The commodity risk premium and neural networks

Abstract

The paper uses linear and nonlinear predictive models to study the linkage between a set of 128 macroeconomic and financial predictors and the risk premium of commodity futures contracts. The linear models use shrinkage methods based on either naive averaging or principal components. The nonlinear models use feedforward deep neural networks (DNN) either as stand-alone or in conjunction with a long short-term memory network (LSTM). Out of the four specifications considered, the LSTM-DNN architecture best captures the risk premium, which underscores the need to estimate models that are both nonlinear and recurrent. The superior performance of the LSTM-DNN portfolio persists after accounting for transaction costs or illiquidity and is unrelated to previously-documented commodity risk factors.