Forecasting

Factor Returns
Chief Investment Officer, Executive Summary
Two Sigma Advisers
Geoff Duncombe In our recent paper, Introducing the Two Sigma Factor Lens, we proposed a parsimonious set of
actionable factors that collectively explains the majority of risk in institutional portfolios.1 This paper
takes the next step, proposing a methodology to estimate the long-term return premium associated
Head of Client with each of these factors. This paper introduces a handful of innovations intended to improve the
Solutions Research accuracy of our long-term return forecasts:
Mike Nigro
• We use new asset class return proxies to extend our analysis much further back than
the daily return histories of most modern indices.
• We separate the most heterogeneous of the prior paper’s factors, Commodities, into
Head of Thematic Research
six sector-based factors for which the long-term premia are individually estimated.
Bradley Kay • We apply (what we believe to be) common sense adjustments to long-term histories —
slightly overweighting recent returns and applying empirically-based shrinkage across
the observed historical Sharpe ratios to generate our forward-looking estimates of
each factor’s premium.

Ultimately, this paper identifies five unique and orthogonal factors across asset classes that we
believe carry a positive historical return premium: Equity, Interest Rates, Credit, Energy, and
Industrial Metals. Our estimates for the long-term Sharpe ratios of these factors (and others in the
Two Sigma Factor Lens) may be found in Exhibit 1. We believe these five compensated factors could
collectively form the basis of an asset allocation strategy with substantial investment capacity.

1For more detail on the construction of these factors and the principles by which they were derived,
please see “Introducing the Two Sigma Factor Lens”, by Duncombe and Kay (2018).

Two Sigma Investments | Two Sigma Advisers The views expressed in this paper reflect those of the authors and
are not necessarily the views of Two Sigma Investments, LP or
www.twosigma.com any of its affiliates. This paper is distributed for informational and
NEW YORK HOUSTON LONDON HONG KONG TOKYO
educational purposes only. Please see Important Disclaimer and
Disclosure Information at the end of this document.
Table of Contents

I. Introduction 3

II. Value of Long Histories: Looking Past the Interest Rates Mountain 5

III. Equities: The Long-Term Returns Champ 6

IV. Credit: From Many Assets, One Factor 8

V. Commodities: Rorschach’s Factor Test 11

VI. Emerging Markets: Risk without Reward? 15

VII. From History to Expectations 17

VIII. Conclusion: Forward-Looking Returns by Factor 18

Appendicies 20

Bibliography 27

1
Exhibit 1 | Factor Descriptions and Expected Premia as of Dec 31, 2018

ESTIMATED
ORTHOGONAL
FACTOR
FACTOR SHARPE RATIO DESCRIPTION

Interest Rates 0.25 Exposure to the time value of money

(inflation risk and future interest rate
changes)
Equity 0.35 Exposure to the long-term economic
growth and profitability of companies
Core Macro

Credit 0.21 Exposure to default and illiquidity

risks specific to developed market
corporate bonds
Energy 0.23 Exposure to changes in prices for hard
assets and any embedded
Industrial 0.19 Exposure to changes in prices for hard
Metals assets
adfja and any embedded futures
premium
Other 0.002 v

Commodities ada
Emerging 0.00 Exposure to the sovereign and
3

Markets economic risks of emerging markets

relative to developed markets
Secondary Macro

Foreign Currency 0.004 Exposure to moves in developed

market currency values versus the
portfolio’s local currency
Local Equity 0.004 Exposure to the “home bias” of
holding disproportionate risk in the
local equity market
Local Inflation ---5 Exposure to inflation-linked rates
relative to fixed nominal rates within
the local currency area

2 Other commodity sectors are found to have no actionable and statistically significant orthogonal factor premium, thus their forward-looking
estimates are fixed to 0.0. See Section V for more details.
3 Emerging Markets assets are found to have no statistically significant orthogonal factor premium, thus the EM factor forward-looking estimate
is fixed to 0.0. See Section VI for more details.
4 We believe the Foreign Currency and Local Equity factors should provide no orthogonal factor premium, thus their forward-looking estimates
are fixed to 0.0. See discussion in Introduction for more details.
5 Although part of the Two Sigma Factor Lens where supported, we can observe only very short histories for the Local Inflation factor due to
the recent introduction of inflation-linked sovereign bonds as an asset class. Hence we do not estimate a long-term expected premium in this
paper.

2
I. Introduction
Our prior paper proposed the Two Sigma Factor Lens, a parsimonious set of factors derived
from the performance of major asset classes that collectively explains the majority of risk
in institutional portfolios. While the academic and practitioner literature includes many fine
efforts to identify risk factors across asset classes,6 we believe our approach contained two key
differentiators.

First, we built our factors directly from traded assets rather than from economic variables like
growth or inflation, which can have a surprisingly tenuous link with the assets themselves.7 The
factors were also each derived from individual asset classes rather than statistically-determined
combinations such as principal components, which eases the interpretability of which assets
in a portfolio are contributing to individual factor exposures and can make it easier to translate
between a factor-based analysis and asset allocations.

Second, we asserted a hierarchy across the factors beginning with the most liquid asset classes,
and statistically separated (i.e., orthogonalized) the unique returns for each set of less liquid
assets from its exposure to the higher-order factors. This helps aggregate common risk factors
across multiple asset classes to better identify concentrations of exposure to, say, the Equity
factor from many different assets also sensitive to economic growth and investor risk aversion.
For investors seeking to implement a factor-based portfolio, this approach can also preferentially
tilt optimal factor exposures toward the lowest-cost factors such as Equity and Interest Rates,8
with factors from less-liquid assets only selected for an optimized portfolio if they appear to
provide significant diversifying return benefits.

A factor lens will suffice to describe a portfolio, providing insight on the attribution of its
historical risk and returns. Choosing a desired allocation of factors or assets is a far trickier issue,
and requires forecasts of risk and return expectations. In this paper, we propose a methodology
using historical data to derive long-term forecasts of the return premia for major asset-class-
based factors. Our methodology for identifying premia-bearing factors and estimating their
respective long-term returns forecast follows five steps:

1. Examine the maximum available return histories, as decades of data provide greater
insight on the likely return premium through different market and economic regimes.
6 Please see the introduction and bibliography of
2. Consolidate individual assets and asset classes wherever feasible, as a single risk factor Duncombe and Kay (2018) for examples of alter-
native approaches to factor identification.
driving multiple assets should carry the same long-term premium regardless of which 7 Ilmanen (2011, Chapter 16) presents correlations
of monthly changes in consensus growth and
individual assets provide a portfolio with exposure to that factor (i.e., there should be no inflation forecasts with a variety of asset and
market factor returns, finding changes in growth
easy arbitrages). expectations have a maximum absolute cor-
relation of 0.45 with listed private equity firms
while changes in inflation expectations have a
3. Estimate the premium of each new factor relative to more liquid factors, to assess maximum absolute correlation of 0.50 with the
S&P GSCI index of commodity futures.
8 We believe these factors may be considered
whether less liquid factors have shown sufficient orthogonal returns to justify their the “lowest cost” in two important but distinct
senses: they have the lowest transaction costs
addition (and inform sizing) in an efficient portfolio. for direct investment as public equities and
developed market sovereign bonds are among
the most liquid marketable assets, and indexed
investments in sovereign bonds and public equi-
ties have among the lowest fees of all managed
investment vehicles.

3
 4. Select factors that exhibit positive long-term premia, backed by empirical and
fundamental evidence. Factors without orthogonal premia may be useful to identify for
risk management purposes, but we believe they should not be included in an efficient
portfolio.

5. Finally, estimate the return premia jointly across all selected factors, overweighting
recent history and shrinking Sharpe ratio estimates toward a reasonable prior to make the
forecasts more robust.

This paper bases its long-term forecasts on historical returns, with only light guidance from
theory, founding its methodology in a belief that human nature is the fundamental force
underlying all risk premia. Whether factors are rational compensation for exposure to “bad
times” or rooted in the common behavioral biases of the marginal human investor, we believe
that typical investors remain as human today as they were in 1900.9 Neither our perceptions of
bad times nor our behavioral reactions are likely to be very different from the investors in past
market cycles, and though the marginal investor may shift over time due to structural changes,
these changes should typically be gradual and minor rather than sudden and cataclysmic.
This inescapably human element of the “price” of risk means that although it may be nearly
impossible to predict next month’s return for a factor, the return over the next few decades
should resemble the price investors demanded for decades past.

We understand that any methodology for forecasting long-term premia, including that proposed
in this paper, will have inevitable flaws. The investor may be tempted to throw up her hands
and assume, for example, that all factors provide equal premia in the face of this uncertainty.
This would even be a reasonably robust assumption,10 and we find in Section VII that shrinking
historically-observed asset class Sharpe ratios toward their cross-sectional average improves
predictions of future premia. However, we find evidence in this paper of significantly differing
premia across our identified factors, including a few that appear to provide no significant long-
term returns. Given that nobody investing for the long run can avoid the implicit inclusion of
some form of forward-looking views in their portfolio, we would rather examine and test each
factor in turn, with more than a century of data behind us, than turn a blind eye to theory and
history.

Before proceeding, we should briefly note a few things this paper does not do. We do not
9 T
 hough we do appreciate the irony of this senti-
provide analysis of returns to the Foreign Currency, Local Equity, or Local Inflation factors, as ment coming from an investment firm employing
almost entirely systematic strategies.
these three factors seem to have neither fundamental nor empirical justification for a premium. 10 R appoport and Nottebohm (2012) analyze sim-
ulated performance of portfolio optimization
Foreign currency exposures across global investors are net-neutral, so static holdings of foreign with forecast uncertainty under a variety of
assumptions, finding that risk parity’s implicit
currency risk should not carry a premium unless they are exposed to systematic style factor assumption of equal Sharpe ratios across asset
classes performs well under conditions of high
risks such as currency carry or momentum.11 For similar reasons, the Local Equity factor, which uncertainty. They also test a methodology
to improve forecasts by combining the equal
captures the orthogonal returns to an investor’s local equity market relative to global equities, Sharpe ratio assumption with a priori estimates
of expected asset class returns.
has a net zero return across all countries in the global portfolio and should not provide a 11 Although Siegel’s paradox suggests that all
investors may have positive excess return ex-
premium beyond exposure to style premia such as cross-country value or momentum. Finally, pectations from holding static foreign currency
exposure due to Jensen’s inequality, Campbell
et al. (2010) show that the expected premium
the Local Inflation factor should theoretically provide a negative return premium on average, as from this mathematical curiosity is negligible
and we will not all get rich by trading currency
exposures with one another.

4
it represents the returns from hedging inflation risk in sovereign bonds by holding securities with
inflation-linked coupon payments rather than their maturity-matched nominal counterparts.

This paper also does not estimate returns for style factors, such as carry strategies or selling
equity volatility, instead focusing on major asset class returns. Styles do play an important part in
understanding the cross-section of returns and risk within asset classes, however, and we plan
to address these in a future report.

The rest of the paper is structured as follows: Sections II through VI provide case studies
analyzing asset class factors for potential risk premia; Section VII illustrates the utility of long-
term historical data and shrinkage toward “average” risk-adjusted returns to prevent overfitting
to the observed history; and Section VIII concludes.

II. Value of Long Histories: Looking Past the Interest

Rates Mountain
Our dive into history begins with the Interest Rates and Equity factors, which are derived from
returns to assets (sovereign bonds and common stock, respectively) that have been trading in
relatively modern forms for over a century in many countries. Recent research has managed to
reconstruct historical returns and pricing data for new markets and much older periods than
previously available, providing useful perspective on whether the commonly-studied asset
returns of recent decades appear truly representative of long-term expectations.

One of the most striking features of these long histories is the “interest rates mountain” of the
1950s through to the present day (see Exhibit 2). Over the course of six decades, long-term bond
yields in most developed markets climbed from the low single digits to peaks around 15%, and
then back down to near zero. This ascent and decline are astonishing in their scale, a treacherous
alpine ridge with sheer drops to each side of 1980. But perhaps the more important part of the
chart is the broad plain of mid single digit rates we see preceding the mountain, extending back to
the pre-Industrial Age origins of modern central banks and sovereign bonds in northern Europe.

This presents a tricky issue for anyone seeking to forecast the Interest Rates premium: 60 years
of data might sound like plenty of historical perspective for any factor, but interest rates appear
to have been in “anomalous” territory for nearly that long. Trickier still, maybe the mountain isn’t
the anomaly after all. The shift by most developed market central banks from the gold standard
to a fiat standard in the mid-20th century (culminating with the breakup of Bretton Woods in
1971) may render any data prior to the mountain obsolete, leaving the mountain as the only
data we have that describes the current fiat regime.

The challenge of identifying the relevant period for estimating the Interest Rates premium puts
in stark relief some of the issues faced in relying on historical data as a lens for the future, and
especially highlights the potential benefit of using longer windows and applying sensible prior
assumptions where possible.

5
Exhibit 2 | Historical Long-Term Bond Yields by Country
Exhibit 2: Historical long-term bond yields for major
developed markets show the pronounced “Interest
Rates Mountain” from the 1950s to today. Previous
to this period, yields tended to hover in the mid-
single-digits outside of extreme periods of sovereign
instability. The shaded region highlights yields after
1970, representing the “modern era” for global
financial markets (see footnote 16 for details). See
Appendix B for data sources and other details.

We ultimately opt for an estimate of the very long-term historical average of the Interest
Rates premium as more representative of our future expectations than the experience of
recent decades. While declining yields have provided a strong tailwind to bond returns since
1980, forward-looking estimates of bond premia have been declining across the world based
on both econometric and survey-based estimation approaches.12 This decline appears tied to 12 The decline of implied long-term bond premia
has been documented by Wright (2011) for
two notable, and likely interlinked, phenomena: inflation uncertainty and volatility appear to G10 sovereign bonds since 1990 (using both
affine curve model estimates and survey-based
have significantly fallen from their 1980 highs, while correlations between stocks and bonds estimates), and by Adrian et al. (2014) in US
Treasuries since 1961 (using affine curve model
have shifted from positive to negative.13 The former suggests a smaller risk premium should be estimates).
13 C ampbell, Pflueger, and Viceira (2018) find
that shocks to US inflation have significantly
priced into long-term bonds due to lower perceived risk of inflation-driven losses, while the decreased in size and changed from positive to
negative correlation with changes in the output
latter suggests that bonds should carry a lower (or even negative) premium from providing a gap since 2000, suggesting that a more stable
macroeconomic environment with better-an-
partial hedge against equity-led asset price shocks. Until we see evidence that expected bond chored inflation expectations can explain
both the lower expected bond premium and
premia have begun to rise from current levels, we choose to base our forecasts upon the lower changing correlation between bonds and eq-
uities in a consumption-based macroeconomic
forward-looking premium estimate derived from very long-run returns.14 model with habit formation. D.E. Shaw (2019)
shows that surveyed inflation expectations
and correlations between bonds and equities
have shifted lower not only in the US, but also
contemporaneously in Japan, Germany, and

III. Equities: The Long-Term Returns Champ

the United Kingdom. The authors also find that
there still appears to be some long-term infla-
tion-related premium priced into US Treasuries
via a positive roll-down yield around 15- to
The value of taking a longer perspective becomes clearer when we add equity markets to 20-year maturities.
14 S ee Appendix B: Exhibits 1 and 14 for more
our historical analysis. To wit, since 1900, risk-adjusted returns on equities are significantly details of the authors’ calculations of the
long-term average returns to the Interest Rates
higher than those on bonds. Yet if we only looked at the five decades since 1970 (i.e., as we factor.
15 C ountries included in analysis since 1970:
rode down the rates mountain and bond prices rose), the historical record would suggest the Australia, Canada, France, Germany, Italy,
Japan, Netherlands, Spain, United Kingdom, and
opposite. Even a question as simple as “which factor has a higher premium” can take over a United States.
16 S everal developments point toward financial
century of data to answer with a reasonable degree of certainty. markets modernizing around 1970, including
but not limited to the end of the Bretton Woods
system in 1971-1973 as many developed
Exhibit 3 shows histograms of the historical risk-adjusted returns by country to equities and markets shifted to floating exchange rates;
the first listing of financial futures (contracts
on foreign currencies) on the International
10-year sovereign bonds over the two different periods. The top panel shows the results for 10 Monetary Market (now part of CME Group)
in 1972; and some of the earliest global stock
major developed markets since 1970.15,16 Note that this corresponds to the grey area in Exhibit market indices being first licensed by Capital
International (now part of MSCI Inc.) in 1969.

6
2, when sovereign yields around the developed world were mostly declining from their peaks.
Over this period, the average 10-year bond outperforms the average equity market on a risk-
adjusted basis.

However, the dataset of historical returns provided by Dimson, Marsh, and Staunton (2016)
allows us to extend this historical analysis to cover the risk-adjusted returns since 1900, for
21 countries.17 This is the bottom panel of Exhibit 3. In contrast to the top panel, here we see
that Equity risk has generated a substantially higher premium than bonds on average over
this period, once the tailwind of falling rates from the past 48 years is sufficiently diluted. This
surprising result shows that taking a longer view can not only enhance results — it can flip them.

Exhibit 3 | Comparisons of Risk-Adjusted Returns for Equities and Bonds

Exhibit 3: Histograms of realized Sharpe ratios for

local equity markets and 10-year sovereign bonds
by country, for a panel of 10 developed markets
since 1970 in the top panel and 21 developed and
emerging markets in the bottom panel. Best-fit normal
distributions are drawn for the cross-section of
realized bond and equity Sharpe ratios in each panel.
Note that the best-fit curves overlap heavily in the
commonly-examined period since 1970, with bonds
showing a higher average Sharpe ratio, while there
is clearer separation of the two distributions and a
much higher average Sharpe ratio for equities in the
longer sample. See Appendix B for data sources and
other details.

17 C
 ountries included in analysis since 1900:
Australia, Austria, Belgium, Canada, Denmark,
Finland, France, Germany, Ireland, Italy, Japan,
Netherlands, New Zealand, Norway, Portugal,
South Africa, Spain, Sweden, Switzerland,
United Kingdom, and United States.

7
Of course, any analysis of long-term equity returns must account for survivorship bias, and
longer histories are more prone to this than shorter ones. For example, neither Russia nor China
are included in the lower panel of Exhibit 3, and both countries experienced complete wipeouts
for equity holders during their respective revolutions in 1917 and 1949. Fortunately, Dimson,
Marsh, and Staunton (2016) provides summary returns for an index of global equity returns
since 1900 that includes the total losses in both Russia and China and also weights all countries
according to their market capitalizations at the start of the period. This index, revolutions
and all, has an arithmetic mean excess return of 5.5% over bills, with a volatility of 17.5%,
corresponding to a Sharpe ratio of 0.31 for equities, while a similarly constructed index for
bonds realized a Sharpe ratio of only 0.12.18 Assuming a near zero correlation, this equity/bond
difference corresponds to a t-stat right around 2.0 given the 115 year history.19

Historical data provide the numbers, but what do we make of them? Why do equities
outperform bonds on a volatility-adjusted basis? As one might expect, markets are hardly
providing a free lunch for equity holders. We believe equities should provide a higher volatility-
adjusted return than bonds because their return stream is more correlated to human capital
and consumption, and because they compose the majority of risk in portfolios of wealthy
stockholders and institutions.20 In simple terms, this means equity downturns are extra painful
because they have a disproportionate impact on portfolio values and happen at times when
people find themselves simultaneously unemployed and short on cash. Self-aware investors
should consider whether this “when it rains it pours” correlation applies to them at the margin (it
probably does), and adjust accordingly.

IV. Credit: From Many Assets, One Factor

Corporate bonds represent a tremendously heterogeneous asset class in terms of geography,
18 The global asset class returns estimated by
issuers, and expected risk of default. Yet, underneath these differences lie two key Dimson, Marsh, and Staunton (2016) are
calculated in a common currency (USD) at spot
commonalities: lower average liquidity per issue than equities or sovereign bonds, and the rates for currency conversion. This inclusion
of currency risk in the asset class returns
negatively-skewed payoff risk from potential default. Both of these commonalities imply that will generate a relatively unbiased estimate
of average return (since currency exposure
credit might provide a premium over and above its embedded exposure to the factor risk of tends to carry zero long-term return, see
Boudoukh et al. (2015)), but will tend to bias
more liquid asset classes like equities. the estimated volatility upward, and thus the
long-term Sharpe ratio downward, relative to
the currency-hedged Equity and Interest Rates
First, we need to survey the credit landscape and condense it into one factor. Corporate factors.
19 U sing the country-level results can provide
bonds across the United States and Europe represent the most liquid of cash credit bonds and greater statistical significance, as the lack of
perfect correlation across countries implies
derivatives (based on the CDX and iTraxx indices of corporate credit default swaps for their that we are picking up some semi-independent
information in observing that each country in
respective geographies), and they form the basis of our combined Credit factor. Fortunately, the Dimson, Marsh, and Staunton (2016) exhibited
a higher risk-adjusted return for local equities
returns to corporate bond indices, after extracting the Equity and Interest Rates factors, appear than for local government bonds.
20 F or models tying the Equity Risk Premium
to explain the residual returns to many other classes of credit instruments as well, as shown (ERP) to consumption risk of concentrated
shareholders, see Guvenen (2009) and Malloy,
Moskowitz, and Vissing-Jørgensen (2009). For
in Exhibit 4. We believe this justifies our parsimonious use of a single orthogonal Credit factor empirical analysis of equity risk in institutional
portfolios, see Duncombe and Kay (2018)
based on the most liquid indices, yet with explanatory power across a wide range of less liquid and Leibowitz, Bova, and Hammond (2010).
For more on the connection of the ERP to
sub-classes of credit bonds. macroeconomic variables and theoretical
motivations, excellent overviews may be found
in Ang (2014), Damodaran (2018), and Norges
Bank Investment Management (2016).

8
Exhibit 4 | Correlation Heatmap of Residualized Credit Indices and Combined Credit Factor

Exhibit 4: After stripping out exposures to the Equity

and Interest Rates factors, residual returns for
corporate bonds in the US and Europe show a high
degree of correlation suggestive of a common risk
factor (represented by “Combined Corp Credit”). Other
US credit sectors show moderate correlations to the
residual returns of corporate credit, while emerging
market bonds have lower correlations due to their
large EM-specific risk (see Section VI, especially
Exhibit 11). See Appendix B for data sources and
other details.

Credit bonds are among the oldest classes of securities,21 yet the modern heterogeneity of the
asset class makes it difficult to assume that past returns say much about today’s Credit factor
premium. For centuries, only the most developed nations and creditworthy corporations were
capable of issuing bonds, and most classes of debt traded today have shifted from residing on
bank balance sheets to securitized bond issues only in the past few decades. High yield bonds
only became acceptable new issues in the mid-1980s in the United States, and they remained
a very niche market in Europe and elsewhere until the late 1990s. Emerging-market sovereign
bonds were also a miniscule market until the 1989 creation of Brady Bonds to help restructure
and offload illiquid sovereign loans from bank balance sheets. Although the common risk
factor underlying all these forms of credit allows us to extend the factor returns back further
than many individual sub-classes of credit bonds, we would still like to ensure that our factor
encompasses the broad range of default risk in today’s credit markets.

For this reason, we have opted to calculate our extended Credit factor only for periods when
both investment grade and high yield indices are available, even if we need to sacrifice some
geographic diversity to achieve a longer-term view. Exhibits 5 and 6 show the results of
extending our Credit factor back to 1990 through the use of US-only investment grade and high
yield indices. 21 The Dutch East India Company issued
corporate bonds as early as the 17th century,
while Giesecke et al. (2011) documents a robust
corporate bond market in the United States
from 1866 onward.

9
It should be acknowledged that our Credit factor does not have a long enough history for its
mean return to appear positive with statistical significance. At an expected Sharpe ratio of
0.15 to 0.35, typical of most factors derived from long-only asset classes, it would take up to a
century of returns to meet the commonly accepted p < 0.05 cut-off for statistical significance.
High yield bonds as a mature asset class have simply not been around long enough for a broad
Credit factor to support this length of analysis.

However, we take some comfort from the longer perspective of researchers who have focused
on the investment grade credit markets. Asvanunt and Richardson (2017) analyzed returns to
US corporate credit bonds going back to 1936, and found their spliced proxy for credit returns
to outperform a combination of duration-matched US Treasuries and the S&P 500 index with
a t-statistic of 2.17, meeting the conventional bar for statistical significance. As they were
analyzing a 79-year history (1936-2014), this is equivalent to a residualized annual Sharpe ratio
of approximately 0.24 for their corporate credit proxy22 — in line with our longer-term historical
and forward-looking estimates.

Exhibit 5 | Cumulative Summed Returns to Orthogonal Credit Factor

Exhibit 6: The exponentially-weighted estimates of

the historical Sharpe ratio use a 20-year half-life, in
line with the findings of Section VII. The historical
performance shown in Exhibit 6 is roughly in line with
the forward-looking estimates in Exhibit 1 for the
longest credit time series, though the shorter-term
exponentially-weighted residualization process used
for Exhibits 5 and 6 may slightly decrease the Sharpe
ratio of the orthogonal Credit factor due to sharp
Exhibit 6 | Performance of Orthogonal Credit Factor shifts in the estimated Equity and Interest Rates
exposures of credit indices around market crisis
periods.
SHARPE RATIO SHARPE RATIO
(EQUALLY WEIGHTED) (EXPONENTIALLY
WEIGHTED)

Broad Credit Factor 0.13 0.11

22 S
 ee Exhibit 8 of Asvanunt and Richardson
(European and US Corporate Credit)
(2017). Approximate annualized Sharpe ratio
is calculated by taking the t-statistic of 2.17
Long History Credit Factor 0.23 0.17 for the intercept from regression III (monthly
excess returns to their spliced corporate
(US only prior to 1999) credit series regressed against excess returns
of duration-matched US Treasuries and the
S&P 500), dividing it by the square root of
the approximate degrees of freedom in the
regression (79×12 - 2), and multiplying by the
square root of 12 to annualize.

10
V. Commodities: Rorschach’s Factor Test
Commodities present an asset class with a much less monolithic risk structure than credit,
requiring us to trade off parsimony (pushing us toward fewer total factors) for holism (pushing us
to incorporate every unique observable risk in some factor). Even if this asset class appears too
heterogeneous to fold into a single cohesive risk factor, we still find it is possible to consolidate
a mere six or so representative factors from dozens of underlying commodities. Unfortunately,
the evidence supporting a risk premium for each of the factors is not so clear cut, especially
due to the relatively short time frame of available data. This leaves the commodities factors
something like the Rorschach inkblots, with suggestions of premia yet still requiring much
interpretation by the researcher.

Exhibit 7 | Correlations of Commodity Futures Returns

Exhibit 7: Correlations of monthly excess returns for

single-commodity S&P GSCI indices from January
1970 (or index inception) to December 2018. Note
the clusters of higher average cross-correlations
within sector groups and relatively low correlations
across commodities in different sectors, suggestive of
a sector-related factor structure. See Appendix B for
data sources and other details.

11
To start identifying potential risk factors that cut across multiple commodities, we show the
long-term correlations of major commodity futures and forwards in Exhibit 7, where discernable
clusters of highly-correlated assets suggest common underpinning risk factors. These statistical
clusters align well with the classic commodity sectors, suggesting a fundamental explanation
that matches the empirical evidence, and motivates our creation of the following risk factors:

• Energy, consisting of oil, natural gas, and distillate products

• Industrial Metals, including copper, aluminum, nickel, and other metals predominately
used as industrial inputs

• Precious Metals, including gold, silver, and the less liquid platinum and palladium

• Grains, including wheat, corn, soybeans, and related products such as soybean meal

• Livestock, including lean hogs and cattle

The remaining “soft” commodities of cocoa, coffee, cotton, and sugar present a harder issue, as
they are each highly idiosyncratic agricultural goods with their own unique supply and demand
characteristics as shown by their very low correlations to other futures. However, we do not
wish to test each of those individual commodities as a separate risk factor.23 We instead opt
to test the four soft commodities collectively for a return premium despite the relative lack of
empirical evidence for a common risk process.

Exhibits 8 and 9 show the historical performance of our orthogonal commodity factors after
extracting their time-varying loadings on the Interest Rates and Equity factors. Of the six
consolidated commodity factors, only Energy and Industrial Metals showed substantial historical
evidence of long-term actionable return premia.24 As with the Credit factor in Section IV, none
of the orthogonal commodity series meet the cut-off for a statistically significant return premium
due to the insufficient history of many futures series, though their historical returns would be
economically significant in a diversified portfolio.

Exhibit 8 | Cumulative Summed Returns to Orthogonal Commodity Factors

23 The decision to consolidate the soft agricultural

commodities as a single factor to test for a
return premium was based on two criteria.
First, each individual risk factor tested for
long-term profitability increases the chance of
false positives. Thus, unless we have a strong
fundamental reason for why assets would have
unique risk premiums, we can mitigate this
risk of false positives by collectively testing
the assets’ returns for a premium. Second,
these contracts have relatively low liquidity, so
the practical limits on long-term investment
capacity of these futures means they would
each individually have a negligible risk weight in
a multi-asset portfolio.
24 L ivestock futures show positive historical
performance but are too illiquid to be seriously
considered as an actionable premium.

12
Exhibit 9 | P
 erformance of Orthogonal Commodity Factors Exhibit 9: The exponentially-weighted estimates of
the historical Sharpe ratio use a 20-year half-life, in
line with the findings of Section VII. The historical
performance shown in Exhibit 9 tends to be lower
SHARPE RATIO SHARPE RATIO than the forward-looking estimates in Exhibit 1 due to
(EQUALLY WEIGHTED) (EXPONENTIALLY WEIGHTED) the cross-sectional shrinkage increasing our forward-
looking estimates for the individual commodity
portfolios underpinning each factor.
Energy 0.15 0.05
Industrial Metals 0.15 0.15
Precious Metals 0.09 0.08
Grains 0.05 0.01
Softs (Other Agricultural) 0.07 -0.04
Livestock 0.12 -0.02

Since the empirical evidence in Exhibit 9 does not meet the typical bar of statistical significance,
we need to find further theoretical and empirical support for the strong assertion that only
specific commodity futures sectors carry a risk premium. Fortunately, recent academic literature
based upon the Theory of Storage points us toward a fundamental explanation and stronger
statistical evidence for the premia found in Energy, Industrial Metals, and Livestock futures.

The Theory of Storage and Commodity Futures Premia

The Theory of Storage posits that the expected premium for long positions in commodity
futures is not constant over time, but varies inversely with the level of inventories for the
underlying commodity.25 When inventories are low, the risk of “stock-out” or scarcity for
consumers of the commodity rises, thus elevating spot prices and inducing higher expected
price volatility in response to any further supply or demand shocks. Risk-averse producers will
then have greater demand for hedges on their future production, being willing to pay some
premium to futures holders as compensation for reducing exposure to the higher price volatility.
In equilibrium models of the Theory of Storage, this time-varying expected premium to futures
positions is known as the “convenience yield”, as it equates to the premium paid by commodity
holders willing to pay (or forgo) elevated spot prices to have inventory on hand immediately.

While recent development of the Theory of Storage leads to predictions of how commodity
futures’ premia change through time in response to inventories and suppliers’ characteristics,
it still says little about how average premia should differ across commodities. The inventory
shock channel posited by the theory suggests that relatively “hard to store” commodities should
be more prone to low inventory levels and more likely to have a premium, but the high cost of
storage for those commodities could wipe out any average level of convenience yield. 25 Although the Theory of Storage was first
outlined in Kaldor (1939), Working (1949),
and Brennan (1958), it is only in recent years
More usefully, the Theory of Storage also suggests that any return premium to commodity that theoretical models have extended to
endogenously derive spot prices, futures prices,
futures should come from the difference in returns to a futures position plus short-term bonds and the risk premium for long-only futures
speculators accounting for the presence of
(as the futures price is discounted in equilibrium by the interest rate to expiry) versus changes stock-out risk. The discussion in this section
relies heavily on the theoretical model and
in the underlying commodity’s spot price. These return differences, in expectation, are equal to empirical results in Gorton, Hayashi, and
Rouwenhorst (2013). Acharya, Lochstoer, and
the unobservable convenience yield minus the cost of storage. Ramadorai (2013) derive and test a similar
model, though they find an additional important
relationship between producers’ time-varying
risk-aversion and firm-level measures of
financial distress.

13
We build off this insight to generate a more statistically powerful test for potential return
premia in commodity futures, by looking at the monthly differences between S&P GSCI single
commodity index total returns (which tracks the returns to a fully-cash-collateralized rolling
futures position) and spot price changes for the underlying commodity. This analysis eliminates
the main source of variance in commodity futures returns, namely spot price changes. When we
pool these monthly realized return differences over time and across the several commodities in
each sector, we can test the estimated sector-level premia (if any) embedded in the historical
futures returns with much greater statistical power. The results of this analysis are shown in
Exhibit 10, and align well with the findings of the residualized time-series returns in Exhibits 8
and 9. Only the Energy, Industrial Metals, and Livestock sectors showed statistically significant
average premia through time.

Exhibit 10 | Results of Pooled Regression on Commodity Futures Returns minus Spot Price Changes

ANNLZD FUTURES T-STATISTIC # OF

RETURN OVER SPOT OBSERVATIONS
Exhibit 10: Estimated sector-level return premia and
Energy 2.0% 3.54 1962 t-statistics after controlling for monthly time effects in
a panel of 24 GSCI single-commodity futures indices
Industrial Metals 4.8% 5.87 1771 over the period January 1970 to December 2018.
Individual contracts’ observations are weighted based
Precious Metals 0.8% 0.95 1042 on an equal blend of their target 2019 weights in
the Bloomberg Commodity Index and the S&P GSCI
Grains 0.0% -0.02 2051 Index to avoid overweighting the estimated premia
of relatively illiquid commodity futures. See Appendix
Softs 0.1% 0.08 1928 B for details of return series used and time periods
available by commodity.
Livestock 3.1% 3.00 1306

This analysis, with its greater statistical power, gives us more comfort in singling out the Energy
and Industrial Metals commodity sectors as having actionable long-term premia, as livestock
futures tend to be too illiquid for large scale investment. But it would be better still to also have
theoretical justification for why Industrial Metals appear to have a risk premium while harder-to-
store commodities such as many agricultural products do not. Although we have been unable to
find pre-existing academic literature addressing this precise question, we do believe that several
recent papers and empirical findings collectively point in a promising direction.

Several papers have studied the fluctuation of rolling futures returns and the futures basis
over the course of the business cycle, finding that the apparent risk premium from holding
commodity futures is correlated with economic conditions and highest around business cycle
peaks.26 This accords with the Theory of Storage’s prediction that risk premia should be driven
by the risk of stock-out in commodities, which would presumably be highest at times of high
economic growth. Assuming that risk-averse speculators are more capacity constrained and 26 In particular, Hong and Yogo (2012) find that
a diversified basket of commodity futures has
demand higher premia when more commodities are close to stock-out, this evidence suggests a procyclical expected returns after controlling
for the lagged average futures basis; Fama and
procyclical common factor driving risk premia across commodities that would be most present French (1988) find that demand shocks for
metals around business-cycle peaks result in
in futures returns for the individual commodities most closely tied to economic growth: Energies negative futures basis and high convenience
yields, supporting predictions of the theory
and Industrial Metals. of storage; and Kucher and Kurov (2014) find
the futures basis for most energy commodities
grows more negative at business cycle peaks
while expected spot returns rise, implying a
higher expected return for futures investors.

14
As the numerous assumptions and caveats highlighted above suggest, this remains a fertile area
for further research. We will defer more detailed tests of the cross-section of commodity futures
premia and a potential procyclical common factor to future publications.

We conclude with one final caveat regarding commodities as an asset class: the exact method
one uses to get exposure matters greatly, as the Theory of Storage predicts that premia may
be available in futures contracts but not by holding commodity inventories. Furthermore, even
the rules one uses for rolling futures can have a material impact on returns.27 Yet the empirical
and theoretical findings above give us some comfort in recommending diversified Energy and
Industrial Metals commodity futures as part of an efficient asset allocation.

VI. Emerging Markets: Risk without Reward?

We continue our process of testing newer asset classes for their marginal returns with emerging
markets (EM). EM assets have grown considerably in size and liquidity over the past couple of
decades, though expected transaction costs for EM bonds, currencies, or stocks typically remain
higher than for their developed market counterparts. As a newer geographic class of investments
exposed to unique political risks and greater illiquidity, EM assets appear to have a reasonable case
for providing a positive factor premium to compensate investors who take on these risks.

Exhibit 11 shows the risk decomposition of individual EM asset class returns since inception,
stripping out the more liquid developed market factors embedded in the classes to isolate the
(sizeable) orthogonal risk associated with EM bonds, currencies, or equities above and beyond
leveraged exposure to global Equity or Credit risk. The orthogonal returns to each EM asset
class do show statistically significant cross-correlations supporting their combination into a
single factor, even if their cross-correlations are not as high as those observed for corporate
credit residual returns in Section IV.28

Exhibit 11 | Risk Decomposition of Emerging Market Asset Class Returns

Exhibit 11: The global Equity factor explains a

significant percentage of time-series risk across
multiple EM asset classes, with smaller contributions
from factors related to global corporate credit and
commodities. The residual, idiosyncratic returns of all
three EM asset classes explain a sizeable portion of
their respective variances, and show significant cross-
correlations supportive of a common EM risk factor.
Resid. EM Risk See Appendix B for data sources and other details.

Commodities
Credit
27 M
 ouakhar and Roberge (2010); Rallis, Miffre,
Equity and Fuertes (2012)
28 C
 orrelations across equally-weighted residual
Interest Rates returns to the EM Credit, Currency, and
Equity proxies used in this paper varied from
0.31 to 0.36, with t-statistics of 3.4 to 7.0.
All correlation estimates were higher when
using exponentially-weighted returns with a
20-year half-life, suggesting that the common
factor across emerging market assets may
have increased in importance over the analysis
period

15
The Emerging Markets factor, however, has thus far fallen short in the most important test for
long-term investment: profitability. Exhibits 12 and 13 show that, once the leveraged exposures
to global Equity, Interest Rates, and Credit risk factors have been stripped out of the EM asset
class returns, the orthogonal factor candidates show negative or near-zero long-term returns, at
least over this relatively short sample.

Despite a reasonable case for a premium, the empirical evidence suggests that the Emerging
Markets factor may provide no extra risk-adjusted return to a globally diversified asset class
portfolio, even without accounting for higher expected transaction costs in these less liquid
markets. This accords with the findings of Dimson, Marsh, and Staunton (2018) that EM equities
have underperformed their developed market counterparts on a cumulative basis since 1900.
Although this may seem surprising given the higher recent economic growth rates of emerging
markets, the link between economic growth and local asset returns actually appears quite
tenuous and has shown little relationship to the returns of EM assets either through time or
across countries.29

Exhibit 12 | Cumulative Summed Returns to Orthogonal Emerging Markets Factors

Exhibit 13 | Performance of Orthogonal Emerging Market Factors

SHARPE RATIO SHARPE RATIO Exhibit 13: The exponentially-weighted estimates of

the historical Sharpe ratio use a 20-year half-life, in
(EQUALLY WEIGHTED) (EXPONENTIALLY WEIGHTED)
line with the findings of Section VII. Despite periods
of outperformance, such as the early 2000s, the
Combined EM -0.12 -0.24 orthogonal EM factors extracted from all three asset
classes show nearly zero long-term premium after
EM Credit 0.00 0.00 accounting for exposure to global Interest Rates,
Equity, Credit, and commodities-related factors.
EM Currency 0.00 -0.07
EM Equity 0.10 0.05

29 Saret (2014)

16
VII. From History to Expectations
Our analysis so far has focused on historical returns to assets and factors, with the implicit
assumption that the future will look like the past. As a conservative start, we do not expect a risk
factor with decades of flat returns to begin providing a premium tomorrow. But we would like
to move beyond simple “zero or non-zero” classifications and provide forward-looking estimates
of the expected premium for each factor. We believe that empirical data from long performance
histories remains our best guide, but consider two additional steps that might make our forecasts
more robust than simply taking the average return over the longest possible history.

First, it seems that it would be a bit silly to claim that returns in 1906 contain just as much
information about today’s expected premium as returns in 2016. Markets change over time —
evolution in the investor base, issuer base, and market structure could all potentially affect the
long-term expectations for factor returns. This provides our motivation to overweight more
recent returns when estimating the long-term premium of any given factor. However, we believe
this rate of change is likely to be relatively modest for major risk factors that have long been
present in investors’ portfolios. A very long-term exponentially-weighted average return, say
with a half-life of decades, may strike a suitable balance between emphasizing more recent data
and capturing a broad sample of performance through longer- and shorter-term market cycles.

Second, we also believe that it may help to “shrink” individual asset returns toward a common
average. This shrinkage toward a common mean could help prevent overfitting our estimates of
long-term factor premia on the returns of the past couple of decades, making them more robust
even after periods of extreme individual asset class returns.

Our empirical analysis of historical returns weighting and shrinkage is in Appendix A. Overall,
the results suggest that the best balance of overweighting recent periods while keeping relevant
historical data for predicting future performance occurs when we use an exponential weighting
with a half-life around twenty years. They also suggest that observed historical Sharpe ratios for
individual asset classes should be shrunk cross-sectionally to better estimate forward-looking
premia, with approximately 50% weight on the asset-specific return premium and 50% weight
on the cross-sectional average. We believe these two adjustments allow for the sensible and
reasonably robust estimate of forward-looking premia from historical returns.

However…30

While we always try to trust in data and follow our models, there are still some occasions when
the historical data appears so clearly atypical that common-sense adjustments could improve
out-of-sample accuracy. We believe the current state of sovereign yields represents one of
these occasions. Although it is impossible to predict the path of global sovereign yields in the 30 W
 ith our deep apologies, as there always
seems to be a “however.” Financial and
next couple decades, we are quite confident they will not decline another 10 percentage points economic data is so noisy that we often
commiserate with President Truman’s
from the current levels of zero to three percent. So for the critical Interest Rates factor, we apocryphal desire for a one-handed economist.
31 F
 or details on the derivation of our estimate
believe that the equal-weighted long-term average Sharpe ratio going back to the early 20th for the long-term average Sharpe ratio of the
Interest Rates factor, please see notes on
century would be a more sensible starting point in estimating future performance than an Exhibit 1 and 14 in Appendix B.

estimate overweighting the past few decades.31

17
VIII. Conclusion: Forward-Looking Returns by Factor
In this paper, we propose a methodology for generating long-term, forward-looking views on
asset-class-based factors. By examining long histories of performance and statistically extracting
the diversifying return of each new asset class, we seek to identify key common factors
and determine which appear to carry a return premium additive to a diversified institutional
portfolio. After a study of major asset classes, we find five factors that appear to provide unique
return premia with high potential investment capacity: Interest Rates, Equity, Credit, Energy
futures, and Industrial Metals futures.

Slightly overweighting recent returns and shrinking historical performance of asset classes
toward the average also appears to improve the explanatory power of past returns in the future.
Our one discretionary exception to the overweighting of recent returns is the Interest Rates
factor, where an unsustainable decline of global yields in the past four decades leads us to
suggest using the longest available returns history as the best guide for the future.

We conclude by recapping our procedure for estimating the long-term risk premia of the factors
underpinning asset class returns:

1. Calculate historical Sharpe ratios for broad asset class indices that reasonably represent
currency-hedged returns to global equities, bonds, credit, and commodities, applying
exponential weights with a half-life of 20 years to overweight more recent returns.

2. Apply haircuts to the global bond and credit indices’ Sharpe ratios, reflecting a go-forward
expected Sharpe ratio for the Interest Rates factor equal to its historical average since
1900.

3. Shrink all post-adjustment asset class Sharpe ratios by 50% toward their cross-sectional
mean.

4. Residualize the shrunken asset classes’ Sharpe ratios to extract the implicit forward-
looking Sharpe ratios for the orthogonal factors.

The historical Sharpe ratio estimates and intermediate steps of our procedure are shown in
Exhibit 14 for illustration purposes, using returns data through December 2018. The second
column provides the long-term observed historical Sharpe ratio for the asset class proxies used
to construct each factor. The third column applies our suggested cross-sectional shrinkage
detailed in Section VII. The fourth column extracts the expected Sharpe ratios we estimated for
the unique, orthogonal premia of each less liquid factor relative to their embedded exposure to
the two most liquid premia: Interest Rates and Equity.

The findings of modest explanatory power for past returns on a three- to five-year forecast
horizon should encourage humility in the application of these forecasts. However, we believe
the “inescapably human element” behind factor premia allows us to draw some careful
conclusions from the historical record.

18
This paper identifies five compensated factors that may collectively form the basis of an asset
allocation strategy with substantial investment capacity, and provides a method for extracting
long-term factor return forecasts from asset class histories. With future research, we will turn to
the next question of forecasting risk and translating those forecasts into suggested long-term
portfolio allocations.

Exhibit 14 | Calculation Steps for Expected Premia as of Dec 31, 2018

WEIGHTED HISTORICAL POST-SHRINKAGE ASSET POST-SHRINKAGE

ASSET CLASS SHARPE RATIO CLASS SHARPE RATIO ORTHOGONAL FACTOR
SHARPE RATIO (FINAL
ESTIMATE)

Interest Rates 0.2032 0.25 0.25

Equity 0.40 0.35 0.35
Credit 0.56 33
0.46 0.21
Energy 0.17 33
0.24 0.23
Industrial Metals 0.28 33
0.29 0.19
Other Commodities ---34 ---34 0.0035
Emerging Markets ---34 ---34 0.0036
Foreign Currency ---34 ---34 0.0037
Local Equity ---34 ---34 0.0037
Local Inflation --- --- ---38

This paper provides only an overview of the subject matter discussed herein. It does not discuss
many important assumptions, methodologies and other aspects of these subjects. All information
herein is subject to change without notice.
32 B ased on long-term, equally-weighted average
returns and risk rather than exponential
weights. See Section VII and Appendix B for
details.
33 H istorical Sharpe ratio estimates for all assets
with shorter return series are adjusted to
reflect the long-term expected Sharpe ratios of
their estimated exposure to Interest Rates and
Equity factors.
34 We believe that other commodity sectors,
emerging markets assets, foreign currency
exposure, and local equity exposure carry
no statistically significant orthogonal factor
premium, and thus are excluded from the
cross-sectional shrinkage calculations applied
to asset classes believed to carry a unique
premium.
35 O ther commodity sectors are found to have
no statistically significant orthogonal factor
premium, thus their forward-looking estimate
is fixed to 0.0. See Section V for more details.
36 E merging Markets are found to have no
statistically significant orthogonal factor
premium, thus their forward-looking estimate
is fixed to 0.0. See Section VI for more details.
37 We believe the Foreign Currency and Local
Equity factors should provide no orthogonal
factor premium, thus their forward-looking
estimates are fixed to 0.0. See discussion in
Introduction for more details.
38 Although part of the Two Sigma Factor Lens
where supported, it is only possible to observe
very short histories for the Local Inflation
factor due to the recent introduction of
inflation-linked sovereign bonds as an asset
class. Hence we do not estimate a long-term
expected premium in this paper.

19
Appendix A: Regression Analysis to Estimate Cross-
Sectional Shrinkage and Historical Returns Estimation
Though we lack the centuries of data across all of our selected factors necessary to rigorously
test these hypotheses, we still have nearly a century of data across three key asset classes
that can be applied to testing long-term returns predictability. By looking at the returns to
global equities, global sovereign bonds, and US credit going back to 1925, we find that both a
multi-decade lookback period and cross-sectional shrinkage appear to provide a modest, but
statistically significant, forecast of asset class returns.

To check whether differences in historical asset class premia persist in the future, we used an
exponentially-weighted moving average of historical returns and volatility for global equities,
global sovereign bonds, and US corporate credit to estimate their historical Sharpe ratio on
a rolling basis for each month starting in December 1940 (so we could begin with at least 15
years of data in the lookback period). We then ran panel regressions of the historical differences
in Sharpe ratio against the three- and five-year forward-looking Sharpe ratios. Two main
specifications were considered:

1. In the top row of each set of results, we tested a model of shrinking historical asset class
Sharpe ratios toward a constant value, persisting through time and across asset classes:

2. In the second row, we tested a model of shrinking historical asset class Sharpe ratios
toward their cross-sectional average at each point in time, allowing the overall return
premium across all assets to fluctuate over the years:

In the equations above, psri,t represents the exponentially-weighted Sharpe ratio of past excess
returns for asset i at time t, psrt represents the cross-sectional average of the exponentially-
weighted past return Sharpe ratios across all assets at time t, and fsri,t represents the equally-
weighted Sharpe ratio of future three- or five-year returns for asset i at time t.

Exhibit 15 shows the results of our regression analyses, with standard errors adjusted to
reflect the large degree of autocorrelation in these slow-moving historical returns. As can be
seen from the tables of r-squared values and t-statistics for the asset-specific betas ( in the
model equations above), an assumption of shrinking historical asset class performance toward
a constant Sharpe ratio provided virtually zero explanatory or statistical power regardless of the
weighting used for historical data or the length of forecast period. Shrinking historical asset class
Sharpe ratios toward their cross-sectional average (in the second row of each set of results)
showed statistically significant predictive power over both forecast horizons, especially when
historical Sharpe ratios were calculated using exponential weights with a 20-year half-life.

20
Exhibit 15 | Results of Long-Term Regressions over Multiple Exhibit 15: Test results of regressions based on
the assumption of shrinking historical asset class
Lookback Periods and Forecast Periods Sharpe ratios toward either a constant Sharpe ratio
or the cross-sectional average. Results across each
row reflect regressions where the independent
variable was based on historical asset class Sharpe
ratios calculated using exponential weights with
the specified half-life. T-statistics were calculated
using Driscoll-Kraay (1998) robust standard errors
accounting for up to 36 months of lagged effects
for the three-year forecast regressions and up to 60
months of lagged effects for the five-year forecast
regressions. See Appendix B for data sources and
other details.

21
Appendix B: Data Sources and Exhibit Details
Exhibits 1 and 14

Historical Estimated Sharpe Ratio for Interest Rates factor is based upon equally-weighted
average returns and risk since 1870/1900, while Historical Estimated Sharpe Ratios for all other
factors with an expected positive risk premium are based upon exponentially-weighted average
returns and volatility with a 20-year half-life following the methodology outlined in Section VII.

Long-term average returns and risk figures for the Interest Rates factor are based on data from
Dimson, Marsh, and Staunton (2014); Jorda et al. (2018); and Piketty and Zucman (2014). The
estimate of 0.20 for the long-term historical Sharpe ratio of the Interest Rates factor is the
authors’ own, based upon two separate estimates derived from the aforementioned papers:

1. Jorda et al. (2018) in Table 3 provide pooled estimates for excess returns (1.53%) and
standard deviation (8.38%) of local-currency bonds across 16 countries for the period
1870-2015. This methodology should eliminate the volatility impact of currency
movements, similar to our currency-hedged Interest Rates factor, and provides a long-
term Sharpe ratio estimate of 0.18.

2. Dimson, Marsh, and Staunton (2014) report 21 country-level real returns and standard
deviations for bills and long-term bonds over the period 1900-2015 in Appendices
1.4 and 1.5, respectively. The authors used these to derive country-level estimates of
the excess return and standard deviation for long-term bonds based on an assumed
average correlation of 0.5 between real returns for bills and bonds within each country.
By assuming a correlation of 0.5 for bond excess returns across countries and applying
historical GDP weights derived from Piketty and Zucman (2014), we come to an estimate
for the Sharpe ratio of a global currency-hedged bond portfolio of 0.21.

Exhibit 2

Annual 10-year government bond yields from Global Financial Data. Developed markets
represented were selected for geographic diversity and length of historical yields available.
Yields for Dutch and French 10-year bonds peaked at 66.67% and 25.64%, respectively, around
the turn of the 18th century.

LABEL GFD SYMBOL

Australia IGAUS10D
Canada IGCAN10D
France IGFRA10D
Japan IGJPN10D
Netherlands IGNLD10D
Norway IGNOR10D
United Kingdom IGGBR10D
USA _TNXD

22
Historical Sharpe ratio estimates for all assets with shorter return series are adjusted to reflect
the long-term expected Sharpe ratios of their estimated exposure to Interest Rates and Equity
factors. This adjustment led to lower estimates of the historical Sharpe ratios for credit assets
and a slightly higher estimate of the historical Sharpe ratio for energy commodity futures.

Exhibit 3

Returns data for January 1970 to December 2018 are total returns series from Global Financial
Data (bonds and bills) and MSCI local currency net return indices from Bloomberg (equities).
The five European countries with the largest 1970 GDP in current USD were selected to avoid
overrepresentation of Europe in the dataset due to greater data availability (GDP data from
World Bank Open Data). Symbols/tickers for the total return indices used for country-level
equity, government bond, and government bills are provided in the table below.

BILLS GFD BONDS EQUITIES

SYMBOL GFD SYMBOL BLOOMBERG
TICKER
Australia TRAUSBIM TRAUSGVM NDDLAS
Canada TRCANBIM TRCANGVM NDDLCA
France TRFRABIM TRFRAGVM NDDLFR
Germany TRDEUBIM TRDEUGVM NDDLGR
Italy TRITABIM TRITAGVM NDDLIT
Japan TRJPNBIM TRJPNGVM NDDLJN
Netherlands TRNLDBIM TRNLDGVM NDDLNE
Spain TRESPBIM TRESPGVD NDDLSP
United Kingdom TRGBRBIM TRGBRGVM NDDLUK
United States TRUSABIM TRUSG10M NDDUUS

Sharpe ratio calculations for 1900 to 2015 are the authors’ own, based upon the country-level
real returns and standard deviations for equities, bonds, and bills provided in Appendices 1.2,
1.4, and 1.5 of Dimson, Marsh, and Staunton (2016). For estimating the standard deviation of
excess returns over bills, real returns for stocks and bills were assumed to be uncorrelated and
the real returns for bonds and bills were assumed to have 0.5 correlation, the latter estimate
based on historical correlation data provided in figure 4 of Jorda et al. (2018).

Exhibit 4

Correlations from index inception through December 2018 of monthly residual returns to credit
sector total return indices after extracting estimated loadings on the global Equity and Interest
Rates factors. “Combined Corp Credit” is an equal-risk-weighted combination of the residual
returns to the four US and European corporate credit indices, mirroring the construction of the
Credit factor in this paper. All returns data from Bloomberg.

23
 INDEX BLOOMBERG FIRST MONTHLY
NAME TICKER RETURN

US IG Corp Bloomberg Barclays LUACTRUU Feb 1973

US Corporate
US HY Corp Bloomberg Barclays LF98TRUU Sep 1983
US Corporate High Yield
Euro IG Corp Bloomberg Barclays Corporate LP05TRUH Feb 1999
Pan-European Aggregate
Hedged to USD
Euro HY Corp Bloomberg Barclays LP01TRUH Feb 1999
Pan-European High Yield
Hedged to USD
Combined Corp Credit Authors’ calculations --- Sep 1983
(see above)
EM USD Credit Bloomberg Barclays EMUSTRUU Feb 1993
Emerging Markets
USD Aggregate
EM LC Credit Bloomberg Barclays EMLCTRUU Jul 2008
Emerging Market
Local Currency Government
US Lvged Loans Credit Suisse Leveraged Loan CSLLLTOT Jan 1992
US Converts Bloomberg Barclays LUCCTRUU Feb 2003
US Convertibles Composite
US ABS Bloomberg Barclays LUABTRUU Jan 1992
US Aggregate ABS

Exhibits 5 and 6

Credit factor returns are an equal-risk-weighted combination of the rolling residual returns
relative to the Equity and Interest Rate factors for US investment grade and high yield
corporate credit and European investment grade and high yield corporate credit, hedged to
USD. All returns data from Bloomberg, using the corporate credit indices listed in the Exhibit 4
data table.

Credit factor returns are calculated daily since January 1990 and monthly to September 1983.
European credit indices are only included in the factor from August 2000, when daily data
becomes available for both. ICE BofAML High Yield Master II returns are used as the US high
yield credit proxy for daily data from January 1990 to August 1998, with Bloomberg Barclays
US Corporate High Yield used for monthly returns before and daily returns after.

Exhibit 7

All returns data from Bloomberg, using the commodity futures indices listed in the Exhibit 10
data table.

Exhibits 8 and 9

Commodity sector returns are an equally-weighted combination of futures index returns for
the commodities in each sector, residualized on a rolling basis to the Equity and Interest Rate
factors. Qualitatively similar results were found when using liquidity-related weights within

24
sectors. All returns data from Bloomberg, using the commodity futures indices listed in the
Exhibit 10 data table.

Exhibit 10

All index data from Bloomberg, with tickers supplied in the table below. Regression was run
controlling for both monthly time-period and sector-level effects, constraining the monthly
time effects to sum to 0.0 so any average futures premium would be present in the sector-level
roll return estimates (rather than the time effect estimates). Regression weights do not sum to
100% due to the lack of Soy Oil futures in our dataset, which are a current constituent of the
Bloomberg Commodity Index but not the S&P GSCI.

FUTURES SPOT REGRESSION START

COMMODITY TICKER TICKER WEIGHT MONTH

Energy WTI Crude SPGSCLTR SPGSCL 17.0% Feb 1987

Brent Crude SPGSBRTR SPGSBR 13.0% Feb 1999
Natural Gas SPGSNGTR SPGSNG 5.7% Feb 1994
Gas Oil SPGSGOTR SPGSGO 4.1% Feb 1999
Unleaded Gasoline SPGSHUTR SPGSHU 3.4% Feb 1988
Heating Oil SPGSHOTR SPGSHO 3.3% Feb 1983

Industrial Copper SPGSICTR SPGSIC 5.9% Feb 1977

Metals
Aluminum SPGSIATR LMAHDY 4.1% Feb 1991
Zinc SPGSIZTR SPGSIZ 2.2% Feb 1991
Nickel SPGSIKTR SPGSIK 1.7% Feb 1993
Lead SPGSILTR SPGSIL 0.4% Feb 1995
Tin SPGSISTR SPGSIS 0.0% Feb 1995

Precious Gold SPGSGCTR SPGSGC 8.0% Feb 1978

Metals
Silver SPGSSITR SPGSSI 2.2% Feb 1973
Palladium SPGSPATR SPGSPA 0.0% Feb 1995
Platinum SPGSPLTR SPGSPL 0.0% Feb 1984

Grains Corn SPGSCNTR SPGSCN 5.1% Jan 1970

Soybeans SPGSSOTR SPGSSO 4.6% Jan 1970
Wheat SPGSWHTR SPGSWH 4.2% Jan 1970
Soybean Meal SPGSSMT SPGSSM 1.7% Feb 1995

Softs Sugar SPGSSBTR SPGSSB 2.3% Feb 1973

Coffee SPGSKCTR SPGSKC 1.6% Feb 1981
Cotton SPGSCTTR SPGSCT 1.4% Feb 1977
Cocoa SPGSCCTR SPGSCC 0.2% Feb 1984

Livestock Live Cattle SPGSLCTR SPGSLC 3.8% Jan 1970

Lean Hogs SPGSLHTR SPGSLH 1.9% Feb 1976
Feeder Cattle SPGSFCTR SPGSFC 0.6% Feb 2002

25
Exhibits 11, 12, and 13

All returns from Bloomberg, with proxy indices used for EM assets in the table below. EM Equity
returns were calculated net of underlying currency movements (i.e. on a currency-hedged basis)
by subtracting daily returns to the MSCI Emerging Markets Currency Index.

Risk decomposition in Exhibit 11 uses an exponentially-weighted regression with 20-year

half-life to determine the variance contribution of the Equity, Interest Rates, Credit, Energy,
and Industrial Metals factors to the monthly returns of each of the EM asset class proxies.
The contributions to variance of the Energy and Industrial Metals factors are combined as
contribution from “Commodities” in the chart.

Exhibits 12 and 13 are based on daily returns to the EM asset proxies residualized on a rolling
basis to the Equity, Interest Rates, Credit, Energy, and Industrial Metals factors.

INDEX NAME BLOOMBERG TICKER

EM Credit JP Morgan EMBI Global JPEIGLBL

EM Currency MSCI Emerging Markets Currency Index MXEF0CX0
EM Equity (after Jan 2001) MSCI Emerging Markets Net Return M1EF
EM Equity (before Jan 2001) MSCI Emerging Markets Gross Return M2EF

Exhibit 15

Monthly returns data from Global Financial Data. Symbols for indices used in this analysis:
_DJCBTD for US corporate credit returns, GFWLDM for global equity price returns, SYWLDYM
for global equity dividend yields, and TRWLDGVM for global government bonds.

26
Bibliography
Acharya, V.V., L.A. Lochstoer, and T. Ramadorai (2013). “Limits to Arbitrage and Hedging: Evidence from Commodity Markets”, Journal
of Financial Economics, 109(2), 441-465.
Adrian, T., R. Crump, B. Mills, and E. Moench (2014). “Treasury Term Premia: 1961-Present”, Liberty Street Economics blog post.
Accessible at https://libertystreeteconomics.newyorkfed.org/2014/05/treasury-term-premia-1961-present.html
Ang, A. (2014). Asset Management. Oxford University Press, New York.
Asvanunt, A. and S. Richardson (2017). “The Credit Risk Premium”, Journal of Fixed Income, 26(3), 6-24.
Brennan, M.J. (1958). “The Supply of Storage”, American Economic Review, 48(1), 50-72.
Campbell, J.Y., K.S. Medeiros, and L.M. Viceira (2010). “Global Currency Hedging”, Journal of Finance, 65(1), 87-121.
Campbell, J.Y., C.E. Pflueger, and L.M. Viceira (2018). “Macroeconomic Drivers of Bond and Equity Risks”, Harvard Business School
Finance Working Paper.
Damodaran, A. (2018). “Equity Risk Premiums (ERP): Determinants, Estimation and Implications – The 2018 Edition”, Working paper,
New York University.
D.E. Shaw Group Market Insights (2019). “Positively Negative: Stock-Bond Correlation and Its Implicaitons for Investors”, DE Shaw
Group White Paper. Accessible at https://www.deshaw.com/assets/articles/DESCO_Market_Insights_20190208_1.pdf
Dimson, E., P. Marsh, and M. Staunton (2016). “Long-Term Asset Returns” in D. Chambers & E. Dimson (Eds.) Financial Market
History: Reflections on the Past for Investors Today (pp. 2-26). CFA Institute Research Foundation.
Dimson, E., P. Marsh, and M. Staunton (2018). Credit Suisse Global Investment Returns Yearbook 2018. Credit Suisse Group AG,
Zurich.
Driscoll, J.C. and A.C. Kraay (1998). “Consistent Covariance Estimation with Spatially Dependent Panel Data”, Review of Economics
and Statistics, 80(4), 549-560.
Duncombe, G. and B. Kay (2018). “Introducing the Two Sigma Factor Lens”, Two Sigma White Paper. Accessible at https://www.
twosigma.com/insights/article/introducing-the-two-sigma-factor-lens/
Fama, E.F. and K.R. French (1988). “Business Cycles and the Behavior of Metals Prices”, Journal of Finance, 43(5), 1075-1093.
Giesecke, K., F.A. Longstaff, S. Schaefer, and I. Strebulaev (2011). “Corporate Bond Default Risk: A 150-Year Perspective”, Journal of
Financial Economics, 102(2), 233-250.
Gorton, G.B., F. Hayashi, and K.G. Rouwenhorst (2012). “The Fundamentals of Commodity Futures Returns”, Review of Finance, 17(1),
35-105.
Guvanen, F. (2009). “A Parsimonious Macroeconomic Model for Asset Pricing”, Econometrica, 77(6), 1711-1750.
Hong, H. and M. Yogo (2012). “What Does Futures Market Interest Tell Us about the Macroeconomy and Asset Prices?”, Journal of
Financial Economics, 105(3), 473-490.
Ilmanen, A. (2011). Expected Returns: An Investor’s Guide to Harvesting Market Rewards, Wiley, Chichester, UK.
Jorda, O., K. Knoll, D. Kuvshinov, M. Schularick, and A.M. Taylor (2018). “The Rate of Return on Everything, 1870-2015”, CESifo
Working Paper.
Kaldor, N. (1939). “Speculation and Economic Stability”, Review of Economic Studies, 7(1), 1-27.
Leibowitz, M.L., A. Bova, and P.B. Hammond (2010). The Endowment Model of Investing: Risk, Return, and Diversification. Wiley
Finance, Hoboken, NJ.
Malloy, C.J., T.J. Moskowitz, and A. Vissing-Jørgensen (2009). “Long-Run Stockholder Consumption Risk and Asset Returns”, Journal of
Finance, 64(6), 2427-2479.
Mouakhar, T. and M. Roberge (2010). “The Optimal Approach to Futures Contract Roll in Commodity Portfolios”, Journal of Alternative
Investments, 12(3), 51-60.
Norges Bank Investment Management (2016). “The Equity Risk Premium”, NBIM Discussion Note. Accessible at https://www.nbim.no/
en/publications/discussion-notes/2016/the-equity-risk-premium/

27
Piketty, T. and G. Zucman (2014). “Capital Is Back: Wealth-Income Ratios in Rich Countries 1700-2010”, Quarterly Journal of
Economics, 129(3), 1255-1310. Online Appendix at http://piketty.pse.ens.fr/fr/capitalisback accessed on Nov 27, 2018.
Rallis, G., J. Miffre, and A-M. Fuertes (2012). “Strategic and Tactical Roles of Enhanced Commodity Indices”, Journal of Futures
Markets, 33(10), 965-992.
Rappoport, P. and N. Nottebohm (2012). “Improving on Risk Parity: Hedging Forecast Uncertainty”, J.P. Morgan Asset Management
research report.
Ross, S.A. (1976a). “The Arbitrage Theory of Capital Asset Pricing”, Journal of Economic Theory, 13(3), 341-360.
Ross, S.A. (1976b). “Risk, Return and Arbitrage” in Risk Return in Finance. I. Friend and J.L. Bicksler (Eds.). Ballinger, Cambridge, Mass.
Saret, J.N. (2014). “Emerging and Frontier Market Equities as Asset Class(es)”, Two Sigma White Paper. Accessible at https://www.
twosigma.com/insights/article/emerging-frontier-markets-asset-classes/
Working, H. (1949). “The Theory of Price of Storage”, American Economic Review, 39(6), 1254-1262.
Wright, J.H. (2011). “Term Premia and Inflation Uncertainty: Empirical Data from an International Panel Dataset”, American Economic
Review, 101(4), 1514-1534.

28
IMPORTANT DISCLAIMER AND DISCLOSURE INFORMATION

This document is distributed for informational and educational purposes only and is not an offer to sell or the solicitation of an offer to buy any
securities or other instruments. The information contained herein is not intended to provide, and should not be relied upon for, investment, accounting,
legal, or tax advice. Exposure to risk factors is not a guarantee of increased performance or of decreased risk. Diversification and asset allocation
may not protect against market risk or loss of principal. Past performance does not guarantee future results. A factor lens is merely a resource
supplementing investment, financial, tax, accounting, legal and other advisors and due diligence. This document does not purport to advise you
personally concerning the nature, potential, value or suitability of any particular sector, geographic region, security, portfolio of securities, transaction,
investment strategy or other matter. No consideration has been given to the specific investment needs or risk-tolerances of any recipient. The recipient
is reminded that an investment in any security is subject to a number of risks including the risk of a total loss of capital, and that discussion herein does
not contain a list or description of relevant risks. The recipient hereof should make an independent investigation of the information described herein,
including consulting its own tax, legal, accounting and other advisors about the matters discussed herein. This document does not constitute any form
of invitation or inducement by Two Sigma to engage in investment activity.

The views expressed herein are not necessarily the views of Two Sigma Investments, LP or any of its affiliates (collectively, “Two Sigma”). Such views
(i) may be historic or forward-looking in nature, (ii) may reflect significant assumptions and subjective judgments of the authors of this document, and
(iii) are subject to change without notice. Two Sigma may use, construct, and maintain factors and methodologies differently than as discussed in this
document, including when making investment decisions on behalf of or when trading for purposes of investable products.

No representation is made as to the accuracy of any information contained herein and the use of information from third-party sources in no way
implies an endorsement of the source of such information or its validity. Any forward-looking statements are inherently subject to significant business,
economic and competitive uncertainties and contingencies, many of which are beyond our control. In addition, these forward-looking statements are
subject to assumptions with respect to future business strategies and decisions that are subject to change. Factors which could cause actual results
to differ materially from those anticipated include, but are not limited to: competitive and general business, economic, market and political conditions
in the United States and abroad from those expected; changes in the legal, regulatory and legislative environments in the markets in which Two Sigma
operates; and the ability of management to effectively implement certain strategies. Words like “believe,” “expect,” “anticipate,” “promise,” “plan,” and
other expressions or words of similar meanings, as well as future or conditional verbs such as “will,” “would,” “should,” “could,” or “may” are generally
intended to identify forward-looking statements. Two Sigma makes no representations, express or implied, regarding the accuracy or completeness of
this information, and the recipient accepts all risks in relying on this document for any purpose whatsoever.

Important Notice to Venn Subscribers: This document is not an offer to, or solicitation of, any subscribers of Venn or otherwise for the provision of
investment management, advisory or any other services. Nothing in this document should be considered as a representation of how any research
or methodology is used on Venn. Importantly, Venn’s use of the Two Sigma Factor Lens and other research or methodologies will differ from the
subject matter discussed herein. This document is subject to change without notice.

This document is being furnished to the recipient on a confidential basis and is not intended for public use or distribution. By accepting this document,
the recipient agrees to keep confidential the existence of this document and the information contained herein. The recipient should not disclose,
reproduce, distribute or otherwise make available the existence of and/or all or any portion of the information contained herein to any other person
(other than its employees, officers and advisors on a need-to-know basis, whom the recipient will cause to keep the information confidential) without
Two Sigma’s prior written consent. This document shall remain the property of Two Sigma and Two Sigma reserves the right to require the return of this
document at any time. Some of the images, logos or other material used herein may be protected by copyright and/or trademark. If so, such copyrights
and/or trademarks are most likely owned by the entity that created the material and are used purely for identification and comment as fair use under
international copyright and/or trademark laws. Use of such image, copyright or trademark does not imply any association with such organization (or
endorsement of such organization) by Two Sigma, nor vice versa.

© 2019 Two Sigma Investments, LP | ALL RIGHTS RESERVED | “Two Sigma,” “2σ” and Venn are trademarks of Two Sigma Investments, LP.