Financial Analysts Journal

ISSN: 0015-198X (Print) 1938-3312 (Online) Journal homepage: https://www.tandfonline.com/loi/ufaj20

Factor-Based Asset Allocation vs. Asset-Class-

Based Asset Allocation

Thomas M. Idzorek & Maciej Kowara

To cite this article: Thomas M. Idzorek & Maciej Kowara (2013) Factor-Based Asset Allocation
vs. Asset-Class-Based Asset Allocation, Financial Analysts Journal, 69:3, 19-29, DOI: 10.2469/
faj.v69.n3.7

To link to this article: https://doi.org/10.2469/faj.v69.n3.7

Published online: 28 Dec 2018.

Submit your article to this journal

Article views: 152

View related articles

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=ufaj20
 Financial Analysts Journal
·
Volume 69 Number 3
©2013 CFA Institute

PERSPECTIVES

Factor-Based Asset Allocation vs.

Asset-Class-Based Asset Allocation
Thomas M. Idzorek, CFA, and Maciej Kowara, CFA

This article addresses the issue of the alleged superiority of risk-factor-based asset allocations over the more
traditional asset-class-based asset allocation. The authors used both an idealized model, capable of precise
mathematical treatment, and optimizations based on different periods of historical data to show that neither
approach is inherently superior to the other. Although the authors appreciate the role of risk models in portfo-
lio management, they urge caution with respect to unwarranted claims of their dominance.

F
or the majority of the last 60 or so years since but we emphasize that we are fans of risk models
the publication of Markowitz (1952), the pri- and advocates of many common risk factors. We
mary building blocks that have been opti- applaud the innovation of using risk factors in the
mized to determine an “asset allocation” have been asset allocation process.
asset classes. More recently, a new type of potential Finally, the information presented here should
building block has emerged: risk factors. Recent not be confused with another popular topic
articles on risk-factor-based asset allocation imply that has the word “risk” in its name: risk par-
that the approach is somehow inherently superior ity. Conceptually, risk parity approaches focus
to asset allocation based on asset classes. If a fair on achieving maximum diversification (however
apples-to-apples comparison is conducted, neither that is defined) as opposed to the more traditional
approach is superior. In the absence of superior approach of maximizing return per unit of risk. If
information, simply reorganizing securities into you believe in risk parity, you can certainly apply a
different building-block clusters does not generate risk parity approach to an opportunity set of either
superior risk-adjusted returns. In a strategic asset risk factors or asset classes.
allocation setting, superior returns come from a
large, granular opportunity set, free from artificial
constraints, coupled with superior skill at fore-
A Brief History of Risk Factors
casting the long-term capital market assumptions We begin with a brief history of asset allocation,
of the opportunity set. In a tactical asset alloca- the needs that lead to risk-factor-based approaches,
tion setting, superior returns relative to the target and examples of recent studies that imply that risk-
come from superior skill at forecasting the short- factor-based asset allocation is inherently superior
term capital market assumptions. In this article, to asset allocation based on asset classes.
we examine whether asset allocation based on risk Harry Markowitz’s work on mean–variance
factors or asset allocation based on asset classes optimization in the 1950s created the crux of mod-
is inherently superior. We mathematically prove ern portfolio theory and the most widely accepted
that neither approach is inherently superior and method for creating portfolios. As practitioners
provide empirical examples to help illustrate the embraced mean–variance optimization throughout
point in a more realistic setting. Our goal is to set the 1970s, 1980s, and 1990s, the investment process
the record straight regarding the two approaches, developed into two distinct processes. First, because
it is nearly impossible to run a single-stage optimi-
Thomas M. Idzorek, CFA, is president and global chief zation that includes all available individual securi-
investment officer at Morningstar Investment Manage- ties, the individual securities are grouped into asset
ment, Chicago. Maciej Kowara, CFA, is portfolio man- classes and the optimization is performed on the
ager and research director at Transamerica Asset Man- asset classes. This first stage is often referred to as
agement, Chicago. the asset allocation, or beta, decision; presumably,

May/June 2013 www.cfapubs.org 19

FAJ_MayJune2013.indb 19 5/9/2013 1:29:07 PM

 Financial Analysts Journal

the asset classes in question have an inherent, non- algorithmic strategies that mimic such strategies as
skill-based return premium in which the primary momentum, distressed debt, and merger arbitrage.
source of that premium is beta, or market risk. In Then, there was “exotic beta,” a Goldman Sachs
the second part of the process, managers that spe- concept that tried to categorize somewhat cheaply
cialize in security selection, typically within each of implementable exposures to the less traditional
the individual asset classes, are hired to implement segments of the market, such as commodities.7
the target asset allocation. This second step is often Some of the latest factors du jour are volatility itself
referred to as the product, or alpha, decision.1 and relative liquidity.8
The asset allocation decision rarely involves For the vast majority of this journey through
more than 20 asset classes, whereas the problem factor development, the primary application of
faced by a fund manager building a portfolio of structural multifactor risk models was the construc-
individual securities within a given asset class can tion of portfolios of individual securities. Only dur-
involve thousands of individual securities. Mean– ing the most recent 20 years or so have structural
variance optimization requires the practitioner to multifactor risk models been used to more closely
estimate expected returns, standard deviations, monitor portfolio managers and perform detailed
and the correlations of each asset relative to all the attribution analyses. Arriving at our current issue,
other assets being optimized. To optimize thou- in the last 10 years or so, asset managers steeped in
sands of individual securities, it is necessary, yet the science of constructing security-level portfolios
impossible, to estimate all the correlations using started using structural multifactor risk models to
time-series data;2 this problem led to the develop- develop multi-asset-class portfolios. The problem
ment of structural multifactor risk models.3 These is not this innovation. The problem is that authors
models attempt to identify a reasonable number of who strongly suggest that risk-factor-based asset
common factors that explain the returns of indi- allocation is inherently superior are at best over-
vidual securities. stating their case and at worst confusing investors
The Sharpe–Lintner–Mossin–Treynor capital with a false value proposition.9
asset pricing model, in which the “market” became Recent papers that strongly suggest that risk-
the risk factor (at least for equities), is the super fac- factor-based asset allocation is inherently superior
tor for most risk-factor models.4 Stephen A. Ross, to asset allocation based on asset classes include
using his arbitrage pricing theory (APT), was the Page and Taborsky (2011) and Bender, Briand,
first to formalize the argument that multiple risk Nielsen, and Stefek (2010). Positive messages from
factors contribute to asset-class returns. Ross (1976) these papers with which we agree are that relaxing
put forth a more general model and left the spe- the long-only investment constraint and expand-
cific factors and robust theory for future research. ing one’s opportunity set to include more potential
That research was undertaken in due course by exposures are often good things (although the latter
academia. Fama and French (1992) supplemented point by itself is rather trivial). Unfortunately, most
the market factor by identifying size and valuation such papers use apples-to-oranges comparisons
factors that could not be explained by the market- that lead all but the most careful reader to believe
risk-factor model. Jegadeesh and Titman (1993), that risk-factor-based asset allocation is inherently
together with Carhart (1997), added momentum to superior to asset allocation based on asset classes; in
the lineup of generally recognized factors.5 These fact, these papers offer neither real proof nor even
original factors were motivated by a desire to a rigorous argument. The comparisons allegedly
explain the market and its anomalies rather than by confirming the superiority of factor-based alloca-
the needs of those involved in portfolio building, tions are typically made between a relatively simple
who sooner or later adopted the factors. That step asset-class set and a risk-factor set that includes
was facilitated by Grinold and Kahn (2000), who many more potential exposures, which is the sense
devoted parts of their standard investment text- in which we consider these to be apples-to-oranges
book to building factor-based portfolios. comparisons. For example, Clarke, de Silva, and
The next set of proposed factors came from a Murdock (2005), early proponents of risk-factor-
combination of two lines of thought: (1) an attempt based asset allocation, compared a five-exposure
to explain manager performance and (2) the real- set (three asset classes and 2 factors) with a much
ization that many seemingly innovative strategies more robust set of 14 factors that includes global
can be cheaply implemented by systematic passive equity and currency exposures, whereas Bender et
exposures. Thus, Fung and Hsieh (2004) proposed a al. (2010) compared a rather simplistic stock/bond
seven-factor model to explain hedge fund returns.6 portfolio with a 10-factor set that includes valua-
Similarly, Jensen, Yechiely, and Rotenberg (2005) tion, momentum, term and credit spreads, currency,
documented how many hedge fund return series and semi-active exposures to such strategies as con-
can be reasonably approximated by purely passive vertible and merger arbitrage. Furthermore, many

20 www.cfapubs.org ©2013 CFA Institute

FAJ_MayJune2013.indb 20 5/9/2013 1:29:07 PM

 Factor-Based Asset Allocation vs. Asset-Class-Based Asset Allocation

of the presumed gains result from the fact that the Above, we emphasized the word “uncon-
comparisons involve different inherent constraints strained.” It is important to realize that in practice,
on the asset classes involved; setting a long-only most optimizations impose a long-only constraint,
portfolio next to a portfolio that may go long and which has different implications for opportunity
short does not make for a meaningful comparison. sets of asset classes and opportunity sets of risk
The often-highlighted fact that the pairwise cor- factors. By construction, many risk factors involve
relations among risk factors are lower than those leverage (e.g., long small caps and short large caps
among asset classes should not be confused with are used to create the size factor); thus, a comparable
the superiority of one approach over the other. “long-only” optimization of risk factors in which
Note that it would be impossible for the world the optimizer is not allowed to short is, in fact, less
constrained than a long-only optimization of asset
as a whole to embrace risk-factor-based asset allo-
classes. Ignoring potential legal and practical con-
cations. In a well-defined asset-class, or “group-
straints associated with leverage extremes, most
ing,” scheme, all individual securities should be
optimization constraints are actually arbitrary and
assigned to an asset class and the various asset
reflect investor preferences. Intuitively, an investor
classes should be mutually exclusive. This process
who is willing to allocate to risk factors—which, by
results in what is known as a macro-consistent construction, require shorting—seemingly would
scheme, which enables all investors to hold the not impose a long-only constraint on an asset-class
same portfolio should they so choose. With risk optimization.
factors, individual securities often live in mul- The realized returns, expected returns, and
tiple factors, such as a bond that is part of both covariance matrices for both asset classes and risk
the duration and credit risk factors, and perhaps factors are, respectively, denoted by
more importantly, it is impossible for all investors
to hold the same portfolio because most factors ra , a , a , r f ,  f , and  f ,
require offsetting long and short positions, such as where the subscripts a and f denote asset classes
valuation (long value, short growth) or size (long and risk factors, respectively.
small cap, short large cap). The entire world can- In this simplified world, the vector of asset-
not simultaneously have a long position in the size class returns is a linear transformation of the fac-
premium, which would require everyone to short tors based on matrix L, where the asset class to risk
large caps. factor exposure-mapping matrix L is square and
Before we go into the details, let us briefly elab- invertible. In other words, we assume that in return
orate on what we mean by “risk factors” because space, asset classes can be fully expressed by fac-
the term can be understood in many different ways. tors and vice versa:
Consistent with the current papers that proclaim ra = Lr f , (1a)
risk-factor superiority, in this article, factors are
essentially long–short portfolios of asset classes. and
We do not attempt to investigate the relative merits
r f = L−1ra . (1b)
of APT or macro-based factors—such as industrial
production, inflation expectations, consumption, The covariance matrix of the asset returns is
and oil prices—for portfolio construction, nor do calculated from the exposure-mapping matrix L
we try to assess the usefulness of models based on coupled with the covariance matrix of risk factors,
security characteristics, such as sector or industry. or alternatively, the covariance matrix of risk factors
is calculated from the exposure-mapping matrix L
coupled with the covariance matrix of asset classes:
Proof of Equality in an Idealized
World a = L f L ', (2a)
The mathematical proof is easiest in a simplified and
world in which the number of factors equals the
number of assets, the asset-class returns are com-  f = L−1a ( L ') −1. (2b)
pletely determined by the risk factors, and the risk Our initial goal is to prove the obvious; when
factors are completely determined by the asset- risk factors perfectly explain the returns of asset
class returns. In such a world, it is easy to demon- classes, there is no inherent advantage from either
strate that the solutions to two unconstrained mean– approach. Stated differently, we can start with
variance optimizations—one in asset-class space either risk factors or asset classes, derive an optimal
and one in risk-factor space—are equivalent. portfolio, and move from one space to the other

May/June 2013 www.cfapubs.org 21

FAJ_MayJune2013.indb 21 5/9/2013 1:29:13 PM

 Financial Analysts Journal

with no gain or loss in efficiency. The solution to should produce a more efficient asset allocation. In
the unconstrained mean–variance maximization fact, both asset classes and risk factors derive their
problem is as follows: returns from individual securities that, in turn, can
be organized into both asset classes and risk factors.
1 −1
w=  µ, (3) It is the somewhat arbitrary construction of the asset
θ classes and risk factors, the selection of apples-to-
where w represents the optimal weights and θ is
oranges opportunity sets, and mean–variance opti-
the risk aversion coefficient. Notice that we have
mization with different inherent constraints that
removed the subscripts because the solution can be can lead to differences—differences that should not
calculated using either asset classes or risk factors. be mistaken for inherent superiority of risk-factor-
After calculating the optimal weights in one space, based asset allocation.
we can move to the other space using

w a = ( L ') −1 w f (4a) The Messy Real World

Having mathematically proven in an idealized
and setting that neither approach offers an inherent
w f = L ' w a . (4b) advantage, we turn to the far messier real world,
for which an empirical example is necessary. Let us
Using Equation 4a, to show that the return and start by defining two opportunity sets that more or
risk of both portfolios are the same is a straightfor- less cover the same underlying universe of individ-
ward exercise. In Appendix A, we derive Equation ual securities; our set is U.S. centric because long-
2a, demonstrate the link between Equations 3 and 4a, term data for the U.S. markets are readily available.
and verify that the returns and risks of asset-class- We include cash as part of the factor opportunity
and risk-factor-based optimal portfolios are the same. set—even though it is not a risk factor—for reasons
We admit that this example is somewhat artifi- that will be obvious later (Exhibit 1).
cial, but it mathematically demonstrates that there The assumptions of our stylized example
is no gain in efficiency from performing the uncon- from the previous section clearly do not apply
strained optimization in risk-factor space rather than here because there is one less factor than there are
in asset-class space. If there is a one-to-one mapping asset classes; hence, there cannot be a one-to-one
between asset-class and risk-factor returns, then mapping between the two. Nevertheless, neither
there is no reason to expect that one set contains any opportunity set would appear to have an inher-
less information than the other, as our calculations ent advantage obtained by including an addi-
indeed confirm. This result is not new. Technically, it tional uncorrelated asset class or risk factor that is
is known as the rotational indeterminacy of factors, unavailable to the other opportunity set.
which means that factors are determined up to a In our asset-class opportunity set, we subdi-
linear transformation. It is, nonetheless, a reminder vide the universe of U.S. stocks along the dimen-
that there is no obvious reason why risk factors sions of size and valuation. On the fixed-income

Exhibit 1.  Opportunity Sets

Asset Classes Proxy Risk Factors Proxy
A. Equity oriented
U.S. large value Russell 1000 Value Index Market Russell 3000 Index – Citigroup 3-month
Treasury bills
U.S. large growth Russell 1000 Growth Index Size Russell 2000 Index – Russell 1000 Index
U.S. small value Russell 2000 Value Index Valuation Russell 3000 Value Index – Russell 3000
Growth Index
U.S. small growth Russell 2000 Growth Index

B. Fixed-income oriented
U.S. mortgage backed Mortgage spread Barclays U.S. MBS – Barclays U.S.
Treasury intermediate
U.S. Treasuries Term spread Barclays U.S. Treasury 20+ year –
(duration) Citigroup 3-month Treasury bills
U.S. credit Credit spread Barclays U.S. credit – Barclays U.S.
Treasury
C. Cash
Cash Citigroup 3-month Treasury bills

22 www.cfapubs.org ©2013 CFA Institute

FAJ_MayJune2013.indb 22 5/9/2013 1:29:16 PM

 Factor-Based Asset Allocation vs. Asset-Class-Based Asset Allocation

side, we include the three biggest components of long in both the Russell 2000 Index and the Russell
the U.S. investment-grade bond universe, which 3000 Value Index, $50 short in both the Russell 1000
are Treasuries, mortgages, and corporate credit. In Index and the Russell 3000 Growth Index, and $100
the risk-factor set, on the equity side, we include long in cash—which amounts to a $100 long posi-
the equity premium (defined as the difference tion. Henceforth, we use this interpretation of fac-
between the broad equity market and cash) and the tor exposures.
size and valuation premiums (defined as the differ- A thought experiment may be useful here as
ences between small and large stocks and between an analogy of how we think about our factors.
value and growth stocks, respectively). On the Imagine that an exchange-traded fund (ETF) pro-
fixed-income side, we include term (duration), vider has created a set of ETFs that captures the
mortgage, and credit spreads. The term spread factors we have just described. Ignoring fees and
is defined as the difference between a 20-year (or tracking errors for the sake of simplicity, the return
longer) Treasury and cash. The mortgage spread is of each ETF would equal the return of cash plus the
defined as the difference between mortgages and return of the factor, or risk premium, in question.
intermediate-term Treasuries because the dura- A portfolio of such ETFs would be subject to the
tion of the mortgage universe typically approxi- same constraint that a typical long-only portfo-
mates that of intermediate Treasuries. Finally, the lio is subject to. In particular, this means that the
credit spread is equal to the difference between the sum of all the investments would add up to 100%,
returns of the credit and Treasury indices. which, in turn—because of how the factors were
Because all the risk factors are derived series constructed—means that the sum of all the long
obtained by subtracting one series from another, positions in the assets that make up the factors can-
they are zero-dollar investments. Interpreting what not exceed 100%. This is how we interpret a factor-
it means to allocate to a zero-dollar risk factor is based portfolio.
tricky. Presumably, the allocation is self-financing Table 1 shows annualized total return and
because the negative (or short) position perfectly excess return arithmetic means and standard devia-
finances the positive (or long) position, with both tions based on historical monthly data starting in
positions summing to zero. Because the exposure January 1979 and ending in December 2011. Recall
nets out the return due to the risk-free rate for each that the asset-class indices used to construct the
factor, the way to think about it is that holding risk-factor series are identified in Exhibit 1.
each factor is equivalent to holding a zero-dollar Table 2 shows the correlations. Note that, con-
position in the factor itself plus a 100% position in sistent with authors who proclaim the superior-
cash. For example, a $100 portfolio that allocates ity of risk-factor allocations, the average pairwise
50% ($50) to the size premium and 50% ($50) to correlation for the risk factors (without cash) is
the valuation premium is equivalent to being $50 considerably lower (0.06) than that for the asset

Table 1.  Historical Returns and Standard Deviations, January 1979–December 2011

Total Return Excess Return
Standard Standard
Asset Class/Factor Arithmetic Mean Deviation Arithmetic Mean Deviation
Barclays U.S. Treasury TR 8.59%* 6.12%* 2.96% 5.79%
Barclays U.S. credit TR 9.05* 7.87* 3.40 7.50
Barclays U.S. MBS TR 8.82* 7.40* 3.18 7.02
Russell 1000 Growth TR 12.22* 19.92* 6.42 18.97
Russell 1000 Value TR 13.04* 16.97* 7.20 16.15
Russell 2000 Growth TR 12.18* 26.47* 6.38 25.24
Russell 2000 Value TR 14.86* 20.40* 8.93 19.44
Market 12.70** 17.85** 6.88 17.00
Size 6.29** 11.02** 0.77 10.49
Valuation 6.44** 10.24** 0.91 9.68
Term spread (duration) 10.64** 13.44** 4.91 12.83
Credit spread 5.93** 3.96** 0.42 3.73
Mortgage spread 5.71** 3.97** 0.22 3.66
Citigroup 3-month Treasury bills 5.48 1.04 0.00 0.00
Note: TR stands for total return.
*Historical capital market assumptions for the asset classes being optimized.
**Historical capital market assumptions for the risk factors being optimized.

May/June 2013 www.cfapubs.org 23

FAJ_MayJune2013.indb 23 5/9/2013 1:29:16 PM

 Financial Analysts Journal

Table 2.  Historical Correlations

Barclays Barclays Barclays Russell Russell Russell Russell
Barclays Barclays Barclays Russell Russell Russell Russell U.S. U.S. U.S. 1000 1000 2000 2000
U.S. U.S. U.S. 1000 1000 2000 2000 Treasury Credit MBS Growth Value Growth Value
Treasury Credit MBS Growth Value Growth Value TR TR TR TR TR TR TR
TR TR TR TR TR TR TR (excess) (excess) (excess) (excess) (excess) (excess) (excess)
Barclays U.S.
Treasury TR 1.000* 0.863* 0.846* 0.072* 0.112* –0.027* 0.030* 0.985 0.844 0.831 0.064 0.104 –0.033 0.023
Barclays U.S.
credit TR 0.863* 1.000* 0.881* 0.271* 0.326* 0.188* 0.255* 0.860 0.991 0.877 0.268 0.323 0.186 0.252
Barclays U.S.
MBS TR 0.846* 0.881* 1.000* 0.172* 0.210* 0.090* 0.133* 0.835 0.868 0.989 0.166 0.204 0.086 0.127
Russell 1000
Growth TR 0.072* 0.271* 0.172* 1.000* 0.843* 0.861* 0.755* 0.067 0.267 0.168 0.998 0.842 0.859 0.753
Russell 1000
Value TR 0.112* 0.326* 0.210* 0.843* 1.000* 0.738* 0.852* 0.105 0.319 0.205 0.840 0.998 0.736 0.849
Russell 2000
Growth TR –0.027* 0.188* 0.090* 0.861* 0.738* 1.000* 0.878* –0.027 0.188 0.090 0.861 0.739 0.999 0.878
Russell 2000
Value TR 0.030* 0.255* 0.133* 0.755* 0.852* 0.878* 1.000* 0.027 0.252 0.130 0.754 0.851 0.877 0.998
Barclays U.S.
Treasury TR
(excess) 0.985* 0.860* 0.835* 0.067* 0.105* –0.027* 0.027* 1.000 0.865 0.845 0.070 0.108 –0.025 0.029
Barclays U.S.
credit TR
(excess) 0.844 0.991 0.866 0.267 0.319 0.188 0.252 0.865 1.000 0.881 0.271 0.325 0.191 0.256
Barclays U.S.
MBS TR
(excess) 0.831 0.877 0.989 0.168 0.205 0.090 0.130 0.845 0.881 1.000 0.171 0.208 0.092 0.133
Russell 1000
Growth TR
(excess) 0.064 0.268 0.166 0.998 0.840 0.861 0.754 0.070 0.271 0.171 1.000 0.843 0.862 0.755
Russell 1000
Value TR
(excess) 0.104 0.323 0.204 0.842 0.998 0.739 0.851 0.108 0.325 0.208 0.843 1.000 0.739 0.852
Russell 2000
Growth TR
(excess) –0.033 0.186 0.086 0.859 0.736 0.999 0.877 –0.025 0.191 0.092 0.862 0.739 1.000 0.879
Russell 2000
Value TR
(excess) 0.023 0.252 0.127 0.753 0.849 0.878 0.998 0.029 0.256 0.133 0.755 0.852 0.879 1.000
Market 0.078 0.301 0.186 0.966 0.945 0.868 0.856 0.083 0.304 0.190 0.967 0.947 0.868 0.857
Size –0.149 –0.038 –0.085 0.156 0.117 0.620 0.590 –0.142 –0.032 –0.079 0.158 0.120 0.621 0.592
Valuation 0.050 0.011 0.013 –0.545 –0.009 –0.469 –0.079 0.047 0.008 0.010 –0.546 –0.010 –0.469 –0.080
Term spread 0.919 0.785 0.730 0.059 0.088 –0.024 0.012 0.937 0.793 0.743 0.063 0.092 –0.021 0.016
Credit spread 0.159 0.636 0.430 0.420 0.467 0.410 0.453 0.176 0.647 0.444 0.425 0.473 0.413 0.458
Mortgage
spread 0.031 0.310 0.559 0.211 0.220 0.211 0.202 0.034 0.311 0.563 0.212 0.221 0.211 0.202
Citigroup
3-month
Treasury
bills 0.128 0.054 0.100 0.029 0.046 0.000 0.020 –0.047 –0.082 –0.045 –0.027 –0.020 –0.042 –0.036
Market (plus
cash) 0.086 0.305 0.192 0.967 0.947 0.868 0.857 0.080 0.299 0.187 0.965 0.946 0.866 0.855
Size (plus
cash) –0.137 –0.033 –0.076 0.159 0.122 0.619 0.591 –0.147 –0.040 –0.084 0.156 0.118 0.617 0.588
Valuation
(plus cash) 0.063 0.017 0.023 –0.538 –0.004 –0.465 –0.076 0.042 0.000 0.005 –0.545 –0.012 –0.470 –0.083
Term spread
(plus cash) 0.931 0.791 0.740 0.061 0.092 –0.024 0.014 0.936 0.788 0.741 0.061 0.091 –0.025 0.013
Credit spread
(plus cash) 0.191 0.643 0.451 0.423 0.474 0.405 0.453 0.161 0.617 0.427 0.414 0.462 0.397 0.443
Mortgage
spread (plus
cash) 0.064 0.315 0.568 0.212 0.225 0.204 0.200 0.020 0.280 0.533 0.198 0.209 0.193 0.186

*Historical capital market assumptions for the asset classes being optimized.
**Historical capital market assumptions for the risk factors being optimized.
(continued)

24 www.cfapubs.org ©2013 CFA Institute

FAJ_MayJune2013.indb 24 5/9/2013 1:29:17 PM

 Factor-Based Asset Allocation vs. Asset-Class-Based Asset Allocation

Table 2.  Historical Correlations (continued)

Citigroup Term
3-Month Market Size Spread Credit Mortgage
Term Credit Mortgage Treasury (plus (plus Valuation (plus Spread Spread
Market Size Valuation Spread Spread Spread Bills cash) cash) (plus cash) cash) (plus cash) (plus cash)
Barclays U.S.
Treasury TR 0.078 –0.149 0.050 0.919 0.159 0.031 0.128 0.086 –0.137 0.063 0.931 0.191 0.064
Barclays U.S.
credit TR 0.301 –0.038 0.011 0.785 0.636 0.310 0.054 0.305 –0.033 0.017 0.791 0.643 0.315
Barclays U.S.
MBS TR 0.186 –0.085 0.013 0.730 0.430 0.559 0.100 0.192 –0.076 0.023 0.740 0.451 0.568
Russell 1000
Growth TR 0.966 0.156 –0.545 0.059 0.420 0.211 0.029 0.967 0.159 –0.538 0.061 0.423 0.212
Russell 1000
Value TR 0.945 0.117 –0.009 0.088 0.467 0.220 0.046 0.947 0.122 –0.004 0.092 0.474 0.225
Russell 2000
Growth TR 0.868 0.620 –0.469 –0.024 0.410 0.211 0.000 0.868 0.619 –0.465 –0.024 0.405 0.204
Russell 2000
Value TR 0.856 0.590 –0.079 0.012 0.453 0.202 0.020 0.857 0.591 –0.076 0.014 0.453 0.200
Barclays U.S.
Treasury TR
(excess) 0.083 –0.142 0.047 0.937 0.176 0.034 –0.047 0.080 –0.147 0.042 0.936 0.161 0.020
Barclays U.S.
credit TR
(excess) 0.304 –0.032 0.008 0.793 0.647 0.311 –0.082 0.299 –0.040 0.000 0.788 0.617 0.280
Barclays U.S.
MBS TR
(excess) 0.190 –0.079 0.010 0.743 0.444 0.563 –0.045 0.187 –0.084 0.005 0.741 0.427 0.533
Russell 1000
Growth TR
(excess) 0.967 0.158 –0.546 0.063 0.425 0.212 –0.027 0.965 0.156 –0.545 0.061 0.414 0.198
Russell 1000
Value TR
(excess) 0.947 0.120 –0.010 0.092 0.473 0.221 –0.020 0.946 0.118 –0.012 0.091 0.462 0.209
Russell 2000
Growth TR
(excess) 0.868 0.621 –0.469 –0.021 0.413 0.211 –0.042 0.866 0.617 –0.470 –0.025 0.397 0.193
Russell 2000
Value TR
(excess) 0.857 0.592 –0.080 0.016 0.458 0.202 –0.036 0.855 0.588 –0.083 0.013 0.443 0.186
Market 1.000 0.205 –0.324 0.072 0.470 0.227 –0.026 0.998 0.203 –0.325 0.070 0.457 0.213
Size 0.205 1.000 –0.131 –0.131 0.153 0.072 –0.044 0.203 0.995 –0.135 –0.135 0.140 0.058
Valuation –0.324 –0.131 1.000 0.034 –0.055 –0.054 0.020 –0.323 –0.129 0.995 0.035 –0.049 –0.047
Term spread 0.072 –0.131 0.034 1.000 0.130 –0.059 –0.069 0.068 –0.138 0.026 0.997 0.110 –0.076
Credit spread 0.470 0.153 –0.055 0.130 1.000 0.559 –0.089 0.464 0.145 –0.063 0.123 0.965 0.517
Mortgage
spread 0.227 0.072 –0.054 –0.059 0.559 1.000 –0.012 0.226 0.070 –0.054 –0.061 0.549 0.965
Citigroup
3-month
Treasury
bills –0.026 –0.044 0.020 –0.069 –0.089 –0.012 1.000 0.036 0.052 0.123 0.012 0.176 0.251
Market (plus
cash) 0.998 0.203 –0.323 0.068 0.464 0.226 0.036 1.000** 0.206** –0.317** 0.071** 0.468** 0.228**
Size (plus
cash) 0.203 0.995 –0.129 –0.138 0.145 0.070 0.052 0.206** 1.000** –0.123** –0.134** 0.157** 0.082**
Valuation
(plus cash) –0.325 –0.135 0.995 0.026 –0.063 –0.054 0.123 –0.317** –0.123** 1.000** 0.036** –0.030** –0.020**
Term spread
(plus cash) 0.070 –0.135 0.035 0.997 0.123 –0.061 0.012 0.071** –0.134** 0.036** 1.000** 0.125** –0.055**
Credit spread
(plus cash) 0.457 0.140 –0.049 0.110 0.965 0.549 0.176 0.468** 0.157** –0.030** 0.125** 1.000** 0.577**
Mortgage
spread (plus
cash) 0.213 0.058 –0.047 –0.076 0.517 0.965 0.251 0.228** 0.082** –0.020** –0.055** 0.577** 1.000**

May/June 2013 www.cfapubs.org 25

FAJ_MayJune2013.indb 25 5/9/2013 1:29:18 PM

 Financial Analysts Journal

classes (0.38). This average correlation difference asset allocations; there is nothing obvious about
should be intuitive. The asset classes are all part of it, and it could go either way, depending on the
the “market” portfolio and, hence, carry with them returns and covariances of the opportunity set in
an element of overall market beta. By construction, question. It is in this sense that we claimed at the
with the exception of the market risk factor, the risk beginning of this section that the superiority (or
factors do not share this common exposure to the otherwise) of a risk-factor-based allocation is an
market. This fact is directly related to a key point empirical question.
from the recent “importance of asset allocation lit- Figure 1 shows the two historical efficient fron-
erature”; most notably, Xiong, Ibbotson, Idzorek, tiers obtained from the mean–variance optimiza-
and Chen (2010) emphasized that one must be tions described above.11 In Figure 1, the empirical
extremely careful when comparing items that question is thus answered for this set of risk fac-
include the market factor with those that do not. tors and this particular time period. In contrast to
In our first optimization comparison, we the current risk-factor papers that highlight much
constrain the weights of the asset classes and of shorter time periods in which equities have gener-
the risk factors to be nonnegative and the alloca- ally performed poorly, the efficient frontier based
tions to sum to 1. The values from Table 1 that are on asset classes dominates the efficient frontier
marked with an asterisk are the historical capital based on risk factors.
market assumptions for the asset classes that are
being optimized, and the values marked with two Figure 1.  Long-Only Constraints: Asset Classes
asterisks are the historical capital market assump- vs. Classic Risk Factors, January
tions for the risk factors that are being optimized. 1979–December 2011
Recall from our zero-dollar investment discussion
that when a dollar is invested in the risk factor, it is Arithmetic Annualized Return (%)
really being invested in Treasury bills plus the zero- 15
dollar (self-financing) risk factor. In Tables 1 and 2, Asset Classes (Long Only)
cash (T-bills) is not marked with an asterisk, but it
is included in both opportunity sets. Factors
As we proceed with our first optimization 10
comparison, let us pause here and reflect on what
we are actually comparing. From a mathemati-
cal point of view, in the asset-class space, our fea-
sible set for weights is a seven-dimensional subset 5
(because of the one linear constraint of unity) of the
eight-dimensional asset-class space, subject to the
constraint of the weights being positive. The risk-
factor space is somewhat more complicated. But for 0
our purposes, what matters is that the feasible set 0 5 10 15 20 25
for risk-factor weights, which are a combination of Annualized Standard Deviation (%)
some asset classes being offset by other asset classes,
is a different subset (fewer dimensions because of
more constraints) of the eight-dimensional asset- This result may look like a decisive score for
class space.10 asset-class-based asset allocation, but that conclu-
The intuition we want to focus on is that sion is wrong. We reran this optimization com-
although the risk-factor weights’ feasible set has parison using a variety of historical capital market
fewer dimensions than its asset-class counterpart, assumptions based on various shorter time peri-
which suggests that it may be slightly handicapped, ods; sometimes, asset classes worked better, and
by construction, the risk factors include underly-
in some cases, risk factors worked better. By cherry
ing asset-class exposures of –100%—something
picking a particular historical time period, almost
that is not allowed in our asset-class set. From this
any desired result can be found. We have omitted
perspective, the “long-only” constraint imposed
these results, except for the somewhat interest-
during the optimization of the risk-factor oppor-
ing results obtained from optimizations based on
tunity set may, in fact, be less constrained and the
capital market assumptions from the most recent 10
“long-only” label for this comparison may be a
years of data (January 2002–December 2011). In this
bit of a misnomer. A priori, one cannot say which
example, neither asset classes nor risk factors con-
opportunity set coupled with its respective explicit
sistently dominated along the whole risk spectrum.
and implicit constraints will produce more efficient

26 www.cfapubs.org ©2013 CFA Institute

FAJ_MayJune2013.indb 26 5/9/2013 1:29:18 PM

 Factor-Based Asset Allocation vs. Asset-Class-Based Asset Allocation

An asset-based approach would have produced stop here? Could we not create an even bigger or
more efficient portfolios for lower risk levels, but better size premium by going short 10,000% large
the opposite is true for riskier portfolios, as shown cap and long 10,000% small cap?
in Figure 2. Moving to another simple risk factor, valua-
tion, which we constructed by shorting growth
stocks (Russell 3000 Growth) and going long value
Figure 2.  Long-Only Constraints: Asset Classes
stocks (Russell 3000 Value), we have observed
vs. Classic Risk Factors, January
2002–December 2011 policy portfolios that strongly favor value stocks
but none that completely ignore growth stocks.
Arithmetic Annualized Return (%) Investing in the valuation risk factor would sug-
15 gest an extremely high level of conviction in the
value premium and a level of conviction that we
have never observed in a traditional asset-class
policy portfolio. Our point is that investors consid-
10 Factors ering the use of a factor-based approach need to be
aware of the rather extreme positions embedded in
the factors. Finally, one can certainly construct an
Asset Classes (Long Only)
asset allocation using asset classes that is designed
5 to capture many of the risk factors.

Conclusion
We mathematically proved that in an idealized
0
world in which risk-factor returns are completely
0 5 10 15 20 25
explained by asset-class returns and vice versa, nei-
Annualized Standard Deviation (%)
ther approach is inherently superior to the other.
Next, in a series of real-world optimizations based
on historical data, we demonstrated that either
We hope we have clearly illustrated that there approach may be superior over a given time period
is nothing obvious about the superiority of asset but it really comes down to the specifics of the com-
allocation based on risk factors. Whether one uses parison and of the composition of the factors. In
historical data, as we have here, or future capital particular, in some cases, the apparent superiority
market assumptions, the realized or expected effi- of the risk factors is a simple result of the fact that
ciency needs to be checked empirically (by which the risk factors are, in a guise, a set of asset classes
we mean the numbers must be run) to determine with the long-only constraint removed. If one truly
which approach may be more appropriate. Neither creates an even playing field, there is no gain in
is simply superior. efficiency, which is hardly surprising. If risk factors
Of course, practical implementation is a whole could be approximated by asset classes, it would
different story. Even if one were able to convince be quite odd if that risk-factor approximation con-
himself that for a given level of risk, a risk-factor- tained more useful information than the totality of
like portfolio is superior, very few institutional information contained in the asset classes that are
investors would be willing to take on the extreme used to approximate the risk factors themselves.
positions implied by most risk factors. For exam- There are a couple practical issues associ-
ple, let us take a simple risk factor—size, which ated with risk-factor-based asset allocation. First,
we constructed earlier by shorting large-cap stocks remember that it would be impossible for the
(Russell 1000) and going long on small-cap stocks world to completely embrace risk-factor-based
(Russell 2000). In the U.S. equity component of a asset allocation because it implies a weighting
strategic asset allocation, a policy portfolio that was scheme that is not macro-consistent. This observa-
allocated equally between large cap and small cap tion puts the risk-factor-based asset allocation on
would be viewed as dramatically small-cap over- the level of a strategy rather than a theoretically
weighted from the typical market-capitalization consistent model. A strategy that is not feasible for
viewpoint. In fact, it is almost impossible to imag- all may nonetheless be feasible for some, and just as
ine a policy portfolio with a U.S. equity split of 0% any strategy, it may or may not result in a more effi-
large cap and 100% small cap, let alone –100% large cient portfolio for a given time period. Second, the
cap and 100% small cap, which is implicit in the largest hurdle to adopting a risk-factor-based asset
typical size factor construction. Furthermore, why allocation is that most institutional investors are

May/June 2013 www.cfapubs.org 27

FAJ_MayJune2013.indb 27 5/9/2013 1:29:19 PM

 Financial Analysts Journal

simply uncomfortable with the extreme positions 1 −1 1

( )
−1
involving leverage that are implied by risk factors. wa =  a = L f L ' L f
θ a θ
Finally, there is nothing wrong with risk-factor- 1 −1
based asset allocation, but it is not a magic bullet that = ( L ' ) −f 1L−1L f
θ
will automatically lead to asset allocations that will
1 −1
dominate asset allocations based on asset classes. = ( L ' ) −f 1 f
θ
−1
= (L ') wf.
Helpful comments were provided by Wai Lee and Philip
Straehl. Also, we can verify that returns of the asset-
class-based and factor-based optimal portfolios are
the same:
This article qualifies for 0.5 CE credit.
−1
w 'a ra = ( L ' ) w f  ' Lr f
 
Appendix A. −1
= w ' f ( L ' )  ' Lr f
 
Following is the derivation of Equation 2a, which −1
= w ' f ( L '' ) Lr f
demonstrates that the asset-class covariance matrix
can be derived from the risk-factor covariance −1
= w ' f (L) Lr f
matrix:
= w ' f rf .
a = E ( ra − a ) ( ra − a ) '  Finally, we show that the risks of the asset-
class-based and factor-based optimal portfolios are
{ (
  ) (
= E L r f −  f  L r f −  f  '
 ) } the same:
= E { L ( r f −  f ) ( r f −  f ) ' L ' }
  

−1
w 'a a w a = ( L ' ) w f ( f  )
 ' L L ' ( L ' )−1 w 
f 

 (
= LE  r f −  f )( )
r f −  f ' L '
 −1 −1
= w ' f ( L ' )  ' L f L ' ( L ' ) w f
 
= L f L '.
−1 −1
= w ' f ( L '' )  L f L ' ( L ' ) w f
Next, we demonstrate the link between  
−1 −1
Equations 3 and 4a and that weights in either asset- = w ' f ( L ) L   f  L ' ( L ' )  w f
class space or risk-factor space can be transformed    
into equivalent weights in the other space: = w'f f w f .

Notes
1. Waring and Siegel (2003) provided an excellent overview of 6. The factors were equity market, small-cap premium, change
this two-phase process. in 10-year U.S. Treasury and credit spreads, and three trend-
2. The number of correlations that must be estimated is N(N following strategies for currencies, bonds, and commodities.
– 1)/2, where N equals the number of securities. Thus, for 7. See, for example, Litterman (2005).
a portfolio of 1,000 securities, 495,000 correlations must be 8. See Ibbotson, Chen, Kim, and Hu (2013).
estimated. Using time series requires a minimum of 1,001 9. A more appropriate value proposition is to claim superior
observations, which corresponds to roughly 4 years of daily skill at forecasting the return of risk factors, although the
return data, 83 years of monthly return data, or 250 years of investing public is rightfully skeptical of this more tradi-
tional value proposition claim.
quarterly return data.
10. Note that, for practical purposes, the asset classes that are
3. Common examples of companies that offer structural
used in the construction of the factors can be expressed as
multifactor risk models are MSCI (Barra) and Northfield combinations of the asset classes from the asset-class space;
Information Services. thus, for example, the Russell 3000 can be expressed as the
4. See Sharpe (1964); Lintner (1965); Mossin (1966); and Treynor weighted sum of the four Russell indices from the asset-class
(1961, 1962). space.
5. Cliff Asness also contributed significantly to the develop- 11. All the optimizations are performed in total return space,
ment of momentum as a factor. with the cash return included.

28 www.cfapubs.org ©2013 CFA Institute

FAJ_MayJune2013.indb 28 5/9/2013 1:29:25 PM

 Factor-Based Asset Allocation vs. Asset-Class-Based Asset Allocation

References
Bender, Jennifer, Remy Briand, Frank Nielsen, and Dan Litterman, Robert. 2005. “Beyond Active Alpha.” Open letter to
Stefek. 2010. “Portfolio of Risk Premia: A New Approach to investors, Goldman Sachs.
Diversification.” Journal of Portfolio Management, vol. 36, no. 2
(Winter):17–25. Markowitz, Harry. 1952. “Portfolio Selection.” Journal of Finance,
vol. 7, no. 1 (March):77–91.
Carhart, Mark M. 1997. “On Persistence in Mutual Fund
Performance.” Journal of Finance, vol. 52, no. 1 (March):57–82. Mossin, Jan. 1966. “Equilibrium in a Capital Asset Market.”
Econometrica, vol. 34, no. 4 (October):768–783.
Clarke, Roger G., Harindra de Silva, and Robert Murdock. 2005.
“A Factor Approach to Asset Allocation.” Journal of Portfolio Page, Sébastien, and Mark A. Taborsky. 2011. “The Myth of
Management, vol. 32, no. 1 (Fall):10–21. Diversification: Risk Factors vs. Asset Classes.” Journal of
Portfolio Management, vol. 37, no. 4 (Summer):1–2.
Fama, Eugene F., and Kenneth R. French. 1992. “The
Cross-Section of Expected Stock Returns.” Journal of Finance, vol. Ross, Stephen A. 1976. “The Arbitrage Theory of Capital
47, no. 2 (June):427–465. Asset Pricing.” Journal of Economic Theory, vol. 13, no. 3
(December):341–360.
Fung, William, and David A. Hsieh. 2004. “Hedge Fund
Benchmarks: A Risk-Based Approach.” Financial Analysts Sharpe, William F. 1964. “Capital Asset Prices: A Theory of
Journal, vol. 60, no. 5 (September/October):65–80. Market Equilibrium under Conditions of Risk.” Journal of
Finance, vol. 19, no. 3 (September):425–442.
Grinold, Richard C., and Ronald N. Kahn. 2000. Active Portfolio
Management. 2nd ed. New York: McGraw-Hill. Treynor, Jack L. 1961. “Market Value, Time, and Risk.”
Unpublished manuscript (8 August).
Ibbotson, Roger G., Zhiwu Chen, Daniel Y.-J. Kim, and Wendy Y.
Hu. 2013. “Liquidity as an Investment Style.” Financial Analysts ———. 1962. “Toward a Theory of Market Value of Risky
Journal, vol. 69, no. 3 (May/June):30–44. Assets.” Unpublished manuscript (Fall).

Jegadeesh, Narisimhan, and Sheridan Titman. 1993. “Returns Waring, M. Barton, and Laurence B. Siegel. 2003. “Understanding
to Buying Winners and Selling Losers: Implications for Stock Active Management.” Investment Insights, vol. 6, no. 1
Market Efficiency.” Journal of Finance, vol. 48, no. 1 (March):65–91. (April):1–42.

Jensen, Greg, Noah Yechiely, and Jason Rotenberg. 2005. “Hedge Xiong, James X., Roger G. Ibbotson, Thomas M. Idzorek, and
Funds Selling Beta as Alpha.” Bridgewater Daily Observations Peng Chen. 2010. “The Equal Importance of Asset Allocation
(24 May). and Active Management.” Financial Analysts Journal, vol. 66,
no. 2 (March/April):22–30.
Lintner, John. 1965. “The Valuation of Risk Assets and the
Selection of Risky Investments in Stock Portfolios and Capital
Budgets.” Review of Economics and Statistics, vol. 47, no. 1
(February):13–37.

May/June 2013 www.cfapubs.org 29

FAJ_MayJune2013.indb 29 5/9/2013 1:29:25 PM