Overfitting and forecasting: linear versus non-linear time series models

摘要: The main purpose of this dissertation is to compare the in-sample estimating and out-of-sample forecasting performance a set non-linear time series models, i.e., threshold autoregressive momentum exponential generalized models bilinear models. First, Monte Carlo simulation used study overfitting forecasting. For AR processes, if AIC SBC criteria are select possibility very high since other linear likely have lower SBC. However, MSPE for one-step ahead forecast can be identify true processes. TAR TAR-C process only difference persistence between two regimes large enough. Underfitting misspecification happen with small regimes. we don't know or process, can't in most cases. Thus, none AIC, model given unknown order. consistently persistency Then, applied term structure interest rates spread wholesale retail pork prices U.S. It shown that there do better than conventional ARMA both estimation forecast. Also, it unlikely dominance results from useful series.