A1553
Title: Machine learning for commodity futures pricing
Authors: Milan Ficura - University of Economics in Prague (Czech Republic) [presenting]
Abstract: The ability of 29 characteristics is tested, previously reported in the literature, to predict the cross-section of 28 US commodity futures returns in the period from March 1996 to January 2024. The characteristics include momentum, volatility, liquidity, basis, relative basis, basis momentum, hedging pressure, speculative crowding, currency betas, inflation betas, stock market betas and commodity market betas, among others. In addition to the standard methodology based on univariate and multi-variate portfolio sorts, the ability of machine-learning methods (Elastic net regression and random forests) is further tested to extract predictive signals from the vectors of individual characteristics and predict the cross-section of commodity futures returns. The constructed models are found to possess significant out-of-sample predictive power and result in sizable economic gains.
