Linear regression versus back propagation networks to predict quarterly stock market excess returns

This paper compares a linear model to predict quarterly stock market excess returns to several backpropagation networks. Research findings suggest that quarterly stock market returns are to some extent predictable, but only marginal attention has been paid to possible nonlinearities in the return generating process. The paper discusses input selection, elaborates on how to generate out-of-sample predictions to estimate generalization performance, motivates the choice for a particular network, examines backpropagation training, and evaluates network performance. The out-of-sample predictions are used to calculate several performance metrics, and to determine added value when applying a straightforward tactical asset allocation policy. A nonparametric test is selected to evaluate generalization behavior, and sensitivity analysis examines the selected network's qualitative behavior. Strong nonlinear effects appear to be absent, but the proposed backpropagation network generates an asset allocation policy that outperforms the linear model.