Recent Developmentsin BootstrappingTime Series

In recentyears,severalnewparametricand nonparametricbootstrapmethods have been proposed for time seriesdata. Which of these methods should applied researchersuse? We provide evidence that for many applicationsin time serieseconometricsparametricmethods are more accurate,and we identify directionsfor future researchon improving nonparametricmethods. We explicitly addressthe important, but often neglectedissue of model selection in bootstrapping. In partictiartwe emphasizethe advantagesof the AIC over other lag order selectioncriteriaand the need to account for lag order uncertainty in resampfing.We also showthat the block size playsan importantrole in determiningthe successof the block bootstrap, andwe propose a data-based block size selectionprocedure. Our discussionalso highlightsthe importance of accountingfor small-samplebias in autoregressionsand some shortcomings of the standardpercentileand percentile-tintervalsin the time seriescontext.