A Note on Prediction with Misspecified Models

Suppose that a time series model is fitted. It is likely that the fitted model is not the true model. In other words, the model has been misspecified. In this paper, we consider the prediction interval problem in the case of a misspecified first-order autoregressive or AR(1) model. We have calculated the coverage probability of an upper one-step-ahead prediction interval for both properly specified and misspecified models through Monte Carlo simulation. It was found that dealing with prediction interval for misspecified model is complicated: the distribution of a future observation conditional on the last observation and the parameter estimator is not identical to the distribution of this future observation conditional on the last observation alone