Commodity Option Return Predictability

Abstract

This paper investigates the predictability of delta-hedged commodity option returns using 103 predictors. We estimate several linear and nonlinear machine learning models and forecast ensembles using futures options data on seven commodities. There is strong evidence of out-of-sample return predictability for horizons of 1 week to 1 month ahead. We show how a machine learning-informed long-short option trading strategy generates positive returns after transaction costs for most commodities. Among the groups of predictors, options-based characteristics are the most informative, but macroeconomic variables typically improve forecasts. A nonlinear ensemble forecast provides the best results, while the best single model is the Random Forest. Some machine learning models perform poorly. Finally, we document strong evidence for increased predictability in periods of high volatility.