THE JOURNAL OF FINANCE • VOL. LXIX, NO.

1 • FEBRUARY 2014

An Anatomy of Commodity Futures Risk Premia

MARTA SZYMANOWSKA, FRANS DE ROON, THEO NIJMAN,
and ROB VAN DEN GOORBERGH∗

ABSTRACT
We identify two types of risk premia in commodity futures returns: spot premia re-
lated to the risk in the underlying commodity, and term premia related to changes
in the basis. Sorting on forecasting variables such as the futures basis, return mo-
mentum, volatility, inflation, hedging pressure, and liquidity results in sizable spot
premia between 5% and 14% per annum and term premia between 1% and 3% per
annum. We show that a single factor, the high-minus-low portfolio from basis sorts,
explains the cross-section of spot premia. Two additional basis factors are needed to
explain the term premia.

FUTURES CONTRACTS ARE ZERO-COST securities, that is, they do not require an
initial investment. Hence, expected futures returns consist only of risk pre-
mia. Understanding these premia is important, as they impact, for example,
the hedging decisions of companies and the investment decisions of financial
institutions. The purpose of this paper is to characterize the cross-sectional
and time-series variation in commodity futures risk premia.1 The cross-section
of commodity futures risk premia has at least two dimensions. First, for each
commodity there are multiple futures contracts that differ in time-to-maturity.
Therefore, analogous to bonds, there is a term structure both of futures prices
and of futures expected returns or risk premia. Second, like stocks, individual
commodity futures differ on characteristics such as the sector to which they
belong (e.g., Energy vs. Metals), as well as on characteristics like momentum
and valuation ratios. The latter also lead to time-series variation in expected
futures returns.

∗ Marta Szymanowska is with Rotterdam School of Management, Erasmus University; Frans de

Roon is with Department of Finance, CentER, Tilburg University; Theo Nijman is with Department
of Finance, CentER, Tilburg University; and Rob van den Goorbergh is with APG. We thank the
Editor (Cam Harvey); the Associate Editor; the referees; Lieven Baele; Hendrik Bessembinder;
Frank de Jong; Michel Robe; Geert Rouwenhorst; Jenke Ter Horst; Chris Veld; Marno Verbeek;
conference participants at the American Finance Association (AFA) 2010 Annual Meeting and
Inquire UK 2009 Autumn Meeting; and seminar participants at the Katholieke Universiteit (KU)
Leuven, Commodity Futures Trading Commission (CFTC), Norwegian School of Management –
BI, Rotterdam School of Management, Erasmus University, and University of Piraeus for helpful
comments.
1 Although we refer to them as risk premia, notice that in futures markets these may be both

negative or positive as futures markets are zero-sum games.

DOI: 10.1111/jofi.12096

453
454 The Journal of FinanceR

The contribution of this paper is threefold. First, we decompose commodity

futures expected returns into spot and term premia that can be identified by
taking long positions in short maturity (nearby) contracts and by combining
long and short (spreading) positions in contracts with different maturities, re-
spectively. These premia, or discounts, show up in different ways in multiperiod
strategies that hold the contract until maturity or that roll over short-term con-
tracts. Whereas rolling over short-term contracts isolates the spot premia in
multiple periods, holding the contract until maturity yields expected returns
that consist of the spot premia plus term premia. This decomposition is im-
portant because the two risk premia are likely to compensate for different risk
factors. For instance, in the case of oil futures, the spot premium reflects oil
price risk, while the term premia mainly reflect the risk present in the con-
venience yield. Like risk premia in the term structure of interest rates, term
premia are also present in the term structure of the futures cost-of-carry or
(percentage) basis.
Second, we show that differences in expected returns on various trading
strategies also result from time-variation in risk premia due to commodity fu-
tures characteristics. As for stocks, the cross-sectional and time-series variation
in commodity futures returns is related not so much to sector (or industry) as
to characteristics like the basis, momentum, volatility, and other instruments.
Finally, just as variation in stock returns can be attributed to a limited
number of factors like the three Fama–French factors, we show that cross-
sectional variation in commodity futures returns can be attributed to a single
basis factor for spot premia and to two additional basis factors for term premia.
A log-linear approximation similar to Campbell and Shiller’s (1988) analysis
of the dividend yield implies that the futures percentage basis contains infor-
mation about expected futures returns or risk premia. This suggests that the
predictive power of valuation ratios such as dividend yield for stocks, forward
premium for bonds, carry trade for foreign exchange, and house price-to-rent
ratio for real estate, among others, also applies to commodity markets.2 This
is also in line with a number of other papers in the commodity literature that
relate futures risk premia to the basis or carry (e.g., Fama (1984), Erb and
Harvey (2006), Gorton and Rouwenhorst (2006), and Liu and Tang (2011)).
Previous literature identifies a number of other variables that lead to pre-
dictable variation in futures risk premia but does not differentiate between spot
and term premia in futures markets. Instruments known to induce time vari-
ation in commodity futures risk premia other than the basis include hedging
pressure and momentum.3 In addition to these instruments, or characteris-
tics, commodity risk premia have been related to futures volatility, inflation,

2 See, for example, Hansen and Hodrick (1980), Fama (1984, 1986), Fama and Bliss (1987),
Campbell and Shiller (1991), Gourinchas and Rey (2007), Piazzesi and Swanson (2008), Koijen
et al. (2012), or Cochrane (2011) for an excellent review and references therein.
3 See, for example, Carter, Rausser, and Schmitz (1983), Fama (1984), Chang (1985), Fama and

French (1987), Bessembinder (1992), de Roon, Nijman, and Veld (1998, 2000), Erb and Harvey
(2006), Miffre and Rallis (2007), and Gorton, Hayashi, and Rouwenhorst (2013).
 Anatomy of Commodity Futures Risk Premia 455

and open interest.4 As we specifically include futures contracts with longer

maturities to analyze the term premia, we also consider the liquidity of the
contracts.
Our results are based on a broad cross-section of 21 commodity futures mar-
kets with as many as four different maturities. Sorting on the percentage basis
of the futures, our center-stage variable, we find in the high-minus-low basis
portfolio that the spot premia are between −8% and −14% per annum, depend-
ing on the maturity of the contracts, and the term premia are of opposite sign,
between 0.5% and 2% per annum. In an in-sample analysis, we show that about
70% of these premia are due to cross-sectional differences in the average basis,
whereas 30% are due to time-series deviations of the basis from its mean.
When sorting on other commonly used predictive variables we also find it is
important to distinguish between spot and term premia. Apart from the basis
sorts, spot premia show up when sorting on momentum, volatility, inflation
β, and liquidity. The resulting spot premia are usually between 8% and 10%
per annum in absolute terms. Term premia, on the other hand, mainly show
up when sorting on the basis, volatility, and inflation β, and marginally when
sorting on hedging pressure and liquidity. The term premia are mostly between
0.5% and 2% per annum, and always of the opposite sign as spot premia.
Our findings thus imply that previously identified forecasting variables affect
expected futures returns in different ways via the spot and term premia. These
findings also contribute to the debate on the existence of time-varying risk
premia in commodity futures markets (e.g., Dusak (1973), Carter, Rausser,
and Schmitz (1983), and more recently, Frank and Garcia (2009), as well as
references therein) as we find spot and term premia to reliably show up when
sorting on the various characteristics.
Although we find many significant spot and term premia among the various
portfolio sorts, standard asset pricing tests show that especially the cross-
sectional patterns in spot premia can be attributed to only a single basis factor.
A factor portfolio that goes long the high basis commodity futures and short the
low basis commodity futures, similar to carry trade for currencies, can explain
most of the other sorted portfolio returns that capture spot premia, leaving
only small unexplained mean returns on the table. A horse race with similar
factor portfolios based on the other characteristics shows that none comes close
to the performance of the basis factor. The basis factor, however, fails to explain
the term premia in the sorted portfolios. This is also the case for single-factor
portfolios based on any of the other characteristics. On the other hand, using
two separate basis factor portfolios, the high basis and the low basis commodity
futures portfolios, explains nearly all of the term premia in our portfolio sorts.
Bessembinder and Chan (1992) find that nearby returns in 12 different futures
markets are driven by two latent factors. Unlike their latent factors, we identify
one observable factor for the spot premia in 21 commodity futures markets, and
two observable factors for the corresponding term premia.

4 See, for example, Erb and Harvey (2006), Dhume (2011), and Hong and Yogo (2012).
456 The Journal of FinanceR

Our findings also add to the literature on cross-sectional predictability across

markets. Papers like Fama and French (1993, 1996), Cochrane and Piazzessi
(2005), Cochrane (2008), and Lustig, Roussanov, and Verdelhan (2011), among
others, find that the cross-section of stocks, bonds, and currency returns, re-
spectively, can be explained by relatively few factors. A paper close to ours is
Lustig, Roussanov, and Verdelhan (2011), who show that the cross-section of
international currency returns can be explained by a single factor, the return
on the highest minus the return on the lowest interest rate currency portfolio.
As high interest rates imply low futures prices, this factor is similar to a futures
carry trade. We contribute to this literature by demonstrating that a similar
phenomenon exists in commodity futures markets.
The rest of the paper is structured as follows. Section I presents a simple
decomposition of futures returns and characterizes the time variation in com-
modity expected returns using a present value relation. Section II describes
the data and analyzes unconditional risk premia. The conditional risk premia
and their implications for asset pricing are discussed in Sections III and IV
respectively. Section V concludes.

I. Theory
A. A Decomposition of Expected Futures Returns
We begin our analysis with a simple decomposition of expected futures re-
turns that highlights the different premia (or discounts if they are negative)
that may be present in futures markets. Denote by St the spot price of the
underlying commodity, and by Ft(n) the futures price for delivery at time t + n,
of a commodity with per-period physical storage costs, Ut(n) , that are a percent-
age of the spot price, and a cash payment, Ct+n. This cash payment is the net
dollar-equivalent income from convenience yield that accrues to the commod-
ity owner (stemming, for instance, from the value of the option to sell out of
storage). We assume that the payment Ct+n occurs at time t + n, but is already
known at time t. The cost-of-carry model (e.g., Fama and French (1988)) then
implies that the futures price equals5
 n  n
Ft(n) = St 1 + RFt(n) 1 + Ut(n) − Ct+n, (1)

where RFt(n) is the n-period risk-free interest rate at time t, matching the
maturity of the futures contract. We can use the same cost-of-carry relation to
define the per-period log or percentage basis, yt(n)
 
Ft(n) = St exp yt(n) × n , (2)

5 This way of expressing the cost-of-carry model, which assumes that storage costs must be paid

upfront, therefore implies financing costs. The expression in Fama and French (1988), equation
(1), differs from ours in that we express storage costs as a fraction of the current spot price. This
representation is more useful for our analysis.
 Anatomy of Commodity Futures Risk Premia 457

with
 n  n C 
1 t+n
yt(n) = ln 1 + RFt(n) 1 + Ut(n) − . (3)
n St

This log basis is also known as the futures (cost of) carry. Thus, yt(n) is the
per-period cost of carry for maturity n, analogous to a bond’s n-period interest
rate. If the cost-of-carry model holds, it consists of the n-period interest rate
(RFt(n) ), and possibly other items, such as storage costs (Ut(n) ) and convenience
yields (Ct+n), depending on the nature of the underlying asset. It is also the
slope of the term structure of (log) futures prices, as follows from solving (2)
for yt(n) . Hereafter we simply refer to yt(n) as the basis. It is important to note
that, although the cost-of-carry model gives an easy interpretation of the de-
composition of futures risk premia, our decomposition is also valid when the
cost-of-carry model does not hold.6
From the one-period expected log spot return, we define the spot risk pre-
mium πs,t as the expected spot return in excess of the one-period basis,
  
Et rs,t+1 = Et ln(St+1 ) − ln(St ) = Et st+1 − st = yt(1) + πs,t , (4)
where we take expectations Et [·] conditional on the information available at
time t and denote log prices using lower case. The spot premium, πs,t , can be
interpreted as the expected return in excess of the short-term basis, in the
manner of stock returns in excess of the short-term interest rate (and adjusted
for the dividend yield).
(n)
Next, we define a term premium π y,t as the (expected) deviation from the
expectations hypothesis of the term structure of the basis,
(n−1)
nyt(n) = yt(1) + (n − 1)Et yt+1 − π y,t
(n)
. (5)

(n)
Note that, without imposing more structure, the term premium π y,t also shows
up in the expected return on a futures contract for delivery at time t + n. This
follows from the log return on such a contract, again using (2).

B. Trading Strategies
To illustrate how spot and term premia can be earned, we consider several
different trading strategies. First, from equation (2) and the fact that the fu-
tures price converges to the spot price at the delivery date, we can identify the
spot premium with a long position in a short-term futures contract, r (1) f ut,t+1 ,
that is, the return on the futures contract that matures at time t + 1,

Et r (1)
f ut,t+1 = Et st+1 − ft
(1)
= Et st+1 − st − yt(1) = πs,t . (6)

6 If the cost-of-carry model does not hold, for instance, because of stochastic interest rates (as in

Cox, Ingersoll, and Ross (1981) or Casassus and Collin-Dufresne (2005)), or because the commodity
is nonstorable, the basis is still defined as the log (or percentage) difference between the futures
price and the spot price.
458 The Journal of FinanceR

(1)
It follows immediately from (6) that π y,t = 0, that is, the short-term futures
contract does not contain a term premium.
Next, consider the return r (n)
f ut,t→t+n, which is simply the holding period return
from buying an n-period futures contract at time t and holding it until the
maturity date t + n. We refer to this as the Holding return, the conditional
expectation of which is

Holding: Et r (n)
f ut,t→t+n = Et st+n − ft
(n)

     
(1) (1) (2) (n−1)
= Et st+n − ft+n−1 + ft+n−1 − ft+n−2 + · · · + ft+1 − ft(n)
n−1 n−1
 (n− j)
= Et πs,t+ j + Et π y,t+ j . (7)
j=0 j=0

Thus, the expected return of the Holding strategy is the sum of expected spot
premia and term premia for all maturities up to n. Note that the expected
return in (7) involves the expectation at time t of the risk premia that show up
in later periods. To the extent that risk premia are time-varying, this will make
the longer term expected returns different from simply adding up one-period
expected returns.
Second, instead of holding an n-period futures contract until maturity, con-
sider investing in one-period futures contracts for n consecutive periods, that
is, rolling them over each period. The returns on those contracts are r (1) f ut,t+ j ,
j = 1, 2, . . . , n, and the expected return on this Short Roll strategy is
⎡ ⎤
n n−1

Short Roll: Et ⎣ r (1)
f ut,t+ j
⎦ = Et πs,t+ j . (8)
j=1 j=0

Naturally, the expected return on this strategy consists of only expected (future)
spot premia. Note that the spot premia in (8) are identical to those in (7), and
again, if risk premia are time-varying, differ from n times the one-period spot
premia in (6).
Comparing the expected returns in (7) and (8), we can isolate the term premia
by going long in the Holding strategy and taking a short position in the Short
Roll strategy, which we refer to as the Excess Holding return, the expectation
of which is
⎡ ⎤
n n−1
Excess Holding: Et ⎣ r (n) ⎦= (n− j)
f ut,t→t+n − r (1)
f ut,t+ j Et π y,t+ j . (9)
j=1 j=0

This is similar to buying a long-term bond and financing this with short-term
loans rolled over until maturity. The Excess Holding expected return consists
of the expected term premia for all maturities up to n, which are identical to
those in (7).
 Anatomy of Commodity Futures Risk Premia 459

The term premia for those maturities can also be earned by taking a portfolio
of one-period spreads r (k) (1)
f ut,t+1 − r f ut,t+1 , for k = 1, 2, ..., n. Using the definitions
(n)
of πs,t and π y,t in (4) and (5), it can be seen that the expected one-period futures
return for a contract that matures at time t + k is

Et r (k) (k−1)
f ut,t+1 = Et ft+1 − ft
(k)
= πs,t + π y,t
(k)
. (10)

Thus, if we combine a long position in a long-term contract with a short

position in a short-term contract, the expected return on the spreading strategy
(k)
is generated by only one term premium π y,t ,

Et r (k) (1)
f ut,t+1 − r f ut,t+1 = π y,t .
(k)
(11)

Note that (9) is simply the multiperiod equivalent of this one-period spread-
ing return. The one-period spreading strategy would yield the same term pre-
mia, but only for period t + 1. Also note that the per-period expected returns in
(9) and (11) are generally not equal, unless the term premia are constant. The
term premia are earned by the spreading strategy in (11) in one period (t + 1),
and by the Excess Holding return in (9) in n consecutive periods (t + 1, ..., t + n).
Buying a portfolio of spreads every period and rolling it over creates a multi-
period Spreading strategy, similar to the Short Roll strategy, the conditional
expected return of which is
⎡ ⎤
n n   n n−1
⎣ 1 (k) (1) ⎦ 1 (k)
Spreading: Et r f ut,t+ j − r f ut,t+ j = Et π y,t+ j . (12)
n n
k=1 j=1 k=1 j=0

Basically, the Spreading strategy earns 1/n of each term premium each period,
while the Excess Holding strategy earns each of the individual term premia
consecutively. If the term structure of the basis is changing over time, or more
generally if risk premia are time-varying, the two strategies have different
types of roll-over risk and different expected returns.

C. Time-Varying Risk Premia

The foregoing decompositions indicate that differences in the various ex-
pected returns (on commodity futures) occur because (i) the different returns
(trading strategies) are exposed to spot and term premia in different ways,
and (ii) both risk premia may be time-varying. Time-varying risk premia, or
expected returns, are by now understood to be a common element across mar-
kets. As noted by Cochrane (2011, p. 1051), “for stocks, bonds, credit spreads,
foreign exchange, sovereign debt, and houses, a basis or valuation ratio trans-
lates one-for-one to expected excess returns [or risk premia].” It is similarly
common in the commodity futures literature to relate expected futures returns
to the (log) basis or carry (see, e.g., Fama (1984), Erb and Harvey (2006), Yang
(2013), Gorton, Hayashi, and Rouwenhorst (2013), and Koijen et al. (2012)).
460 The Journal of FinanceR

As for stocks and other markets, the use of the basis can be motivated by
a present value relation, as in Campbell and Shiller’s (1988) analysis of the
dividend yield. To see this, we start from the cost-of-carry model in (1). Basically,
we interpret Ct+1 /St as a valuation ratio, and use a log-linear approximation
to relate the basis to expected returns.
Using (1) and assuming for ease of exposition that the risk-free rate and
storage costs are constant over time and across maturities, the return on a
one-period futures contract is

St+1 St+1
R(1)
Fut,t+1 = = . (13)
Ft(1) St (1 + RF ) (1 + U ) − Ct+1

Relative to the stock return that underlies the Campbell and Shiller lineariza-
tion, (13) looks unusual: the cash payment occurs in the denominator instead
of the numerator and the current spot price St is compounded at the risk-free
rate and storage costs. Both adjustments follow from the fact that the return
is calculated from the futures Ft(1) instead of spot St price and reflects the
cost-of-carry. Taking logs of (13) gives
 
St+1
r (1)
f ut,t+1 = ln = ln St+1 − ln ( St (1 + RF ) (1 + U ) − Ct+1 )
Ft(1)
 
Ct+1
= st+1 − st − ln (1 + RF ) (1 + U ) − .
St

From (5) the expectation of this is πs,t . In Appendix A we show that log-
linearizing the last
 term around  the mean (log) basis c − s − r f − u, and defin-
ing θ = 1 − exp c − s − r f − u , we obtain
⎡ ⎤
∞
κ  
yt(1) ≈ + Et ⎣ θ j
ct+ j+1 − πs,t+ j ⎦ , (14)
1−θ
j=0

where κ contains constants that follow from the linearization. As shown in the
Appendix A, for 0 < θ < 1 we need the average cash yield to be strictly positive
and not exceed the current spot price of the commodity compounded at the risk-
free rate plus storage costs. The equivalent assumption for stock prices would
be that the average dividend payment does not exceed the current stock price
compounded at the risk-free rate. These are mild assumptions. If the average
cash yield does go to zero, the basis will be constant and naturally not contain
any information about either risk premia.
Equation (14) shows that the current basis contains information about future
cash yield growth and future spot premia. It follows that yt(1) is a natural
predictor of spot risk premia. Performing the same analysis for longer term
contracts, Appendix A shows that yt(n) contains information about future cash
yield growth and both spot and term premia:
 Anatomy of Commodity Futures Risk Premia 461

yt(n) ≈ st + n (r f + u) − ct+n
⎡ ⎤
∞  
κn n−1 n−1
= + Et ⎣ θnj ct+( j+1)n − πs,t+i − (i)
π y,t+i ⎦ . (15)
1 − θn i=0 i=0
j=0

We use subscripts n on κn and θn in equation (15) to emphasize that these

parameters do depend on the maturity n chosen.
Thus, similar to dividend yields for stocks, the yield curve for bonds, and
the interest rate spread for currency returns, equation (15) suggests that the
commodity futures basis predicts commodity (excess) returns. The basis is
therefore a natural candidate for explaining time-variation in commodity risk
premia. The extent to which basis reflects changing risk premia or growth in
cash flow yields is an empirical question.

II. Futures Data and Summary Statistics

A. Futures Data
We use bimonthly returns constructed from data obtained from the Com-
modity Research Bureau (CRB) on 21 commodity futures contracts. Data are
available for different sample periods, depending on the contract. We use March
1986 as the starting date for our sample to ensure that we have at least three
commodities per portfolio when sorting returns for each maturity series into
four portfolios. From this date onwards we can also construct hedging pressure
data as one of our predictive instruments. The end of our sample is December
2010.
As futures contracts are unevenly spread over the calendar year in terms of
available delivery dates, with available delivery dates varying between 5 and 12
months per year, the use of bimonthly data allows us to construct more evenly
distributed maturity contracts. We construct 2-month (which is one period)
returns for nearest-to-maturity contracts as the short maturity contracts and
holding period returns for 4, 6, and 8 months until the delivery date. We take
for each bimonthly date the nearest-to-maturity contract as the spot contract,
the second nearest-to-maturity contract as the futures contract with one period
to maturity, and so forth.
As commodity spot markets are known to be illiquid, we use the nearest-
to-maturity futures price as the spot price, similar to most other studies on
commodity futures. Although this gives rise to some irregularities in delivery
date, given that we use bimonthly observations, the resulting errors will be
small. The 21 commodities were chosen with an eye to minimizing the irregu-
larities in delivery dates. Prices of futures observed a month prior to and during
the delivery month are excluded from the analysis to avoid irregular price be-
havior close to the delivery date. Although it is common in the literature to roll
over to the next nearest contract at the end of the month prior to delivery month
T , we observe for many contracts in our sample low open interest during the
last 6 weeks, and thus we roll over 1 month earlier, that is, just before month
462 The Journal of FinanceR

T − 1, to avoid thinly traded prices. Moreover, for many contracts traders often
start rolling over their contracts from 4 to 6 weeks before the delivery date,
implying that we can expect to observe erratic price behavior, this long before
the maturity date.
We divide the data into seven commonly used categories: Energy (3), Meats
(3), Metals (3), Grains (4), Oilseeds (3), Softs7 (3), and Industrial Materials (2).8
These markets have relatively large trading volumes and provide a broad cross-
section of commodity futures contracts. In the Internet Appendix we describe
our data set in detail.9
For each of the seven categories, we construct equally weighted (EW) “sector-
maturity” indices of the futures contracts as the EW average of log returns.
The average index returns (and later portfolio returns) should therefore be
interpreted as average log returns, not real portfolio returns (which would have
rebalancing returns in them). Indices are created for the nearest-to-maturity
contracts (referred to as “nearby” indices) and for the next three farther-to-
maturity contracts. In addition to the seven sector indices, we create EW indices
by taking the simple average of the log returns over all 21 contracts.

B. Unconditional Expected Returns

Table I contains summary statistics for the seven sector indices and the
EW index of 21 commodities. The first panel shows average returns and stan-
dard deviations for the Short Roll returns that isolate the spot premia. Except
for Metals and Meats, Short Roll returns show clear downward- or upward-
sloping patterns. Recall that the difference between expected Short Roll re-
turns across maturities is due to time-variation in spot premia. Thus, these
patterns in the average Short Roll returns are indicative of time-varying spot
premia. The t-statistics indicate that approximately one-third of individual
sector-maturity indices have average Short Roll returns significantly different
from zero. Between-sector variation is quite high, with average spot premia
ranging from around 10% per annum for Energy contracts to around −6.5%
for Grains and Softs. The resulting average of the EW index for the 21 futures
contracts is close to and indistinguishable from zero.
The average Excess Holding returns in the second panel isolate the term
premia and show them to be an order of magnitude smaller than the spot
premia (except for Industrial Materials), never exceeding 2% per annum. For
individual sectors, t-statistics confirm the average Excess Holding returns to
be mostly indistinguishable from zero, except for Industrial Materials. The
EW index shows average returns to be significantly different from zero though,
implying that average term premia across sectors are reliably different from
zero.

7 The category “Softs” as used by the CRB consists of Coffee, Orange Juice, and Cocoa.
8 The classification we use is similar to that used by the Institute for Financial Markets (IFM).
9 The Internet Appendix may be found in the online version of this article.
 Table I
Summary Statistics
The table contains summary statistics for the seven sector indices as well as for the EW (overall) commodity index. The table presents mean returns,
standard deviations, and t-statistics for the various sector indices for the nearest-to-maturity contracts, second nearest-to-maturity contracts, and so

Anatomy of Commodity Futures Risk Premia

on. The first panel shows summary statistics for the Short Roll returns, and the second one for the Excess Holding returns. t-statistics are based on
Newey–West corrected standard errors. The returns are quoted bimonthly for a sample period between March 1986 and December 2010.

Annualized Mean Returns Annualized Standard Deviations t-Statistics

n=1 n=2 n=3 n=4 n=1 n=2 n=3 n=4 n=1 n=2 n=3 n=4

Short Roll Energy 10.83% 9.83% 8.96% 8.56% 32.88% 29.06% 28.80% 28.5% (1.64) (1.68) (1.55) (1.49)
Meats 4.20% 4.03% 3.93% 3.89% 13.03% 11.98% 12.08% 12.6% (1.60) (1.67) (1.62) (1.53)
Metals 5.42% 5.02% 4.76% 4.63% 17.16% 15.42% 15.10% 15.0% (1.57) (1.62) (1.57) (1.54)
Grains −6.10% −6.24% −6.50% −6.74% 18.96% 18.56% 18.11% 17.8% (−1.60) (−1.67) (−1.78) (−1.88)
Oilseeds 1.86% 1.61% 1.46% 1.26% 20.96% 18.44% 17.65% 16.8% (0.44) (0.43) (0.41) (0.37)
Softs −6.58% −6.57% −6.58% −6.70% 18.48% 16.72% 15.52% 14.7% (−1.77) (−1.95) (−2.11) (−2.27)
Ind materials −4.82% −4.62% −4.83% −4.87% 19.56% 18.21% 17.94% 18.6% (−1.22) (−1.26) (−1.34) (−1.30)
EW 0.65% 0.38% 0.11% −0.07% 11.70% 11.32% 11.42% 11.3% (0.27) (0.16) (0.05) (-0.03)
Excess Holding Energy 0.19% 0.43% 0.56% 1.45% 2.32% 3.2% (0.65) (0.92) (0.86)
Meats 0.10% −0.06% −0.16% 2.08% 3.52% 4.7% (0.24) (−0.08) (−0.16)
Metals 0.03% 0.00% −0.06% 0.49% 0.87% 1.3% (0.29) (0.02) (−0.24)
Grains 0.82% 1.68% 1.49% 4.90% 7.64% 10.0% (0.83) (1.09) (0.74)
Oilseeds 0.22% 0.40% −0.82% 1.10% 1.91% 4.9% (0.97) (1.03) (−0.83)
Softs 0.32% 0.52% −1.11% 1.20% 1.80% 21.1% (1.32) (1.43) (−0.26)
Ind materials 1.41% 2.88% 3.52% 2.91% 6.91% 16.4% (2.40) (2.07) (1.07)
EW 0.73% 1.08% 2.77% 1.20% 2.07% 4.3% (3.01) (2.58) (3.21)

463
464 The Journal of FinanceR

We report the results for the Holding and Spreading returns in the Inter-
net Appendix. We find that Holding returns are similar to Short Roll returns,
although differences between maturities are usually larger for Holding than
Short Roll returns. As Holding returns are the sum of Short Roll and Ex-
cess Holding returns, these differences are due to the term premia that are
more distinct in longer maturity contracts. We also observe that Spreading
returns, although also mostly indistinguishable from zero, are different from
Excess Holding returns, suggesting that there is time-variation in these term
premia.

III. Analysis of Conditional Expected Returns

The patterns in the different return strategies are indicative of time-
variation in both spot and term premia. We use portfolio sorts as a way to
capture time-variation in risk premia. Extensively used in studying stock mar-
ket returns, the portfolio sorting approach has been adopted in recent papers
on commodity futures (e.g., Dhume (2011), Gorton, Hayashi, and Rouwenhorst
(2013)). We sort 21 commodities into four portfolios based on the quartiles of
the instruments described in detail later. We choose four portfolios to (i) reduce
return variance by balancing a sufficient number of commodities per portfo-
lio, and (ii) be able to detect monotonic increasing or decreasing patterns in
estimated premia across sorts. For each sort, we consider maturities of 2, 4,
6, and 8 months for Short Roll and Excess Holding returns. The results for
the Holding and Spreading returns are similar to those reported here and are
tabulated in the Internet Appendix.

A. Sorting on the Basis

Table II, which presents our first main result, shows the different types of
mean returns and standard deviations (Short Roll and Excess Holding) when
futures contracts are sorted on the short maturity (log) basis. The table is
structured the same as for the sector returns presented in Table I.
Panel A of Table II shows clear patterns in the portfolio returns result-
ing from the sorts. The Short Roll returns provide a direct estimate only of
the spot premia. Looking at these returns, we see that for all holding peri-
ods (n = 1, 2, 3, and 4) mean returns always decrease as the basis increases.
The resulting spread in the high-minus-low basis portfolios (P4–P1) decreases
from −8.3% to −14.5% per annum across the holding periods. Thus, sorting on
the basis results in a spread of about −10% for the high versus low portfolio,
which is both economically and statistically highly significant. Commodities
with the lowest basis, and thus highest convenience yield, have the highest
mean returns, which increase from 4.8% to 9.9% per year as the maturity of
the Short Roll return increases. For the highest basis portfolio, mean returns
are all between −3.5% and −5.6%. Total spreads between the high and low
basis portfolios are comparable to those reported in Dhume (2011) and Gor-
ton, Hayashi, and Rouwenhorst (2013), who find (absolute) spreads of 9.7%
 Anatomy of Commodity Futures Risk Premia 465

Table II
Sorts Based on the Basis
The table contains mean returns and standard deviations (for Short Roll and Excess Holding
returns) when futures contracts are sorted on the basis in Panel A and mean and de-meaned
basis in Panel B. In Panel B for each of the returns, the first row, “mono,” indicates whether the
underlying mean returns on the four portfolios show a monotonic pattern across the sort. The
next two rows show mean returns and t-statistics for the spread in mean return across the four
portfolios. t-statistics are in parentheses and are based on Newey–West corrected standard errors.
The returns are quoted bimonthly for a sample period between March 1986 and December 2010.

Panel A. Basis

Annualized Mean Returns Annualized Standard Deviations

n=1 n=2 n=3 n=4 n=1 n=2 n=3 n=4

Short Roll Low 4.82% 7.00% 7.89% 9.92% 16.97% 15.39% 16.83% 18.3%
P2 4.68% 4.68% 3.46% 5.46% 14.25% 12.83% 11.18% 13.1%
P3 −2.93% −3.71% −2.01% −0.89% 13.46% 12.93% 12.78% 13.9%
High −3.47% −4.35% −5.61% −4.62% 15.98% 13.16% 11.95% 15.0%
P4–P1 −8.29% −11.35% −13.51% −14.53% 17.15% 12.96% 12.76% 16.1%
t(P4–P1) (−2.40) (−4.33) (−5.22) (−4.45)
Excess Low 0.32% 0.07% −0.30% 1.61% 2.33% 3.2%
Holding P2 0.20% 0.35% 0.59% 0.98% 1.57% 2.0%
P3 0.47% 0.84% 0.88% 0.95% 1.41% 1.9%
High 0.93% 1.51% 1.53% 1.37% 2.20% 2.2%
P4–P1 0.61% 1.44% 1.84% 1.75% 2.43% 3.0%
t(P4–P1) (1.72) (2.91) (3.00)

Panel B. Cross-Section of Basis

B.1 Mean Basis B.2 De-Meaned Basis

Short Roll Mono y y y y y y y y

P4–P1 −15.42% −15.36% −14.86% −18.02% −3.82% −5.27% −6.66% -10.22%
t-Stat (−4.37) (−4.40) (−4.17) (−3.91) (−1.12) (−2.01) (−2.74) (−3.27)
Excess Mono y y
Holding P4–P1 1.12% 1.59% 2.37% 0.23% 0.80% 1.15%
t-Stat (3.55) (3.20) (3.22) (0.62) (1.56) (1.64)

and 10.0%, respectively (although their studies neither distinguish between

maturities nor differentiate between spot and term premia). Erb and Harvey
(2006) use a slightly different strategy, going long in commodities that are
backwardated (i.e., have a negative basis) and short in commodities that are
contangoed (i.e., have a positive basis), and obtain an excess return of 8.2%
relative to a long-only strategy.
The Excess Holding returns isolate the term premia. Except for n = 2, we also
see a monotonic pattern in the term premia, which now increase as a function of
the basis. The resulting spreads for the high-minus-low basis portfolios range
from 0.6% to 1.8% per annum. Although the term premia are much smaller
466 The Journal of FinanceR

than the spot premia, their spreads are significantly different from zero, and
the standard deviations of the Excess Holding returns are also modest between
1.0% and 3.2% across all portfolios.
The Internet Appendix reports, as a robustness check, tables similar to
Table II but for different sample periods. We first construct a sample that
starts at the same date of January 1986, but ends in November 2008, before
the start of the financial market crisis. We then construct two samples that start
at earlier dates. One begins in July 1967 that consists of only 11 commodities
and no energy contracts. Another, begins, at least for the shortest maturities
contracts, in August 1978 and has 18 different contracts. The results of sorting
on basis are similar across these samples.
As many commodities show seasonal patterns, at least in the basis, the In-
ternet Appendix also reports sorting results when correcting for seasonalities.
Sorting returns and seasonally adjusted returns on the seasonally adjusted ba-
sis again give results very similar to those we report in Table II. Finally, given
that the sorting based on the basis is motivated by the cost-of-carry model, we
report in the Internet Appendix the extent to which the sorting results from the
basis are driven by interest rates rather than the convenience yield, and find
that there is no meaningful effect from the interest rate on the sorted portfolio
returns.

B. Sorting on the Cross-Section of the Basis

Sorting futures on the current level of the basis produces clear patterns in
the cross-section of commodity futures returns, with significant spot and term
premia. To see the extent to which the resulting spreads (premia) are due to the
fact that the average level of the basis is high or low, Panel B.1 of Table II sorts
the commodity futures on their mean basis. Notice that this sort is done on the
total sample mean of the basis and therefore, unlike the results in Panel A, does
not represent an investable strategy. For each of the two returns (Short Roll
and Excess Holding) the first row, “mono,” indicates whether the underlying
mean returns on the four portfolios show a monotonic pattern across the sort.
The next two rows show the average return spread for the high-minus-low
portfolio (P4–P1) and the corresponding t-statistic.
Sorting on the mean basis in Panel B.1 produces monotonic return patterns
in all cases except one, and the resulting spreads in both spot and term premia
are highly significant and especially for the spot premia even higher than the
sorts for the basis itself. Although these returns do not represent an investable
strategy, Panel B.1 suggests that most if not all of the results from sorting on
the basis come from the cross-section of the mean basis. However, Panel B.2
shows the results when sorting commodities on the deviation of the current
basis from its sample mean. Thus, if we take the mean basis as given, the
portfolio goes long in commodities whose basis is currently high relative to
its mean and short in commodities whose basis is currently low relative to its
mean. The resulting returns are again monotonic and yield a significant spread
for the Short Roll returns, except for the shortest maturity. For Excess Holding
 Anatomy of Commodity Futures Risk Premia 467

returns the patterns are weaker and the implied term premia are significant
for the longest maturities only. Although the implied spot and term premia are
smaller, they still represent about 50% of the basis premia in Panel A. If we
add the two premia in Panels B.1 and B.2, on average about 70% of the spot
and term premia is due to sorting on the (in-sample) mean, and the remaining
30% is due to sorting on deviations from the mean.

IV. Explaining the Cross-Section of Commodity Expected Returns

Although from the analysis of the valuation ratio in Section I the basis is a
natural predictor for both spot and term premia, the literature on commodity
futures identifies many other variables that may predict commodity futures
returns. We first sort our commodity futures according to a number of fore-
casting variables and characterize the resulting portfolios in terms of spot and
term premia. We then attempt to answer the question of whether the differ-
ent sorts capture different types of risk or can be explained by one factor or
a limited number of factors. Bessembinder and Chan (1992) find, for a set of
eight commodity and four currency futures, that the (unconditional) returns on
the nearest-to-maturity contracts (reflecting spot premia in our terminology)
are driven by two latent (unobservable) factors. We construct observable factor
portfolios from the basis sort and use standard asset pricing tests to analyze
whether the resulting basis factors explain the cross-sectional patterns in the
various portfolio returns. Again, given that we observe consistent results across
the different types of trading strategies that capture spot and term premia, we
only report here the results for the Short Roll and Excess Holding returns,
leaving the other results for the Internet Appendix.

A. Alternative Sorts
Similarly to the analysis of the basis mentioned above, every 2 months we
sort our 21 commodities into four portfolios based on a forecasting variable and
then analyze the different types of returns that capture spot and term premia. A
detailed description of the way we construct the different forecasting variables
is given in Appendix B. The set of forecasting variables we use is not meant to
be exhaustive, but rather to be representative of earlier studies.
We use the following forecasting variables. First, similar to other asset
classes, there is momentum in commodity futures returns ((Erb and Harvey
(2006), Miffre and Rallis (2007), Gorton, Hayashi, and Rouwenhorst (2013),
and Asness, Moskowitz, and Pedersen (2013)). Second, reflecting the fact
that high risk induces high expected returns, commodities with high spot
price volatility (measured by the coefficient of variation) are known to have
higher expected futures returns (Dhume (2011)). Third, commodity returns are
positively correlated with inflation (Greer (2000), Erb and Harvey (2006), and
Gorton and Rouwenhorst (2006)) and commodity’s unexpected inflation β’s are
highly correlated with roll returns (Erb and Harvey (2006)). Fourth, somewhat
related, as most commodity markets are denominated in U.S. dollars, this
468 The Journal of FinanceR

implies that commodity markets are likely exposed to currency risk. Erb and
Harvey (2006) find a significant negative exposure of commodities with respect
to changes in the U.S. dollar versus a basket of foreign currencies. Fifth,
an extensive literature relates expected futures returns to the net (long vs.
short) positions of hedgers in the futures market, known as hedging pressure.
Markets in which hedgers are net short (long) are found to have positive (neg-
ative) expected futures returns (Carter, Rausser, and Schmitz (1983), Chang
(1985), Bessembinder (1992), and de Roon, Nijman, and Veld (2000)). More
recently, next to hedging pressure, Hong and Yogo (2012) show open interest
in a futures market to (positively) predict commodity, currency, stock, and
bond prices. Their model, supported by empirical findings, implies that, owing
to hedging demand and downward-sloping demand curves in futures markets,
open interest is an informative signal of future price inflation, which we use
as a sixth instrument. Finally, as liquidity may differ widely between different
commodity futures and between different maturities, expected futures returns
may reflect the liquidity of the contract. We use the Amivest measure (Amihud,
Mendelson, and Lauterbach (1997)) as suggested by Marshall, Nguyen, and
Visaltanachoti (2012) as our last forecasting variable.
Using the same format as in Panel B of Table II, Table III summarizes the
results for the alternative portfolio sorts. For comparison, the first panel of
Table III also summarizes the results from sorting on the basis as reported
in Table II. Spot premia in the short-maturity Short Roll returns show up
reliably when sorting on the (percentage) basis, momentum, volatility, infla-
tion β, and liquidity, but not in the other sorts. For the shortest maturity
contracts, the (absolute) spreads in the high-minus-low portfolios vary be-
tween 8.1% per annum for the volatility sorts to 9.6% per annum for the
(unexpected) inflation β sorts. It is only for the basis and inflation β that
the Short Roll returns are also monotonic and show significant spreads for
longer maturities, whereas for momentum and liquidity the sorted returns be-
come nonmonotonic and/or the spreads become insignificant as the maturity
of the contract increases. The results for volatility sorts are somewhat mixed
in this respect. The high-minus-low spreading returns for the Short Roll re-
turns are relatively stable across maturities for sorts on inflation β, indicating
that there is no additional time-variation in these spot premia (unlike for basis
spreads).
Term premia, as measured by the Excess Holding returns, show up reliably
when sorting on the basis, volatility, and inflation β. They show up marginally
in the longest-maturity hedging pressure and liquidity-sorted portfolios, based
on the Spreading returns as reported in the Internet Appendix. Term premia
are always of the opposite sign as spot premia, and on an order of magnitude
of 0.5% to 1.5% per annum with little cross-sectional variation between the
sorts.
It is only for sorts on β with respect to changes in the U.S. dollar and for sorts
on open interest that we observe neither reliable spot nor term premia in the
various portfolio returns. Also, although for hedging pressure and open interest
previous studies show regression-based evidence for a significant relation with
 Table III
Alternative Sorts
The table contains summary results for mean returns (Short Roll and Excess Holding returns) when futures contracts are sorted on different
instruments. For each of the returns, the first row, “mono,” indicates whether the underlying mean returns on the four portfolios show a monotonic
pattern across the sort. The next two rows show mean returns and t-statistics for the spread in mean return across the four portfolios. We report
the results for portfolios sorted on the basis, momentum, coefficient of variation (CV), inflation β, dollar β, hedging pressure (HP), open interest,

Anatomy of Commodity Futures Risk Premia

and liquidity. t-statistics are in parentheses and are estimated using Newey–West correction. The returns are quoted bimonthly for a sample period
between March 1986 and December 2010.

Annualized Mean Returns Annualized Mean Returns

n=1 n=2 n=3 n=4 n=1 n=2 n=3 n= 4

Panel A. Returns Sorted on Basis Panel B. Returns Sorted on Momentum

Short Roll Mono y y y y y y y

P4–P1 −8.29% −11.35% −13.51% −14.53% 9.00% 6.57% 4.68% 2.11%
t-Stat (−2.40) (−4.33) (−5.22) (−4.45) (2.02) (1.90) (1.35) (0.51)
Excess Holding Mono y y
P4–P1 0.61% 1.44% 1.84% −0.47% −0.63% −0.26%
t-Stat (1.72) (2.91) (3.00) (−1.12) (−1.01) (−0.40)

Panel C. Returns Sorted on CV Panel D. Returns Sorted on Inflation β

Short Roll Mono y y y y y

P4–P1 8.13% 8.67% 9.27% 9.28% 9.56% 9.60% 8.60% 10.04%
t-Stat (2.37) (2.94) (3.18) (2.56) (1.99) (2.19) (1.87) (1.86)
Excess Holding Mono y y y y y
P4–P1 −1.00% −1.25% −0.79% −0.60% −1.15% −1.46%
t-Stat (−3.09) (−2.69) (−1.16) (−1.53) (−1.76) (−1.67)

(Continued)

469
 470
Table III—Continued

Annualized Mean Returns Annualized Mean Returns

n=1 n=2 n=3 n=4 n=1 n=2 n=3 n= 4

Panel E. Returns Sorted on Dollar β Panel F. Returns Sorted on HP

The Journal of FinanceR

Short Roll Mono y y
P4–P1 −1.86% −1.41% −0.91% −1.81% 5.58% 5.75% 4.17% 5.09%
t-Stat (−0.35) (−0.30) (−0.21) (−0.34) (1.66) (1.77) (1.31) (1.64)
Excess Holding Mono y
P4–P1 0.91% 1.24% 0.87% −0.50% −0.57% −0.93%
t-Stat (2.48) (2.10) (1.05) (−1.30) (−0.89) (−1.34)

Panel G. Returns Sorted on Open Interest Panel H. Returns Sorted on Liquidity

Short Roll Mono y y

P4–P1 5.78% 5.33% 6.35% −5.08% −9.40% −7.47% −5.89% −6.92%
t-Stat (1.71) (1.77) (1.83) (−1.39) (−2.22) (−2.05) (−1.85) (−1.82)
Excess Holding Mono
P4–P1 −1.01% −1.38% 0.69% 0.49% 0.66% 1.26%
t-Stat (−2.49) (−2.05) (0.52) (1.57) (1.55) (2.04)
 Anatomy of Commodity Futures Risk Premia 471

commodity futures returns, this shows up only marginally, if at all, in our

sorted portfolios.10
In sum, except for the U.S. dollar and open interest, sorting on other fore-
casting variables yields similar patterns in spot and term premia as in sorting
on the basis—the order of magnitude of the premia is often very similar, with
term premia being of the opposite sign and much smaller in absolute value
than spot premia.

B. A Factor-Model for Commodity Returns

Our next task is to investigate whether the different sorts capture different
types of risk factors or can be explained by one factor or a limited number of
factors. We therefore proceed with formal asset pricing tests to identify the
factor(s) that may price the various sorted portfolios.11

B.1. A Basis-Based Factor Model

Our starting point is again the basis sorts, from which we first construct a
factor portfolio based on the Holding returns for the two highest basis portfolios
(P3 + P4) minus the two lowest basis portfolios (P1 + P2). We start with the
Holding returns, as these consist of both spot and term premia and thus may
be able to capture all types of returns. We go long in an EW portfolio of the
10 commodities with the highest basis and short in an EW portfolio of the 10
commodities with the lowest basis. Using this factor portfolio, with Holding
return r H ML(n)t→t+n, we then test whether this portfolio can explain the risk
premia on the sorted portfolios using the regressions
(n)
rit→t+n = αi(n) + βi(n) rHML(n) (n)
t→t+n + εit→t+n, i = 1, . . . , 4, (16)
(n)
where i is the indicator for the four portfolios within each sort and rit→t+n is the
(n)
return on sorted portfolio i with maturity n. Note that for rit→t+n we use Short
Roll returns and Excess Holding returns. If the factor portfolio can explain the
portfolio sorts, standard asset pricing tests imply that αi(n) equals zero. We use
a Wald test estimated using Newey–West corrected standard errors to jointly
test whether the four αi(n) ’s in each sort are zero.12
Table IV reports the test results for the basis factor. The first two columns
(n)
present the results of the tests for Short Roll returns rit→t+n based on all the
10 Gorton, Hayashi, and Rouwenhorst (2013), using similar sorting techniques as we do here,
also cannot confirm the regression-based evidence for hedging pressure effects.
11 We also investigate a possible factor structure in the different sorts by analyzing the return

variance explained by their principal components. We find that the spot premia are related to one
factor (the first principal component of Short Roll and Spreading returns), whereas term premia
are related to one or two separate factors (the second and third principal components). We discuss
these results in detail in the Internet Appendix.
12 The reader may argue that this test relates commodity risk premia only to commodity fac-

tors, whereas asset pricing models like the CAPM or Consumption CAPM imply that these pre-
mia should be explained by the market factor or consumption risk. Many papers, however, show
472 The Journal of FinanceR

Table IV
Asset Pricing Tests for Basis Factor from Holding Returns
The table reports asset pricing tests for Short Roll and Excess Holding returns when futures
contracts are sorted on different instruments (basis, momentum, coefficient of variation (CV),
inflation β, hedging pressure (HP), and liquidity). We construct a single factor from Holding returns
on basis-sorted portfolios by forming a long-short portfolio, r H ML(n)
t→t+n, from the two highest basis
portfolios minus the two lowest basis portfolios, and estimate the following regression:
(n)
rit→t+n = αi(n) + βi(n) rHML(n) (n)
t→t+n + εit→t+n, i = 1, . . . , 4.

The first column gives the average absolute αi(n) of the four portfolios within a sort, and the second
column gives the p-values for the Wald test that these αi(n) ’s are zero in parentheses. Standard
errors are estimated using Newey–West correction. The returns are quoted bimonthly for a sample
period between March 1986 and December 2010.

Short Roll Excess Holding Short Roll Excess Holding

α(abs) p α(abs) p α(abs) p α(abs) p

Panel A. Returns Sorted on Basis Panel B. Returns Sorted on Momentum

n=1 0.60% (0.993) 1.28% (0.853)

n=2 0.63% (0.988) 0.48% (0.102) 0.95% (0.935) 0.49% (0.072)
n=3 2.27% (0.390) 0.82% (0.031) 2.26% (0.538) 0.81% (0.073)
n=4 1.24% (0.745) 0.71% (0.007) 1.95% (0.191) 0.61% (0.099)
Panel C. Returns Sorted on CV Panel D. Returns Sorted on Inflation β

n=1 1.92% (0.546) 2.09% (0.688)

n=2 2.35% (0.258) 0.76% (0.000) 1.53% (0.798) 0.74% (0.001)
n=3 3.03% (0.024) 1.12% (0.001) 3.35% (0.676) 1.15% (0.010)
n=4 2.63% (0.277) 1.11% (0.017) 2.13% (0.181) 1.13% (0.070)
Panel E. Returns Sorted on HP Panel F. Returns Sorted on Liquidity

n=1 2.06% (0.179) 2.11% (0.655)

n=2 2.51% (0.104) 0.50% (0.025) 1.73% (0.598) 0.75% (0.000)
n=3 2.24% (0.670) 0.87% (0.021) 3.00% (0.278) 1.01% (0.000)
n=4 2.01% (0.390) 0.70% (0.126) 2.62% (0.230) 1.09% (0.000)

sorts discussed earlier save those on dollar β and open interest, for which we
do not report any meaningful results in Table III. The first column gives the
average absolute αi(n) of the four portfolios within a sort, and the second column
gives the p-value for the Wald test that these αi(n) ’s are zero. By way of example,
the first four lines show that, when confronting the basis-sorted portfolios with
the basis factor, the average (absolute) αi(n) varies between 0.6% and 2.3% per

commodity returns to be basically unrelated to such marketwide factors. See, for example, Dusak
(1973), Black (1976), Carter, Rausser, and Schmitz (1983), Jagannathan (1985), Bessembinder
(1992), de Roon, Nijman, and Veld (2000), and Erb and Harvey (2006). We thus believe that, at
this stage, in order to obtain a better understanding of the structure of these premia within the
commodity markets, it is more useful to try to characterize the commodity risk premia in terms of
commodity factors.
 Anatomy of Commodity Futures Risk Premia 473

Figure 1. Average returns and basis factor β’s. This figure plots the average returns on 24
portfolios sorted on basis, momentum, coefficient of variation, inflation, hedging pressure, and
liquidity, and their β with respect to the basis factor.

annum across maturities, and the p-values of the Wald test show these αi(n) ’s to
be indistinguishable from zero.
As can be seen from the p-values of the Wald tests as well as from the
αi(n) ’s, the basis factor can explain almost all portfolio Short Roll returns for the
other sorted portfolios. The hypothesis of zero intercepts is rejected for only
one individual portfolio sort at the 5% level. The (absolute) αi(n) is about 2%
per annum for most sorted portfolios, and exceeds 3% per annum in only two
out of 24 cases. Overall, the basis factor does a good job explaining the sorted
portfolio Short Roll returns in our sample.
Figure 1 graphically shows the explanatory power of the basis factor for the
portfolio returns. For each maturity, the four panels show the relation between
βi(n) and the mean return for each of the four portfolios in every sort, resulting
in 24 portfolios. These graphs show that the mean returns line up with their
β with respect to the basis factor. The (absolute) correlations between the mean
returns and the β’s are all about 0.80.
474 The Journal of FinanceR

This is quite different from the story told by the next two columns in each
panel of Table IV, which show the test results for the Excess Holding returns
on the various portfolios. These returns, which capture the term premia on
the various sorts, are virtually unexplained by the Holding returns from the
basis factor. The Wald tests reject the zero intercepts in almost all sorts for all
maturities, and the αi(n) ’s are of the same order of magnitude as the mean
sorted portfolio returns. Thus, the basis factor (from Holding returns) ex-
plains almost all of the spot premia but cannot explain the term premia in our
sample.

B.2. Explaining Term Premia

Since the basis factor from Holding returns explains spot premia well but can-
not explain term premia, we first check whether term premia can be captured
by basing the factor portfolio r H ML(n) t→t+n on the Excess Holding or Spreading
returns, which are directly related to term premia. Although we might use
either to construct the factor portfolio, we prefer the Spreading returns, as
they contain all term premia for n = 1, 2, ... each period, whereas the Excess
Holding returns contain only one in each period, and all of them only in the n
consecutive periods. Having deemed them more informative about the different
term premia, we create the factor portfolio based on the Spreading returns for
the two highest basis portfolios (P3 + P4) minus the two lowest basis portfolios
(P1 + P2), which implies that we go long in the spreads, as in (11), for the 10
commodities with the highest basis, and short in the spreads for the 10 com-
modities with the lowest basis. Depending on the maturity n, we then roll the
spreads forward for n periods as in (12).
The first two columns in each panel of Table V clearly indicate that this
factor portfolio does not improve upon the factor portfolio based on the Holding
returns presented in Table IV. The Wald tests reject the hypothesis that the
αi(n) ’s are zero for all sorts and across all maturities, and the αi(n) ’s are themselves
similar in magnitude to the term premia estimated in Table III. Thus, one basis
factor cannot explain any of the term premia.
The last two columns in each panel of Table V report the results of similar
tests, but with two factors. That is, we do not create a high-minus-low basis
(n)
portfolio of spreads, but use the two portfolios separately: r Ht→t+n is the EW
average of the Spreading returns for the commodities with the highest ba-
sis; r L(n)
t→t+n is the EW average of the Spreading returns for the lowest basis
commodities. The tests are now based on the regression
(n)
rit→t+n = αi(n) + β Hi
(n) (n)
r Ht→t+n (n)
+ β Li r L(n) (n)
t→t+n + εit→t+n, i = 1, . . . , 4. (17)

The results of this two-factor model are very different, with the two basis factors
now able to capture almost all term premia across the sorts and maturities
save for the sorts on liquidity. The average absolute α’s are usually less than 40
basis points per year, with the exception of the sorts on liquidity, where almost
 Anatomy of Commodity Futures Risk Premia 475

Table V
Asset Pricing Tests for Basis Factor from Spreading Returns
The table reports asset pricing tests for Excess Holding returns when futures contracts are sorted
on different instruments (basis, momentum, coefficient of variation (CV), inflation β, hedging
pressure (HP), and liquidity). We use either one long-short factor, r H ML(n)
t→t+n, constructed from
Spreading returns on the two highest basis portfolios minus the two lowest basis portfolios, or the
(n)
two portfolios as two factors, r Ht→t+n and r L(n)
t→t+n. We estimate the following regressions:

(n)
rit→t+n = αi(n) + βi(n) rHML(n) (n)
t→t+n + εit→t+n, i = 1, . . . , 4,
(n)
rit→t+n = αi(n) + β Hi
(n)
rH (n) (n) (n) (n)
t→t+n + βLi rLt→t+n + εit→t+n, i = 1, . . . , 4.

The first column gives the average absolute αi(n) of the four portfolios within a sort, and the
second column gives the p-values for the Wald test that these αi(n) ’s are zero. Standard errors are
estimated using Newey–West correction. The returns are quoted bimonthly for a sample period
between March 1986 and December 2010.

One Factor Two Factors One Factor Two Factors

α(abs) p α(abs) p α(abs) p α(abs) p

Panel A. Returns Sorted on Basis Panel B. Returns Sorted on Momentum

n=2 0.50% (0.053) 0.07% (0.937) 0.49% (0.059) 0.09% (0.910)

n=3 0.75% (0.053) 0.15% (0.757) 0.75% (0.048) 0.18% (0.814)
n=4 0.79% (0.008) 0.40% (0.193) 0.74% (0.073) 0.25% (0.536)
Panel C. Returns Sorted on CV Panel D. Returns Sorted on Inflation β

n=2 0.75% (0.000) 0.21% (0.434) 0.73% (0.004) 0.08% (0.964)

n=3 1.08% (0.004) 0.24% (0.356) 1.09% (0.019) 0.14% (0.900)
n=4 1.24% (0.003) 0.38% (0.490) 1.35% (0.013) 0.38% (0.579)
Panel E. Returns Sorted on HP Panel F. Returns Sorted on Liquidity

n=2 0.52% (0.015) 0.18% (0.601) 0.73% (0.000) 0.30% (0.097)

n=3 0.81% (0.037) 0.20% (0.531) 1.01% (0.000) 0.75% (0.000)
n=4 0.85% (0.036) 0.25% (0.477) 1.28% (0.000) 0.89% (0.000)

all average absolute α’s exceed 50 basis points per annum and are highly
significant. But note from Table III that sorting on liquidity in itself does not
yield a clear pattern of term premia. We therefore interpret the failure of the
basis factors to explain the liquidity portfolios as a pure liquidity effect, rather
than as unexplained risk premia.
Thus, save for the liquidity sorts, two basis factors from Spreading returns
capture most of the cross-sectional variation in the term premia. These factors
are different from the basis factor that explains the spot premia, implying that
we need in total three factors to explain both spot and term premia. In the
Internet Appendix we find that the term premia cannot be explained from two
factors based on Holding returns, which would imply only two factors to explain
both spot and term premia.
476 The Journal of FinanceR

B.3. Alternative Factors

At this point, the reader may wonder whether only the basis factor can
explain the spot and term premia, or whether factors based on other forecasting
variables explain the various portfolio sorts as well. Because sorting on the
basis is only one way to capture time-variation in commodity risk premia, and
Table III shows sorting on other variables to result in meaningful risk premia
as well, we can also construct factors based on these alternative sorts.
(n)
Table VI addresses the question of whether the Short Roll returns rit→t+n
(which capture spot premia) for the various sorts can be explained by Holding
returns on factor portfolios r H ML(n) t→t+n that come from sorts other than the
basis. The table presents the average absolute αi(n) for all sorted portfolio Short
Roll returns and for different factor portfolios as well as the Wald test ( p-values)
that the four αi(n) ’s in each sort are zero. The columns in Table VI can thus be
compared to the first two columns in each panel of Table IV for the basis factor.
The horse race presented in Table VI shows that the various factor portfolios
can explain the own-portfolio sorts well, as well as the sorts on momentum,
inflation, and hedging pressure, but generally fail to explain the portfolios
sorted on basis and most of the volatility-sorted portfolios. The factor portfolios
also have difficulties with the sorts on liquidity, especially for the longer ma-
turities where liquidity is likely to play a more important role. Overall, none
of the factor portfolios come close to the performance of the basis factor (in
Table IV). The Wald tests reject the factor models in many more cases, and the
αi(n) ’s show much more unexplained return to be left on the table than in the
case of the basis factor. We conclude that spot premia are better characterized
by the basis factor than by any one of the other factors.
Finally, Table VII shows similar results for the term premia. To save space,
we only report whether two factors based on the various sorts are able to
explain the term premia from the basis sorts. The Internet Appendix shows
the explanatory power for the alternative factors for the other portfolio sorts as
well. The results in Table VII clearly show that, none of the alternative factors
are able to explain the term premia from sorting on the basis. In all cases but
one the hypothesis of zero intercepts is rejected at least at the 5% level and
the α’s themselves vary between 0.50% and 1% per year, double those from the
basis factors in Table V. We thus conclude again that none of the factors based
on the other forecasting variables come close to the explanatory power of the
two basis factors.

V. Summary and Conclusions

This paper analyzes the various risk premia present in commodity futures
markets that can be identified on the one hand when sorting commodity fu-
tures on characteristics such as the basis, volatility, and momentum, and on
the other hand by distinguishing contracts according to their maturity. A sim-
ple decomposition of futures returns shows futures expected returns to con-
sist of two risk premia: spot premia related to the risk in the underlying
 Anatomy of Commodity Futures Risk Premia 477

Table VI
Asset Pricing Tests for Alternative Factors from Holding Returns
The table reports asset pricing tests for Short Roll returns when futures contracts are sorted on
different instruments (basis, momentum, coefficient of variation (CV), inflation β, hedging pressure
(HP), and liquidity). We construct a single factor from Holding returns on portfolios sorted on each
of the instruments by forming a long-short portfolio, r H ML(n)
t→t+n, from the two highest portfolios
minus the two lowest portfolios within each sort, and estimate the following regression:
(n)
rit→t+n = αi(n) + βi(n) r H ML(n) (n)
t→t+n + εit→t+n, i = 1, . . . , 4.

The first column gives the average absolute αi(n)

of the four portfolios within a sort, and the second
column gives the p-values for the Wald test that these αi(n) ’s are zero in parentheses. Standard
errors are estimated using Newey–West correction. The returns are quoted bimonthly for a sample
period between March 1986 and December 2010.

Mom Factor CV Factor Infl Factor HP Factor Liquidity Factor

α(abs) p α(abs) p α(abs) p α(abs) p α(abs) p

Panel A. Returns Sorted on Basis

n=1 2.88% (0.105) 2.68% (0.219) 2.73% (0.134) 3.89% (0.018) 3.23% (0.077)
n=2 4.53% (0.000) 3.68% (0.007) 3.24% (0.009) 5.01% (0.000) 3.74% (0.003)
n=3 4.40% (0.000) 3.50% (0.001) 3.16% (0.000) 4.76% (0.000) 3.68% (0.000)
n=4 5.07% (0.000) 4.57% (0.009) 3.68% (0.009) 5.38% (0.000) 4.77% (0.001)

Panel B. Returns Sorted on Momentum

n=1 0.16% (1.000) 3.21% (0.278) 2.86% (0.295) 3.21% (0.270) 3.49% (0.187)
n=2 1.07% (0.739) 1.96% (0.627) 1.75% (0.735) 2.26% (0.384) 1.71% (0.540)
n=3 0.62% (0.933) 1.34% (0.797) 1.74% (0.862) 1.76% (0.611) 1.31% (0.820)
n=4 2.50% (0.728) 3.30% (0.821) 0.85% (0.957) 2.54% (0.741) 1.46% (0.992)

Panel C. Returns Sorted on Coefficient of Variation

n=1 2.12% (0.380) 2.00% (0.598) 2.75% (0.293) 2.48% (0.160) 2.37% (0.149)
n=2 2.30% (0.092) 2.19% (0.060) 2.94% (0.042) 2.81% (0.020) 2.54% (0.014)
n=3 2.85% (0.024) 2.59% (0.087) 2.53% (0.026) 3.30% (0.013) 3.05% (0.006)
n=4 3.95% (0.045) 2.88% (0.250) 3.02% (0.147) 4.65% (0.004) 3.92% (0.028)

Panel D. Returns Sorted on Inflation β

n=1 2.15% (0.610) 3.13% (0.372) 1.52% (0.952) 3.37% (0.296) 2.85% (0.385)
n=2 2.70% (0.233) 3.23% (0.122) 1.85% (0.890) 3.40% (0.124) 2.88% (0.199)
n=3 2.66% (0.198) 2.91% (0.172) 1.82% (0.558) 3.18% (0.133) 2.92% (0.158)
n=4 3.72% (0.070) 3.53% (0.014) 0.70% (0.727) 3.79% (0.079) 3.78% (0.061)

Panel E. Returns Sorted on Hedging Pressure

n=1 1.78% (0.221) 2.51% (0.192) 3.21% (0.142) 1.80% (0.215) 2.50% (0.216)
n=2 1.72% (0.299) 2.86% (0.075) 3.49% (0.020) 1.75% (0.178) 2.62% (0.072)
n=3 1.17% (0.750) 2.06% (0.441) 2.67% (0.250) 1.36% (0.562) 1.67% (0.444)
n=4 2.70% (0.348) 3.47% (0.150) 2.37% (0.299) 2.79% (0.325) 2.13% (0.349)

Panel F. Returns Sorted on Liquidity

n=1 1.96% (0.690) 3.20% (0.238) 1.95% (0.574) 2.52% (0.444) 1.12% (0.143)
n=2 1.67% (0.509) 3.35% (0.109) 1.70% (0.578) 2.16% (0.338) 1.19% (0.157)
n=3 1.84% (0.321) 3.70% (0.038) 2.48% (0.135) 2.52% (0.207) 1.88% (0.005)
n=4 3.27% (0.141) 4.93% (0.033) 2.57% (0.163) 3.25% (0.194) 2.92% (0.068)
478 The Journal of FinanceR

Table VII
Asset Pricing Tests for Alternative Factors from Spreading Returns
The table reports asset pricing tests for Excess Holding returns when futures contracts are sorted
on the basis. We construct two factors using Spreading returns on portfolios sorted on basis,
momentum, coefficient of variation, inflation β, hedging pressure, and liquidity. The first factor is
(n)
the return on the two highest basis portfolios r Ht→t+n and the second one is the return on the two
(n)
lowest basis portfolios r Lt→t+n. We estimate the following regression:
(n)
rit→t+n = αi(n) + β Hi
(n) (n)
r Ht→t+n (n)
+ β Li r L(n) (n)
t→t+n + εit→t+n, i = 1, . . . , 4.

The first column gives the average absolute αi(n)

of the four portfolios within a sort, and the second
column gives the p-values for the Wald test that these αi(n) ’s are zero in parentheses. Standard
errors are estimated using Newey–West correction. The returns are quoted bimonthly for a sample
period between March 1986 and December 2010.

α(abs) p α(abs) p

Basis Factor Mom Factor

n=2 0.07% (0.937) 0.16% (0.659)

n=3 0.15% (0.757) 0.47% (0.034)
n=4 0.40% (0.193) 0.68% (0.001)
CV Factor Infl Factor

n=2 0.29% (0.189) 0.29% (0.159)

n=3 0.48% (0.051) 0.45% (0.033)
n=4 0.60% (0.038) 0.74% (0.006)
HP Factor Liquidity Factor

n=2 0.26% (0.236) 0.19% (0.259)

n=3 0.52% (0.012) 0.52% (0.093)
n=4 0.69% (0.001) 0.72% (0.037)

commodity, and term premia related to the changes in basis. We show how
these different premia can be isolated using simple trading strategies. We find
that, in most cases, spot and term premia have opposite signs and are highly
predictable. Sorting on the futures basis, momentum, volatility, inflation, and
liquidity results in sizable spot premia in the high-minus-low portfolios be-
tween 5% and 14% per annum and term premia between 1% and 3% in absolute
value.
We also find that the cross-sectional patterns in spot premia based on these
characteristics can be captured by one basis factor, whereas two additional
factors are needed to explain term premia. Thus, for asset pricing models to
explain commodity futures risk premia, the challenge is to explain the basis-
sorted high-minus-low Holding portfolio for spot premia, and the high and low
Spreading portfolios for term premia.

Initial submission: September 22, 2011; Final version received: June 21, 2013
Editor: Campbell Harvey
 Anatomy of Commodity Futures Risk Premia 479

Appendix A: Relating Basis to Expected Futures Returns

We start by writing (13) for the n-period return for an n-period contract (the
Holding return),
St+n St+n
R(n)
Fut,t→t+n = = . (A1)
Ft(n) St (1 + RF ) (1 + U )n − Ct+n
n

Taking logs gives

 
St+n  
r (n)
f ut,t→t+n = ln = ln St+n − ln St (1 + RF )n (1 + U )n − Ct+n
Ft(n)
  
Ct+n
= ln St+n − ln St (1 + RF )n (1 + U )n −
St
 
Ct+n
= st+n − st − ln (1 + RF )n (1 + U )n − .
St

Note that the last term would be yt(n) in our setting. Proceeding with log returns,
we write this as
  
(n) n n Ct+n/St
r f ut,t→t+n = st+n − st − ln (1 + RF ) (1 + U ) 1 − n n
(1 + RF ) (1 + U )
= st+n − st − n (r f + u) − ln (1 − exp (ct+n − st − n (r f + u))) .
Following Campbell and Shiller (1988), the last term on the right-hand side,
n (r f + u) − nyt(n) , can be approximated using a first-order Taylor series expan-
sion,
ln (1 − exp (ct+n − st − n (r f + u)))
 
   exp cn − s − n (r f + u)
≈ ln 1 − exp cn − s − n (r f + u) +  
1 − exp cn − s − n (r f + u)
 
× ct+n − st − cn − s .
  
Defining ρn = 1/ 1 − exp cn − s − n (r f + u) , the log futures return can be
written as

r (n)
f ut,t→t+n ≈ κn + st+n − st − n (r f + u) + (1 − ρn) (ct+n − st − n (r f + u))

= κn + st+n − ρnst − ρnn (r f + u) + (1 − ρn)ct+n,

θnr (n)
f ut,t→t+n ≈ κn + θn (st+n − n (r f + u)) + (1 − θn) (ct+n − n (r f + u)) − st . (A2)

Here, κn contains all the constant terms (including r f and u) and θn = 1/ρn.
As in Campbell and Shiller, we can now solve forward

κn
∞  
st = + θnj (1 − θn) ct+n+ jn − r (n) − n (r f + u) .
1 − θn f ut,t+ jn→t+( j+1)n
j=0
480 The Journal of FinanceR

Ct+n/St
Note that for 0 < θn < 1, we need the average value of (1+RF) n
(1+U )
n to be be-

tween zero and one. This means that the average cash yield must be strictly
positive, and that, on average, the cash yield cannot exceed the current spot
price compounded at the risk-freerate and storage costs. Taking expectations
and rewriting gives
⎡ ⎤
κn
∞  
st − ct+n = + Et ⎣ θnj ct+( j+1)n − r (n)
f ut,t+ jn→t+( j+1)n − n (r f + u)
⎦.
1 − θn
j=0

(A3)

For our purposes, it is useful to subtract n (r f + u) from both sides and use the
definition of the spot and term premia

yt(n) ≈ st + n (r f + u) − ct+n
⎡ ⎤
∞  
κn ⎣
n−1 n−1
⎦.
= + Et θn ct+( j+1)n −
j
πs,t+i − π (i) (A4)
1 − θn i=0 i=0 y,t+i
j=0

For n = 1 this simplifies to (14).

Appendix B: Forecasting Variables

Momentum: We sort on momentum by sorting on the cumulative log return
from month t − 12 to t − 1.
Coefficient of Variation: As in Dhume (2011), we use the coefficient of vari-
ation as a measure of volatility, that is, variance scaled by mean return. We
calculate the coefficient of variation over the period t − 36 to t − 1. Scaling the
variance by the mean return can be interpreted as correcting the volatility
effect for a momentum effect.
Inflation β: We use commodities inflation β from a 60-month rolling regres-
sion of monthly commodity futures returns on unexpected inflation, measured
by the change in one-month CPI inflation. In the Internet Appendix, we use
two additional measures of unexpected inflation, namely inflation minus the
risk-free interest rate, and inflation minus its prediction from an ARIMA model.
Dollar β: We use commodities dollar β from a 60-month rolling regression
of monthly commodity futures returns on changes in the U.S. dollar versus a
basket of foreign currencies.
Hedging Pressure: The hedging pressure variable in a futures market is
defined as the difference between the number of short and the number of long
hedge positions by large traders relative to the total number of hedge positions
by large traders in that market,

# of short hedge positions − # of long hedge positions

hpt = ,
total # of hedge positions
 Anatomy of Commodity Futures Risk Premia 481

where positions are measured by the number of contracts in the market. Hedg-
ing pressure is calculated using data published in the Commitment of Traders
reports issued by the Commodity Futures Trading Commission (CFTC).
Open Interest: Following Hong and Yogo (2012), we use the total open interest
in a futures market.
Liquidity: Following Marshall, Nguyen, and Visaltanachoti (2012), we use
the Amivest measure (Amihud, Mendelson, and Lauterbach (1997)) for liq-
uidity, which divides the volume on a trading day by the absolute return on
that trading day. The bimonthly measure is the average of the daily Amivest
measures over the 2-month period.

REFERENCES
Amihud, Yakov, Haim Mendelson, and Beni Lauterbach, 1997, Market microstructure and securi-
ties values: Evidence from the Tel Aviv Stock Exchange, Journal of Financial Economics 45,
365–390.
Asness, Cliff, Tobias Moskowitz, and Lasse Heje Pedersen, 2013, Value and momentum everywhere,
Journal of Finance 68, 929–985.
Bessembinder, Hendrik, 1992, Systematic risk, hedging pressure, and risk premiums in futures
markets, Review of Financial Studies 5, 637–667.
Bessembinder, Hendrik, and Kalok Chan, 1992, Time-varying risk premia and forecastable returns
in futures markets, Journal of Financial Economics 32, 169–193.
Black, Fischer, 1976, The pricing of commodity contracts, Journal of Financial Economics 3, 167–
179.
Campbell, John Y., and Robert J. Shiller, 1988, Stock prices, earnings and expected dividends,
Journal of Finance 43, 661–676.
Campbell, John Y., and Robert J. Shiller, 1991, Yield spreads and interest rate movements: A bird’s
eye view, Review of Economic Studies 58, 495–514.
Carter, Colin A., Gordon C. Rausser, and Andrew Schmitz, 1983, Efficient asset portfolios and the
theory of normal backwardation, Journal of Political Economy 91, 319–331.
Casassus, Jaime, and Pierre Collin-Dufresne, 2005, Stochastic convenience yield implied from
commodity futures and interest rates, Journal of Finance 60, 2283–2331.
Chang, Eric C., 1985, Return to speculators and the theory of normal backwardation, Journal of
Finance 40, 193–208.
Cochrane, John H., 2008, Decomposing the yield curve, Working paper, Chicago Booth.
Cochrane, John H., 2011, Discount rates, Journal of Finance 66, 1047–1108.
Cochrane, John H., and Monika Piazzesi, 2005, Bond risk premia, American Economic Review 95,
138–160.
Cox, John, Jonathan Ingersoll, and Stephen Ross, 1981, A reexamination of traditional hypotheses
about the term structure of interest rates, Journal of Finance 36, 321–346.
Dhume, Deepa, 2011, Using durable consumption risk to explain commodities returns, Working
paper, Harvard University.
Dusak, Katherine, 1973, Futures trading and investor returns: An investigation of commodity
market risk premiums, Journal of Political Economy 81, 1387–1406.
Erb, Claude B., and Campbell R. Harvey, 2006, The strategic and tactical value of commodity
futures, Financial Analysts Journal 62, 69–97.
Fama, Eugene F., 1984, Forward and spot exchange rates, Journal of Monetary Economics 14,
319–338.
Fama, Eugene F., 1986, Term premiums and default premiums in money markets, Journal of
Financial Economics 17, 175–196.
Fama, Eugene F., and Kenneth R. French, 1987, Commodity futures prices: Some evidence on
forecast power, premiums, and the theory of storage, Journal of Business 60, 55–73.
482 The Journal of FinanceR

Fama, Eugene F., and Kenneth R. French, 1988, Business cycles and the behavior of metal prices,
Journal of Finance 43, 1075–1093.
Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in the returns on stocks and
bonds, Journal of Financial Economics 33, 3–56.
Fama, Eugene F., and Kenneth R. French, 1996, Multifactor explanations of asset pricing anoma-
lies, Journal of Finance 51, 55–84.
Fama, Eugene F., and Robert R. Bliss, 1987, The information in long maturity forward rates,
American Economic Review 77, 680–692.
Frank, Julieta, and Philip Garcia, 2009, Time-varying risk premium: Further evidence in agricul-
tural futures markets, Applied Economics 41, 715–725.
Gorton, Gary B., Fumio Hayashi, and K. Geert Rouwenhorst, 2013, The fundamentals of commodity
futures returns, Review of Finance 17, 35–105.
Gorton, Gary, and Geert K. Rouwenhorst, 2006, Facts and fantasies about commodity futures,
Financial Analysts Journal 62, 47–68.
Gourinchas, Pierre-Olivier, and Helene Rey, 2007, International financial adjustment, Journal of
Political Economy 115, 665–703.
Greer, Robert J., 2000, The nature of commodity index returns, Journal of Alternative Investments
3, 45–52.
Hansen, Lars Peter, and Robert J. Hodrick, 1980, Forward exchange rates as optimal predictors
of future spot rates: An econometric analysis, Journal of Political Economy 88, 829–853.
Hong, Harrison, and Motohiro Yogo, 2012, What does futures market interest tell us about the
macroeconomy and asset prices? Journal of Financial Economics 105, 473–490.
Jagannathan, Ravi, 1985, An investigation of commodity futures prices using the consumption-
based intertemporal capital asset pricing model, Journal of Finance 40, 175–191.
Koijen, Ralph S.J., Tobias J. Moskowitz, Lasse Heje Pedersen, and Evert B. Vrugt, 2012, Carry,
Working paper, Chicago Booth.
Liu, Peng, and Ke Tang, 2011, The stochastic behavior of commodity prices with heteroskedasticity
in the convenience yield, Journal of Empirical Finance 18, 211–224.
Lustig, Hanno, Nikolai Roussanov, and Adrien Verdelhan, 2011, Common risk factors in currency
markets, Review of Financial Studies 24, 3731–3777.
Marshall, Ben R., Nhut H. Nguyen, and Nuttawat Visaltanachoti, 2012, Commodity liquidity
measurement and transaction costs, Review of Financial Studies 25, 599–638.
Miffre, Joelle, and Georgios Rallis, 2007, Momentum strategies in commodity futures markets,
Journal of Banking and Finance 31, 1863–1886.
Piazzesi, Monika, and Eric Swanson, 2008, Futures prices as risk-adjusted forecasts of monetary
policy, Journal of Monetary Economics 55, 677–691.
de Roon, Frans A., Theo E. Nijman, and Chris H. Veld, 1998, Pricing term structure risk in futures
markets, Journal of Financial and Quantitative Analysis 33, 139–157.
de Roon, Frans A., Theo E. Nijman, and Chris H. Veld, 2000, Hedging pressure effects in futures
markets, Journal of Finance 55, 1437–1456.
Yang, Fan, 2013, Investment shocks and the commodity basis spread, Journal of Financial Eco-
nomics, 110, 164–184.

Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s web site:
Appendix S1: Internet Appendix