A new technique for postsample model selection and validation

Abstract

The model selection and Granger-causality literatures have focused on insample rather than postsample hypothesis testing. This is because postsample model validation periods are usually short, whereas the correlations in typical postsample forecast errors require large-sample methods. A resampling-based postsample inference procedure is described which explicitly estimates the uncertainty which its large-sample approximation induces in the inference significance levels it produces. The procedure is applied to postsample forecasting errors from the Ashley et al. (1980) study examining Granger-causation between US consumption and advertising expenditures. Postsample model validation is apparently feasible, but requires longer postsample periods than past studies have allocated.