Commodity Market Interest and Asset Return. Predictability

Transcription

1 Commodity Market Interest and Asset Return Predictability Harrison Hong Motohiro Yogo March 25, 2010 Abstract We establish several new findings on the relation between open interest in commodity markets and asset returns. High commodity market activity, as measured by high open-interest growth, predicts high commodity returns and low bond returns. Openinterest growth is a more powerful and robust predictor of commodity returns than other known predictors such as the short rate, the yield spread, the basis, and hedging pressure. Although positively correlated with commodity returns, open-interest growth contains information for future asset returns beyond contemporaneous commodity prices. Open-interest growth also predicts changes in inflation and inflation expectations. These findings suggest that open-interest growth contains information about future inflation that gets priced into commodity and bond markets with delay. Our findings are consistent with recent theories of gradual information diffusion and have implications for macroeconomic forecasting models. This paper subsumes our earlier work titled Digging into Commodities. For comments and discussions, we thank Erkko Etula, Hong Liu, David Robinson, Nikolai Roussanov, Allan Timmermann, and seminar participants at Boston College, Centre de Recherche en Economie et Statistique, Dartmouth College, Federal Reserve Bank of Chicago, Fordham University, Imperial College London, Ohio State University, PanAgora Asset Management, Stockholm School of Economics, University of California San Diego, University of Minnesota, University of Pennsylvania, University of Southern California, University of Texas at Austin, Washington University in St. Louis, the 2008 Economic Research Initiatives at Duke Conference on Identification Issues in Economics, and the 2010 Annual Meeting of the American Finance Association. We thank Jennifer Kwok, Hui Fang, Yupeng Liu, James Luo, Thien Nguyen, and Elizabeth So for research assistance. Hong acknowledges a grant from the National Science Foundation. Yogo acknowledges a grant from the Rodney L. White Center for Financial Research at the University of Pennsylvania. Princeton University and NBER ( hhong@princeton.edu) University of Pennsylvania and NBER ( yogo@wharton.upenn.edu)

2 1. Introduction We analyze how open interest in commodity markets is related to commodity and bond returns. Our analysis is motivated by the recent volatility in commodity prices and the renewed interest in the behavior of these markets, which have not been seen since the energy crisis of the 1970s. Once largely ignored by the investment community, commodities have emerged as an important asset class. By some estimates, index investment in this asset class increased from $13 billion at the end of 2003 to $317 billion in July 2008, just prior to the financial crisis (Masters and White, 2008). During the same period, the influx of new investors led to elevated levels of activity as measured by open interest in commodity futures, which grew from $103 billion to $509 billion. This elevated activity has led to inquiries about how trading affects asset price formation in these markets, underscored by the recent Congressional hearings on the impact of excessive speculation. Hence, understanding the link between open interest and commodity price fluctuations is important not only for investors but also for public policy. Ouranalysiscovers30commoditiesacrossfour sectors (agriculture, energy, livestock, and metals) over the period of 1965 through Using hand-collected data on open interest from the Commitments of Traders since 1965, we establish several new findings on the relation between open interest and returns in commodity markets. Figure 1 summarizes our main finding. The first series is the percentage change in open interest over the previous 12 months, averaged across all commodities. The second series is the return on fully collateralized commodity futures over the previous 12 months, averaged across all commodities. During the recent commodity boom from 2003 to 2008, open interest grew at a persistently high rate, more so than in any other period over the previous thirty years. Only the energy crisis of the 1970s witnessed higher activity. During these two historic periods and also more generally, open-interest growth and commodity returns are highly correlated. But the most interesting finding in this plot is that open-interest growth seems to lead commodity returns. In other words, high commodity market activity appears to predict subsequent appreciation 2

3 of commodity prices. To formalize our observation, we regress the monthly excess returns on a portfolio of commodity futures onto lagged 12-month open-interest growth. We find that a standard deviation increase in open-interest growth increases expected commodity returns by 0.64% per month. Similarly, we find that a standard deviation increase in open-interest growth increases expected spot-price growth by 0.41% per month. Both of these estimates are economically large and statistically significant. Open-interest growth is a more powerful and robust predictor than a number of other variables that are known to predict commodity returns. These include common predictors such as the short rate and the yield spread and commodity-specific predictors such as aggregate basis (i.e., the ratio of futures to spot price averaged across commodities) and aggregate hedging pressure (i.e., the net short position of hedgers averaged across commodities). 1 Open-interest growth is a more robust predictor than these other variables in two important ways. First, aggregate open-interest growth predicts returns on sector portfolios, in contrast to other variables that predict returns for only particular sectors. Second, open-interest growth is the only variable that continues to demonstrate forecasting power in the most recent period since 1987, when there are the greatest number of commodities in the database. Open-interest growth is most closely related to 12-month commodity returns. We find that past aggregate commodity returns forecast the subsequent month s return. In other words, there is momentum in the time series of aggregate commodity returns. However, in a horse race between these variables, open-interest growth entirely drives out the forecasting power of past commodity returns. This means that open-interest growth contains information about future returns that is not fully captured by past commodity prices. A potential 1 Bessembinder and Chan (1992) are the first to establish that the same variables that predict bond and stock returns (such as the short rate, the default spread, and the dividend yield) also predict commodity returns. There is mixed evidence that basis predicts returns on commodity futures. Fama and French (1987) are the first to establish that basis predicts returns for some commodities. They emphasize that there is more consistent evidence for the theory of storage. A number of other studies have documented mixed evidence for the theory of backwardation, controlling for systematic risk and using an empirical proxy for hedging pressure (Carter, Rausser, and Schmitz, 1983; Chang, 1985; Bessembinder, 1992; de Roon, Nijman, and Veld, 2000). 3

4 interpretation of these findings is that open interest in commodity markets rises in response to news, which get impounded into commodity prices with delay. In the 1970s, for example, there were news about supply shocks to oil. In the most recent period, there were news about strong demand for commodities from the emerging economies. To test this hypothesis, we examine whether open-interest growth predicts inflation and excess bond returns. Consistent with the hypothesis, we find that high open-interest growth predicts rising inflation and also a rising nominal short rate. In addition, high open-interest growth predicts low bond returns with a t-statistic over 3. A standard deviation increase in open-interest growth decreases expected bond returns by 0.32% per month. Open-interest growth is the only predictor that survives in the most recent period since 1987, when the short rate and the yield spread fail to predict bond returns. Hence, open-interest growth not only contains powerful information about future commodity returns, but also important information about future bond returns. To summarize, our novel finding is that commodity market activity contains information about inflation news and bond returns that is not fully captured by commodity prices. As we discuss in the paper, our findings are most consistent with recent theories of gradual information diffusion in asset markets (see Hong and Stein, 2007, for a review). These theories suggest that when market prices under-react to news, trading activity emerges as a useful additional predictor of future returns. Moreover, the fact that open interest can be useful for predicting economic activity like inflation expectations has important implications for macroeconomic forecasting models. This paper is part of a set of recent studies that have emerged in response to renewed interest in commodity markets. These studies build on earlier important work, most notably by Fama and French (1987) and Bessembinder and Chan (1992). In a very interesting paper, Gorton and Rouwenhorst (2006) emphasize that commodities have a high Sharpe ratio and low correlations with other asset classes. They argue that this evidence is consistent with the theory of backwardation in particular and market segmentation more generally. de Roon 4

5 et al. (2009) find that returns on commodity futures are predictable by both macroeconomic predictors and the basis. Acharya, Lochstoer, and Ramadorai (2009) find that producers hedging demand, as captured by their default risk, predicts commodity returns. Etula (2009) finds that the supply of speculator capital, as captured by changes in broker-dealer balance sheets, predicts commodity returns, especially in energy. Relative to these studies, we share the view that market segmentation is a key driver of predictability in commodity markets. However, our focus on the implications of gradual information diffusion is quite different from these other studies that focus on limited risk-bearing capacity in commodity markets. More closely related is a group of studies that document momentum in the cross section of commodity returns (Erb and Harvey, 2006; Gorton, Hayashi, and Rouwenhorst, 2007; Miffre and Rallis, 2007; Asness, Moskowitz, and Pedersen, 2009). We find momentum in the time series of aggregate commodity returns, which interacts with open interest in commodity markets. The rest of the paper proceeds as follows. Section 2 describes the commodity market data and the construction of the key variables used in our empirical analysis. Section 3 reports summary statistics for commodity returns, spot-price growth, and the predictor variables. Section 4 presents our main finding that open-interest growth predicts commodity returns. We also present evidence that open-interest growth predicts inflation news and bond returns to illuminate the economic mechanism behind our findings. Section 5 concludes. 2. Commodity Market Data and Definitions 2.1. Commodity Market Data Our data on commodity prices are from the Commodity Research Bureau, which has daily prices for individual futures contracts as well as spot prices for many commodities beginning in December Gorton and Rouwenhorst (2006) also use this database, and additional details can be found in the appendix to their paper. As they point out, the database mostly 5

6 contains data for contracts that have survived until the present or that were in existence for an extended period between 1965 and the present. Many different types of contracts fail to survive because of lack of interest from market participants, and they are consequently not recorded in the database. Consequently, the computed returns on commodity futures may be subject to survivorship bias. Following Gorton and Rouwenhorst (2006), we work with a broad set of commodities contained in the database. Table 1 is a list all our commodities, together with the date of the first recorded futures price for each commodity. We categorize commodities into four broad sectors. Agriculture consists of 15 commodities and tends to contain the oldest contracts. Energy consists of five commodities. Heating oil is the oldest contract in energy, which starts in November Data for crude oil are available only since March Livestock consists of five commodities, and metals consists of six commodities. A potential concern with using a broad set of commodities is that not all contracts are liquid. In results that are not reported here, we have confirmed our main findings on a subset of 17 relatively liquid commodities that are in the AIG Commodities Index. Following Gorton and Rouwenhorst (2006), we exclude futures contracts with one month or less to maturity. These contracts are typically illiquid because futures traders do not want to take delivery of the underlying physical commodity. We therefore rule out investment strategies that require holding futures contracts to maturity. While Gorton and Rouwenhorst (2006) isolate the contract that is closest to maturity for each commodity, we include all contracts with more than one month to maturity. We also use data on open interest (i.e., the number of futures contracts outstanding) as well as the long and short positions of noncommercial traders (or hedgers ) for each commodity. Since January 1986, the data are available electronically from the Commodity Futures Trading Commission. Prior to that date, we hand-collected data from various volumes of the Commitments of Traders in Commodity Futures. Data for December 1964 through June 1972 are from the Commodity Exchange Authority ( ). Data for 6

7 July 1972 through December 1985 are from the Commodity Futures Trading Commission ( ). There is a 11 month gap from January through November of 1982, during which the Commodity Futures Trading Commission did not collect data due to budgetary reasons. Figure 2 shows the share of total dollar open interest that each sector represents. The figure shows that agriculture dominates the early part of the sample, while energy becomes the biggest sector later in the sample. The relative size of the four sectors is much more balanced in the second half of the sample starting in These stylized facts have two important implications for our empirical analysis. First, we construct the aggregate commodity portfolio as an equal-weighted portfolio of the four sectors, which ensures that the portfolio composition is consistent throughout the sample. Second, we examine the predictability of commodity returns by sector and in subsamples to check the robustness of our main results. The subsample since 1987 is perhaps more representative of what we can expect from commodity markets going forward because it has a more balanced representation across the four sectors Aggregate Commodity Returns To construct aggregate commodity returns, we first compute the return on a fully collateralized position in commodity futures as follows. Let R f,t be the monthly gross return on the 1-month T-bill in month t, which is assumed to be the interest earned on collateral. Let F i,t,t be the price of a futures contract on commodity i at the end of month t, which matures at the end of month T. The monthly gross return on a fully collateralized long position in commodity i with maturity T t is R i,t,t = F i,t,t R f,t F i,t 1,T. (1) We sort the universe of commodity futures into four sectors and two levels of maturity. 7

8 We define short maturity contracts as those with more than one but no more than three months to maturity. Long maturity contracts are those with more than three months to maturity. We then construct eight equal-weighted portfolios of commodity futures, corresponding to two levels of maturity for each of the four sectors. For each portfolio, we compute its monthly gross return as an equal-weighted average of returns on fully collateralized commodity futures. Finally, we construct an aggregate commodity portfolio as an equal-weighted portfolio of the these eight portfolios. The aggregate commodity portfolio that results from this construction is consistently balanced with respect to sector and maturity. For some of our analysis, it is useful to look at commodity returns separately by sector and maturity. Using the eight portfolios, we construct four sector portfolios as an equalweighted portfolio of the short-maturity and long-maturity portfolio for each sector. For example, the agriculture portfolio is an equal-weighted portfolio of the short-maturity and the long-maturity portfolio for agriculture. Using the eight portfolios, we also construct two maturity-sorted portfolios as an equal-weighted portfolio of the four sector portfolios for each level of maturity. For example, the short-maturity portfolio is an equal-weighted portfolio of the short-maturity portfolio for agriculture, energy, livestock, and metals. Our construction of the aggregate commodity portfolio differs from that of Gorton and Rouwenhorst (2006), who construct their portfolio by equal-weighting all commodities that exist at each point in time. The advantage of our approach is that the sectors are always equal-weighted, and hence no sector dominates even as the number of commodities within each sector changes over time. Despite the differences in the construction, the summary statistics for our aggregate commodity portfolio (reported in Section 3) are very close to those reported by Gorton and Rouwenhorst (2006) Aggregate Spot-Price Growth We construct aggregate spot-price growth in analogy to our construction of aggregate commodity returns. Let S i,t be the spot price of commodity i at the end of month t. The monthly 8

9 spot-price growth for commodity i is G i,t = S i,t S i,t 1. (2) We compute spot-price growth for each sector as an equal-weighted average of all commodities in that sector. We then compute aggregate spot-price growth as an equal-weighted average spot-price growth across the four sectors. The reason we examine spot-price growth separately from futures price movements is that spot-price growth is economically different from the return on commodity futures for two important reasons. First, unlike investment in physical commodities, commodity futures are financial investments like bonds and stocks. Therefore, the results for commodity futures are more meaningful than those for the spot price from a pure investment or asset allocation perspective. Second, as we discuss below, different theories have different implications for mean reversion in futures versus spot prices. Therefore, the way in which a variable predicts futures versus spot prices can help us discriminate among different theories Predictor Variables Our key predictor variable is aggregate open-interest growth. To construct this variable, we first compute the dollar open interest for each commodity as the spot price times the number of contracts outstanding. We then aggregate dollar open interest within each sector and compute its monthly growth rate. Finally, we compute aggregate open-interest growth as an equal-weighted average of open-interest growth across the four sectors. 2 Because monthly open-interest growth is noisy, we smooth it by taking a 12-month geometric average in the time series. A variable that is closely related to open-interest growth is the 12-2 We have tried an alternative construction that uses only the number of contracts outstanding and does not involve the spot price. We first compute the growth rate of open interest for each commodity. We then compute the median of open-interest growth across all commodities within each sector. Finally, we compute aggregate open-interest growth as an equal-weighted average of open-interest growth across the four sectors. This alternative construction leads to a time series that is very similar to our preferred construction of open-interest growth. 9

10 month geometric average of aggregate commodity returns. We use this variable to test for momentum in aggregate commodity returns. In addition to open-interest growth, we consider a variety of predictor variables that are known to predict commodity returns, which can grouped into two categories (see de Roon et al., 2009, for a similar list). The first category consists of aggregate market predictors, which are motivated by theories like the (I)CAPM that view commodity markets as being fully integrated. According to this view, commodity prices are driven by aggregate market predictors that influence portfolio allocation decisions across different asset classes. An implication of this theory is commodities may be useful for hedging time-varying investment opportunities in other asset classes. Hence, commodity prices may be high, or its expected returns may be low, when such hedging motives are important. Since an investor can hedge market fluctuations by either entering futures contracts or by holding physical commodities, this theory implies that the same aggregate market predictors should predict returns on commodity futures and spot-price growth with similar sign and magnitude. We focus on two aggregate market variables that are known to predict the common variation in bond and stock returns: the short rate and the yield spread (Fama and Schwert, 1977; Campbell, 1987; Fama and French, 1989). These variables are also known to predict commodity returns (Bessembinder and Chan, 1992; Bjornson and Carter, 1997). The short rate is the monthly average yield on the 1-month T-bill. The yield spread is the difference between Moody s Aaa corporate bond yield and the short rate. In analysis that is not reported here, we have also experimented with other predictor variables. In particular, we have examined the dividend yield for a value-weighted portfolio of NYSE, AMEX, and Nasdaq stocks. We have also examined the default spread (i.e., the difference between Moody s Baa and Aaa corporate bond yields) and measures of aggregate stock market volatility (i.e., both realized volatility and the VIX). Although these variables predict returns individually, their incremental contribution is weak once we control for the short rate and the yield spread. The second category consists of those variables that are motivated by the view that 10

11 commodity markets are segmented to some degree. The basis, or the ratio of futures to spot price, emerges as a particularly important variable in these theories. In the theory of backwardation, producers take a (long or short) position in commodity futures to hedge their position in the underlying spot (Keynes, 1923; Hicks, 1939). Risk averse speculators demand a risk premium for providing insurance. The excess demand for insurance, or hedging pressure, drives a wedge between the futures price and the expected future spot price. In the theory of storage, a low inventory causes the spot price to be temporarily high, which subsequently mean reverts as supply-demand imbalances correct (e.g., Deaton and Laroque, 1992). Note that the theory of backwardation and storage have very different implications for mean reversion in futures versus spot prices. The theory of backwardation implies that a low basis predicts high returns on commodity futures, but it remains silent about spot-price movements. In contrast, the theory of storage implies that a low basis predicts low spot-price growth, but it remains silent about futures-price movements. We construct aggregate basis in analogy to our construction of aggregate commodity returns. We first compute the basis for each commodity i with maturity T t as Basis i,t,t = ( Fi,t,T S i,t ) 1 T t 1. (3) While the Commodity Research Bureau has a reliable record of spot prices, a spot price is not always available on the same trading day as a recorded futures price. In instances where the spot price is missing, we first try to use an expiring futures contract to impute the spot price. If an expiring futures contract is not available, we then use the last available spot price within 30 days to compute the basis. For example, if we have a futures price on December 31, but the last available spot price is from December 30, we compute the basis as the ratio of the futures price on December 31 to the spot price on December 30. We then compute the median of basis within each of eight portfolios, corresponding to four sectors and two levels of maturity. We use the median, instead of the mean, because it 11

12 is less sensitive to outliers in the basis for individual futures contracts. Finally, we compute aggregate basis as an equal-weighted average of the basis across the eight portfolios. Recall that the basis is the implied net convenience yield derived from the cost-of-carry relation. The net convenience yield is defined as the riskless interest rate, plus additional storage costs, minus the convenience usage earned from owning the spot. Hence, aggregate basis identifies the common variation in the net convenience yield across all commodities. In addition to the basis, the theory of backwardation implies that a direct measure of hedging pressure should be correlated with the basis and also predict returns on commodity futures. We construct an aggregate version of hedging pressure as defined by de Roon, Nijman, and Veld (2000). We first construct hedging pressure for each sector as the ratio of two objects. The numerator is the dollar value of short minus long positions held by noncommercial traders in the Commitments of Traders, summed across all commodities in that sector. The denominator is the dollar value of short plus long positions held by noncommercial traders, summed across all commodities in that sector. Finally, we compute aggregate hedging pressure as an equal-weighted average of hedging pressure across the four sectors. 3. Summary Statistics of Commodity Markets 3.1. Commodity Returns Panel A of Table 2 reports the summary statistics for monthly excess returns over the 1- month T-bill rate. The aggregate commodity portfolio has a mean of 0.58% and a standard deviation of 4.05%. This corresponds to an annualized average excess return of 6.96% and an annualized standard deviation of 14.03%. During the same period, the 10-year Treasury bond has an annualized average excess return of 2.04% and an annualized standard deviation of 8.00%. The CRSP value-weighted market portfolio for NYSE, AMEX, and Nasdaq stocks has an annualized average excess return of 4.32% and an annualized standard deviation of 12

13 15.66%. The Sharpe ratio for the aggregate commodity portfolio was higher than that for stocks in this sample period, which is emphasized by Gorton and Rouwenhorst (2006). The table also reports the autocorrelation and the cross correlation of excess returns. The first-order autocorrelation for aggregate commodity returns is 0.08, which is comparable to that for bond and stock returns. Aggregate commodity returns have a correlation of 0.11 with bond returns and a correlation of 0.07 with stock returns. Because commodities have a high Sharpe ratio and low correlation with bonds and stocks, it is an attractive asset class from the perspective of diversifying the investment portfolio. The table also reports the summary statistics for the four sector portfolios. Both the mean and the standard deviation of excess returns have the same ordering across the four sectors. Agriculture has the lowest average excess return at 0.27% per month and the lowest standard deviation at 4.25% per month. Livestock has an average excess return of 0.47% per month and a standard deviation at 4.87% per month. Metals have an average excess return of 0.59% per month and a standard deviation of 7.40% per month. Energy has the highest average excess return at 0.91% per month and the highest standard deviation at 8.08% per month Predictor Variables Table 3 reports the summary statistics for the predictor variables. Aggregate basis has a mean of 0.03% and a standard deviation of 0.84%. Its autocorrelation is 0.69, which is lower than that for the short rate and the yield spread. This suggests that aggregate basis is a predictor variable that operates at a higher frequency than the aggregate market predictors. Aggregate basis has a positive correlation of 0.22 with the short rate and a negative correlation of 0.11 with the yield spread. Recall that the net convenience yield is defined as the riskless interest rate, plus additional storage costs, minus the convenience usage earned from owning the spot. Because aggregate basis is the average net convenience yield across commodities, the positive correlation between aggregate basis and the short rate 13

14 is unsurprising (Fama and French, 1988; Bailey and Chan, 1993). Figure 3 shows aggregate basis together with the spot-price index for an equal-weighted portfolio of commodities. The spot-price index is the cumulative growth rate of aggregate spot-price growth, deflated by the consumer price index. Note that movements in the spotprice index tend to be inversely related to movements in aggregate basis. When the spot-price index rises, aggregate basis tends to fall. This is due to mean reversion in the spot price. An increase in the spot price, due to a transitory demand shock for instance, will not lead to a one-for-one movement in the futures price because the futures market anticipates mean reversion. Since 2003, there is a break in this relation between the spot price and the basis, which interestingly coincides with the period of high open interest in commodity markets. Although spot prices have appreciated considerably, aggregate basis has risen if anything. This implies that futures prices have responded at least (if not more than) one-for-one with spot-price movements. This striking movement in futures prices is unprecedented, even if we consider the energy crisis of the 1970s. A potential explanation for the recent experience is that investors believe that there were permanent, or highly persistent, shocks to the demand for commodities. Another explanation is the conventional wisdom that more capital flowed into commodity futures than into the spot market in this period, as commodity index investors chased returns. 3 This story provides additional motivation for us to study the relation between open-interest growth and returns in commodity markets. Open-interest growth has a mean of 1.47% and a standard deviation of 2.07%. Its autocorrelation is 0.90, which arises from the fact that open-interest growth is smoothed as a 12-month moving average. Open-interest growth is essentially uncorrelated with aggregate basis, but it has a positive correlation of 0.31 with hedging pressure. Unsurprisingly, 12- month open-interest growth has a high contemporaneous correlation of 0.50 with 12-month 3 Pindyck and Rotemberg (1990) find excess co-movement among commodity prices and conjecture that this is due to speculators moving in and out of commodities as an asset class. In related work, Tan and Xiong (2009) find that commodities have co-moved more with stocks in the recent period, which they attribute to the indexation of commodity markets. 14

15 commodity returns. This imperfect correlation suggests that open-interest growth contains information that is not entirely reflected in commodity prices. In fact, Figure 1 shows that open interest in commodity markets tends to lead a subsequent rise in commodity prices. 4. What Does Commodity Market Interest Predict? 4.1. Predictability of Commodity Returns Table 4 presents our main findings on the predictability of aggregate commodity returns and spot-price growth. In column (1), we first examine the predictability of commodity returns by the short rate, the yield spread, and aggregate basis. This specification allows us to establish a benchmark to see the incremental forecasting power of open-interest growth, which is our key variable of interest. All coefficients are standardized so that they can be interpreted as the percent change in monthly expected returns per one standard deviation change in the predictor variable. The short rate enters with a coefficient of 0.47 and a t-statistic of This means that a standard deviation increase in the short rate decreases expected returns by 0.47% per month. Hence, the short rate explains an economically important magnitude of predictability in commodity returns. The fact that the short rate, or the inflation rate, predicts asset returns with a negative coefficient is well known (Fama and Schwert, 1977). However, there is no consensus on its economic interpretation. In our view, a more interesting and novel finding is the coefficient for the yield spread. The yield spread predicts commodity returns with a coefficient of 0.45 and a t-statistic of This means that a standard deviation increase in the yield spread decreases expected returns by 0.45% per month. The fact that the yield spread predicts commodities with a negative coefficient is in sharp contrast to the positive coefficient for bonds and stocks, reported in previous studies (Campbell, 1987; Fama and French, 1989). The usual interpretation for bonds and stocks is that when the yield spread is high, which tends to coincide with reces- 15

16 sions, the risk premia for all risky assets are high (due to high risk aversion or high quantity of risk). In contrast, expected commodity returns are low, or alternatively commodity prices are high, when the yield spread is high. This finding suggests that commodities are a good hedge for time-varying investment opportunities in bond and stock markets. Aggregate basis predicts commodity returns with a coefficient of 0.52 and a t-statistic of This means that a standard deviation increase in aggregate basis decreases expected returns by 0.52% per month. The fact that low basis (i.e., a low futures price relative to the spot price) predicts high returns on being long commodity futures is consistent with the theory of backwardation. Overall, the R 2 of the forecasting regression is 3.29%. While the specification in column (1) follows earlier work and is not our main focus, we obtain much stronger results than previously reported. The primary reasons are that we have access to a longer sample period and that we use of a broader cross section of commodities in constructing aggregate basis. In column (2), we introduce open-interest growth to examine its incremental power for predicting aggregate commodity returns. The coefficients for the other three predictor variables are virtually unchanged from column (1) because open-interest growth is essentially uncorrelated with these other predictor variables. Open-interest growth enters with a coefficient of 0.64 and a t-statistic of This means that a standard deviation increase in open-interest growth increases expected returns by 0.59% per month. In this sample, openinterest growth explains a larger share of the variation in expected returns than the other predictor variables. Importantly, the introduction of open-interest growth increases the R 2 of the forecasting regression from 3.29% to 5.29%. In columns (3) and (4) of Table 4, we examine whether the same four variables predict aggregate spot-price growth. In column (3), we predict spot-price growth using only the short rate, the yield spread, and aggregate basis. The short rate enters with a coefficient of 0.68 and a t-statistic of The yield spread enters with a coefficient of 0.40 and a t-statistic of These results are similar to those for aggregate commodity returns 16

17 in column (1). Aggregate basis enters with a coefficient of 0.37 and a t-statistic of Note that the sign of the coefficient is the opposite of that for aggregate commodity returns in column (1). The fact that high basis predicts high spot-price growth is consistent with the theory of storage. A standard deviation increase in aggregate basis increases expected spot-price growth by 0.37% per month. The R 2 of this forecasting regression is 2.50%, which is comparable to that for aggregate commodity returns in column (1). In column (4), we introduce open-interest growth to examine its incremental power for predicting aggregate spot-price growth. Open-interest growth enters with a coefficient of 0.41 and a t-statistic of Importantly, the inclusion of open-interest growth increases the R 2 of the forecasting regression from 2.50 to 3.16%. This is similar to our findings for aggregate commodity returns. In Table 5, we examine additional predictor variables that are potentially related to open-interest growth. For the purposes of comparison, column (1) repeats our main specification from column (2) of Table 4. In column (2), we predict commodity returns with past 12-month commodity returns instead of open-interest growth. Past returns enter with a coefficient of 0.42 and a t-statistic of This means that a standard deviation increase in past returns increases expected returns by 0.42% per month, which is comparable to the economic magnitude of open-interest growth in column (1). Our finding boils down to a strong momentum effect in the time series of aggregate commodity returns. Our finding is different from the well known momentum effect in stock and commodity returns, which is a cross sectional phenomenon. In column (3), we run a horse race between open-interest growth and past returns. We find that open-interest growth crowds out past returns. Open-interest growth enters with a coefficient of 0.59 and a t-statistic of 1.84, while past returns enter with a statistically insignificant coefficient of On the one hand, this result shows that the forecasting power of open-interest growth is closely related to the momentum effect in aggregate commodity returns. On the other hand, open-interest growth and past returns, while highly correlated, 17

18 contain different information about future commodity returns. As revealed by Figure 1, open-interest growth tends to lead returns slightly and can therefore be more informative. Our finding is not only novel for the commodity market literature, but it connects more generally to evidence for slow information diffusion in asset markets as we discuss below. In column (4), we predict commodity returns with hedging pressure instead of openinterest growth. Hedging pressure enters with a coefficient of 0.28 and a t-statistic of The sign of the coefficient is consistent with the theory of backwardation, which implies that high hedging pressure should predict high returns. However, hedging pressure is only marginally significant. In column (5), we run a horse race between open-interest growth and hedging pressure. We find that open-interest growth entirely eliminates the forecasting power of hedging pressure. Open-interest growth enters with a coefficient of 0.61 and a t-statistic of 2.50, while hedging pressure enters with a coefficient of 0.11 and a t-statistic of Recall that the correlation between open-interest growth and hedging pressure is Hence, a plausible interpretation of these results is that hedging pressure is just a noisy proxy for open-interest growth. We summarize our main findings in Tables 4 and 5 as follows. The fact that the short rate and the yield spread predict changes in both futures and spot prices is consistent with the asset allocation view of commodities, that commodity prices are partly driven by aggregate market predictors. The fact that aggregate basis predicts changes in both futures and spot prices with opposite signs is consistent with the segmented markets view, that the theory of backwardation and the theory of storage are partly responsible for commodity price fluctuations. The results for hedging pressure lends further support to the theory of backwardation. More importantly, these tables demonstrate that our findings for open-interest growth are unrelated to these traditional theories. First, the forecasting power of open-interest growth is unaffected by these other known predictors based on these theories. Second, the fact that open-interest growth predicts changes in both futures and spot prices with the same 18

19 sign suggests that our findings are unrelated to the theory of backwardation or hedging pressure in futures markets. A more likely explanation for our findings is that open interest in commodity markets contains important information about commodity prices in general (e.g., about the future supply or demand for commodities), instead of being a proxy for risk. As we discuss in more detail below, the fact that open-interest growth is contemporaneously correlated with changes in commodity prices and also forecasts future returns suggests the following interpretation: open-interest growth contains information about fundamentals that is not entirely impounded in commodity prices Dissecting the Predictability of Commodity Returns We consider a number of additional analyses to check the robustness of our findings and to build towards an interpretation. In Table 6, we examine the predictability of commodity returns by contract maturity and sector. We first examine the predictability of a portfolio of short-maturity contracts separately from a portfolio of long-maturity contracts. Short maturity refers to futures contracts with three months or less to maturity, and long maturity refers to futures contracts with more than three months to maturity. The results for these two portfolios are very similar to those for the aggregate commodity portfolio in Table 4. The coefficient for the yield spread is 0.37 for short maturity and 0.45 for long maturity. The coefficient for aggregate basis is 0.50 for short maturity and 0.47 for long maturity. Finally, the coefficient for open-interest growth is 0.50 for short maturity and 0.79 for long maturity, both of which are statistically significant. Overall, these results show that our main findings are not driven by long-maturity contracts that are potentially less liquid than short-maturity contracts. We next examine whether the same aggregate variables that were used in Table 4 predict returns on the four sector portfolios. For agriculture, the short rate enters with a coefficient of 0.44 and a t-statistic of 1.75, similar to that obtained for the aggregate commodity portfolio. The yield spread has a much weaker forecasting power for the agriculture portfolio, 19

20 compared to the aggregate commodity portfolio. The yield spread enters with a statistically insignificant coefficient of Aggregate basis has virtually no forecasting power for the agriculture portfolio. The coefficient is only 0.09 with a t-statistic of Relative to these other predictor variables that have weak forecasting power, open-interest growth has some forecasting power for the agriculture portfolio. Open-interest growth enters with a coefficient of 0.44 and a t-statistic of We next examine energy, which is a sector of particular interest in light of our introduction. Note that energy contracts are available only since 1978, so the sample is shorter than the other three sectors. The short rate enters with a statistically insignificant coefficient of However, the yield spread enters with a coefficient of 1.36 and a t-statistic of This coefficient is over three times the magnitude of that for the aggregate commodity portfolio. This implies that energy is responsible for much of the forecasting power of the yield spread for the aggregate commodity portfolio. We also find that aggregate basis is a strong predictor of returns on the energy portfolio. A high aggregate basis predicts low returns on being long energy futures with a coefficient of 1.65 and a t-statistic of The magnitude of this coefficient is larger than that for any other sector. Importantly, open-interest growth is a strong predictor with a coefficient of 1.34 and a t-statistic of Again, the magnitude of this coefficient is larger than that for any other sector. For livestock, neither the short rate or the yield spread are statistically significant. The short rate enters with a coefficient of 0.17 and a t-statistic of The yield spread enters with a coefficient of 0.14 and a t-statistic of However, aggregate basis is a strong predictor of returns on the livestock portfolio with a coefficient of 0.88 and a t-statistic of Finally, open-interest growth enters with a coefficient of 0.52 and a t-statistic of The short rate has the highest forecasting power for metals. The short rate enters with acoefficientof 1.04 and a t-statistic of The magnitude of this coefficient is more than two times that for any other sector. The yield spread also works quite well for metals, although its forecasting power is less than that for energy. The yield spread enters with a 20

21 coefficient of 0.71 and a t-statistic of Aggregate basis enters with a statistically insignificant coefficient of However, open-interest growth predicts returns on the metal portfolio with a coefficient of 0.70 and a t-statistic of We now summarize our findings for the sector portfolios. Open-interest growth is the most robust variable in the sense that it consistently predicts returns across all four sectors. While open-interest growth is only marginally significant in agriculture and livestock, the magnitude of the coefficients is economically important across all four sectors. In contrast, the forecasting power of the other three predictor variables is confined to particular sectors. The short rate works best for metals and to some extent for agriculture. The yield spread works best for energy and to a lesser extent for metals. Aggregate basis has the highest forecasting power for energy and livestock and virtually no power for agriculture and metals. These patterns for aggregate basis can be interpreted in light of the theory of backwardation. For example, fluctuations in the convenience yield are apparently more important drivers of expected returns in energy than they are for metals. In the first two columns of Table 7, we examine the predictability of aggregate commodity returns by subsample. We split our sample into two halves, and The short rate, the yield spread, and aggregate basis have virtually no forecasting power for aggregate commodity returns in the second half of the sample. For example, the short rate enters with a coefficient of 0.90 and a t-statistic of 2.98 in the first half, while it enters with a coefficient of 0.07 and a t-statistic of 0.17 in the second half. The yield spread is not statistically significant in either subsample. Its coefficient of 0.35 in the first half is larger than its coefficient of 0.20 in the second half. Aggregate basis enters with the same coefficient of 0.40 in both subsamples. However, it is not statistically significant in either subsample. Overall, our results for the sample are consistent with the classic literature on commodities that did not find a strong role for the yield spread or the basis (Fama and French, 1987; Bessembinder and Chan, 1992). In contrast to these other predictor variables, open-interest growth has consistent fore- 21

22 casting power in both subsamples. Open-interest growth enters with a coefficient of 0.69 and a t-statistic of 1.92 for the first half, while it enters with a coefficient of 0.73 and a t-statistic of 2.60 for the second half. These results show that our findings for open-interest growth are not driven by a particular period in the sample, nor any peculiarity related to our hand-collected data prior to In the last two columns of Table 7, we examine the predictability of aggregate spot-price growth by subsample. The forecasting power of the short rate and the yield spread again comes mostly from the first half of the sample. The short rate enters with a coefficient of 0.93 and a t-statistic of 3.38 in the first half, while the coefficient is nearly zero in the second half. The yield spread enters with a coefficient of 0.28 in the first half, which is larger than the coefficient of 0.16 in the second half. Moreover, the yield spread is not statistically significant in either subsample. Aggregate basis retains its forecasting power in the second half of the sample, which can be interpreted as evidence for the theory of storage. Its coefficient is 0.81 with a t-statistic of Open-interest growth predicts spot-price growth somewhat better in the second half of the sample compared to the first half. Its coefficient is 0.34 with a t-statistic of 1.24 in the first half, while its coefficient is 0.77 with a t-statistic of 2.06 in the second half. Having established that open-interest growth is a robust predictor of commodity returns, we next examine whether it predicts the volatility of commodity prices. The motivation for this analysis is that open-interest growth may be proxying for an omitted risk factor that is not captured by the yield spread, aggregate basis, or hedging pressure. In Table 8, we regress the monthly volatility of spot-price growth onto lagged predictor variables that include open-interest growth. As shown in column (1), open-interest growth enters with a statistically insignificant coefficient of In column (3), we include past 12-month returns as an additional predictor variable. Interestingly, high 12-month returns predicts the volatility of spot-price growth with a coefficient of 0.14 and a t-statistic of This means that a standard deviation increase in 12-month returns increases the absolute value of 22

23 spot-price growth by 0.14% relative to its mean. However, open-interest growth still enters with a statistically insignificant coefficient of These results show that open-interest growth does not seem to be related to the volatility of commodity prices in a way that is consistent with a risk story Predictability of Inflation News and Bond Returns Having established the robustness of our main results, we now focus on building an economic interpretation of our findings. A natural interpretation is that open interest in commodity markets contain information about fundamentals that is not fully incorporated into commodity prices. This view of delayed reaction to fundamental news has some anecdotal support. For example, the 1970s experienced supply shocks to oil, which eventually led to high inflation. The most recent period starting around 2003 experienced strong demand for commodities from emerging economies like China and India, which again led to worries about inflationary pressures. This episode ended abruptly in 2008 with the onset of the financial crisis. As highlighted by Figure 1, each of these periods with inflation worries experienced high activity, followed immediately by a run-up of commodity prices. In Table 9, we examine whether open-interest growth does indeed contain information about inflation in several ways. In column (1), we test whether open-interest growth predicts the monthly change in the annual inflation rate. We regress the change in inflation onto its own lag as well as 1-month lags of the yield spread and open-interest growth. This specification assumes that inflation is integrated of order one, so that we do not include the level of inflation or the short rate in the regression (see Stock and Watson, 1999, for a similar specification). Open-interest growth enters with a coefficient of 0.05 and a t-statistic of 2.99, consistent with our intuition that it contains fundamental news about inflation. In column (2), we add past 12-month commodity returns to see which of the two variables has more forecasting power for inflation. We find that past 12-month returns is a more powerful predictor of inflation. It enters with a coefficient of 0.07 and a t-statistic of The 23