Bootstrap prediction intervals for autoregressive models of unknown or infinite lag order

Recent studies on bootstrap prediction intervals for autoregressive (AR) model provide simulation findings when the lag order is known. In practical applications, however, the AR lag order is unknown or can even be infinite. This paper is concerned with prediction intervals for AR models of unknown or infinite lag order. Akaike's information criterion is used to estimate (approximate) the unknown (infinite) AR lag order. Small-sample properties of bootstrap and asymptotic prediction intervals are compared under both normal and non-normal innovations. Bootstrap prediction intervals are constructed based on the percentile and percentile-t methods, using the standard bootstrap as well as the bootstrap-after-bootstrap. It is found that bootstrap-after-bootstrap prediction intervals show small-sample properties substantially better than other alternatives, especially when the sample size is small and the model has a unit root or near-unit root.