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 subsequent commodity futures returns. The linear models use shrinkage methods based on naive averaging and principal components. The nonlinear models use feedforward deep neural networks either as stand-alone (DNN) or in conjunction with LSTM, a recurrent long short-term memory network. Out of the four specifications considered, the LSTM-DNN architecture is the most successful at transforming the 128 predictive variables into profitable investment strategies. The risk premium then modelled is unrelated to, and exceeds, those earned on previously-published characteristic-sorted portfolios. Our analysis is robust to the presence of transaction costs and illiquidity.