|
Abstract:
|
We investigate the role of significance levels in selecting models for forecasting when using binary decisions to retain or drop variables as it controls retention frequencies under both null and alternative. Analysis identifies the best selection significance level in a bivariate model facing location shifts near the forecast origin. The trade-off for selecting variables in forecasting models in a stationary world, retain variables if their non-centralities exceed 1, applies in non-stationary settings with structural breaks. The results confirm the optimality of the Akaike Information Criterion for forecasting in very different settings. Empirically forecasting UK inflation demonstrates the applicability of the analysis. Simulation explores regressor selection for 1-step ahead forecasts in larger models with unknown location shifts for a range of scenarios and significance levels, using the multipath tree search algorithm, Autometrics. The costs of model selection are small and provide support for model selection at looser than conventional settings, with the caveat that retaining irrelevant variables subject to location shifts can worsen forecast performance.
|
