Forecast Model Evaluation in Small-Sample Persistent Processes: A Simulation Study.

摘要

This paper examines the effect of highly persistent processes on methods of evaluating the out-of-sample predictions of nested models. The non-parametric bootstrap method of Hubrich and West (Journal of Applied Econometrics, 2010) is modestly undersized for stationary processes, but I find that their method becomes oversized as the process examined approaches the unit root boundary. I also find size distortions in several other leading forecast evaluation procedures [e.g., Giacomini and White (Econometrica, 2006), Hansen (Journal of Business & Economic Statistics, 2005), Harvey and Newbold (Journal of Applied Econometrics, 2000), and White (Econometrica, 2000)]. I use simulation-based techniques to demonstrate that the Maximized Monte Carlo (MMC) method of Dufour (Journal of Econometrics, 2006) corrects for the over-rejection of the null even with highly persistent processes and small sample sizes. The MMC method exhibits good power properties, although in this study, the MMC procedure becomes more conservative as sample sizes increase.