Time Series Cross Validation : Theoretical Properties and Empirical Performance

Under the framework of Shao (1993), I provide a theoretical analysis of the cross validation-based model selection for time series proposed by Racine (2000). I also consider a variant of the method that does not require the size of validation samples to be the same and find that it often improves the model selection performance. Various time series cross validation methods as well as the standard BIC are compared in a simulation study. A theoretical insight about the choice of validation sample sizes is explored