays its part in pushing financial
affected by the number of factors targeted and the cor- research toward the use of large-scale experiments using
relation between the factors. many candidate factors, rather than the current focus on
Our findings are highly relevant for all those particular combinations of one or two factors among the
engaged in factor investing; in particular, our findings hundreds available.
pose a challenge for practitioners of single-factor smart
98 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
Exhibit 10
IR of Top-Down versus Bottom-Up Portfolio Construction
Number of Factors = 2 Number of Factors = 3
2.5 2.5
2
2
1.5
1.5
Bottom–Up
Bottom–Up
1
1
0.5
0.5
0
–1 –0.5 0 0.5 1 1.5 2 2.5
–0.5 0
–0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
–1 –0.5
Top–Down Top–Down
2
2
1.5
1.5
Bottom–Up
Bottom–Up
1
0.5
0.5
0
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
Top–Down Top–Down
Amenc, N., F. Ducoulombier, M. Esakia, F. Goltz, and Freyberger, J., A. Neuhierl, and M. Weber. “Dissecting Char-
S. Sivasubramanian. 2017. “Accounting for Cross-Factor acteristics Nonparametrically.” NBER Working Paper No.
Interactions in Multifactor Portfolios without Sacrificing 23227, 2017.
Diversification and Risk Control.” The Journal of Portfolio
Management 43 (5): 99–114. Ghayur, K., R. Heaney, and S. Platt. “Constructing Long-
Only Multi-Factor Strategies: Portfolio Blending versus
Asness, C. 1997. “The Interaction of Value and Momentum Signal Blending.” Working paper, Goldman Sachs, 2016.
Strategies.” Financial Analysts Journal 53 (2): 29–36.
Kelly, B., S. Pruitt, and Y. Su. “Characteristics Are Covari-
Beck, N., J. Hsu, V. Kalesnik, and H. Kostka. 2016. “Will ances: A Unified Model of Risk and Return.” Working paper,
Your Factor Deliver? An Examination of Factor Robustness 2018.
and Implementation Costs.” Financial Analysts Journal 72 (5):
58–82. Liu, D. 2017. “Pure Quintile Portfolios.” The Journal of
Portfolio Management 43 (5): 115–129.
Bender, J., and T. Wang. 2016. “Can the Whole Be More
than the Sum of the Parts? Bottom-Up versus Top-Down
Multi-Factor Portfolio Construction.” The Journal of Portfolio To order reprints of this article, please contact David Rowe at
Management 42 (5): 39–50. d.rowe@pageantmedia.com or 646-891-2157.
100 On the Theory and Practice of Multifactor Portfolios Quantitative Special Issue 2019
Implementation Matters:
Relaxing Constraints
Can Improve the Potential
Returns of Factor Strategies
Jack Davies, Dave Gibbon, Sara Shores,
and Josephine Smith
M
Jack Davies uch of the asset pricing lit- or proprietary investment signals. At the
is a vice president and erature has examined the other end of the spectrum, private funds often
portfolio manager for the returns of factor strategies have few constraints. Private funds may take
Factor Based Strategies
Group at BlackRock, Inc.
by constructing hypothetical on leverage, take more frequent rebalancing
in London, UK. portfolios formed by ranking or optimizing decisions, and selectively disclose informa-
jack.davies@blackrock.com securities based on stock attributes (e.g., past tion. Between these extremes, institutional
returns, valuation ratios) associated with investors may have access to a host of dif-
Dave Gibbon excess returns.1 This voluminous academic ferent vehicles, such as collective investment
is a managing director and
literature, however, tends to study return funds, separately managed accounts, funds of
head of investment strategy
in EMEA for the Factor strategies that are usually not directly invest- one, and mutual funds.
Based Strategies Group ible. In the real world, investors face a variety We illustrate the effects of the con-
at BlackRock, Inc. in of investment constraints largely governed by straints implied by the selection of different
London, UK. the choice of investment vehicle. Each invest- investment vehicles on the potential return
david.gibbon@blackrock.com ment vehicle carries different restrictions that stream available to an investor using two
Sara Shores affect the returns earned by an investor. equity factor strategies—momentum and
is a managing director and The goal of this article is to quantify value—following Jegadeesh and Titman
the head of investment the impact of the investment constraints by (1993) and Fama and French (1993), respec-
strategy for the Factor showing how well-known investment strate- tively, among many others. In particular, we
Based Strategies Group at gies could potentially fare in representative investigate the return patterns of two families
BlackRock, Inc. in
examples of each vehicle type. At one end of momentum and value strategies: market
San Francisco, CA.
sara.shores@blackrock.com of the spectrum of investment vehicles are neutral (or absolute return) and benchmark
exchange-traded funds (ETFs), which are relative.
Josephine Smith most often based on a low-turnover, fully We start with a hypothetical long–short,
is a director and senior transparent third-party index. ETFs are market-neutral factor portfolio, with a con-
researcher for the Factor traded on an exchange—like a stock—but struction that is most suited to a private fund
Based Strategies Group at
BlackRock, Inc. in
most indexes are long only and, because they investment vehicle. These investment vehi-
San Francisco, CA. are publicly disseminated, rarely ref lect novel cles are unregistered under the Investment
josephine.smith@blackrock.com Company Act of 1940 (henceforth 40-Act)
1
See Ang (2014) for a comprehensive literature and tend to have the fewest constraints on
summary on factor investing.
102 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
Exhibit 1
Factor Strategies and Vehicle Types
7\SLFDO0LQLPXP 1RWHVRQ5LVN
3RUWIROLR %HQFKPDUN 7\SLFDO,QYHVWRU ,QYHVWPHQW /HYHUDJH7UDGLQJ
3ULYDWH)XQG &DVK ,QVWLWXWLRQDO ([DQWHULVNWDUJHWRI
4XDOLILHG,QYHVWRUV
8&,76 &DVK (8,QVWLWXWLRQDO ([DQWHULVNWDUJHWRI
,QGLYLGXDO,QYHVWRUV JURVVOHYHUDJHOLPLWRIî
0DUNHW $Q\RQH 9DULHV *URVVOHYHUDJH
OLPLWRIî
/RQJ2QO\ 0DUNHW $Q\RQH 9DULHV 1ROHYHUDJH
/RQJ2QO\(7) 0DUNHW $Q\RQH ± 1ROHYHUDJHVHPLDQQXDO
UHEDODQFLQJ
Notes: This exhibit is for illustration purposes only. Characteristics listed are for typical portfolios. The last column reflects the specific constraints applied
to the hypothetical strategies herein; actual constraints may differ significantly depending on the specific investment or account.
104 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
important feature across the investment vehicles, which Exhibit 2
we separately investigate in the fourth subsection. Factor Strategy Performance in Hypothetical
Market-Neutral Private Fund Strategies
Market-Neutral Factor Strategies
3ULYDWH)XQG 8&,76
Exhibit 2 reports portfolio performance statistics 3DQHO$0RPHQWXP6W\OH)DFWRU
for the hypothetical market-neutral momentum and 5HDOL]HG5HWXUQ
value strategies. We analyze a variety of metrics: real- 7DUJHW5LVN
ized return, target and realized risk, Sharpe ratio, rela- 5HDOL]HG5LVN
6KDUSH5DWLR
tive turnover (turnover per unit gross leverage), gross
5HODWLYH7XUQRYHU î î
leverage, maximum drawdown, and skewness. All risk $YHUDJH*URVV/HYHUDJH î î
and return numbers are in average annualized percent- 0D['UDZGRZQDW5LVN ± ±
ages and are stated in excess of cash, relative turnover and 0D['UDZGRZQ5DQJH -XQH± -XQH±
gross leverage are averages over the simulation period, 6HSWHPEHU 6HSWHPEHU
and the maximum drawdown is calculated assuming 6NHZQHVVDW5LVN ±
±
returns are scaled to 7% ex post risk.8 We also compare 3DQHO%9DOXH6W\OH)DFWRU
the maximum drawdown statistics of the factor strate- 5HDOL]HG5HWXUQ
gies scaled to 7% risk with the maximum drawdown of 7DUJHW5LVN
5HDOL]HG5LVN
a Monte Carlo simulation of 10,000 normally distrib-
6KDUSH5DWLR
uted return series over the same sample size, where the
5HODWLYH7XUQRYHU [ [
mean and variance are chosen to match the documented $YHUDJH*URVV/HYHUDJH [ [
Sharpe ratio and 7% risk. This allows us to gauge, along 0D['UDZGRZQDW5LVN ± ±
with the skewness, how close these strategy returns are 0D['UDZGRZQ5DQJH -XO\± $XJXVW±
to normally distributed returns. As an example, for 7% )HEUXDU\ )HEUXDU\
annualized risk and a Sharpe ratio of 0.5, the normal 6NHZQHVVDW5LVN ± ±
distribution has an annualized mean assumption of 3.5%
Notes: The simulation period is January 1998 through June 2017.
and an annualized standard deviation of 7%. Returns and risk are in excess of the cash rate. Return and risk num-
Market-neutral momentum. In Panel A of bers (both realized and drawdown) are reported in average annualized
Exhibit 2, the realized returns and risk of the private percentage. Target risk reflects the optimization constraint of the strategy
achieving a given level of target ex ante risk each period. The Sharpe ratio
fund momentum strategy are 13.9% and 14.1%, respec- is calculated as the ratio of the realized return of the strategy divided by
tively, implying a Sharpe ratio of 0.99. Relative turnover the realized risk of the strategy. The relative turnover is the average turn-
is 3.3×, and gross leverage averaged 5.3×. over per unit of leverage. Average gross leverage is calculated as the average
The observed maximum drawdown of the private sum of the absolute holdings of the strategy. The maximum drawdown
represents the peak-to-trough drawdown and is calculated after scaling each
fund momentum strategy at 7% risk is -24.3%, which strategy to ex post risk of 7%. The skewness statistic is computed on the
occurs from June 2008 through September 2009. This 7% ex post risk-scaled returns for each strategy. Past performance is not
is nearly twice the maximum drawdown of -12.3% for indicative of future results. Returns do not account for any fees or transac-
tion costs. If fees were included, returns would be lower. For illustrative
a simulated normal distribution. This is not surprising; purposes only using hypothetical strategies and not representative of an
momentum strategies have been well documented to actual investment or account. The modeled performance is calculated with
exhibit sharp drawdowns and significant downside risk the benefit of hindsight and knowledge of factors that may have positively
affected its performance and has inherent limitations. This analysis cannot
(see, for example, Daniel and Moskowitz 2016). The account for risk factors that may affect actual portfolio performance.
downside risk of momentum is also ref lected in the *** denotes rejection of the null hypothesis that the risk-scaled returns
skewness statistic of -0.8, which is statistically different have skewness of a corresponding normal distribution at the 1% level.
from zero at the 1% level.
Moving to the more constrained UCITS momen-
8
We scale to 7% ex-post risk for the drawdown analysis to be tum strategy, still in Panel A, the realized return and
able to compare the private fund and UCITS strategies at similar risk numbers fall to 9.5% and 9.9%, respectively, which
risk levels. We perform a similar exercise for the benchmark-relative
are both lower than the private fund implementation.
strategies.
106 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
Exhibit 3
Cumulative Returns of Hypothetical Market-Neutral Private Fund Strategies
3DQHO$0RPHQWXP6W\OH)DFWRU
&XPXODWLYH/RJ5HWXUQV
±
3DQHO%9DOXH6W\OH)DFWRU
&XPXODWLYH/RJ5HWXUQV
±
3ULYDWH)XQG 8&,76
Notes: Past performance is not indicative of future results. Returns do not account for any fees or transaction costs. If fees were included, returns would be
lower. For illustrative purposes only using hypothetical strategies and not representative of an actual investment or account. The modeled performance is
calculated with the benefit of hindsight and knowledge of factors that may have positively affected its performance and has inherent limitations. This analysis
cannot account for risk factors that may affect actual portfolio performance.
3DQHO%9DOXH6W\OH)DFWRU
/RQJ2QO\ /RQJ2QO\(7)
7RWDO5HWXUQV
5HDOL]HG5HWXUQ
5HDOL]HG5LVN
6KDUSH5DWLR
5HODWLYH7XUQRYHU î î î
0D['UDZGRZQDW5LVN ± ± ±
0D['UDZGRZQ5DQJH 2FWREHU± 2FWREHU± 0D\±
)HEUXDU\ )HEUXDU\ )HEUXDU\
6NHZQHVVDW5LVN ±
±
±
5HWXUQVLQ([FHVVRI&DS:HLJKWHG%HQFKPDUN
5HDOL]HG5HWXUQ
5HDOL]HG5LVN
6KDUSH5DWLR
Notes: The simulation period is January 1998 through June 2017. We analyze performance statistics for both the total portfolio in excess of cash and the
active portfolio in excess of returns of a cap-weighted benchmark of the asset universe. Return and risk numbers (both realized and drawdown) are reported
in average annualized percentage. Each optimization has a risk aversion parameter of 1. The Sharpe ratio is calculated as the ratio of the realized return
of the strategy divided by the realized risk of the strategy. The relative turnover is the average turnover per unit of leverage. The maximum drawdown rep-
resents the peak-to-trough drawdown and is calculated after scaling each strategy to ex post risk of 15%. The skewness statistic is computed on the 15%
ex post risk-scaled returns for each strategy. Past performance is not indicative of future results. Returns do not account for any fees or transaction costs. If fees
were included, returns would be lower. For illustrative purposes only using hypothetical strategies and not representative of an actual investment or account.
The modeled performance is calculated with the benefit of hindsight and knowledge of factors that may have positively affected its performance and has
inherent limitations. This analysis cannot account for risk factors that may affect actual portfolio performance.
*** denotes rejection of the null hypothesis that the risk-scaled returns have skewness of a corresponding normal distribution at the 1% level.
is approximately the same volatility as the market port- maximum drawdowns with a normal distribution, with
folio during the simulation period.) The market-neutral Monte Carlo simulations similar to those examined for
strategies have insignificant market direction as a result the market-neutral strategies, but with volatility cali-
of the long–short portfolio construction. In contrast, brated to 15% risk. For the portfolio characteristics in
the major component of the benchmark-relative factor excess of the cap-weighted benchmark, we report real-
returns is the market return. We compare the realized ized return, risk, and Sharpe ratios.
108 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
Benchmark-relative momentum. Panel A of Sharpe ratios range from 0.50 for 130/30, to 0.44 for
Exhibit 4 focuses on the benchmark-relative momentum long-only, and finally to 0.41 for the long-only ETF.
strategies. With a modest degree of leverage, the 130/30 Turnover falls significantly in the long-only ETF, at
strategy has the highest return of all three benchmark- 0.6×, compared to the 1.3× turnover for the 130/30 and
relative momentum strategies across the sample, posting long-only vehicles. Consistent with the results for the
a return, risk, and Sharpe ratio of 6.9%, 14.5%, and market-neutral strategies, turnover is lower for the value
0.48, respectively. Relative turnover is 1.9×, and the factor (Panel B) than for the momentum factor (Panel
maximum drawdown of the strategy scaled to 15% A). As with their momentum counterparts, the active
risk is -54.4%. The maximum drawdown period is performance falls precipitously when moving from the
during the global financial crisis from October 2007 130/30 strategy to the long-only ETF strategy. Although
to February 2009 and is the same for the long-only the 130/30 strategy has an active Sharpe ratio of 1.27,
and long-only ETF vehicles. This 130/30 drawdown is in line with the market-neutral counterparts, the long-
substantially higher than for a comparable normal dis- only and long-only ETF portfolios exhibit sharply lower
tribution, which would have a maximum drawdown active Sharpe ratios of 0.83 and 0.39, respectively.
of -35.9%, which is consistent with the large downside One interesting result is that the benchmark-relative
risk of the momentum factor (see Exhibit 2). The skew- value strategies exhibit significantly more downside risk
ness of this momentum strategy is also statistically sig- than the market-neutral value strategies (see Exhibit 2).
nificant at -1. Upon removing the benchmark from the The maximum drawdown at 15% risk for the bench-
calculations, the active performance exhibits a Sharpe mark-relative value strategies is around -54%.9 This is
ratio of 0.79, highlighting both the skill of the 130/30 much higher than a normal distribution, which would
strategy in allocating around a cap-weighted benchmark have a maximum drawdown of -35.9%. The skewness
and the benefits of limited leverage. of the benchmark-relative value strategies is between
The two long-only strategies may appear similar at -0.8 and -0.9, and the tests overwhelmingly reject the
face value, given that neither takes on leverage, but there null hypothesis of a normal distribution. In contrast,
are important differences. We expect that both strategies the market-neutral value strategies in Exhibit 2 exhibit
should have lower Sharpe ratios (0.43 and 0.40 for the no pronounced skewness.10 This implies that limited
long-only and long-only ETF, respectively) than the leverage, in this context, has not decreased risk. That
Sharpe ratio of 0.48 for the 130/30 strategy. Although is, we find that using more unconstrained leverage can,
Sharpe ratios decline from 130/30 to long-only to the in the case of a value strategy, result in more consistent
long-only ETF—consistent with there being more portfolio returns.
restrictions on the portfolio—realized risk is approxi- Cumulative returns of benchmark-relative
mately the same. There is a large drop in the turnover of momentum and value. Exhibit 5 graphs the log
the long-only ETF strategy, which has a turnover of 0.9× cumulative returns of the hypothetical benchmark-
compared to 1.9× for both the 130/30 and long-only relative momentum and value strategies in excess of the
funds. This is potentially desirable for certain taxable cap-weighted benchmarks in Panels A and B, respec-
investors. Panel A shows that the tighter constraints have tively. Although the 130/30 portfolios outperformed
not significantly affected the downside risk of imple- their long-only counterparts throughout most of the
menting the momentum factor in these vehicles, with sample period, there are periods of drawdown com-
similar drawdowns and skewness for all the benchmark- monality. The momentum strategies exhibit their largest
relative offerings. When moving to active return space,
there is a larger decrease in the Sharpe ratio from the 9
The drawdown range for the long-only ETF value strategy
130/30 strategy at 0.79, to the long-only strategy at 0.50, occurs earlier, in May 2007, compared to October 2007 for the
and finally to the long-only ETF strategy at 0.38. 130/30 and long-only value strategies. This is because of the limited
Benchmark-relative value. Panel B of Exhibit 4 semiannual rebalancing for the long-only ETF.
10
reports summary statistics for the benchmark-relative Part of this is because the benchmark-relative value strate-
gies have inherited the negative skewness present in the market itself
value strategies. Like the momentum strategies in (see French, Schwert, and Stambaugh 1987). Note that the skewness
Panel A, Sharpe ratios decrease as leverage decreases of the 130/30 strategy is slightly less in absolute value, at –0.8, than
and rebalancing decisions are taken less frequently. that of the other two long-only strategies.
&XPXODWLYH/RJ5HWXUQV
±
3DQHO%9DOXH6WUDWHJLHV
&XPXODWLYH/RJ5HWXUQV
±
Notes: Past performance is not indicative of future results. Returns do not account for any fees or transaction costs. If fees were included, returns would be
lower. For illustrative purposes only using hypothetical strategies and not representative of an actual investment or account. The modeled performance is
calculated with the benefit of hindsight and knowledge of factors that may have positively affected its performance and has inherent limitations. This analysis
cannot account for risk factors that may affect actual portfolio performance.
110 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
drawdowns starting in mid-2008 through the end of We start with Panel A of Exhibit 6 and the
2009, whereas the value strategies exhibit their largest momentum strategies. For the market-neutral strate-
drawdowns through the late 1990s, with the 130/30 gies, the correlation exceeds 0.99 most of the time. The
strategy exhibiting a significantly shorter drawdown correlations fall during two periods that coincide with
than the long-only strategies. The long-only portfo- outperformance of the private fund strategy: November
lios, which rebalance every month, outperformed the 2005 to January 2008 and October 2014 to the end of
semiannually rebalanced ETFs. The more frequent the sample in June 2017.
rebalancing allows more optimal exposure to the factors. Although the correlations of the benchmark-relative
momentum strategies are lower than those of the
HOLDINGS AND LEVERAGE market-neutral strategies, they are also high, above 0.90.
Overall, the positions in the long-only and the long-
In the first subsection of this part of the article, we only ETF strategies are most similar. The correlations
show that although all of the representative implemen- for the 130/30 and the long-only ETF are lowest, as
tations—private fund, UCITS, 130/30, long-only, and would be expected, given the differences in leverage
long-only ETF—pursue the same momentum and value and rebalancing frequency.
factors, the specific holdings of the hypothetical strate- The cross-sectional correlations of the value strat-
gies vary significantly across the different investment egies, in contrast, can be fairly low. In Panel B, the
constraints. In the second subsection, we discuss how market-neutral private fund and UCITS have an average
leverage varies over time for the factor implementations, correlation of 0.92, but there is significant variation,
which allow shorting. ranging from a minimum of 0.82 to a maximum of
more than 0.99. The final holdings correlation is 0.93.
Comparing Factor Strategy Holdings Interestingly, correlations tend to fall during bull mar-
kets and rise during bear markets. The low correlation
We plot cross-sectional holdings correlations in periods also tend to coincide with times in which the
Exhibit 6. For each of the factors, we compare the hold- private fund strategy outperformed the UCITS strategy.
ings of the two hypothetical market-neutral strategies In summary, although the risk and return charac-
and the three hypothetical benchmark-relative strategies. teristics of the momentum and value strategies can be
We compute one correlation series for the market-neu- very similar, the actual holdings generating those returns
tral private fund and UCITS strategies and three correla- may sometimes differ markedly.
tion series for the benchmark-relative strategies (130/30
and long-only mutual fund, long-only mutual fund and Time-Varying Leverage
long-only ETF, and 130/30 and long-only ETF).11
Exhibit 7 examines the time-varying leverage of
11
The cross-sectional holdings correlations are computed as the momentum and value strategies for the hypothetical
follows. At each month t, there are two portfolios, X = [x(1, t), market-neutral strategies. The strategies are calibrated to
x(2, t), … , x(n, t), … , x(N, t)] and Y(t) = [y(1, t), y(2, t), … , y(n, t), … , y
achieve an ex ante risk of 10% (7%). Note that we place
(N, t)], each consisting of a set of holdings in n = 1,2, … ,N equities.
Let x (t ) and y (t ) denote the sample means of X(t) and Y(t), and let a constraint on UCITS leverage of 4× per side (gross
sx(t) and sy(t) be the corrected sample standard deviations of X(t) leverage of 8×), which is almost always binding for the
and Y(t). The means and standard deviations are computed at time value UCITS strategy.
t in the cross section, n = 1,2, … ,N. The cross-sectional holdings First, the leverage of the value factor is higher
correlation between these two portfolios at time t is calculated as than or equal to the leverage of the momentum strategy
∑
N
( x(n, t ) − x (t ))( y(n, t ) − y (t )) to achieve the same ex ante volatility. This is because
rXY (t ) = n =1
. momentum strategies have higher levels of risk than
( N − 1)σ x (t )σ y (t )
do value strategies. As a result, momentum strategies
For each pair of portfolios (X(t), Y(t)), Exhibit 6 plots the time require less leverage to hit a given ex ante risk target.
series rXY (t). For correlations relative to the long-only ETF portfolio,
Second, given the constant risk target of these
we plot a 12-month exponentially weighted moving average of the
correlation to control for the semiannual rebalancing dates. If the portfolios, the leverage of each strategy is inversely
holdings are identical, then the holdings correlations are equal to related to the volatility of the overall equity market.
one.
+ROGLQJV&RUUHODWLRQ
3ULYDWH)XQG8&,76 /RQJ2QO\
/RQJ2QO\/RQJ2QO\(7) /RQJ2QO\(7)
3DQHO%9DOXH6W\OH)DFWRU
+ROGLQJV&RUUHODWLRQ
3ULYDWH)XQG8&,76 /RQJ2QO\
/RQJ2QO\/RQJ2QO\(7) /RQJ2QO\(7)
Note: For illustrative purposes only using hypothetical strategies and not representative of an actual investment or account.
112 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
Exhibit 7
Gross Leverage across Style Factors and Hypothetical Market-Neutral Implementation
*URVV/HYHUDJH
0RPHQWXP3ULYDWH)XQG 0RPHQWXP8&,76
9DOXH3ULYDWH)XQG 9DOXH8&,76
Note: For illustrative purposes only using hypothetical strategies and not representative of an actual investment or account.
Times of high market volatility also tend to coincide We examine how implementation matters through
with market drawdowns. Given this relation with vola- the lens of two well-known factor investment strate-
tility, the leverage of both momentum and value tend to gies: momentum and value. Eligible investors can choose
be positively correlated; the gross leverage numbers of to implement these in relatively unconstrained private
the private fund momentum and value strategies have a funds—which take long and short positions—or the most
correlation of 0.81 over the simulation period. constrained ETFs, of which the vast majority are index-
Third, leverage varies over time, which carries based, fully transparent, and long only. In between,
strong ramifications for risk management capabilities of investment constraints vary across European UCITS,
factor investors using leverage. The average gross leverage mutual funds or unit trusts, funds of one, or separately
of the private fund momentum strategy is 5.3×, but it managed accounts. We take constraints on leverage and
varies from a minimum of 3.4× to a maximum of 8.6×. Its turnover typical for each vehicle and examine the effect
leverage also has a negative correlation of -0.72 with the on returns, risk, holdings, and other portfolio charac-
risk of the overall strategy. The private fund value strategy teristics induced by those constraints.
exhibits higher average leverage of 8.9×. The minimum We find that, in many cases, more unconstrained
and maximum leverage values are 5.0× in April 2001 hypothetical implementations with more f lexible use of
and 11.7× in August 2007, respectively. Like momentum, leverage result in higher Sharpe ratios and better perfor-
the private fund value leverage has a negative correlation mance in terms of downside risk statistics, such as max-
(-0.74) with the risk of the strategy itself. imum drawdowns. Higher leverage does not directly
correspond to higher risk or higher turnover. Indeed,
CONCLUSION some of the additional f lexibility with higher leverage is
employed by an optimizer to help reduce risk.
In the real world, the practical decisions of invest- We focus only on hypothetical momentum and
ment vehicle and the resulting constraints may lead to value strategies in equities. Implementation consider-
meaningful differences in the returns of factor strategies. ations also affect, potentially to an even higher degree,
Cheng, M., and A. Madhavan. 2009. “The Dynamics of Madhavan, A. Exchange-Traded Funds and the New Dynamics of
Leveraged and Inverse-Exchange Traded Funds.” Journal of Investing. New York: Oxford University Press, 2016.
Investment Management 7 (4).
Qian, E. E., R. H. Hua, and E. H. Sorensen. Quantitative
Clarke, R., H. de Silva, and S. Sapra. 2004. “Towards More Equity Portfolio Management: Modern Techniques and Applications.
Information-Efficient Portfolios.” The Journal of Portfolio Man- Boca Raton, FL: Chapman & Hall/CRC, 2007.
agement 31 (1): 34–63.
Scharfman, J. Hedge Fund Governance: Evaluating Oversight,
Daniel, K., and T. J. Moskowitz. 2016. “Momentum Crashes.” Independence, and Conflicts. San Diego, CA: Academic Press,
Journal of Financial Economics 122 (2): 221–247. 2015.
114 Relaxing Constraints Improves Returns of Factor Strategies Quantitative Special Issue 2019
Trade-Off in Multifactor Smart
Beta Investing: Factor Premium
and Implementation Cost
Feifei Li and Joseph (Yoseop) Shim
F
Feifei Li actor investing, which requires sense, a strategy that requires stock selection
is director, head of choosing securities based on attri- based on multiple measures is often viewed
investment strategy at
butes associated with high expected as complex in terms of product features and
Research Affiliates, LLC,
in Newport Beach, CA. returns, is becoming increasingly the construction process. Brightman et al.
li@rallc.com popular as investors realize they can harvest (2017) answered two important questions
the excess return over the market-capital- for multifactor strategy investors: (1) Which
Joseph (Yoseop) Shim ization-weighted benchmark by a simple, factors should be included in a multifactor
is vice president of smart transparent, and rules-based approach. Some portfolio, and (2) how to allocate assets across
beta at Research Affiliates,
LLC, in Newport Beach,
of the most commonly used factors in investor factors over time. The authors reviewed the
CA. portfolios are value, momentum, quality, low theory and evidence in support of the value,
shim@rallc.com risk, and size, although many other factors are momentum, profitability, investment, low
used. These common five factors are based volatility, and size factors and then demon-
on the well-established academic literature of strated that a combination of those six factor-
Fama and French (1993, 2012, 2016), Carhart based smart betas can outperform the market
(1997), and Frazzini and Pedersen (2014). with their much lower tracking error (TE)
Because some factors are not highly corre- and downside risk.
lated, combining them can further benefit In this follow-up article, we focus more
investors through the diversification effect.1 on the implementation issues of multifactor
A multifactor strategy, which is a one-stop investing, including portfolio concentra-
solution for investors who are seeking the tion, turnover, trading cost, and capacity, all
excess returns associated with various fac- of which are crucial elements in real-world
tors and which can give investors the ability product design. We address two questions
to weather the storm during adverse market in the presence of implementation costs:
conditions, seems to be a natural path for Which factors should be included in a mul-
factor-investing product proliferation. tifactor portfolio, and what is the best way to
When investors start down the path of construct a single-factor portfolio? We pro-
multifactor investing, they typically raise a vide the detailed analysis that supports our
number of important questions. Even though recommendations.
at a conceptual level combining factors makes
FACTOR PERFORMANCE
1
Among others, Asness, Moskowitz, and AND CORRELATION
Pedersen (2013) and Asness et al. (2015) documented
the diversification benefit of combining multiple fac- Consistent with Brightman et al. (2017),
tors with low correlation to each other. we include six factor-based smart beta
$EVROXWH3HUIRUPDQFH 5HODWLYH3HUIRUPDQFH
6WUDWHJ\ 5HWXUQ 9RO 6KDUSH5DWLR 9DOXH$GG 7( ,5
0DUNHW
9DOXH
/RZ%HWD
3URILWDELOLW\
,QYHVWPHQW
0RPHQWXP
6L]H
$YHUDJHRI6L[)DFWRUV
strategies in our tests: value, low beta, profitability, make the strategy a helpful diversifier in the aggregate
investment, momentum, and size. Each of these factors portfolio of factor-based smart beta strategies.
is widely deemed robust in the academic literature (Fama We first review the correlations between the fac-
and French 1993, 2012, 2016; Carhart 1997; Frazzini tors to understand how they may interact with each
and Pedersen 2014). To illustrate the opportunities pre- other. The correlations across a factor’s excess return
sented by investing in real-world strategies, we construct over market are mostly low or negative, implying their
simple long-only, investable portfolios in accordance performance drivers have orthogonal components.
with widely accepted academic and investment practice. On average, the pairwise cross-factor correlation,
We start with the large-cap universe of US stocks, reported in Exhibit 2, is very close to zero (i.e., 0.01).
except for the small-size strategy. For each of these Among the factors we study, the only factors that appear
five factor portfolios, we use the top 30% based on to have a correlation above 0.4 are investment, value,
NYSE breakpoints to select stocks based on the cor- and low beta. This observation suggests that combining
responding characteristics. For example, we construct those factors can create diversification benefits.
the value portfolio from stocks above the 70th per- In the absence of trading costs, including momentum
centile on the NYSE by book-to-market ratio. Our and size factors in a multifactor strategy seems a no-
small-size strategy consists of all the available stocks in brainer, given the significant diversification benefit these
the small-cap universe. We then capitalization weight factors bring to the table. The momentum factor has a
the selected stocks, except for low beta, which we negative correlation with three of the five other factors
weight by beta ranking. The portfolios are rebalanced (value, low beta, and investment) and a slight positive cor-
annually each July with the exception of momentum relation with the remaining two (profitability and size).
and low beta, which are rebalanced quarterly. Details The size factor seems to be a strong diversifier when com-
about our factor construction method are provided in bined with the profitability factor because the two have a
the Appendix. correlation of −0.60, and it has relatively low correlation
On average, these six factor-based smart beta with the other factors. Without considering implementa-
strategies deliver economically signif icant returns. tion issues, both the momentum and size factors seem to
The average annualized excess return over our study be desirable from a diversification perspective.
period, July 1973 to June 2017, is 1.96%, as illustrated in
Exhibit 1. As noted by Brightman et al. (2017), profit- IMPLEMENTATION COSTS OF FACTORS
ability, on its own, generates the lowest return. As we
will show, however, profitability’s low or negative corre- Chow et al. (2017) documented that the cost of
lations with other factors, as well as its low trading cost, executing certain factor-based smart beta strategies can
116 Trade-off in Multifactor Smart Beta Investing: Factor Premium and Implementation Cost Quantitative Special Issue 2019
Exhibit 2
Cross-Correlation between Factor Excess Returns, United States, July 1973–June 2017
be nontrivial. Following Chow et al., we estimate the another strategy that spreads out trades across many
implementation costs of these strategies by applying the holdings.
Aked and Moroz (2015) market-impact model. Imple-
mentation cost is the movement in a security’s price as Intuitively, low portfolio volume, high tilt, high
a result of trading execution: When a large buy order turnover, and high turnover concentration would lead
of a particular security is placed, it typically drives up to a high market-impact cost. By using these attributes,
the price of the security. The aggregate implementa- we can explain how different aspects of a strategy are
tion cost of a portfolio rebalancing is the summation of accountable for its implementation costs.
all such costs across all stocks traded to accomplish the The average turnover across these six strategies
rebalancing. is 62.4%, and the average estimated trading cost of the
With simplifying assumptions, we can attribute strategies is 13 bps, given the assumption of $1 billion in
the market impact of a rebalancing on the performance assets under management (AUM).2 The costs of imple-
of a smart beta strategy to the following characteristics: menting a strategy with high turnover, strong tilt, or
low volume can be quite large. Using the momentum
• Portfolio volume—the aggregate of the median daily strategy as an example, and as reported in Exhibit 3,
trading volume of all stocks in a portfolio. A strat- annualized one-way turnover is 160.4%, assuming
egy’s cost is inversely proportional to its portfolio quarterly rebalancing. This extraordinarily high turn-
volume; in other words, a strategy is more costly over creates a high trading cost (28 bps at $1 billion
to implement if it requires holding illiquid stocks. AUM) for the strategy even though its portfolio volume
• Tilt—a measure of illiquidity that represents the ($35.6 billion) and tilt (1.29) are at reasonable levels.
degree to which portfolio weights deviate from At $5 billion AUM, trading costs erode 138 bps of
the weights proportional to trading volume. The return. If we increase AUM to $10 billion, the trading
smallest possible tilt value is 1, which is achieved cost (277 bps) more than offsets the entire estimated
by a volume-weighted portfolio, theoretically the value-add (164 bps) of the strategy reported in Exhibit 1.
most liquid combination of a given set of holdings. For this reason, we do not recommend that investors use
• Turnover—a measure of how frequently assets in the naively constructed momentum factor as a stand-
a portfolio are bought and sold by the strategy alone investment strategy.
over a 12-month period, typically a calendar year.
In general, a strategy that requires a higher rate of
trading incurs higher market-impact costs. 2
We report the recent five-year average of the estimated
• Turnover concentration—the degree to which trades trading costs instead of the historical average based on the entire
are spread across the holdings in a portfolio. Highly sample. The trading volume has been increasing exponentially since
the explosion of electronic trading in the late 1990s. Therefore, the
concentrated trades are more costly to execute. current estimate is a much better representation of the expected
A strategy that demands high liquidity from only trading cost than the historical average, which would be unreason-
a few names causes a higher market impact than ably high by today’s standards.
In contrast, the profitability strategy is quite inex- a multifactor strategy, however, because of their nega-
pensive to trade (2 bps at $1 billion AUM) because the tive or low positive correlations with other factors. So,
profitability factor tends to favor larger and more-liquid how should investors measure this diversification benefit
stocks. This characteristic is ref lected by its high port- versus the increased costs of implementation?
folio volume ($52.2 billion) and low tilt (1.27) and by We consider four portfolios to illustrate how
the more stable portfolio holdings indicated by its low including the momentum and size factors affects the
turnover (16.7%). The value strategy also incurs rela- performance characteristics and implementation costs of
tively low trading costs (6 bps at $1 billion AUM) by a multifactor strategy. We combine the factors in each
virtue of its relatively low turnover and tilt. portfolio by allocating an equal weight to each factor
Interestingly, the size strategy, which focuses pri- and rebalancing every quarter:
marily on stocks with a small market capitalization and
which is often believed to be a high-cost strategy, actu- • Portfolio 1: Value, low beta, profitability, and
ally has a below-average trading cost (10 bps at $1 billion investment
AUM). Although the size factor does tend to trade small, • Portfolio 2: Four factors in portfolio 1 plus
illiquid stocks, as indicated by its low portfolio volume momentum
($12.4 billion) and high tilt (2.38) relative to the other • Portfolio 3: Four factors in portfolio 1 plus size
factors, it has low turnover (26%) as a result of its broad • Portfolio 4: Four factors in portfolio 1 plus
coverage—2,232 names as of June 2017. On balance, the momentum and size
higher coverage of the size strategy almost offsets the
cost of trading smaller companies. The performance and implementation cost of
these four portfolios are reported in Exhibits 4 and 5,
INCLUSION OF MOMENTUM AND SIZE respectively.
IN A MULTIFACTOR STRATEGY The implementation cost of the multifactor
strategy is much lower than the simple average of
In this section, we study the impact of including the implementation costs of each of the underlying
the supposedly high-cost strategies of momentum factors. As reported in Exhibit 3, the average cost of
and size on the performance and cost of a multifactor implementing the six single-factor portfolios is 127
strategy. Momentum and size can be good additions to bps at $10 billion AUM. By contrast, the trading cost
118 Trade-off in Multifactor Smart Beta Investing: Factor Premium and Implementation Cost Quantitative Special Issue 2019
Exhibit 4
Performance of Multifactor Smart Beta Strategies, United States, July 1973–June 2017
$EVROXWH3HUIRUPDQFH 5HODWLYH3HUIRUPDQFH
1HWRI&RVW65 1HWRI&RVW,5
9DOXH
6WUDWHJ\ 5HWXUQ 9RO 65 ELOOLRQ ELOOLRQ ELOOLRQ $GG 7( ,5 ELOOLRQ ELOOLRQ ELOOLRQ
0DUNHW
)DFWRUV
)DFWRUV0RP
)DFWRUV6L]H
)DFWRUV
0RP 6L]H
Exhibit 5
Implementation Cost of Multifactor Smart Beta Strategies, United States, July 1973–June 2017
7XUQRYHUDQG7UDGLQJ&RVW
&XUUHQW 3RUWIROLR
:$0& 9ROXPH 7UDGLQJ&RVWDW
&XUUHQW 7XUQRYHU &DSDFLW\
6WUDWHJ\ 1 ELOOLRQ 7XUQRYHU ELOOLRQ &RQFHQWUDWLRQ 7LOW ELOOLRQ ELOOLRQ ELOOLRQ ELOOLRQ
0DUNHW
)DFWRUV
)DFWRUV0RP
)DFWRUV6L]H
)DFWRUV
0RPDQG6L]H
of a multifactor strategy that equally weights the six in the multifactor strategy as a result of diversification.
factors and rebalances every quarter (portfolio 4) is This observation appears to be informative regarding the
only 25 bps, as Exhibit 5 shows. decision an investor pursuing a multifactor strategy must
The huge reduction in trading costs is partly due make: To capture the desired factor premiums, should
to the cancelling trades initiated by uncorrelated factors the investor hire a different manager for each factor
at the headline portfolio level. For example, in Exhibit 5, strategy or hire one manager of a multifactor strategy?
the one-way turnover of an equally weighted six-factor We demonstrate that the latter choice is more cost effec-
strategy including momentum and size is 55.7%, which tive from a trading perspective.
is lower than the simple average turnover (62.4%) of We now examine portfolio 2, the impact of
the six individual factors, as reported in Exhibit 3. adding a momentum sleeve to the four-factor strategy.
In addition, by including all six factors, the multifactor As expected, before applying the trading cost model,
strategy has a portfolio volume quite close to that of the because of negative or low positive correlations with
entire market. other factors, the addition of momentum to an equally
Moreover, the turnover concentration of the mul- weighted four-factor strategy (value, low beta, profit-
tifactor strategy is smaller than the simple average turn- ability, and investment) reduces TE by 83 bps (3.46%
over concentration of the six underlying factors, which versus 4.29%) and leads to a much higher information
implies that trades are more spread out across holdings ratio (IR), 0.60 versus 0.49.
120 Trade-off in Multifactor Smart Beta Investing: Factor Premium and Implementation Cost Quantitative Special Issue 2019
Exhibit 6
Net-of-Cost Performance versus Concentration Level, United States, July 1973–June 2017
3DQHO$1HWRI&RVW6KDUSH5DWLRYHUVXV6HOHFWLYHQHVVPXOWLIDFWRU 3DQHO%1HWRI&RVW,QIRUPDWLRQ5DWLRYHUVXV6HOHFWLYHQHVVPXOWLIDFWRU
$IWHU&RVW,QIRUPDWLRQ5DWLR
$IWHU&RVW6KDUSH5DWLR
6HOHFWLYHQHVV 6HOHFWLYHQHVV
to the 25% concentration level and remains more or less concentration level and then f lattens out at $1 billion
unchanged after that. The IR also increases from 0.53 to AUM. At $5 billion and $10 billion AUM, however,
0.70 from the 50% to the 25% level but declines there- the net-of-cost Sharpe ratio declines rapidly beyond the
after, even before considering trading costs. 25% concentration level. For example, at $10 billion
The trade-off between performance and cost AUM, the highly concentrated portfolio, with a 15%
becomes much more important at the higher concen- cutoff, yields a lower net-of-cost IR (0.53) than the less-
tration levels because the impact of trading costs on the concentrated portfolio at a 25% cutoff (0.61).
Sharpe ratio and the IR is much larger in more-con- To summarize, the sweet spots for the best per-
centrated multifactor strategies. As we make the under- formance in the presence of implementation costs seem
lying factor portfolios more concentrated, the turnover, to reside around the 25% concentration level, based on
turnover concentration, and tilt of the multifactor our sample. We acknowledge, however, that if we only
strategy all tend to increase, whereas the portfolio focus on empirical evidence, the optimal cutoff level
volume decreases, leading to higher implementation may differ slightly across individual factor portfolios.
costs. The goal of our analysis, however, is to effectively cap-
Additionally, the mix with more-concentrated ture the factor premium at a reasonable cost through a
portfolios has fewer offsetting trades that cancel each simple, transparent, and rules-based approach. Lacking
other out, and thus the turnover is reduced by a lesser a strong reason for the individual factor portfolios to
amount. As a result, the reduction in performance be constructed using different concentration levels, we
from implementation costs is much higher for more- suggest investors avoid overfitting risk and use consistent
concentrated strategies, especially under the assumption coverage across all factors included in the multifactor
of larger AUM. For instance, the net-of-cost Sharpe portfolio.
ratio monotonically increases from the 50% to 25%
3DQHO%,PSOHPHQWDWLRQ&RVWV
7XUQRYHUDQG7UDGLQJ&RVW
&XUUHQW 3RUWIROLR
:$0& 9ROXPH 7UDGLQJ&RVWDW
&XUUHQW 7XUQRYHU &DSDFLW\
6WUDWHJ\ 1 ELOOLRQ 7XUQRYHU ELOOLRQ &RQFHQWUDWLRQ 7LOW ELOOLRQ ELOOLRQ ELOOLRQ ELOOLRQ
0DUNHW
0XOWLIDFWRU
0XOWLIDFWRU
0XOWLIDFWRU
0XOWLIDFWRU
0XOWLIDFWRU
0XOWLIDFWRU
0XOWLIDFWRU
0XOWLIDFWRU
122 Trade-off in Multifactor Smart Beta Investing: Factor Premium and Implementation Cost Quantitative Special Issue 2019
We find that by using more-concentrated under- Exhibit A1
lying factor portfolios in constructing a multifactor Signals Used to Sort Factor-Based Smart Beta
strategy, investors can improve its performance. These Portfolios
benefits come, however, at the expense of higher vola-
tility, TE, turnover, and trading costs. Our study sug- )DFWRU 6LJQDO 'HILQLWLRQ
gests that the sweet spots for the best risk-adjusted 9DOXH %RRNWR3ULFH5DWLR %RRN9DOXH0DUNHW&DS
performance in the presence of implementation costs /RZ%HWD 0DUNHW%HWD )UD]]LQLDQG3HGHUVRQ>@
reside around the 25% concentration level. GHILQLWLRQLQZKLFK
Neither of the extremes of (1) maximizing paper FRUUHODWLRQLVHVWLPDWHGZLWK
ILYH\HDUVRIGDLO\UHWXUQV
portfolio performance while ignoring the trading
DQGYRODWLOLW\ZLWKRQH\HDU
costs that reduce such performance in practice nor (2) RIGDLO\UHWXUQV
focusing on low-cost implementation while missing
3URILWDELOLW\ 2SHUDWLQJ $QQXDOUHYHQXHVPLQXVFRVW
opportunities for better performance will produce an 3URILWDELOLW\ RIJRRGVVROGLQWHUHVW
optimal result for multifactor smart beta investors. We H[SHQVHDQGVHOOLQJJHQHUDO
strongly advocate the thoughtful design of a multi- DQGDGPLQLVWUDWLYHH[SHQVHV
factor strategy, which requires a conscious and delib- GLYLGHGE\ERRNHTXLW\IRU
erate decision to find the most advantageous balance WKHODVWILVFDO\HDU
between effectively harvesting the factor premium and ,QYHVWPHQW &KDQJHLQ$VVHWV <HDURYHU\HDUSHUFHQWDJH
implementation costs. FKDQJHLQWRWDODVVHWV
0RPHQWXP íWRí0RQWK 3ULRUPRQWKUHWXUQV
5HWXUQ VNLSSLQJPRVWUHFHQWPRQWK
Appendix 6L]H 0DUNHW&DS 0DUNHW&DS
124 Trade-off in Multifactor Smart Beta Investing: Factor Premium and Implementation Cost Quantitative Special Issue 2019
Factor Investing from Concept
to Implementation
Eduard van Gelderen, Joop Huij, and Georgi Kyosev
M
Eduard van utual funds following factor To better understand how investors
Gelderen investing strategies based on dynamically allocate to factor funds, we
is the chief investment
equity asset pricing anoma- study the f low-performance relation for
officer at PSP Investments
in Montreal, QC, Canada. lies, such as the small cap, these funds. Although factor funds have
evangelderen@investpsp.ca value, and momentum effects, earn signifi- attracted signif icant fund f lows over our
cantly higher alphas than traditional actively sample period, it appears that fund f lows
Joop Huij managed mutual funds. This effect is unre- have been driven by factor funds earning
is an associate professor lated to other fund characteristics such as age, high past returns and not by the funds pro-
at the Rotterdam
School of Management
expenses, and turnover; is robust to a global viding factor exposures. Similar to Zheng
in Rotterdam, the sample of mutual funds and bootstrapped (1999), we f ind very little evidence of a
Netherlands, and the confidence intervals; and is stronger for funds “smart money” effect in the sense that
head of factor investing at that are exposed to multiple factors simul- f lows predict future fund performance. In
Robeco in Rotterdam, the taneously. Although excess returns earned fact, consistent with the recent findings of
Netherlands.
by factor funds net of fees are significantly Cornell, Hsu, and Nanigian (2017), we do
jhuij@rsm.nl
smaller than the theoretical premiums of the not observe a positive relationship between
Georgi Kyosev asset pricing anomalies, they are still positive fund f lows and future performance.
is a PhD candidate and statistically and economically significant. We argue that rather than timing fac-
at the Rotterdam For example, if investors would randomly tors and factor managers, investors would be
School of Management select a factor fund and would apply a buy- better off by using a buy-and-hold strategy
in Rotterdam, the
Netherlands, and a
and-hold strategy, these investors would earn and selecting a multifactor manager. For
researcher at Robeco 110 bps per annum in excess of the return that example, if an investor would randomly
in Rotterdam, the is earned by the average traditional actively select a factor fund that is exposed to two
Netherlands. managed mutual fund. factors simultaneously and would apply a
kyosev@rsm.nl Nevertheless, the actual returns that buy-and-hold strategy, this investor would
investors earn by investing in factor mutual earn 190 bps per annum in excess of the
funds appear to be significantly lower than return that is earned by the average tradi-
this number because investors do not follow tional actively managed mutual fund. This
buy-and-hold strategies, but rather dynami- number would be 240 bps per annum if the
cally reallocate their funds both across factors investor would have selected a manager that
and factor managers. By attempting to time is exposed to three factors simultaneously,
across factors, investors lose a large portion of and even 270 bps per annum if the manager
the return they could earn with a buy-and- would be exposed to four or more factors
hold strategy. simultaneously.
126 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
Methodology To measure fund performance, we use the inter-
cept from the following one-factor model:
Our empirical analyses consist of three main
sections: respectively, evaluating the performance of Ri ,t = α i + βi ⋅ ( RM , t − R f , t ) + εi ,t (2)
factor fund managers; computing the actual returns
earned by investors in factor funds; and investigating We use CAPM alpha instead of the intercept from
the f low-performance relation for factor funds. Equation 1 as our main performance measure because we
Factor fund classification and performance want to measure the excess return coming from exposures
evaluation. In the first empirical section of our article to one of the six factors. We limit the effect of outliers
we investigate whether mutual funds following factor by calculating the Z-score of fund alphas, winsorizing it
investing strategies based on equity asset pricing anoma- at -2 and 2:
lies, such as the small cap, value, and momentum effects,
earn higher alphas than traditional actively managed α − µα
mutual funds. For these analyses we employ three sta- z _ Alphai = min 2, max −2, i (3)
σ α
tistical techniques: return-based style analysis to classify
factor funds; cross-sectional regressions to evaluate the
where ai is the alpha of fund i from the one-factor
performance of factor funds; and bootstrap analyses to
model, ma is the average alpha across all funds in the
test the robustness of our results.
sample, and sa is the cross-sectional standard deviation
Our fund classification method closely follows
of all fund alphas.
the methodology employed by van Gelderen and
We use the following cross-sectional regression to
Huij (2014). We download monthly factor returns
evaluate the performance of factor funds:
from Kenneth French’s data library. For each fund
we run the five-factor Fama and French model, aug- z _ Alphai = α i + b1 ⋅ Low _ beta + b2 ⋅ Small _ cap
mented with momentum, using all available return
observations + b3 ⋅Value + b4 ⋅ Momentum
+ b5 ⋅ Profitability + b6 ⋅ Investments + εi (4)
Ri ,t = α i + βi ⋅ ( RM ,t − R f ,t ) + si ⋅SMBt + hi ⋅ HML t
where Low_beta, Small_cap, Value, Momentum, Profit-
+ w i ⋅WML t + ri ⋅ RMWt + c i ⋅ CMAt + εi ,t (1)
ability, and Investments are 1 if the fund is classified as a
where R i,t is the excess return of mutual fund i in low beta, small cap, value, momentum, profitability, or
month t; R f,t is the risk-free return in period t; ai is investments factor fund, or 0 otherwise.
the alpha of fund I; R M,t is the return on the market We also run an augmented version of this
portfolio in period t; and SMB, HML, WML, RMW, regression:
and CMA are returns of long–short factor-mimicking
z _ Alphai = α i + b1 ⋅ Low _ beta + b2 ⋅ Small _ cap
portfolios for the size, value, momentum, profitability,
and investments factors, respectively. b, s, h, w, r, and c + b3 ⋅Value + b4 ⋅ Momentum + b5 ⋅ Profitability
are the estimated fund-specific factor coefficients, and + b6 ⋅ Investments + b7 ⋅ log age + b8 ⋅ log size
ei,t is the residual return of fund i in month t, under + b19 ⋅ exp _ ratio + b10 ⋅ turn _ ratio + εi
the assumption of i.i.d. Similar to van Gelderen and
(5)
Huij (2014), we classify a fund as being a factor fund
if the regression coefficient on the respective factor is where log age is the natural logarithm of fund age, cal-
positive and statistically significant. For example, if the culated as the number of months with available return
SMB beta coefficient of fund i is higher than 2, we observations; log size is the natural logarithm of fund
identify fund i as a small cap fund. A fund is consid- size, measured as its average total net assets; exp_ratio
ered to be a low-beta fund when its b is smaller than is the average total expense ratio; and turn_ratio is the
0.8. Funds can have multiple factor fund classifications average turnover ratio. For our global markets sample,
simultaneously. we do not include exp_ratio and turn_ratio in our regres-
sions because the underlying data are unavailable.
128 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
Exhibit 2
Distribution of Fund Alphas
Notes: This exhibit shows distributions of annualized fund alphas across all US funds in the CRSP Mutual Fund Database with total assets exceeding
$5 million. Alphas are calculated per fund as the intercept from CAPM regressions over all available observations during the sample period of January 1990
to December 2015. “< -4” shows the percentage of funds with annualized alphas less than -4%; “-4:-2” shows the percentage of funds with annualized
alphas between -2% and -4%; and “% > 0” shows the number of funds with positive alphas.
(or past 12-month CAPM alpha in some of our regres- test whether this return is different from the return that
sion specifications) and top, middle, and bottom are cal- a buy-and-hold investor obtains by randomly selecting
culated as follows: a factor fund. Finally, in the third section, to better
understand how investors dynamically allocate to these
rankibottom
,t = min(ranki ,t ,0.2) funds, we study the f low-performance relation for factor
mutual funds.
rankimiddle
,t = min(ranki ,t − rankibottom
,t ,0.6)
Do Factor Funds Earn Higher Alphas?
rank top
i ,t = min(ranki ,t − rank bottom
i ,t − RANK middle
i ,t ,0.2) In our first analyses we consider the distribution
of fund alphas for various fund classifications. Exhibit 2
where ranki,t is the rank of fund i in month t based on the shows that factor funds earn significantly higher alphas
measure of past performance, which is past 12-month than traditional actively managed mutual funds. Only
outperformance or past 12-month CAPM alpha, 17% of the traditional actively managed mutual funds
depending on the regression specification. earn positive alphas after fees in the long run, but this
number is substantially larger for factor funds: 52% for
EMPIRICAL RESULTS low-beta funds; 53% for small cap funds; 52% for value
funds; 40% for momentum funds; 57% for profitability
This section describes the results of our empirical funds; and 60% for investments funds.
analyses along three research questions: (1) Do factor To test whether differences in performance are
premiums exist in mutual fund returns? (2) Do investors statistically significant and independent, we perform
in factor funds successfully harvest factor premiums? and a regression analysis in which we regress fund perfor-
(3) What drives the allocation decision of investors in mance on fund classifications. The results of this regres-
mutual funds? sion analysis are presented in Panel A of Exhibit 3 and
In the first section of our empirical analysis, we indicate that factor funds earn significantly higher alphas
investigate whether mutual funds following factor than traditional actively managed mutual funds. Specifi-
investing strategies earn higher alphas than traditional cally, when we consider the results in our most parsimo-
actively managed mutual funds. In the second section, nious specification (Exhibit 3, Panel A, Regression 7),
we calculate what returns investors earn by investing in we find that funds with exposure to the low-beta, small
mutual funds that follow factor investing strategies and cap, value, momentum, profitability, and investments
Notes: This exhibit shows univariate and multiple regression results of all US funds during the sample period of January 1990 to December 2015 in the
CRSP Mutual Fund Database with total assets exceeding $5 million. The winsorized (at -2 and 2) Z-score of CAPM alphas is regressed on dummies,
indicating funds belonging to a specific factor group. Ln(age) is the natural logarithm of the fund’s age, calculated as the number of months the fund has been
in our sample. Ln(size) is the natural logarithm of the average fund’s total assets in US dollars. exp_ratio and turn_ratio are the average expense ratio and
turnover ratio per mutual fund in our sample.
130 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
factors have alpha that is, respectively, 0.34, 0.48, 0.20, Exhibit 4
0.12, 0.35, and 0.30 standard deviations higher than Multifactor Exposures and Outperformance
that of traditional actively managed mutual funds. The
t-values of these coefficient estimates are larger than 3 'HS9DULDEOH ]BDOSKD ]BDOSKDFRQWUROV
in all cases, indicating that our results are statistically ,QWHUFHSW ± ±
significant. WVWDW >±@ >±@
In Panel B of Exhibit 3 we extend the analysis by IDFWRU
controlling our regressions for fund characteristics such WVWDW >@ >@
as fund age, size, total expense ratio, and turnover ratio. IDFWRUV
WVWDW >@ >@
Our results appear to be robust to controlling for these
IDFWRUV
fund characteristics because the coefficient estimates and
WVWDW >@ >@
their t-values remain very similar to our first results. IDFWRUV
To further understand the effect of factor expo- WVWDW >@ >@
sures on mutual fund performance, we classify funds OQDJH
according to the number of factors to which they are WVWDW >@
exposed, presented in Exhibit 4. Groups are mutually OQVL]H
exclusive and contain funds with significant loading WVWDW >@
to one, two, three, and four or more factors. Results H[SBUDWLR ±
WVWDW >±@
provide convincing evidence that a larger number of
WXUQBUDWLR ±
factor exposures leads to higher risk-adjusted mutual
WVWDW >±@
fund returns even after transaction costs and taxes are
5
taken into account. The first column shows that funds
with one, two, three, and four or more exposures have Notes: This exhibit shows multiple regression results of all US funds
0.37, 0.60, 0.97, and 1.36 standard deviations higher during the sample period of January 1990 to December 2015 in the
alpha than funds with no factor exposures. CRSP Mutual Fund Database with total assets exceeding $5 million.
A winsorized (at -2 and 2) Z-score of CAPM alphas (z_alpha) is
Results remain intact after controlling for fund- regressed on dummies, indicating funds belonging to a specific factor
specific characteristics, as shown in column 2. group. Ln(age) is the natural logarithm of the fund’s age, calculated as the
number of months the fund has been in our sample. Ln(size) is the natural
logarithm of the average fund’s total assets in US dollars. exp_ratio and
Luck versus Skill in Mutual Fund Returns turn_ratio are the average expense ratio and turnover ratio per mutual
fund in our sample.
In this section we take a critical look at our pre-
vious findings. Fama and French (2010) showed that, to
a large extent, the performance of mutual funds can be actual distribution of fund returns is skewed to the left,
attributed to luck. This is a strong and valid argument it shows that fund managers’ skills do not cover for their
against the skill level of outperforming mutual funds; expenses and, on average, mutual funds underperform
even if the true alpha is zero in specific periods, it can the market portfolio. If the distribution is skewed to the
be higher or lower than zero just by chance. In the pre- right, mutual funds have skills exceeding the fees they
vious sections we show that funds that incorporate aca- charge and generate added value for their investors.
demic insights in their investment process and provide Exhibit 5 compares the distribution of fund alphas
exposure to proven factor premiums deliver higher net across all style groups to the simulated distribution.
alpha relative to the control group. In this section we Consistent with Fama and French (2010), we show
take a more conservative approach and test whether the that in the right tail of the distribution, managers have
observed performance exceeds what could have been enough skill to deliver higher returns than their capm
generated simply by chance. beta predicts. Specifically, at the 90th percentile the
Our simulated distribution of mutual fund returns actual distribution has higher t(a) 56% of the time,
possesses the important property that true net alpha is compared to the 5,000 simulated runs. At the 99th
known to be zero, which assumes that all managers have percentile this percentage increases to 84 with a t(a) of
enough skills to cover for the fees they charge. If the 2.63, compared to a simulated t(a) of 2.29, indicating
1R /RZ 6PDOO
6LPXODWHG $OO)XQGV ([SRVXUH %HWD &DS 9DOXH 0RPHQWXP 3URILWDELOLW\ ,QYHVWPHQWV
SHUF ± ± ± ± ± ± ± ± ±
DFW
SHUF ± ± ± ± ± ± ± ± ±
DFW
SHUF ± ± ±
DFW
SHUF ±
DFW
SHUF
DFW
SHUF
DFW
SHUF
DFW
Notes: This exhibit shows actual fund performance over simulated performance for different percentiles. Performance is measured by the t-statistic of fund
CAPM alpha t(a). “Simulated” is the average t(a) at the respective percentile over all 5,000 simulated runs. The remaining columns show the actual t(a)
of the respective mutual fund style groups. “% < act” is the percentage of runs out of all 5,000 runs with lower t(a) relative to the actual.
that the returns of top-performing managers are sig- because they show that they are much more robust,
nificantly higher than the returns that could have been surviving even the most conservative tests of our boot-
generated by luck, even after adjusting for the fees they strapping method. Furthermore, the newly documented
charge. profitability and investments factors seem to be at least
This picture significantly changes if we look at equally robust compared with the well-known value,
the control group “No exposure.” Even at the 99th size, and low-risk premiums.
percentile, t(a) is only 2.03, which is higher than a In Exhibit 6 we graphically show the cumula-
randomly simulated one in only 24% of the cases. This tive density function of funds with low beta, small
result indicates that funds with no factor exposures cap, value, momentum, profitability, and investments
systematically fail to deliver positive net alphas that exposures. The horizontal axis shows the value of t(a),
cannot be attributed to luck. On the contrary, the net and the vertical axis shows the percentile values. The
performance of all style groups does not seem to be percentile at which the actual line is below the simu-
attributable to chance. The net alphas of all groups lated line indicates that from this percentile onward,
(except momentum) are significantly higher relative fund managers from the respective group generate
to those based on our simulated distribution at most of net benchmark-adjusted returns beyond what can be
the percentile levels. Namely, low-beta and value funds expected by chance.
generate positive luck-adjusted net returns in 50% of
the cases and small cap, profitability, and investments Global Markets
in 75% of the cases. Momentum produces positive luck-
adjusted returns in only the top one percentile, which In this section we extend the scope of our
might be a result of the higher turnover and total costs research. We conduct our main analysis on a global
of momentum managers. markets universe, including long-only equity mutual
These results strengthen the previously docu- funds from all developed countries. In Exhibit 7 we
mented positive performance of factor investing funds show the distribution of fund alphas. It strengthens
132 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
the conclusions of our US analysis because all groups for the momentum factor stand out; unlike in the US
of factor funds have a higher probability of earning a universe, momentum managers have an alpha that is 0.21
positive alpha compared to traditionally active global (t-statistic of 6.37) standard deviations higher than that of
mutual funds. non-momentum funds. These results are not explained
In Panel A of Exhibit 8 we show that our US results by controlling for other style exposures. In our regres-
spill over to global markets. Funds belonging to all our sion specification 7, in which we include all style dum-
style groups—low beta, size, value, momentum, prof- mies simultaneously, we see that with the exception of
itability, and investments—deliver higher beta-adjusted value (t-statistic of 1.38), all factor funds have significantly
returns compared to the average mutual fund. Results higher alphas than funds with no factor exposure.
Exhibit 6
Simulated and Actual Cumulative Density Function of CAPM t(a) Factor Funds
/RZ%HWD 6PDOO&DS
± ± ± ± ± ±
9DOXH 0RPHQWXP
± ± ± ± ± ±
6LPXODWHG $FWXDO
(continued)
± ± ± ± ± ±
6LPXODWHG $FWXDO
Notes: This exhibit shows cumulative density function of actual fund performance over simulated performance. Performance is measured by the t-statistic
of the fund CAPM alpha t(a). “Simulated” is the average t(a) at the respective percentile over all 5,000 simulated runs.
Exhibit 7
Distribution of Fund Alphas—Global Markets
Notes: This exhibit shows the distribution of annualized fund alphas across all global developed market funds in the Morningstar Mutual Fund Database
with total assets exceeding $5 million. Alphas are calculated per fund as the intercept from the CAPM regressions over all available observations during
the sample period of January 1991 to December 2015. “< -4” shows the percentage of funds with annualized alphas less than -4%; “-4:-2” shows
the percentage of funds with annualized alphas between -2% and -4%; and “% > 0” shows the number of funds with positive alphas.
Panel B extends the analysis by controlling for robust, indicating that despite the higher turnover, the
fund-specific characteristics and shows that the supe- momentum premium can be harvested in practice. The
rior performance of factor funds cannot be attributed newly documented factors profitability and investments
to their age or size. Only the coefficient on our value seem to be some of the strongest factors; their coeffi-
group becomes negative but insignificant (t-statistic cients remain positive and highly significant in all our
of -0.77). On the other hand, momentum results remain tests.
134 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
Exhibit 8
Factor Exposures and Outperformance—Global Markets
Notes: This exhibit shows univariate and multiple regression results of all global developed market funds during the sample period of January 1991 to
December 2015 in the Morningstar Mutual Fund Database with total assets exceeding $5 million. A winsorized (at -2 and 2) Z-score of CAPM alphas
is regressed on dummies, indicating funds belonging to a specific factor group. Ln(age) is the natural logarithm of the fund’s age, calculated as the number
of months the fund has been in our sample. Ln(size) is the natural logarithm of the average fund’s total assets in US dollars.
Do Investors in Factor Funds Successfully be able to harvest these premiums. However, the actual
Harvest Factor Premiums? returns that investors earn by investing in factor mutual
funds appear to be significantly lower because inves-
In the previous section we provided evidence that tors do not seem to follow buy-and-hold strategies but,
factor premiums survive even the most robust research rather, dynamically reallocate their funds across both
specifications and that mutual fund managers seem to
1R
$OO)XQGV ([SRVXUH /RZ%HWD 6L]H 9DOXH 0RPHQWXP 3URILWDELOLW\ ,QYHVWPHQWV
D7KHRUHWLFDO)DFWRU5HWXUQV
E0XWXDO)XQGV%X\DQG+ROG(:
F0XWXDO)XQGV%X\DQG+ROG9:
G0XWXDO)XQGV'ROODU:HLJKWHGSHU)DFWRU
'LIIHUHQFHG±E ± ± ± ± ± ± ±
39DOXH'LIIHUHQFH
Notes: This exhibit shows returns of all US funds during the sample period of January 1991 to December 2015 in the CRSP Mutual Fund Database with total
assets exceeding $5 million. “EW” is the equally weighted, geometrically calculated, annualized return; “VW” is the total asset-weighted, geometrically calculated,
annualized return; and “Dollar-Weighted” is the geometrically annualized internal rate of return (IRR), or the rate of return that makes the sum of discounted
ending total assets and the present value of monthly distributions equal to the initial total assets. The IRR calculation is done per factor level. It calculates the IRR
on an aggregate level because distributions are calculated based on the sum of all assets per fund style and value-weighted fund returns in the same style group.
Theoretical factor returns are downloaded from Kenneth French Data Library. Value, momentum, profitability, and investments are based on six portfolio sorts as
the average of small attractive and big attractive portfolios. For example, value is the average of small value and big value portfolios. Size is the average of the small
value, small growth, and small middle portfolios based on six size-value sorted portfolios. Low beta is the lowest quintile based on a past market beta sort.
factors and factor managers. By attempting to time level by summing the dollar amount of all assets and
across factors, investors lose a large portion of the return calculating the IRR according to the methodology pro-
they could earn with a buy-and-hold strategy, and in this posed by Dichev (2007). The resulting return of 7.9%
section we quantify this loss. To do so we calculate the per annum captures the amount of equity timing or
magnitude of factor premiums in three settings, gradu- fund f lowing in and out of our sample of US long-only
ally reducing the level of abstraction. First, we calculate domestic equity funds. For example, if a fund has strong
the long-only premiums of Fama and French (2015) and returns in the first year of its existence but a very low
the low beta factors. Second, we calculate the return of asset base, very few investors benefit from it. If later,
mutual funds with exposure to these factors. Finally, we because of good performance, it attracts inf lows, but
estimate the returns realized by investors in these funds. subsequent returns are lower, the return of the fund will
This analysis extends work by Hsu (2016), who showed be higher than the return of the investors in this fund
that investors in value and small cap funds underper- over the sample period.
formed the S&P 500, whereas the value and small cap The remaining columns of Exhibit 9 show the
funds themselves outperformed the benchmark. same analysis per different groups of funds. Row (d)
Exhibit 9 describes the main results of this section. shows that size, value, momentum, profitability, and
For each factor group we show the return of the long- investments fund investors lose, respectively, 2.5%,
only academic factor return, equally and value-weighted 1.2%, 3.0%, 1.3%, and 0.7% as a result of factor timing.
mutual fund returns, and dollar-weighted returns at The highest loss is incurred by investors in momentum
an aggregation level per factor. Starting with the first funds. A possible reason is the intrinsic trend following
column we see that the market portfolio has earned a the nature of momentum, which stimulates investors
buy-and-hold return of 10.1% per annum, compared to to allocate to momentum funds after a period of good
9.6% for the average mutual fund. Moving from top to performance and which lowers their subsequent realized
bottom in the exhibit reduces the level of abstraction in returns compared to the fund returns.
calculating returns and gets closer in approximating the Our results have strong implications for mutual
return to the end investors. A value-weighted return of fund investors. Even though factor funds deliver posi-
9.2% implies that larger mutual funds have generated tive alpha, their investors have not been able to capture
lower returns than their small counterparts, consistent it because of their allocation decisions. For example, if
with studies such as that by Chen et al. (2004). Next, we investors believe in the value and momentum premiums
calculate the dollar-weighted return on an aggregated but they only invest in value or momentum funds, after
136 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
Exhibit 10
Multifactor Mutual Fund Returns and Investor Returns
Notes: This exhibit shows returns of all US funds during the sample period of January 1991 to December 2015 in the CRSP Mutual Fund Database
with total assets exceeding $5 million. “EW” is the equally weighted, geometrically calculated, annualized return; “VW” is the total asset-weighted,
geometrically calculated, annualized return; and “Dollar-Weighted” is the geometrically annualized internal rate of return (IRR), or the rate of return
that makes the sum of discounted ending total assets and the present value of the monthly distributions equal to the initial total assets. The IRR calculation
is done per factor level. It calculates the IRR on an aggregate level because distributions are calculated based on the sum of all assets per fund style and
value-weighted fund returns in the same style group.
a period of strong performance they lose a significant decisions on returns. It appears that despite one-factor
portion of the factor premium because of the cyclicality mutual funds having a sizable return premium of 1.1%
in factor returns. A potential solution would be to buy over traditionally actively managed mutual funds, the
both funds and hold on to them instead of moving assets investors in those same funds underperform the control
across them. group with 0.5% as a result of poor timing decisions.
Exhibit 10 shows the annualized returns of funds The right three bars of the exhibit present the returns
with one, two, three, and four or more factor expo- investors could have had if they allocated to funds with
sures. The average buy-and-hold investor would have two, three, or four factors and held on to them instead
earned 9.1%, 9.9%, 10.3%, and 10.6%, compared to of timing across factors. The buy-and-hold premium is
8.0% in funds with no factor exposures. Similar to the 1.9%, 2.4%, and 2.7%, respectively.
single-factor funds, the dollar-weighted returns of 7.4%,
7.6%, 8.9%, and 8.9% are lower than the time-weighted What Drives Allocation Decisions
returns for the same group of funds, indicating that even of Mutual Fund Investors?
if they invest in multifactor funds, investors still make
allocation decisions that cost them a significant portion In this section we unveil the drivers behind
of performance. investor allocation decisions. This is crucial in having
Exhibit 11 graphically illustrates our main points. a full understanding of why investors consistently lose
The bars represent the return premium over tradition- returns even when selecting the right funds. The nat-
ally actively managed mutual funds, or “no exposure” ural starting point is to follow the insights of Sirri and
funds. The average academic long-only factor of Fama Tufano (1998), who show that because of the complexity
and French (2015) outperforms our control group with of a full understanding of the methodology of each
4.7% per annum. The second bar shows the return strategy, investors just buy funds with high past returns,
of mutual funds with one-factor exposure. It is based assuming that past returns are an accurate proxy for
on net asset values and, as such, includes the negative manager skills. We follow their piecewise linear regres-
effect of taxes and trading costs on returns. This brings sion specification to control for the different degree of
down the premium to 1.1%, which is the return gener- sensitivity of f lows to performance in the tails of the
ated by investors invested in this group of funds at the performance distribution.
beginning of our sample and holding on to the invest- Our article offers convincing evidence that funds
ment until the end of the sample; however, investors exposed to factors outperform in the long term, and
often make active allocation decisions based on their investors who strategically allocate to them can benefit
views on which factor is going to outperform going from those premiums. As such, we test the hypothesis of
forward. The third bar incorporates the effect of these whether investors allocate to factor funds strategically or
± ±
±
±
±
7KHRUHWLFDO)DFWRU 0XWXDO)XQG ,QYHVWRU5HWXUQV )XQGVZLWK )XQGVZLWK )XQGVZLWK
5HWXUQV 5HWXUQV)DFWRU )DFWRU )DFWRU )DFWRUV )DFWRUV
%X\DQG+ROG %X\DQG+ROG 'ROODU:HLJKWHG %X\DQG+ROG %X\DQG+ROG %X\DQG+ROG
Notes: This exhibit is based on returns of all US funds during the sample period of January 1991 to December 2015 in the CRSP Mutual Fund Data-
base with total assets exceeding $5 million. All returns are calculated in excess of the returns of traditional actively managed mutual funds ( fund with no
positive factor exposures). Mutual fund returns are equally weighted, geometrically calculated, annualized returns. Investor returns are the dollar-weighted,
geometrically annualized internal rate of return (IRR), or the rate of return that makes the sum of discounted ending total assets and the present value of
monthly distributions equal to the initial total assets. The IRR calculation is done per factor level. It calculates the IRR on an aggregate level because the
distributions are calculated based on the sum of all assets per fund style and value-weighted fund returns in the same style group. Styles are funds with one,
two, three, or four or more factor exposures. The theoretical return is the average long-only Fama and French (2015) factor returns and the lowest quantile
based on low beta sorts. Value, momentum, profitability, and investments are based on six portfolio sorts as the average of small attractive and big attractive
portfolios. For example, value is the average of small value and big value portfolios. Size is the average of the small value, small growth, and small middle
portfolios based on six size-value sorted portfolios. Low beta is the lowest quintile based on past market beta sorts.
simply end up investing in factor funds because of their Most importantly, it seems that past performance is
good past performance. Toward this goal, we extend the main driver of investor decisions. The coefficients
the Sirri and Tufano (1998) f low-performance model on size, value, and momentum dummies are negative,
with dummies that indicate whether funds belong to a indicating that investors tend to avoid those funds. The
certain factor group. coefficient on profitability is insignificant, and only the
Exhibit 12 contains the main results of this sec- coefficients on low beta and investments are positive and
tion. In regression 1 we show that relative f lows are significant, showing that investors invest in low beta and
significantly higher for funds with high past 12-month investments funds more than their past 12-month returns
outperformance over the market (a coefficient of 0.02, would suggest. In regression 2 we extend the analysis
with a t-statistic of 12.02). This effect is highly non- and test whether the allocation based on past returns is a
linear; the top (bottom) group has a significant coef- good timing decision in terms of future returns. As such,
ficient of 0.04 (-0.03), meaning that funds belonging we include the future 12-month returns in the equation.
to this group exhibit abnormally high (low) f lows. The coefficient on future performance is virtually zero
138 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
Exhibit 12
Flow-Performance Relationship
Notes: This exhibit shows Fama and Macbeth (1973) multiple regression results of all US funds during the sample period of January 1991 to December
2015 in the CRSP Mutual Fund Database. Each month, relative fund flows are regressed on dummies, indicating funds belonging to a specific factor group
and measures on performance. Ln(age) is the natural logarithm of the fund’s age, calculated as the number of months the fund has been in our sample at
each point in time. Ln(size) is the natural logarithm of the most recent fund’s total assets in US dollars, and exp_ratio and turn_ratio are the most recent
expense ratio and turnover ratio per mutual fund at each point in time. Presented coefficients are the average coefficients over the full sample, and t-statistics
are calculated as in Fama and Macbeth (1973).
(0.00, with a t-statistic of 0.09), meaning that investing positive coefficient on the low beta dummy becomes
in funds with high past performance has no predictive zero, indicating that investors allocate to low beta funds
power for future performance. In regression 3 we sub- just as much as their past alpha implies.
stitute the past return with the past CAPM alpha, and This section presents evidence that fund f lows
results remain intact. The only difference is that the have been driven by factor funds earning high past
140 Factor Investing from Concept to Implementation Quantitative Special Issue 2019
Extending Fama–French Factors
to Corporate Bond Markets
Demir Bektić, Josef-Stefan Wenzler, Michael Wegener,
Dirk Schiereck, and Timo Spielmann
D
Demir Bektić o equity market factors explain Analyzing the link between equity and
is a senior portfolio manager,
corporate bond return dynamics? corporate bond factors is a tempting proposi-
fixed income at Deka
Investment GmbH, a member
This empirical relationship tion for the following reasons. First, over the
of IQ-KAP, and an assistant is of substantial importance to years evidence has mounted that alternative
professor at Darmstadt investors (interested in return and diversifica- risk premiums beyond the traditional asset
University of Technology in tion characteristics), economists (who want class risk premiums (e.g., equity and term
Darmstadt, Germany. to know the mechanisms that connect these premium) do indeed exist. Harvey, Liu, and
demir.bektic@gmail.com
markets), and not least policymakers (con- Zhu (2016) provided an excellent summary
Josef -Stefan Wenzler cerned about the stability of the financial of factors in the equity space and recount
is a senior portfolio manager, system).1 Although the relevant empirical more than 300 papers on cross-sectional
fixed income at Deka studies can be traced back to the seminal work return patterns published in various journals.
Investment GmbH and a by Fama and French (1993), there is limited According to structural credit risk models,
member of IQ-KAP in
direct evidence in the literature concerning both equity and corporate debt are driven by
Frankfurt, Germany.
josef-stefan.wenzler@deka.de
the pricing of risk across stock and bond the fundamentals of the same underlying cor-
markets. Although Merton (1974) provided poration, implying that stock prices and credit
M ichael Wegener an intuitive approach to describe the link spread changes must be related to ensure the
is a managing director and between equities and corporate debt based absence of arbitrage. Consequently, risk pre-
head of quant products on contingent claims, we are still lacking miums in equity and corporate bond mar-
equities and fixed income at
relevant factors to describe the dynamics of kets should be related.2 Second, although the
Deka Investment GmbH and
a general manager of IQ-KAP credit markets. We aim to fill this gap by relationship between firms’ default risk and
in Frankfurt, Germany. examining whether the four risk factors in equity risk premiums has been analyzed in
michael.wegener@deka.de the Fama and French (2015) framework—size numerous studies (see Vassalou and Yuhang
(SMB), value (HML), profitability (RMW), 2004 or, more recently, Friewald, Wagner,
Dirk Schiereck
and investment (CMA)—are priced in the and Zechner 2014), there is little evidence
is a professor at Darmstadt
University of Technology in
corporate bond market.
Darmstadt, Germany.
schiereck@bwl.tu-darmstadt.de 2
Kapadija and Pu (2012) argued that there is
1
As long as some investors have access to both cross-sectional variation in the correlation of equity
T imo Spielmann equity and corporate bond markets, the absence of and credit markets and that short-horizon pricing
is a portfolio manager at arbitrage opportunities imposes cross-market restric- discrepancies across equities and corporate bonds are
Deka Investment GmbH and tions on the stochastic discount factor (SDF). The SDF common and predominantly anomalous because of
a member of IQ-KAP in is sometimes referred to as the pricing kernel and ref lects limited arbitrage activity. However, these authors also
Frankfurt, Germany. the fact that the price of an asset can be computed by suggested that the Merton (1974) model is appropriate
timo.spielmann@deka.de discounting the future cash f low by a stochastic factor. for longer time scales.
142 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
RETURN DYNAMICS BETWEEN insight from the formalized Merton model. Therefore,
EQUITY AND DEBT equity market factors are relevant for pricing corpo-
rate debt only if they capture changes in firm value or
Relating equity and corporate bond returns is not changes in risk-neutral probabilities.
trivial but represents a natural starting point. Rational
asset-pricing models suggest that risk premiums in the
DATA AND METHODOLOGY
equity market should be consistent with those of the cor-
porate bond market, assuming that the two markets are Data
integrated. The earliest formalized structural credit risk
model developed by Merton (1974) provides important Similar to Israel, Palhares, and Richardson (2018),
intuition on why changes in equity and corporate bond we use monthly data from the Bank of America Mer-
returns should be related because they represent con- rill Lynch (BAML) for this analysis. Prices are provided
tingent claims against the assets of the same company. by BAML traders and are used as the primary pricing
The only state variable in the model is the value of source. The dataset includes monthly data of all senior
the firm V, and one of the main assumptions is that the US HY, US IG, and European IG corporate bond issues
value of a company’s assets follows a geometric Brownian rated by at least one of the three major rating agencies
motion W under the risk-neutral martingale measure Q (S&P, Moody’s, and Fitch), issued in euros or US dollars.
in which m and s are the drift and volatility, respectively. The employed BAML indexes only include bonds with
a minimum amount outstanding of 250 million for IG
dVt = Vt µdt + Vt σdWtQ (1) and 100 million for HY in local currency terms,7 a fixed
coupon schedule, and a minimum remaining time to
It is assumed that the company issues only a single maturity of one year. Newly issued bonds must exhibit
zero-coupon bond with face value F payable at T in a time to maturity of at least 18 months.8
which the payoff to the creditors at date T is European HY bonds are not included in this anal-
ysis because of the insufficient size of the European HY
D(VT ,T ) = min(VT , F ) = F − (F − VT )+ (2) market until 2013. As done by Elton et al. (2001), put-
table bonds are excluded. We further eliminate subor-
To relate equity and debt in the Merton model, dinated and contingent capital securities (cocos) as well
equity is valued as a call option on the value of assets, as taxable and tax-exempt US municipal, equity-linked,
and applying the accounting identity or put–call parity securitized, dividends-received deduction (DRD)
equates the value of debt D and equity E: eligible,9 and legally defaulted securities because these
have distinctly different payout characteristics compared
E(Vt , t ) = CallBS (Vt , F ,µ,T − t, σ ) (3) to standard senior coupon bonds.
The dataset covers the period from December 1996
D(Vt , t ) = P (t,T ) − PutBS (Vt , F ,µ,T − t, σ ) (4) to December 2016 for US HY and IG bonds and from
December 2000 to December 2016 for European IG
The model makes it clear that the spread between bonds.10 Since the adoption of the euro in 1999, BAML
risky credit debt and risk-free debt is the value of the put
option.6 Consequently, possible determinants of credit 7
This is similar to the equity market anomaly literature, in
spreads, and hence the key factors that inf luence credit which too small stocks are typically removed to ensure that results
spreads, are the company’s business risk of the assets s, are not driven by market microstructure or liquidity.
8
maturity of the debt T, and the leverage F. Removing bonds that have less than one year to maturity
The theoretical link that equities and corporate is applied to all major corporate bond indexes, such as Citi Fixed
Income Indices and Barclays Capital Corporate Bond Index as well
bonds of a company are connected through their expo- as the BAML Corporate Master Index. The 18-month cutoff for
sure to the underlying company value is an important newly issued bonds is a standard choice of BAML.
9
A DRD is a tax deduction received by a corporation on the
dividends paid by companies in which it has an ownership stake.
6
CallBS (Vt, F, m, T - t, s) denotes the value of a call option 10
Trade Reporting and Compliance Engine (TRACE) trans-
and PutBS (Vt, F, m, T - t, s) is the value of a put option according action data are available for US bonds only and do not start before
to Black and Scholes (1973). July 2002.
144 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
Exhibit 1
Summary of Universe Statistics
Avg. No. of Avg. No. of Avg. No. of Avg. No. of Avg. Modified Avg. Option
Year Firms Public Firms Private Firms Bonds Duration Adjusted Spread Avg. Rating
Panel A: US HY Universe
1997 333 151 182 324 4.18 285 13.65
1998 381 173 208 337 4.38 405 13.65
1999 465 220 245 406 4.37 579 13.73
2000 485 239 246 503 4.2 660 13.81
2001 501 259 242 611 3.89 1,110 14.06
2002 511 293 218 779 3.92 1,369 14.12
2003 615 361 255 980 4.13 802 14.14
2004 719 410 309 1,036 4.28 411 14.04
2005 745 420 325 1,021 4.16 360 14.00
2006 757 424 334 965 4.15 329 13.97
2007 709 404 305 798 4.25 330 13.76
2008 819 460 359 890 4.09 921 13.84
2009 782 477 305 981 3.68 1,639 14.26
2010 876 529 346 1,106 3.94 662 14.12
2011 1,054 619 435 1,278 4.18 605 13.95
2012 1,124 677 447 1,412 3.84 682 14.04
2013 1,231 742 489 1,570 3.89 559 14.24
2014 1,314 799 514 1,722 3.89 494 14.21
2015 1,347 869 478 1,963 3.95 761 14.19
2016 1,292 876 415 2,053 3.74 946 14.29
Panel B: US IG Universe
1997 816 563 253 2,516 5.61 62 6.72
1998 909 619 290 3,026 5.65 101 6.86
1999 860 609 251 3,004 5.61 139 6.90
2000 738 560 178 2,595 5.29 181 6.91
2001 736 578 158 2,785 5.19 201 7.11
2002 756 620 136 2,939 5.21 208 7.25
2003 752 621 131 3,019 5.38 151 7.38
2004 799 647 152 3,186 5.48 100 7.44
2005 700 548 152 2,399 5.68 91 7.22
2006 717 559 158 2,484 5.70 101 7.21
2007 759 594 165 2,267 6.06 124 7.40
2008 809 649 160 2,751 5.64 332 7.46
2009 789 640 149 3,036 5.41 435 7.58
2010 855 700 155 3,396 5.69 192 7.60
2011 928 765 163 3,894 5.86 194 7.67
2012 994 820 175 4,388 5.94 219 7.78
2013 1,117 922 195 5,021 6.08 174 7.86
2014 1,220 1,000 220 5,545 6.00 142 7.85
2015 1,294 1,055 239 6,055 6.00 172 7.85
2016 1,306 1,057 249 6,501 5.89 186 7.87
(continued)
Avg. No. of Avg. No. of Avg. No. of Avg. No. of Avg. Modified Avg. Option
Year Firms Public Firms Private Firms Bonds Duration Adjusted Spread Avg. Rating
Panel C: European IG Universe
2000 221 142 79 289 4.73 104 5.73
2001 266 173 93 402 4.39 113 6.01
2002 320 225 96 551 4.09 128 6.33
2003 319 230 90 591 3.95 90 6.41
2004 340 245 95 624 4.01 56 6.59
2005 325 233 92 569 4.3 46 6.48
2006 324 237 87 602 4.46 51 6.4
2007 350 263 87 567 4.84 64 6.46
2008 388 293 95 725 4.17 199 6.4
2009 383 292 91 918 3.81 288 6.47
2010 385 293 92 1,034 3.98 154 6.63
2011 386 292 94 1,069 3.96 174 6.82
2012 387 295 92 1,125 3.88 181 7.19
2013 395 300 95 1,187 4.32 121 7.35
2014 436 322 113 1,270 4.72 102 7.41
2015 486 362 124 1,415 5.17 109 7.44
2016 545 411 134 1,595 5.18 121 7.54
Notes: Average monthly number of total firms, public firms, private firms, and bonds and the average duration, spread, and rating for each year. Rating
description: AAA = 1, AA+ = 2, AA = 3, AA- = 4, A+ = 5, A = 6, A- = 7, BBB+ = 8, BBB = 9, BBB- = 10, BB+ = 11, BB = 12, BB- =
13, B+ = 14, B = 15, B- = 16, CCC+ = 17, CCC = 18, CCC- = 19.
the additional risk (see Banz 1981). The behavioral bias higher returns for the additional risks inherent to smaller
argument for a size risk premium is given by limited issuers. As firm size decreases, equity volatility increases
investor attention to smaller companies. and so does the probability of default (see Merton 1974).
Because empirical evidence for size premiums in From a bondholder’s perspective, these firms are likely
credit markets is relatively sparse and its definition has less to be more risky and should therefore offer higher risk
consensus than that for the other three factors analyzed in premiums.
this article,12 we follow Banz (1981) and Fama and French To harvest the size premium in credit markets, we
(2015) and define size in the corporate bond space as the construct the decile portfolio containing the bonds of
market capitalization of the company’s equity: the smallest 10% of the companies as measured by equity
market capitalization and rebalance it on a monthly basis.
Sizet = SOt × PPSt (5) Value. Fama and French (1992) used the book-
to-market ratio (BE/ME) as a measure of equity value.
in which SOt denotes the number of shares outstanding A high BE/ME is indicative of a cheap stock in relative
and PPSt the price per share in month t. The same terms, whereas a low BE/ME suggests the opposite.
arguments for a size premium in the equity space apply to According to Zhang (2005), a rationale for the value
the realm of credit markets; an investor should demand premium is based on costly reversibility of invest-
ments. Hence, companies with high sensitivity to eco-
12
Houweling and van Zundert (2017) used the index weight nomic shocks in bad times should offer higher returns.
of each company from the Barclays US Corporate IG index and A behavioral-based explanation suggests that investors
the Barclays US Corporate HY index, suggesting that the size of
overreact to bad news and extrapolate recent price move-
a company’s public debt in a given index represents the size factor
in credit markets. ments into the future, which results in underpricing.
146 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
We use the Fama and French (2015) definition measured by operating profitability and rebalance it on
of value and adjust it according to Asness and Frazzini a monthly basis.
(2013), using the last available share price:13 Investment. Fama and French (2015) suggested
the existence of an investment premium in equity mar-
BEt −6 kets using the dividend discount model and defined the
Valuet = (6)
MEt investment factor as the change in total assets, TA, from
the fiscal year ending in t - 18 to the fiscal year ending
in which BEt−6 and MEt denote book equity and market in t - 6, divided by total assets in t - 18. A risk-based
equity in month t − 6 and t, respectively. Fama and explanation of the investment factor is linked to the fact
French (1995) and Chen and Zhang (1998) showed that that corporations with low levels of investment tend to
firms with a high book-to-market equity ratio are more have higher costs of capital and therefore only engage
likely to exhibit persistently low earnings, more earn- in projects that are most likely to lead to future earnings
ings uncertainty, and high financial leverage and are and ultimately to higher returns for equity investors.
more likely to cut their dividend, which in turn leads to The behavioral explanation is based on the premise that
an increased volatility. High equity volatility implies a investors misprice low-investment companies because of
higher probability of default in the Merton (1974) model; expectation errors. According to prospect theory, inves-
thus, value should have the same directional impact on tors prefer and overprice firms that engage in excessive
equity and bond prices. investment strategies because of their lottery-like pay-
We construct the decile portfolio containing the offs, which could explain the low average returns of such
bonds of the most valuable 10% of the issuers as mea- lotteries. Conversely, they tend to underprice companies
sured by book-to-market value and rebalance it on a that pick their investments wisely.
monthly basis. Based on the definition of Fama and French (2015),
Profitability. Fama and French (2015) used the we define the investment factor, Inv, for credit markets
dividend discount model to show that firms with high in month t as
earnings relative to book equity have higher expected
returns. The behavioral explanation suggests that TAt −6 − TAt −18
Invt = (8)
investors do not differentiate between high and low TAt −18
profitability in growth firms, and therefore bonds of
companies with high profitability have higher returns. We construct the decile portfolio every month.
Chordia et al. (2017) noted that highly profitable compa- The portfolio contains the bonds of the 10% of compa-
nies generate healthy cash f lows and thus have a smaller nies with the lowest investment and is rebalanced on a
probability of default. monthly basis.
We measure operating profitability, OP, using
accounting data available at end of month t - 6 and Comparing Corporate Bond Factor
define it as earnings before taxes, EBT (revenues minus Portfolio Returns
cost of goods sold minus selling, general, and admin-
istrative expenses minus interest expenses), divided by To compare factor portfolios, we compute risk-
book equity, BE (see Fama and French 2015). adjusted returns for all portfolios. Furthermore, we
regress factor portfolio returns on corporate bond market
EBTt −6 excess returns (DEF) representing the credit default pre-
OPt = (7)
BEt −6 mium and the equity returns of the Fama–French factors
market (MKT), size (SMB), value (HML), profitability
We construct the decile portfolio containing (RMW), and investment (CMA)14 as well as the bond
the bonds of the most profitable 10% of the firms as factor TERM, representing the term premium that
can be harvested by investing in interest rate securities
13 14
See Correia, Richardson, and Tuna (2012); Houweling and The data on MKT, SMB, HML, RMW, and CMA are
van Zundert (2017); or Israel, Palhares, and Richardson (2018) for obtained from Kenneth French’s website: http://mba.tuck.dart-
other definitions of value in credit markets. mouth.edu/pages/faculty/ken.french/data_library.html.
148 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
Exhibit 2
Performance Summary of Factor Portfolios
(continued)
Notes: Results for market, size, value, profitability, and investment factor portfolios for the US HY and the US and European IG corporate bond markets.
At the beginning of each calendar month, equal-weighted portfolios are constructed from the 10% issuers with the highest factor exposure to small size, high
value, high profitability, and low investment. Section A summarizes risk and return statistics. Section B shows the excess return statistics. Section C dis-
plays the CAPM alpha and beta. In section D, alphas are estimated from the time-series regression using DEF as the credit default premium, MKT (minus
the risk-free rate, RF) as the equity market premium in addition to the Fama–French factors SMB (size) and HML (value), RMW (profitability), CMA
(investment), and TERM in the seven-factor model. Means, volatilities, SRs, excess returns, tracking errors, maximum drawdowns, information ratios, and
alphas are annualized. We test whether the outperformance of a factor compared to the market is larger than 0 (section B, t-test), and whether the alphas of
a factor portfolio are larger than 0 (t-test in sections C and D).
*, **, and *** denote statistical significance corresponding to the 90%, 95% and 99% confidence levels, respectively.
150 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
largest in size and value for US HY and EUR IG corpo- profitability) to 7% (low size) for US HY, 0.88% (high
rate bonds, whereas size and investment have the largest profitability) to 2.01% (high value) for US IG, and 0.94%
drawdowns in US IG markets. SRs are up to 30% higher (high profitability) to 1.87% (high value) for European
compared to market. For instance, in US HY markets, IG corporate bond factor portfolios and thus are quite
the SR ranges from 0.46 (value) to 0.62 (profitability) large compared to the market volatilities of 9.47%, 3.59%,
compared to 0.45 for the market. The SR ranges from and 2.36%, respectively. The information ratios range
0.24 (profitability) to 0.33 (size) in US IG markets and from 0.28 (high value) to 0.52 (low investment) in the
from 0.27 (value) to 0.68 (investment) in European IG US HY market, from -0.21 (high profitability) to 0.29
markets. For completeness, bottom-decile factor portfolio (small size) in the US IG market, and from -0.23 (high
returns are also included in Exhibit 2 and are described profitability) to 0.58 (low investment) in the European
in more detail in the robustness section in the context of IG market. The SRs in combination with the relatively
long–short portfolios. high active risk imply that factor portfolios may under-
Panel B reports the excess returns over benchmark perform the benchmark on shorter investment horizons
returns. In the US HY segment, all factors exhibit statis- (see Houweling and van Zundert 2017). Hence, single-
tically significant premiums in both US credit segments. factor portfolios are rather unattractive for portfolio man-
However, for US IG markets, only size is significant at agers and for investors looking for benchmark-oriented
the 10% level. For European IG credit markets, only portfolio management. Instead, investors who consider
the investment factor generates economically and sta- factor investing with corporate bonds should strategically
tistically significant returns. Interestingly, profitability allocate to factors to harvest risk premiums on a consistent
is negatively priced in both IG credit markets, and this basis (see Ang, Goetzmann, and Schaefer 2009).
is in line with the findings of Chordia et al. (2017) and
Campbell, Chichernea, and Petkevich (2016).15 Multifactor Performance
Panel C reports the excess returns versus DEF
(CAPM alpha), and Panel D reports the excess return It is well documented that a portfolio constructed
statistics after controlling for Fama–French equity fac- of different assets (here, factors) will, on average, gen-
tors MKT, SMB, HML, RMW, and CMA and the bond erate higher risk-adjusted returns than any individual
market factors TERM and DEF. Here, for US HY mar- security within the investment universe (only true if the
kets, profitability and investment are still significant fac- assets or factors in the portfolio are not perfectly corre-
tors, but for US IG markets no factor exhibits statistical lated). Exhibit 3 shows the outperformance correlations
significance. In European IG markets, investment gener- among our four analyzed factors (small size, high value,
ates economically and statistically significant returns also high OP, and low Inv) for US HY and US and European
after controlling for DEF and the seven-factor model. IG credit markets; the original Fama–French equity fac-
In addition, it is not surprising that DEF carries the tors of MKT, SMB, HML, RMW, and CMA; and the
majority of the explanatory power for long-only factor bond factors TERM and DEF. For US HY markets, the
portfolios considering the common market exposure corporate bond factor correlations range from -0.15 (OP
incorporated in long-only portfolios. and DEF) to 0.56 (size and Inv). For US IG markets, the
However, single-factor tracking errors suggest corporate bond factor correlations range from -0.44 (OP
that investing in factor portfolios can be risky in rela- and DEF) to 0.58 (size and DEF). Finally, for European
tive terms.16 Tracking errors range from 3.45% (high IG markets, the corporate bond factor correlations range
from -0.67 (OP and DEF) to 0.51 (value and DEF).
15
In contrast, Choi and Kim (2016) reported that profitability Low and negative correlations between equity and bond
has no explanatory power for the US IG credit market. factors suggest that the single factors capture different
16
Information ratio (IR i) is defined as the active return of aspects across asset classes.17 However, these low and
a portfolio i divided by tracking error, where active return is the
difference between the return of the portfolio (Ri) and the return of 17
Interestingly, CMA and HML exhibit a relatively high
a selected benchmark index (Rb), and tracking error is the standard
positive correlation of 0.63 for the period from December 1996 to
deviation of the active return (sRi−Rb).
December 2016, and this is probably why Fama and French argued
that the HML factor of the three-factor model becomes redundant
Ri − Rb (12)
IRi = for describing average returns in the US stock market when incor-
σ iRi − Rb porating CMA and RMW, respectively.
6L]H 9DOXH 23 ,QY 0.75) 60% +0/ 50: &0$ 7(50 '()
3DQHO$86+<
6L]H
9DOXH
23 ±
,QY ±
0.75) ± ±
60%
+0/ ± ± ±
50: ± ± ±
&0$ ± ±
7(50 ± ± ± ± ± ±
'() ± ± ± ±
3DQHO%86,*
6L]H
9DOXH ±
23 ± ±
,QY ± ±
0.75) ± ±
60% ± ±
+0/ ± ± ± ±
50: ± ± ± ± ±
&0$ ± ± ± ± ±
7(50 ± ± ± ± ±
'() ± ± ± ± ±
3DQHO&(85,*
6L]H
9DOXH
23 ± ±
,QY ± ± ±
0.75) ± ±
60% ± ±
+0/
50: ± ± ± ±
&0$ ± ±
7(50 ± ± ± ± ±
'() ± ± ± ± ±
Notes: Return correlations between the Fama–French factors MKT, SMB, HML, RMW, and CMA and the bond market factors TERM and DEF as
well as the outperformance of the analyzed bond factors size, value, profitability, and investment for the US HY and IG universe over the period December
1996 to December 2016 (December 2000 to December 2016 for European IG bonds).
negative correlations suggest that a combination of two rtMultiFactor = 0.25rtSize + 0.25rtValue + 0.25rtOP + 0.25rtInv (13)
or more factors offers significant diversification benefits.
Therefore, we also construct equal-weighted long- where r t denotes the return of each corresponding
only multifactor portfolios by combining small size, single-factor portfolio and the multifactor portfolio in
high value, high profitability, and low investment decile month t. These equal-weighted portfolios are another
factor portfolios, as described by way to examine the efficacy of the four analyzed fac-
tors across markets and segments. Exhibit 4 reports the
152 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
Exhibit 4
Performance Summary of Long-Only Multifactor Portfolios
Notes: Results for the US HY and the US and European IG benchmarks (BM) as well as long-only multifactor corporate bond portfolios (MF).
At the beginning of each calendar month equal-weighted portfolios are constructed from the 10% issuers with the highest factor exposure to small size, high
value, high profitability, and low investment. Panel A summarizes risk and return statistics. Panel B shows the excess return statistics. Panel C displays
the CAPM alpha and beta. In Panel D, alphas are estimated from the time-series regression using DEF as the credit default premium; MKT (minus the
risk-free rate, RF) as the equity market premium; and the Fama–French factors SMB (size), HML (value), RMW (profitability), CMA (investment),
and TERM in the seven-factor model. Means, volatilities, SRs, excess returns, tracking errors, maximum drawdowns, information ratios, and alphas are
annualized. We test whether the outperformance of a factor compared to the market is larger than 0 (Panel B, t-test) and whether the alphas of a factor
portfolio are larger than 0 (t-test in Panels C and D).
*, **, and *** denote statistical significance corresponding to the 90%, 95%, and 99% confidence levels, respectively.
multifactor portfolio statistics. The multifactor portfolios is that multifactor portfolios are more diversified because
deliver annual average excess returns of 2.33% in the US more securities in the cross section have nonzero weights
HY market (t-stat of 3.12), 0.23% in the US IG market and the weights are less extreme than in single-factor port-
(t-stat of 1.63), and 0.21% in the European IG market (t-stat folios. However, when controlling for the Fama–French
of 1.39). Moreover, the equal-weighted combination of equity factors MKT, SMB, HML, RMW, and CMA as
size, value, profitability, and investment within the dif- well as the bond market factors TERM and DEF, the
ferent markets and segments generates higher SRs than alphas of multifactor portfolios are still economically and
the equal-weighted market index. SRs are still up to 30% statistically significant for US HY markets; however, US
higher compared to the market. One possible explanation and European IG markets cannot withstand the test.
&XPXODWLYH2XWSHUIRUPDQFH
3HUIRUPDQFH
±
'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF
These findings suggest that the combination of turnover by setting a rebalancing threshold at which
all four factors leads to diversification benefits. The a portfolio is not rebalanced until two-way turnover
equal-weighted multifactor portfolios demonstrate an reaches 70% in a quarter. The main idea is to avoid
annualized SR of 0.59% for US HY, 0.31% for US IG, rebalancing when new weights deviate from the cur-
and 0.45% for European IG corporate bonds; the SRs rent weights by a relatively small amount. By applying
of their corresponding markets are 0.45%, 0.23%, and this rule, portfolio turnover and corresponding trading
0.42%, respectively. Maximum drawdowns are compa- costs can be reduced. Novy-Marx and Velikov (2016)
rable to those of the market; thus, similar to the results confirmed that most factor-based strategies continue
of Lettau, Maggiori, and Weber (2014), we find that to generate statistically significant alphas despite low
downside risk is not able to explain corporate bond turnover. Nevertheless, we estimate transaction costs as
returns. Exhibits 5, 6, and 7 show the multifactor port- a function of issue rating, maturity and total turnover
folio performance versus the benchmark and the cumu- associated with each factor portfolio, similar to Chen,
lative outperformance. Lesmond, and Wei (2007).
Our results remain unchanged after accounting
Robustness Checks for transaction costs (the results are available from the
authors). Thus, the factors studied here are not only
Corporate bonds are typically traded less frequently properly motivated and theoretically sound but can also
than stocks. Therefore, most academic research focuses be implemented. This implies that factor-based investing
on low turnover strategies to avoid high transaction costs. in credit markets does indeed offer additional benefit to
Edwards, Harris, and Piwowar (2007) found that bid– corporate bond investors. However, the specific imple-
ask spreads for US corporates vary significantly through mentation design18 has an impact on performance and
time and depend on the bond’s quality and transaction explains why two portfolios based on the same factor
size. However, most of the literature either ignores costs may perform differently.
completely or assumes transaction costs to be a fixed
amount of 13 bps (Gebhardt, Hvidkjaer, and Swamina-
than 2005) or 12 bps ( Jostova et al. 2013). According 18
Investment universe, rebalancing frequency, weighting
to Amenc et al. (2012), it is possible to achieve lower scheme, and definition of portfolio configuration (e.g., decile,
quintile).
154 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
Exhibit 6
Cumulative US IG Multifactor Long-Only Portfolio Returns (December 1996–December 2016)
2XWSHUIRUPDQFH
0XOWL)DFWRU86,*
86,*%HQFKPDUN
&XPXODWLYH2XWSHUIRUPDQFH
3HUIRUPDQFH
±
±
'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF
Exhibit 7
Cumulative European IG Multifactor Long-Only Portfolio Returns (December 2000–December 2016)
2XWSHUIRUPDQFH
0XOWL)DFWRU(XURSHDQ,*
(XURSHDQ,*%HQFKPDUN
&XPXODWLYH2XWSHUIRUPDQFH
3HUIRUPDQFH
±
±
±
'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF 'HF
DISCUSSION AND INTERPRETATION reward for bearing additional risk. The relationship
OF THE RESULTS between equities and corporate bonds in general is
explained by Merton (1974), who claimed that a cor-
Frazier and Liu (2016) showed that risk aversion porate bond, in essence, consists of a default-free bond
is still a fundamental concept in theory and in practice and a short put on the issuer’s equity. As such, each
and stated that, all else equal, investors require additional corporate bond is described by a risk-free component
156 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
REFERENCES Chordia, T., A. Goyal, Y. Nozawa, A. Subrahmanyam, and
Q. Tong. 2017. “Are Capital Market Anomalies Common to
Amenc, N., F. Goltz, A. Lodh, and L. Martellini. 2012. Equity and Corporate Bond Markets? An Empirical Inves-
“Diversifying the Diversifiers and Tracking the Tracking tigation.” Journal of Financial and Quantitative Analysis 52:
Error: Outperforming Cap-Weighted Indices with Limited 1301–1342.
Risk of Underperformance.” The Journal of Portfolio Manage-
ment 38: 72–89. Cochrane, J. H. 2011. “Presidential Address: Discount Rates.”
The Journal of Finance 66: 1047–1108.
Ang, A., W. Goetzmann, and S. Schaefer. “Evaluation of
Active Management of the Norwegian Government Pension Correia, M., S. Richardson, and I. Tuna. 2012. “Value
Fund—Global.” Working paper, Columbia Business School, Investing in Credit Markets.” Review of Accounting Studies
Yale School of Management, and London Business School, 17 (3): 572–609.
2009.
Crawford, S., P. Perotti, R. Price, and C. J. Skousen.
Asness, C., and A. Frazzini. 2013. “The Devil in HML’s “Accounting-Based Anomalies in the Bond Market.”
Details.” The Journal of Portfolio Management 114: 49–69. Working paper, University of Houston, University of Bath,
and Utah State University, 2015.
Baker, M., and J. Wurgler. 2012. “Comovement and Predict-
ability Relationships between Bonds and the Cross-Section Edwards, A., L. Harris, and M. S. Piwowar. 2007. “Cor-
of Stocks.” Review of Asset Pricing Studies 2: 57–87. porate Bond Market Transaction Costs and Transparency.”
The Journal of Finance 62: 1–21.
Banz, R. 1981. “The Relationship between Return and
Market Value of Common Stocks.” Journal of Financial Elton, E., M. Gruber, D. Agrawal, and C. Mann. 2001.
Economics 9: 3–18. “Explaining the Rate Spread on Corporate Bonds.”
The Journal of Finance 56: 247–277.
Bessembinder, H., K. Kahle, W. Maxwell, and D. Xu. 2009.
“Measuring Abnormal Bond Performance.” The Review of Fama, E., and K. French. 1992. “The Cross-Section of
Financial Studies 22: 4220–4258. Expected Stock Returns.” The Journal of Finance 47: 427–465.
Bhojraj, S., and B. Swaminathan. 2009. “How Does the ——. 1993. “Common Risk Factors in the Returns on Stocks
Corporate Bond Market Value Capital Investments and and Bonds.” Journal of Financial Economics 33: 3–56.
Accruals?” Review of Accounting Studies 14: 31–62.
——. 1995. “Size and Book-to-Market Factors in Earnings
Black, F., and M. Scholes. 1973. “The Pricing of Options and and Returns.” The Journal of Finance 50: 131–155.
Corporate Liabilities.” Journal of Political Economy 81: 637–654.
——. 2015. “A Five-Factor Asset Pricing Model.” Journal of
Campbell, T., D. Chichernea, and A. Petkevich. 2016. “Dis- Financial Economics 116: 1–22.
secting the Bond Profitability Premium.” Journal of Financial
Markets 27: 102–131. Frazier, D., and X. Liu. 2016. “A New Approach to Risk–
Return Trade-Off Dynamics via Decomposition.” Journal of
Chen, L., D. Lesmond, and J. Wei. 2007. “Corporate Yield Economic Dynamics & Control 62: 43–55.
Spread and Bond Liquidity.” The Journal of Finance 62: 119–149.
Friewald, N., C. Wagner, and J. Zechner. 2014. “The
Chen, N. F., and F. Zhang. 1998. “Risk and Return of Value Cross-Section of Credit Risk Premia and Equity Returns.”
Stocks.” The Journal of Business 71: 501–535. The Journal of Finance 69: 2419–2469.
Choi, J., and Y. Kim. “Anomalies in the Joint Cross-Section Gebhardt, W., S. Hvidkjaer, and B. Swaminathan. 2005.
of Equity and Corporate Bond Returns.” Working paper, “Stock and Bond Market Interaction: Does Momentum Spill
University of Illinois at Urbana-Champaign, 2016. Over?” Journal of Financial Economics 75: 651–690.
Harvey, C., Y. Liu, and H. Zhu. 2016. “... and the Cross- Merton, R. 1974. “On the Pricing of Corporate Debt: The
Section of Expected Returns.” The Review of Financial Studies Risk Structure of Interest Rates.” The Journal of Finance 29:
29: 5–68. 449–470.
Houweling, P., and J. van Zundert. 2017. “Factor Investing Novy-Marx, R., and M. Velikov. 2016. “A Taxonomy of
in the Corporate Bond Market.” Financial Analysts Journal Anomalies and Their Trading Costs.” The Review of Financial
73: 100–115. Studies 29: 104–147.
Ilmanen, A., R. Byrne, H. Gunasekera, and R. Minikin. Vassalou, M., and X. Yuhang. 2004. “Default Risk in Equity
2004. “Which Risks Have Been Best Rewarded?” The Journal Returns.” The Journal of Finance 59: 831–868.
of Portfolio Management 30: 53–57.
Zhang, L. 2005. “The Value Premium.” The Journal of Finance
Israel, R., D. Palhares, and S. Richardson. 2018. “Common 60: 67–103.
Factors in Corporate Bond Returns.” Journal of Investment
Management 16: 17–46.
Disclaimer
Jostova, G., S. Nikolova, A. Philipov, and C. Stahel. 2013. Views expressed in this article are those of the authors and do not necessarily
ref lect those of Deka Investment or its employees.
“Momentum in Corporate Bond Returns.” The Review of
Financial Studies 26: 1649–1693.
To order reprints of this article, please contact David Rowe at
d.rowe@pageantmedia.com or 646-891-2157.
Kapadija, N., and X. Pu. 2012. “Limited Arbitrage between
Equity and Credit Markets.” Journal of Financial Economics 105:
542–564.
158 Extending Fama–French Factors to Corporate Bond Markets Quantitative Special Issue 2019
Factors in Time: Fine-Tuning
Hedge Fund Replication
Joseph Simonian and Chenwei Wu
T
Joseph Simonian he primary utility of factor models risk is overfitting, in which the calculation of
is director of quantitative is that they can help simplify the beta coefficients may depend so precisely on
research at Natixis
analysis of portfolios consisting of the training data that a given model fails to
Investment Managers in
Boston, MA. many assets by explaining their generalize on test data, producing poor pre-
joseph.simonian@natixis.com behavior using a parsimonious set of drivers. dictive results. This issue has historically been
This is true of linear regression–based factor somewhat of a secondary concern for those
Chenwei Wu models and those constructed using prin- engaged in building hedge fund replication
is a quantitative analyst cipal component analysis, as well as non- models because they have been primarily
at Natixis Investment
Managers in Boston, MA.
linear factor models, including those based on interested in building frameworks that maxi-
chenwei.wu@natixis.com machine learning algorithms.1 By providing mize in-sample explanatory power. Never-
a more transparent view of the systemic risks theless, predictive power, our focus in this
to which a portfolio is exposed, factor models article, is necessarily the central concern of
can be employed in both risk management practitioners who want to apply hedge fund
and alpha generation. Asset allocators, for replication ex ante in portfolio management.
example, use factor models both to track the Another potential risk with replication strate-
amount and type of systemic risk carried by gies is multicollinearity, where predictors are
different asset classes and to build factor- linearly correlated to a degree that renders
tracking portfolios in which specific factor them problematic for use in an ordinary least
exposures (loadings) are targeted during squares (OLS) regression. The latter risk is
portfolio construction to more precisely ever-present in hedge fund replication, in
express their investment views or to replicate which practitioners generally use publicly
strategies with specific factor profiles. available vehicles, such as exchange-traded
One application of factor-tracking port- funds (ETFs), as factor proxies. A final risk is
folio construction is hedge fund replication, that, because of the dynamic nature of hedge
where liquid investment vehicles are used to funds, models built using a single sample may
build portfolios that are expected to mimic fail to capture the evolution of hedge fund
the investment behavior of different hedge exposures over time.
fund styles. In regression-based frameworks, As a remedy to the foregoing shortcom-
the construction of factor-tracking portfolios ings of the standard approach to factor-based
entails a number of risks. Perhaps the major hedge fund replication, in this article the
authors demonstrate how to use the regu-
1
See Simonian et al. (2019) for an example of larization technique known as ridge regression
the latter approach. (RR) to both mitigate the impact of factor
160 Factors in Time: Fine-Tuning Hedge Fund R eplication Quantitative Special Issue 2019
APPLYING RIDGE REGRESSION and test it on a set of seven hedge fund strategies. Our
TO HEDGE FUND REPLICATION factors are the MSCI USA Value and Momentum
indices, the Bloomberg High Yield Total Return
Because hedge fund replication is generally exe- Index, and the ICE Core Treasury Bond Index. The
cuted using publicly available investment vehicles, the hedge fund strategies we test are merger arbitrage,
factors in hedge fund replication models are generally convertible arbitrage, long/short equity, global macro,
correlated to a sufficient degree to disqualify them as equity market neutral, S&P 500 PutWrite, and S&P
true factor models, which by definition need to consist 500 BuyWrite.4 We show the full-sample correlations
of a (parsimonious) set of uncorrelated predictor vari- of our factors in Exhibit 1.
ables. From a statistical standpoint, multicollinearity Although the matrix in Exhibit 1 is full rank,
negatively affects a regression’s output by inducing and thus valid for a regression, three of the factors
inaccurate estimates of regression coefficients, inf lating show fairly high correlations to each other (value,
standard errors, generating false p-values, and signifi- momentum, and high yield). Given this, RR can still
cantly reducing the predictive utility of a given model. play a useful role in mitigating some of the potentially
Accordingly, to counter the risk of heightened variance deleterious modeling impact of the positive co-move-
between training and test samples, it may be desirable ment of these factors.
to inject some bias into an OLS regression to generalize The central consideration in any application of RR
the model to a greater degree. is what value the penalty parameter l should take. Here
RR 2 is a regression framework that provides a way we calibrate (“fine-tune”) the value of l by testing a
of formally addressing multicollinearity in a standard range of values through time and choosing the l value
OLS regression by adjusting the OLS regression coef- that exhibits the best predictive power over a set of rolling
ficient estimates using a penalty parameter (so called windows. By evaluating different values for the penalty
because it penalizes less inf luential predictors), that parameters across evolving sample sets, we ultimately
injects varying degrees of bias into the regression. When arrive at the penalty value that minimizes prediction
the penalty parameter is applied to the diagonal of the error on test data. Given that our primary goal is to
regression matrix, it forces the columns of the matrix correct for overfitting, it is more prudent to calibrate l
to be linearly independent. The expectation is that by over a multitude of samples to avoid tying the value of
applying the penalty parameter, the regression standard l to a single set of data; doing the latter would risk the
errors will be reduced and lead to the generation of same type of overfitting we are trying to avoid with the
more predictively useful models. Formally, an RR is regression coefficients.5 This approach is also adaptive to
described as follows: new information because the optimal l may change over
time as additional rolling samples are evaluated.
βˆ = argmin β∈R p ∑ i =1( yi − xTi β)2
n
In Exhibit 2, we show the values of l that provide
the best predictive results over two different test hori-
+ λ ∑ j =1 β 2j = argmin β∈R p || y − Xβ ||22 + λ||β ||22 (5)
p
%ORRPEHUJ%DUFOD\V86
06&,86$ 06&,86$ ,&(86 &RUSRUDWH+LJK<LHOG
9DOXH1HW7RWDO 0RPHQWXP86'1HW 7UHDVXU\&RUH 7RWDO5HWXUQ,QGH[
5HWXUQ86',QGH[ 7RWDO5HWXUQ,QGH[ %RQG75,QGH[ 9DOXH8QKHGJHG86'
06&,86$9DOXH1HW ±
7RWDO5HWXUQ86',QGH[
06&,86$0RPHQWXP ±
86'1HW7RWDO5HWXUQ
,QGH[
,&(867UHDVXU\&RUH ± ± ±
%RQG75,QGH[
%ORRPEHUJ%DUFOD\V86 ±
&RUSRUDWH+LJK<LHOG
7RWDO5HWXUQ,QGH[
9DOXH8QKHGJHG86'
Exhibit 2
Hedge Fund Replication—OLS versus Ridge Regression (monthly data, 1/2006–12/2018)
Strategy Optimal λ Average SSE Batting Average In-Sample R2 OLS Average SSE OLS In-Sample R2
Rolling 12-Month Windows; 1-Month-Ahead Replication Horizon
Merger Arbitrage 1.4 6.52E-05 77% 0.35 1.01E-04 0.51
Convertible Arbitrage 0.1 1.95E-04 78% 0.70 2.22E-04 0.73
Long/Short Equity 0.5 7.88E-05 73% 0.70 1.03E-04 0.77
Global Macro 1.4 1.47E-04 90% 0.40 2.51E-04 0.58
Equity Market Neutral 0.2 3.89E-05 85% 0.54 5.21E-05 0.59
S&P 500 PutWrite 0.4 4.26E-04 84% 0.71 6.32E-04 0.76
S&P 500 BuyWrite 0.3 2.65E-04 75% 0.86 4.22E-04 0.89
Rolling 36-Month Windows; 12-Month-Ahead Replication Horizon
Merger Arbitrage 2.4 5.08E-04 71% 0.27 6.03E-04 0.41
Convertible Arbitrage 0.1 7.84E-04 52% 0.68 8.11E-04 0.70
Long/Short Equity 0.2 6.34E-04 63% 0.75 6.81E-04 0.77
Global Macro 0.3 1.21E-03 73% 0.42 1.31E-03 0.45
Equity Market Neutral 1.5 3.62E-04 67% 0.25 4.40E-04 0.46
S&P 500 PutWrite 0.3 2.48E-03 65% 0.73 2.70E-03 0.74
S&P 500 BuyWrite 0.1 1.70E-03 50% 0.87 1.73E-03 0.88
162 Factors in Time: Fine-Tuning Hedge Fund R eplication Quantitative Special Issue 2019
month. The second test horizon is a “strategic” horizon, poor predictive performance when tested on new
where we derive beta coefficients over a 36-month observations. As a remedy to the latter maladies, the
window and attempt to match the performance of a authors employ ridge regression, a technique that is
strategy over the following 12 months. Our measure designed to alleviate multicollinearity and produce
of predictive error is the sum of squared errors (SSE) more reliable estimates of predictor variables. Ridge
over 1-month and 12-month horizons, respectively. For regression uses a penalty parameter that penalizes less
example, in the tactical study, given the generation of inf luential predictors, resulting in an increase in the
beta coefficients at time t (e.g., end-of-month 12/2005), bias and a decrease in the variance of the regression.
we multiply the latter coefficients by the factor returns The decrease in variance typically results in enhanced
at the end of time t + 1 (e.g., end-of-month 1/2006) generalizability of a given model and, accordingly, an
and compare the weighted sum of the latter values to increase in its predictive power. An outstanding ques-
the actual hedge fund strategy returns for the month. tion in any application of ridge regression is what the
The first statistic we show in Exhibit 2 is the value of the penalty parameter should be. The authors
average SSE over the entire set of rolling windows. show how to calibrate the value of the penalty param-
Next, we look at what we call the batting average of eter dynamically by choosing the value that exhibits
the RR replication framework versus a standard OLS the strongest performance across a set of rolling sam-
approach. The batting average is simply the percentage ples. By testing different values for the penalty param-
of rolling windows in which the RR SSE is lower than eter across the training data, they are able to arrive at
the OLS SSE. This is an important statistic because it the parameter value that most consistently minimizes
indicates the generalizability of the ridge framework the prediction error on different sets of test data. The
over evolving sets of data. authors moreover show that the method described pro-
As Exhibit 2 shows, in both the tactical and stra- duces superior predictive results without significantly
tegic studies, the RR-based replication models improve sacrificing a model’s backward-looking explanatory
significantly upon the standard OLS models in terms of power. Thus, ridge regression shows itself to be an
their average SSEs over the set of rolling windows. Per- enhancement over the standard OLS-based approach
haps the most important result, however, is that the bat- to hedge fund replication because it can be reliably
ting average of the ridge approach is generally very high employed in both ex post risk decomposition and ex
(the results for the convertible arbitrage and BuyWrite ante portfolio management.
replication models in the strategic study being the sole
exceptions), indicating that the predictive superiority REFERENCES
of the RR models is consistent through time. Finally,
it is notable that the deterioration in the in-sample R 2 Agarwal, V., and Naik, N. 2000. “Generalised Style Analysis
when using RR is generally minimal, indicating that of Hedge Funds.” Journal of Asset Management 1 (1): 93–109.
the significant improvement in predictive power gained
through the use of RR-based models is not made at the Amenc, N., L. Martellini, J. Meyfredi, and V. Ziemann. 2010.
“Passive Hedge Fund Replication—Beyond the Linear Case.”
expense of a commensurate reduction in ex post explana-
European Financial Management 16 (2): 191–210.
tory power.
Ennis, R., and Sebastian, M. 2003. “A Critical Look at the
CONCLUSION Case for Hedge Funds: Lessons from the Bubble.” The Journal
of Portfolio Management 29 (4): 103–112.
Regression-based factor models are often
employed in hedge fund replication because of their Foster, M. 1961. “An Application of the Wiener-Kolmogorov
simplicity and familiarity. In practice, however, these Smoothing Theory to Matrix Inversion.” Journal of the Society
models are often misapplied because the investment for Industrial and Applied Mathematics 9 (3): 387–392.
vehicles used to replicate a given strategy are typically
correlated to a sufficient degree to disqualify them Fung, W., and D. A. Hsieh. 1997. “Empirical Characteristics
as true factor models. More importantly, OLS-based of Dynamic Trading Strategies: The Case of Hedge Funds.”
Review of Financial Studies 10 (2): 275–302.
models are often overfitted to test data, resulting in
Hasanhodzic, J., and A. W. Lo. 2007. “Can Hedge-Fund ——. 1992. “Asset Allocation: Management Style and Per-
Returns Be Replicated?: The Linear Case.” Journal of formance Measurement.” The Journal of Portfolio Management
Investment Management 5 (2): 5–45. 18 (2): 7–19.
Hastie, T., R. Tibshirani, and J. Friedman. The Elements Simonian, J., C. Wu, D. Itano, and V. Narayanam. 2019.
of Statistical Learning: Data Mining, Inference and Prediction. “A Machine Learning Approach to Risk Factors: A Case
Springer, 2009. Study Using the Fama–French–Carhart Model.” The Journal
of Financial Data Science 1 (1): 32–44.
Hoerl, A. E., and R. Kennard. 1970. “Ridge Regression:
Biased Estimation for Nonorthogonal Problems.” Technometrics Tikhonov, A. N. 1943. “On the Stability of Inverse Prob-
12 (1): 55–67. lems.” Doklady Akademii Nauk SSSR 39 (5): 195–198.
Jaeger, L. Alternative Beta Strategies and Hedge Fund Replication. ——. 1963. “Solution of Incorrectly Formulated Problems
Chichester, UK: Wiley Finance, 2008. and the Regularization Method.” Soviet Mathematics Doklady
4 (4): 1035–1038.
Larson, S. C. 1931. “The Shrinkage of the Coefficient of
Multiple Correlation.” Journal of Educational Psychology 22 (1):
45–55. Disclaimer
This article ref lects the current opinions of the authors, which are not those
of Natixis Investment Managers.
Markov, M., I. Muchnik, V. Mottl, and O. Krasotkina.
“Dynamic Analysis of Hedge Funds.” In Proceedings of the 3rd
To order reprints of this article, please contact David Rowe at
IASTED International Conference on Financial Engineering and
d.rowe@pageantmedia.com or 646-891-2157.
Applications. Cambridge, MA: October 2006.
164 Factors in Time: Fine-Tuning Hedge Fund R eplication Quantitative Special Issue 2019
A Factor- and Goal-Driven Model
for Defined Benefit Pensions:
Setting Realistic Benefits
John M. Mulvey, Lionel Martellini, Han Hao,
and Nongchao Li
T
John M. Mulvey hroughout the world, many In 1998, most DB pensions were either
is a professor at Princeton defined-pension plans are seri- close to fully funded or overfunded. Over
University in the
ously underfunded and unlikely the past 20 years, the funding ratios have
Operations Research and
Financial Engineering to support their future long-term trended downward. Part of the plunge can
Department in Princeton, liabilities without massive contributions and be attributed to the decrease in interest
NJ. possible bailouts. This unfortunate situation rates over this period and to two equity
mulvey@princeton.edu has been caused by a convergence of poor crashes—2001–2002 and 2008–2009. Nev-
contribution decisions, demographic head- ertheless, equities have rebounded from these
Lionel M artellini
is a professor of finance
winds, and inadequate attention to setting events. On the asset side, DB pensions in the
at EDHEC Business realistic benefits. In this article, we focus on United States have had a strong tilt to high
School and the director of the benefit decisions. returning assets, including equities (around
EDHEC–Risk Institute From the perspective of capital, the 50%) and alternative investments (28%), with
in Nice, France. United States has the largest pension endow- solid asset performance.3 In comparison, the
lionel.martellini@edhec-risk.com
ment in the world—$25.4 trillion at the end Netherlands, for example, has 33% equi-
H an H ao of 2017,1 or 131.2% of gross domestic product ties and 17% alternatives and possesses the
is a doctoral student (GDP) (Willis Towers Watson 2017). In com- largest endowment relative to GDP at 193.8%
in the Operations parison, France has an endowment of 6.5% (Willis Towers Watson 2017). Hence, it is
Research and Financial of GDP, and Germany has 12.9% of GDP.2 difficult to blame asset performance. Instead,
Engineering Department It would appear, therefore, that the US pen- we attribute the underfunding to two pri-
at Princeton University
in Princeton, NJ.
sion system is much better capitalized than mary sources. First, historical contributions
hhao@princeton.edu that in most countries. Unfortunately, many have been too low to generate adequate cap-
defined benefit (DB) pensions in the United ital relative to the plan promises. Exhibit 1
Nongchao Li States are severely underfunded, even with shows the contributions over time for one of
is a doctoral student high discount rates by public sector plans: the largest plans within the California Public
in the Operations
7% to 8%. Employees’ Retirement System (CalPERS):
Research and Financial
Engineering Department 15% of salary with much variability. Second,
at Princeton University plan benefits (liabilities) have been too gen-
in Princeton, NJ. 1
The United States has 61% of the total pen- erous, relative to the average contributions
nongchao@princeton.edu sion assets among the 22 countries with the largest and demographic conditions. Exhibit 2 shows
pension plans. the long-term growth of the present value of
2
The second and third largest global pensions
are the United Kingdom and Japan, with roughly
3
$3.1 trillion assets under management each (Willis For example, the average annual performance
Towers Watson 2017). by CalPERS has been 8.5% over the past 30+ years.
Exhibit 2
CalPERS Liabilities over Time
+LVWRULFDO)XQGLQJ6WDWXVRI&DO3(56±
166 A Factor- and Goal-Driven Model for Defined Benefit Pensions: Setting R ealistic Benefits Quantitative Special Issue 2019
liabilities for CalPERS. The present value of liabilities representatives, when aggregated as a whole group, are
has exceeded inf lation and average salaries.4 able to achieve the desired goals at retirement as stand-
To address the mismatch of contributions and alone entities. This achieves two purposes. First, it helps
liabilities, we propose a factor-consistent and goal-driven with the goal-setting process. If the promises are too
framework for managing a DB pension system. The deci- high relative to a majority of individuals, the pension
sions regarding the size of the benefit promises should plan will be unable to support itself. Of course, if the
be evaluated in conjunction with realistic assumptions demographics are positive in the sense that workers sig-
about the investment returns and the ability of the spon- nificantly outnumber retirees, the system can maintain
sors to make adequate contributions. Otherwise, there a generous retiree payment system; conversely, if the
will be an eventual crisis. Although the situation will demographics are negative (as they are in most of the
be most difficult to reverse in the United States,5 there world), there will need to be extra contribution or lower
are opportunities to establish realistic pension systems promises. Second, the approach provides a consistent
in fast-emerging countries such as China (1.5% of GDP) linkage between the factor risks and the critical invest-
and India. Furthermore, individuals will face similar ment, contribution, and benefit decisions. These issues
issues as they manage their own defined contribution can be evaluated by means of the factor and goal-based
(DC) plans. micro–macro system.
The usual order of business for a DB pension is to
compute its funding ratio6 annually with a prespecified MULTIREGIME SCENARIO GENERATOR
discount rate, amortize any deficits over multiple years,
and render required contributions. This process has been In this section, we take up the critical task of
unable to preserve the long-term health of DB pension modeling the performance of the asset categories. First,
systems in the United States: It leads to undercontribu- we employ the traditional approach for setting capital
tions and a slow realization that the survivability of the market assumptions for macro risk factors, taking into
pension system is threatened. Furthermore, the volatility account historical performance and current conditions
of contributions can be high (Exhibit 1) and thereby (Ilmanen 2011; Bogle and Nolan 2015). We employ five
difficult to manage from the standpoint of the sponsor. macro risk factors: (1) global equities, (2) US government
Today, there are DB plans that must contribute over bonds, (3) high-yield bonds, (4) inf lation protection,
50% of employee salary to catch up with previous low and (5) currency protection. Drawing from the factor
contribution levels. investing framework, we generate scenarios and param-
We propose that a micro–macro simulation eter values for all asset classes that are consistent with
be conducted to assess the factor/asset risks that the the scenarios generated for the five risk factors. In an
planned contributions and investment returns achieve effort to obtain more robust estimates, the factor load-
the target promised benefits. An innovation in this ings are determined by means of a regularized regression
article is to evaluate the factor and goal-based analysis algorithm with a LASSO penalty. Blyth, Szigety, and
from the standpoint of individuals (micro level). Here, Xia (2016) applied a similar approach to managing the
we evaluate a set of representative individuals to see if Harvard University endowment.
there is an equilibrium and a reasonable chance that the The output is a set of stochastic scenarios: 10,000
in this study. Each scenario depicts a plausible path across
4
the planning period. Scenario generators similar to ours
The population of California has roughly doubled since have been in widespread use over the past 30+ years for
1970 (from 20 million to 40 million), along with inf lation that has
gone up 6.5 times. In concert, the number of state employees has
pension plans, insurance companies, and wealth man-
almost doubled as well—from 183,000 to 351,000 over the same agement firms.7
period. Liability growth has much exceeded these values. To generate realistic scenarios relative to history,
5
Modifying pension benefits ex post is most difficult in a we employ a two-regime Markov switching approach.
number of high-profile plans, such as the Illinois State pension The idea is to address contraction/crash periods as
systems, because the benefits are virtually guaranteed by the Illinois
State constitution, even though the plan has an approximately 30%
7
funding ratio. Mulvey et al. (2004, 2009) and Mulvey, Bae, and Kim
6
The funding ratio is equal to the market value of assets (2013) have provided details about the underlying equations and
reduced by the present value of expected liabilities. the methodology.
–10.00%
–20.00%
–30.00%
Drawdown
–40.00%
–50.00%
–60.00%
–70.00%
distinct from normal economic growth periods. Abun- and crash. This breakdown is consistent with historical
dant research shows that both volatility and correlation performance (Mulvey and Liu 2016).
increase considerably during crash regimes. Exhibit 3 The underlying Markov model generates scenarios
shows the large drawdowns in equity that have occurred based on a random transition matrix. At each period, we
since the 1920s in the United States. A single multi- fix the regime for the next month based on the transi-
normal distribution cannot generate scenarios that depict tion probabilities: If we are in a normal regime, the
this type of performance. Furthermore, DB pension probability of remaining so for the next period (month)
plans with a tilt to equities and equity-like assets display is 98%. If we are in a crash regime, the probability of
similar characteristics (see Exhibit 4). remaining so is 87.5%. Thus, within a given scenario,
There are advantages to identifying and employing there should be a mixture of normal and crash periods.
economic regimes within a consistent framework. The On average, the steady state probability of a normal
multiple regime approach improves the realism of the regime is 86.2%, which is also consistent with history.
forward projections.8 In addition, machine learning
approaches can be applied to identifying regimes AN OVERVIEW OF MICRO–MACRO
(Tibshirani 1996, 2014; Zou and Hastie 2005; Hastie, ANALYSIS
Tibshirani, and Friedman 2009).
The summary statistics of our scenario generator This section considers the basic framework for
are shown in Exhibit 5. First, we list the expected returns conducting factor-driven asset and liability management
and volatilities for the asset categories over the long (ALM) studies, with attention to the internal factor con-
horizon needed for pension planning. Second, we sepa- sistency of the contribution decisions to the projected
rate the expected returns for the two regimes—normal benefit obligations for a DB pension plan. The output
of an effective ALM system should aim at measuring
8 the impact of asset allocation and other major deci-
For example, see Kyle and Xiong (2001); Ang and Bekaert
(2002); Bauer, Haerden, and Molenaar (2004); Guidolin and sions, such as benefit selection, on portfolio priorities,
Timmermann (2007); Bilgili (2009); Tu (2010); Song (2014); including ensuring the survivability of the plan. As men-
Goldstein and Razin (2015); Nystrup et al. (2015); and Mulvey tioned, we are primarily interested in setting realistic
and Liu (2016). pension benefits.
168 A Factor- and Goal-Driven Model for Defined Benefit Pensions: Setting R ealistic Benefits Quantitative Special Issue 2019
Exhibit 4
Historical Performance for a Large DP Pension Plan
&DO3(56+LVWRU\RI,QYHVWPHQW5HWXUQV±
5HWXUQ
±
±
±
<HDU
'LVWULEXWLRQRI+LVWRULFDO5HWXUQRI&D,3(56±
)UHTXHQF\
± ± ±
5HWXUQ
To start, we make several assumptions within a contributions by the employees and employer plus invest-
steady-state environment. The first assumption is that ment returns should equal expected benefit payments.
across time, a pension plan should balance the sum of This macro-level consistency is required for a pension
expected contributions plus investment returns with the plan to stay solvent. We also assume that the annual
expected outf lows to the retirees. Thus, on average, the contributions will be a fixed proportion of salary under
stochastic inf lation. This assumption is implementable The cash-f low variables indicate the payments to
by the plan sponsors and much less disruptive than large the retirees, xtpay cont
,s , and the contributions, x t ,s , at each
changes in contributions (Exhibit 1). time period under each scenario. This framework
At the micro-level, any individual, say Ms. Kim, allows for a wide variety of investment, contribution,
should have the same consistency—on average. For and payment policies. In this article, our goal is to
example, suppose that an employee begins to work evaluate the ability of the pension system to support
for a sponsoring organization (X-Big) at the begin- the promised payments.
ning of her career age 25. She expects to work 40 Let’s look at a trivial example. Suppose that
years—until age 65, when she desires to retire. Let’s Ms. Kim puts aside $20,000 each year, with adjustments
assume that Ms. Kim remains alive and working for for annual inf lation. Thus, she contributes $20,000 ×
X-Big until age 65, whereupon she retires and lives 40 years = $800,000 real in total. For this highly simpli-
for the next 20 years. Then, in a steady state whereby fied example, we assume that the investment returns are
the number of employees remains approximately con- equal to inf lation. During retirement, again, we assume
stant, the actuarial fair value should enforce the con- that the investment returns are equal to the inf lation
straint: Future values of contribution and investment rate (through inf lation-linked government bonds, for
performance equal the future value of future benefits. example). Thus, she is able to spend $40,000 adjusted
Importantly, individuals in a DC plan cannot violate for inf lation per year, or $800,000, given that the real
this consistency rule. discount rate is equal to 0%. A higher level of spending
The basic equation for the micro–macro simula- or lower contributions would suggest that Ms. Kim is
tion over the long-term horizon involves the capital Ct,s getting more than she deserves, at least from an actu-
at the beginning of time t = [1, 2, …, T ] under each arial fair-value basis and with respect to the significant
scenario s e {S}. At each period, we model uncertainties inf lation risk factor.
via a set of stochastic scenarios {S}. In our example, the We build on this idea by computing the fair value
horizon T will depict a long time in the future. Thus, of a population of individuals at the macro level by simu-
at each time step t, we update the capital by adding lating the fair values for a set of representative individ-
contributions from the population that is alive and uals—micro level—with explicit longevity, investment,
working and subtract the payments to the retirees who and inf lation risks included (Exhibit 6). This type of
remain alive. Furthermore, capital will be adjusted by analysis fits within the realm of agent-based modeling
the asset performance during the previous period: rt–1,s. (Schelling 1971; Bonabeau 2002; Macal and North 2011;
The equation is defined as follows: Lychkina and Morozova 2014). The scenarios depict
uncertainties for inf lation, investment returns, and lon-
Ct ,s = Ct −1,s ∗ (1 + rt −1,s ) + xtcont
,s − x t ,s
pay
for all t,s gevity risks across the long time period.
170 A Factor- and Goal-Driven Model for Defined Benefit Pensions: Setting R ealistic Benefits Quantitative Special Issue 2019
Exhibit 6
Consistent Linkage between Fair-Value of Individuals and Survivability of Pension System
:RUNIRUFH :RUNIRUFH :RUNIRUFH
VHOI VHOI VHOI
VXVWDLQLQJ VXVWDLQLQJ VXVWDLQLQJ
7LPHOLQH
0LFUR±0DFUR&RQVLVWHQF\
Clearly, for actual individuals, fair value will be tency. It fits a conservative liability-driven investment
violated under random circumstances. Ms. Kim may live (LDI) framework.
longer than age 85 and thus reap greater benefits than
her contributions. Or there might be a larger number of STYLIZED EXAMPLE: THE US SOCIAL
workers than retirees, and again the benefits can exceed SECURITY SYSTEM
the costs, as has been the case for DB pensions since their
inception in the late 1880s with Siemens and other firms This analysis extends the previous example for
in Germany.9 Unfortunately, we are moving toward the Ms. Kim. We approximate the contribution and the
opposite environment, wherein the number of retirees benefits at the maximum Social Security level, then
is beginning to exceed the number of workers. In this apply this rule across a homogeneous population using
case, a fair-value calculation can assist with setting rea- current longevity tables and the inf lation risk factor.
sonably attainable benefits because it provides an upper The purpose is to illustrate the process and provide an
bound on benefits. example with approximate micro–macro consistency.
Stylized illustrations of a micro–macro analysis Next, we highlight the primary assumptions and issues.
based on our experiences with actual DB pension The stylized micro–macro model of a small subset
plans are discussed in this section. These examples are of the Social Security Trust Fund10 focuses on the
most appropriate for public sector pensions, in which matching of net contributions and benefits. Exhibit 7
employees are likely to work within the same spon- provides high-level information on the current status
soring organization over their career, and for individual of the Trust Fund. We simulate individuals with some
DC plans. wage distribution and life expectancy estimates by fol-
The first case study shows that the US Social lowing the assumptions mentioned in the old age, sur-
Security system is reasonably fair valued, at least for vivors, and disability insurance (OASDI) annual report
individuals with higher than average incomes, average
longevity risks, stochastic inf lation risks, and a steady- 10
The Social Security program started in 1935 and has been
state population. A primary reason for micro–macro
paying on a timely schedule since then. The program is funded by
consistency is the relatively low level of benefits for contribution from both employers and employees in the form of pay-
the target population: about 25% of final years’ salary roll taxes or self-employment taxes. It pays for retirement and other
as defined by the Social Security maximum amount benefits, such as disability. Beginning in 1941, the Social Security
and under stochastic inf lation. The level of benefits is Board of Trustees is required to present to the Congress a financial
commensurate with the low returns (and safety) of the report and detailed actuarial estimates of the fund. According to
Goss (2010), the board has six members, including the Secretary of
investment mix: 100% government inf lation-linked the Treasury as the managing trustee, the Secretary of Labor, the
bonds. The micro–macro analysis confirms this consis- Secretary of Health and Human Services, and the Commissioner of
Social Security, plus two public trustees appointed by the president
9
The earliest pension plans were created by Otto von and confirmed by the Senate. Exhibit 7 shows the current financial
Bismarck in the late 1880s. statement of the fund.
Notes: aLess than $50 million. Totals do not necessary equal the sums of rounded components.
Source: OASDI Board of Trustees (2018).
172 A Factor- and Goal-Driven Model for Defined Benefit Pensions: Setting R ealistic Benefits Quantitative Special Issue 2019
Exhibit 8
Sensitivity Analysis with 0.25% and 0.5% Real Returns on TIPS
3DQHO$3UREDELOLW\RI5HDFKLQJ*RDO5HDO5HWXUQV 3DQHO&3UREDELOLW\RI5HDFKLQJ*RDO5HDO5HWXUQV
3DQHO%0HVK3ORWIRU3UREDELOLWLHVRI5HDFKLQJ*RDO 3DQHO'0HVK3ORWIRU3UREDELOLWLHVRI5HDFKLQJ*RDO
5HDO5HWXUQV 5HDO5HWXUQV
3U5HDFK*RDO
3U5HDFK*RDO
JH
$JH
H$ H
H WLU WLU
&RQWULE
XWLRQ5
5 &RQWULE
XWLRQ5 5H
DWH,Q3 DWH,Q3
HUFHQWD HUFHQWD
JH JH
now is 12.4%. This means, in real terms, each person sustainability. Exhibit 8 shows that the probability of
in the new workforce (together with the employer) is reaching macro-consistency is 44.2% at age 65, 98%
paying a maximum fixed amount of $16,479 per year. at age 67, and virtually 100% at ages 69 and 71. In the
In addition, the maximum benefit a person can receive United States, the full retirement age for Social Secu-
at full retirement age is $2,861 per month in today’s rity benefits is moving toward 67, which corresponds
dollars. We will assume this amount to increase at our to a 98% probability of micro–macro consistency for
simulated stochastic COLA values. the target population. Thus, we see that the current
Based on all of these assumptions, we generate contribution and benefits have micro–macro consis-
10,000 scenarios for each of the retirement age ranges tency from the standpoint of these 100,000 individuals
from 65 to 75 and calculate the probability of reaching the under steady-state assumptions. In addition, Exhibit 8
goal of sustainability by calculating the probability that the provides sensitivity analyses along two lines: changes in
simulated present values of net contribution are larger than the contribution amounts (from 12% to 13%) and two
zero across the 100,000-member population. The simula- return values (0.25% and 0.5%) above inf lation for the
tion algorithm appears in the supplemental materials. TIPS. Clearly, the chance of achieving the modest goal
(25% of maximum social security salary) improves with
Baseline incremental increases in contribution and higher returns.
Moreover, as we observe, given a fixed contribution
In this section, we first show the baseline results rate and TIPS real return (0.25% per annum), the later
with the previous model specifications. Assuming the people retire, the higher the probability of reaching the
contribution rate is 12.4% (current level), under different consistency goal. The probability increases from 44.2%
retirement ages from 65 to 71 with an increment of to 98% when the age of retirement increases from 65
2, we calculate the probability of reaching the goal of
3DQHO%0HVK3ORWIRU3UREDELOLWLHVRI5HDFKLQJ*RDO5HWLUHDW 3DQHO'0HVK3ORWIRU3UREDELOLWLHVRI5HDFKLQJ*RDO5HWLUHDW
3U5HDFK*RDO
3U5HDFK*RDO
H H
DW DW
Q 5 H
Q 5 H
LR J LR J
XW WD XW WD
3D\RXW
5DWH, W U LE UFHQ 3D\RX
W5DWH W U LE UFHQ
Q3HUFH Q H ,Q3H Q H
QWDJH &R ,Q3 UFHQWD
JH &R ,Q3
to 67. The approach shows the benefits of LDI with a contribution per year, with a max cap to retire at age 65.
stochastic inf lation risk factor. Extending retirement to age 70 is much more attainable
for contributions in the 15%–20% range.
EMPIRICAL TEST #2: EXAMPLES WITH In the next analysis, we introduce a portfolio11 that
A PORTFOLIO OF RISKY ASSETS is closer to those that larger DB pension plans employ
with a mix of asset categories, including world equity,
In this section, we expand the previous analysis government and high-yield bonds, real estate, com-
to address the consistency issue with two portfolios of modities, and corporate bonds. As mentioned, the factor
risky assets. approach12 estimates the projected performance based on
First, we employ a 60/40 mix, in which 60% is the five macro factors. For simplicity, we assume that the
invested in US equities and 40% in US long-duration portfolio is rebalanced annually. As expected as shown
government securities. Given the higher expected values in Exhibit 10, this portfolio assumes higher expected
and the long-term horizon, the probability of reaching a returns and thus achieves its goals with larger probabili-
consistent plan is higher than that by investing in TIPS ties, such as a 70% chance of retiring with 15% annual
alone—compare Exhibits 8 and 9. Even in this case, contribution at age 65 versus a 63% chance. However, at
however, the benefits must be restricted to approxi- the higher levels of contribution, the traditional 60/40
mately the 40% threshold under 15% contribution if mix provides more consistency and thereby generates
we want to achieve the goals with 90% confidence.
In comparison, financial planners suggest that retirees 11
Our sample portfolio is obtained via robust optimization
should aim for 60%–80% of final salary during retire- method. See Fabozzi et al. (2007).
ment. Clearly, this goal would require at least 20%–25% 12
See Simonian et al. (2018).
174 A Factor- and Goal-Driven Model for Defined Benefit Pensions: Setting R ealistic Benefits Quantitative Special Issue 2019
Exhibit 10
DB Pension Plan (10,000 scenarios), Stylized Pension Asset Mix (world equity 25%, US equity 35%,
US Treasury 10%, high yield 5%, real estate 10%, commodity 5%, corporate bonds 10%)
3DQHO$3UREDELOLW\RI5HDFKLQJ*RDO5HWLUHDW 3DQHO&3UREDELOLW\RI5HDFKLQJ*RDO5HWLUHDW
5HWLUHDW 5HWLUHDW
3DQHO%0HVK3ORWIRU3UREDELOLWLHVRI5HDFKLQJ*RDO5HWLUHDW 3DQHO'0HVK3ORWIRU3UREDELOLWLHVRI5HDFKLQJ*RDO5HWLUHDW
3U5HDFK*RDO
3U5HDFK*RDO
H H
DW DW
Q 5 H Q 5 H
LR J LR J
XW WD XW WD
3D\RXW
5DWH, W U LE UFHQ 3D\RX
W5DWH W U LE UFHQ
Q3HUFH Q H ,Q3H Q H
QWDJH &R ,Q3 UFHQWD
JH &R ,Q3
slightly higher probabilities. If the age of retirement is increased above the Social Security 25% level if higher-
delayed to age 70, the 60/40 strategy largely dominates returning assets are part of the asset portfolio. How-
the second broader asset mix at age 65. It seems wise, ever, even by increasing contribution to 15% per year
therefore, to expect increased retirement ages in the from 12.4%, we are confident of reaching this goal with
future, even in the case of a steady-state population. 90% only if the benefits are set at 40% of the final max
salary at age 65. Of course, higher contributions and
CONCLUSIONS later retirement will generate greater levels of attainable
benefits. Additional real-world considerations,13 such as
A critical objective for any liability- or goal-driven goal-risk measures and dynamic investment strategies
asset owner is to evaluate the size of the benefits (prom- (Martellini, Malhau, and Mulvey 2018, 2019), can be
ises or goals) relative to the chance of meeting these included in the micro–macro modeling approach.
promises and with respect to the underlying major risk In summary, the determination of appropriate
factors—longevity and inf lation, in our study. This pension benefits should be evaluated on several fronts.
objective is attainable, in our opinion, for a DB pension First, the benefits need to be fairly attainable for the
system within the developed factor- and goal-driven sponsoring organization and the employees. Under most
framework. The first example suggests that even a circumstances, a long-term contribution rate above a
highly conservative asset allocation for the Social Secu-
rity system, 100% TIPs, can achieve a baseline goal for 13
The average years of service for all retirees in some of the
a portion of the steady-state population. In the second
plans in CalPERS is 19.8 years (Fox 2017). The micro–macro simu-
set of examples, we show that the selected benefit can be lation can include real-world considerations of this type.
Ang, A., and G. Bekaert. 2002. “International Asset Allo- Hastie, T., R. Tibshirani, and J. Friedman. The Elements of
cation with Regime Shifts.” The Review of Financial Studies Statistical Learning: Data Mining, Inference, and Prediction. 2nd
15 (4): 1137–1187. ed. New York: Springer, 2009.
Bauer, R., R. Haerden, and R. Molenaar. 2004. “Asset Allo- Ilmanen, A. Expected Returns: An Investor’s Guide to Harvesting
cation in Stable and Unstable Times.” The Journal of Investing Market Rewards. Chichester: John Wiley & Sons, 2011.
13 (3): 72–80.
Kyle, A., and Xiong, W. 2001. “Contagion as a Wealth
Bilgili, M. “Clustering Techniques and Multi Regime Sto- Effect.” The Journal of Finance 56 (4): 1401–1440.
chastic Optimization with Applications in Finance.” Ph.D.
dissertation, Princeton University, 2009. Lychkina, N., and Y. Morozova “Dynamic Simulation of
Pension System Development Processes.” 32nd Interna-
Blyth, S., M. C. Szigety, and J. Xia. 2016. “Flexible Indeter- tional Conference of the System Dynamics Society, Delft,
minate Factor-Based Asset Allocation.” The Journal of Portfolio the Netherlands, July 24–28, 2014.
Management 42 (5): 79–93.
Macal, C., and M. North. “Introductory Tutorial: Agent-
Bogle, J., and M. Nolan Jr. 2015. “Occam’s Razor Redux: Based Modeling Simulation.” Proceedings of the 2011 Winter
Establishing Reasonable Expectations for Financial Market Simulation Conference, 2011.
Returns.” The Journal of Portfolio Management 42 (1): 119–134.
Martellini, L., V. Malhau, and J. Mulvey. “Securing Replace-
Bonabeau, E. 2002. “Agent-Based Modeling: Methods and ment Income with Goal-Based Retirement Income Strate-
Techniques for Simulating Human Systems.” Proceedings of gies.” Working paper, 2018.
the National Academy of Sciences of the United States of America
99 (3): 7280–7287. ——. 2019. “Flexicure Retirement Solutions: A Part of
the Answer to the Pension Crisis?” The Journal of Portfolio
CalPERS. “California Public Employees’ Retirement System Management (forthcoming).
Investments.” 2019, https://www.calpers.ca.gov/page/
investments. Mulvey, J., G. Bae, and W. Kim. 2013. “Dynamic Asset Allo-
cation for Varied Financial Markets under Regime Switching
CalPERS. “CalPERS State Actuarial Valuation Report.” 2017, Framework.” European Journal of Operational Research 234 (2):
https://www.calpers.ca.gov/docs/forms-publications/2017- 450–458.
state-valuation.pdf.
Mulvey, J., F. Fabozzi, W. Pauling, and K. Simsek. 2004.
Fabozzi, F., P. Kolm, D. Pachamanova, and S. Focardi. Robust “Modernizing the Defined Benefit Pension System.” The
Optimization and Management. Hoboken: John Wiley, 2007. Journal of Portfolio Management 31 (2): 73–82.
14
Similar to the Social Security system, CalPERS recently
added a retirement cap to new employees.
176 A Factor- and Goal-Driven Model for Defined Benefit Pensions: Setting R ealistic Benefits Quantitative Special Issue 2019
Mulvey, J. M., and H. Liu. 2016. “Identifying Economic Tibshirani, R. 1996. “Regression Shrinkage and Selection
Regimes: Reducing Downside Risks for University Endow- via the LASSO.” Journal of the Royal Statistical Society. Series B
ments and Foundations.” The Journal of Portfolio Management (Methodological) 58 (1): 267–288.
43 (1): 100–108.
——. 2014. “Adaptive Piecewise Polynomial Estimation via
Mulvey, J., K. Simsek, Z. Zhang, and F. Fabozzi. 2009. Trend Filtering.” The Annals of Statistics 42 (1): 285–323.
“Assisting Defined-Benefit Pension Plans.” Operations Research
56 (5): 1066–1078. Tu, J. 2010. “Is Regime Switching in Stock Returns Impor-
tant in Portfolio Decisions?” Management Science 56 (7):
Nystrup, P., B. W. Hansen, H. Madsen, and E. Lindstrom. 1198–1215.
2015. “Regime-Based versus Static Asset Allocation: Letting
the Data Speak.” The Journal of Portfolio Management 42 (1): Willis Towers Watson. “Global Pension Asset Study.”
103–109. 2017, https://w w w.wil listowerswatson.com/en-US/
insights/2017/01/global-pensions-asset-study-2017.
OASDI Board of Trustees. “The 2018 OASDI Trustees
Report.” 2018, https://www.ssa.gov/OACT/TR/2018/ Zou, H., and T. Hastie. 2005. “Regularization and Variable
tr2018.pdf. Selection via the Elastic Net.” Journal of the Royal Statistical
Society. Series B (Statistical Methodology) 67 (2): 301–320.
Schelling, T. C. 1971. “Dynamic Models of Segregation.”
Journal of Mathematical Sociology 1 (2): 143–186.
To order reprints of this article, please contact David Rowe at
Simonian, J., O. Sosa, S. Chandrashekar, and D. Nielson. d.rowe@pageantmedia.com or 646-891-2157.
2018. “LDI-Sensitive Equity Factor Portfolios: The ALM
Perspective to Smart Beta Investing.” The Journal of Portfolio
Management 44 (4): 106–112.
A
ll submitted manuscripts must be original work that hasn’t been published in another form such
as another journal, magazine, or book chapter. Upon submitting a manuscript for consideration
by Institutional Portfolio Research Journals (IPR Journals), an author is restricted from submitting
their manuscript elsewhere until an editorial decision on their work has been made. Submissions are
only accepted by uploading to our online article submission system. We do not require any specific style
or format for initial submission. However, upon acceptance of an article for publication, the author will
be expected to bring their manuscript into conformity with the stylistic and mechanical specifications
described in the full submission guidelines. Also at that time, the author will be expected to sign the IPR
Journals copyright agreement and to remove all prior versions of the article from everywhere that they
have been posted, including SSRN and all other websites.
Target Length: The target length for articles is 4,000 words, Footnotes: IPR Journals publications now use footnotes rather
not including the abstract, exhibits, and references. How- than endnotes. Footnotes generally should not be used for
ever, the editor of a journal may accept articles that are references but rather for peripheral or tangential commen-
substantially longer or shorter. Articles shorter than 2,500 tary that would not be suitable for inclusion in the main text.
words or longer than 7,500 words are rarely accepted.
References: IPR Journals follows the “author-date” reference
Title: You should give your article a short descriptive title. In format of The Chicago Manual of Style (CMS), 17th Edition,
general, an article’s title should be no more than 12 words. Chapter 15 both for references included in the text of an arti-
You should try to avoid including words that carry little or cle and for the list of references at the end of an article. The
no meaning. list of references should immediately follow the main text.
Abstract: You should write a strong abstract for an article Style, Usage, and Grammar: IPR Journals follow CMS with
appearing in an IPR Journals publication. The target length respect to matters of style, usage, and grammar. This includes
for abstracts is 160 words. The abstract should be non-tech- standards for hyphenation, abbreviations, and capitalization.
nical, should not contain any references, and should define
Exhibits: Please call all tabular or graphical material Exhibits.
any acronyms on first use. An article’s abstract is the only
Number exhibits with Arabic numbers consecutively in order
part of the article that appears on the unrestricted portion
of appearance in the text. Do not call them Table 1 and
of the related journal’s website. In the printed version of a
Figure 1. An article’s text should refer to all of its exhibits. In
journal, the abstract appears in the table of contents (as
general, the text should clearly explain the point supported
if it is written by the editorial team, not the authors). It does
or demonstrated by each exhibit. Additionally, except for
not appear directly before the article’s main text.
the simplest tables and charts, the text should explain how
Introductory Material: Please write strong introductory para- to read each exhibit. For tables that report statistical results,
graphs. The introductory material is an important part of an the text should direct readers to the key values and clearly
article. It serves as a substitute for the abstract at the start explain what they mean. For an equation, the text should
of the article. Because an abstract is not presented right describe and explain the relationship represented by the
before its associated article, a reader will not necessarily equation and clearly define all terms used in the equation.
have read the abstract before reading the main text. An In many cases, it is better to place content that is heavy with
article’s introductory material should describe the article’s equations in an appendix. Within the IPR Journals family of
main points and briefly explain the reasoning or analysis journals, The Journal of Fixed Income and The Journal of
that supports them. The introductory material should explain Derivatives generally allow for the highest proportions of
why the findings, results, or conclusions are important and quantitative material in the main text.
how they add to prior research in the area. If applicable,
Exhibit Presentation: Authors are encouraged to submit arti-
the introductory material should convey how an investment
cles with exhibits in color. Although IPR Journals publications
professional can use the findings, results, or conclusions in a
are generally printed in black and white, the online version
practical way. In most cases it is helpful to include a “road-
can show color exhibits. Also, upon article acceptance, if
map” paragraph at the end of the introductory material.
you wish to have your exhibits printed in color, you may do
The purpose of the roadmap paragraph is to inform readers
so at a cost of $1,200. Please make sure that all categories
about how the article is organized into different sections and
in an exhibit can be distinguished from each other and
what is covered in each one.
match the categories in any legend connected with an
Headings: Articles in IPR Journals publications can have up to exhibit. If you use color, please make sure that each exhibit
three levels of headings. The first heading in an article should will be legible if printed in black-and-white. This is important
appear after the introductory material. Articles in IPR Journals for all articles because a reader may print an article on a
publications do not include any heading before an article’s black-and-white printer.
introductory material.
Copyright Agreement: IPR Journals’ copyright agreement
Conclusion: Please end the main text of an article with a short form must be signed prior to publication. Only one author’s
conclusion that briefly summarizes the article’s main point. signature is necessary. Should your article not be accepted
The conclusion may resemble the article’s abstract, though the copyright agreement will be considered null and void.
it need not omit references and technical terms used in the Upon acceptance of the article, no further changes are
article. The conclusion should not introduce new material. allowed, except with the permission of the editor.
Submit Your Article to JPM
jpm.iprjournals.com
1
Go to jpm.iprjournals.com
and choose Submit an article
2
Click to SUBMIT AN ARTICLE and
follow the prompts to register with
Editorial Manager
3
Scroll down to download
the complete version of the
Submission Guidelines
We’re proud to support and
contribute to the advancement
of investment management research.
aqr.com