Bootstrap prediction intervals for threshold autoregressive models

This paper examines the performance of prediction intervals based on bootstrap for threshold autoregressive models. We consider four bootstrap methods to account for the variability of estimates, correct the small-sample bias of autoregressive coefficients and allow for heterogeneous errors. Simulation shows that (1) accounting for the sampling variability of estimated threshold values is necessary despite super-consistency, (2) bias-correction leads to better prediction intervals under certain circumstances, and (3) two-sample bootstrap can improve long term forecast when errors are regime-dependent.