We propose tests for the null that the best model of a class produces as good forecasts as the best model of another one. Forecasts are evaluated using a loss function. Thus, causality can be tested if only the models in one class use a certain input. This is applied to the unemployment/inflation and industrial orders/production relationships. We find causality for the USA, but neither for France nor Spain.
Palabras clave / Key words: Evaluating forecasts, Loss function, Model selection, Causality, Bootstrap, Monte Carlo.Testing the predictive ability of two classes of models (Pdf 214 KB)