The sampling distributions of the predictor for an autoregressive model under misspecifications (Q1066595)

Asymptotic approximations are presented for the sampling distributions of the h-step (h\(\geq 1)\) ahead predictions of a non-zero mean ARMA(1,1) process, when the predictor is misspecified as a zero-mean AR(1) model. The approximate distributions include terms of the order 1/T, where T denotes the sample size. Both the conditional (on the last value of the available sample) and the unconditional distributions are obtained. Furthermore, the dependent and independent cases are treated; here dependent (independent) refers to the situation where the sample used for estimation is dependent (independent) of that used for prediction. Numerical examples verifying the presented theoretical results are also included.