In evaluating prediction models, many researchers flank comparative ex-ante prediction experiments by significance tests on accuracy improvement, such as the Diebold-Mariano test. We argue that basing the choice of prediction models on such significance tests is problematic, as this practice may favor the null model, usually a simple benchmark. We explore the validity of this argument by extensive Monte Carlo simulations with linear (ARMA) and nonlinear (SETAR) generating processes. For many parameter constellations, we find that utilization of additional significance tests in selecting the forecasting model fails to improve predictive accuracy.
forecasting time series predictive accuracy model selection!