Are aggregate commodity futures returns predictable based on prices across the maturity curve and/or on the state of the global economy? In her January 2017 paper entitled “Commodity Return Predictability”, Regina Hammerschmid investigates aggregate commodity futures return predictability based on variables incorporating information from the term structure of futures prices and several global economic variables. She includes commodity futures series spanning five sectors (energy, grains/oilseeds, livestock, metals and softs). She considers three groups of predictive variables: (1) commodities spot and futures prices; (2) aggregate OECD economic data (industrial production, total exports and imports, the composite leading indicator and business confidence index); and, (3) for comparison tests, commodities trading volume, open interest and hedging pressure (net difference between short and long positions of hedgers). She uses returns for fully collateralized long positions in commodity futures contracts with 1, 2, 3 and 4 months to maturity, rolled at the end of each month. She aggregates returns by first averaging within each sector and then averaging sector averages (all equally weighted). She considers forecast horizons of 1, 3, 6, 9 or 12 months. For out-of-sample regression testing, she uses an inception-to-date window of at least 10 years of data. Using daily spot and commodity futures settlement prices as available, monthly economic data and monthly S&P-GSCI levels since January 1975, and associated monthly trading volume, open interest and hedging pressure data as available since January 1986, all through August 2015, she finds that:
- Average monthly excess returns of aggregate commodity futures portfolios range from 0.23% to 0.28% (depending on contract maturity), with standard deviations about 3.6%. Correlations with S&P GSCI monthly returns are around 0.80.
- Basis (difference between futures and spot price) for a given futures maturity has a weak negative relationship with associated future aggregate commodity return, but only for 1-month and 2-month forecast horizons.
- More complex predictors employing full-sample multiple regression or principal component analysis to combine information from the entire futures curve significantly improve in-sample predictability of aggregate commodity futures returns across maturities and forecast horizons. These methods indicate:
- Data for maturities 1, 2, 3 and 4 months contribute about equally to predictive power.
- Curvature of the commodity futures curve conveys more predictive power than level or slope.
- Predictive power concentrates during economic expansions.
- Out-of-sample predictive power is weak, confirming that the use of the full sample to construct the complex predictors introduces look-ahead bias. In fact, simple basis works better than the complex predictors out-of-sample.
- The specified economic data relate positively to future aggregate commodity futures returns across maturities, most strongly for short forecast horizons.
- Monthly changes in the composite leading indicator and the business confidence index have the strongest predictive powers.
- Using economic data jointly with the predictors combining futures curve information significantly boosts predictive power.
- Imposing a lag of one month in availability of economic data, most results remain statistically significant. With a lag of three months, only monthly change in the composite leading indicator significantly predicts returns.
- Expected returns for commodity futures are pro-cyclical. When economic activity is high, expected returns are high.
- For example, in-sample, a 1% increase in aggregate industrial production (indicating elevated demand for commodities) predicts positive next-month returns over 0.7%.
- Out-of-sample, only the composite leading indicator and business confidence index consistently predict returns (outperform the historical average return as a predictor). However, simple basis plus any of the economic variables significantly predict returns.
- Out-of-sample predictive power concentrates around the end of the financial crisis in 2009.
- Overall, findings for the aggregate commodity futures portfolio specified above apply also to the S&P GSCI.
- After adding spot return, spot return volatility, trading volume, open interest and hedging pressure to an in-sample multiple regression including both commodity futures curve data and economic data:
- Change in open interest dominates all other predictor variables, significantly predicting aggregate commodity futures returns for all maturities and subsuming much of the predictive power of curve data.
- Spot return or volatility of spot returns significantly improves overall predictability
for an annual forecast horizon. - Hedging pressure is never a significant contributor to predictive power.
- Findings are somewhat similar, but weaker, for the five commodity sectors tested separately, with predictability concentrating in sectors related to industrial production (energy and metals) rather than food.
In summary, evidence suggests some potentially exploitable ways to predict commodity futures returns, with change in open interest most promising.
Cautions regarding findings include:
- The study is largely academic, offering no commodity futures timing strategy tests.
- Exploitability as indicated by out-of-sample tests and economic data lagged (realistically) by several months appears very limited.
- Testing many predictors and multiple modeling approaches on the same commodity futures data introduces data/model snooping bias, such that the best-performing predictors overstate expectations.
- Economic data series may have retroactive revisions, such that they impound look-ahead bias that confounds backtesting.
- Concentration of out-of-sample predictive power around the end of the financial crisis in 2009 is disconcerting because such a crisis may not recur.
