The Serial Dependence of the Commodity Futures Returns: A Machine Learning Approach
53 Pages Posted: 8 Mar 2020 Last revised: 17 Aug 2020
Date Written: January 11, 2020
This paper uses machine learning tools to study the serial dependence (lead-lag relations) of commodity futures returns during the post financialization period (January 2004 – December 2019). We use LASSO (Least Absolute Shrinkage and Selection Operator) to select the predictors as the number of predictors is large relative to the number of observations. We find significant full-sample and out-of-sample predictability. In the full sample, we find that LASSO can identify a sparse set of predictors that come from economically linked commodities or are likely driven by excessive speculative trading. The out-of-sample forecasts based on the LASSO generate statistically and economically large gains. When we separate the indexed futures from the non-indexed futures and replicate the above analysis, we find that the out-of-sample performance exists in the indexed futures but disappears in the non-indexed futures. The lead-lag relations are also more significant after the advent of ETF or ETNs that track the broad futures indices such as S&P GSCI and BCOM indices, indicating that index trading due to financialization drives the excessive comovement among the commodity futures. Overall, we find that serial dependence generates significant predictability during the sample period when the performance of the long-only commodity index futures is poor.
Keywords: Commodity Futures, LASSO, Machine Learning, Predictability, Serial Dependence, Financialization, Comovement
JEL Classification: C22, C58, G11, G12, G14
Suggested Citation: Suggested Citation