In linear time-series regression analysis, there is typically uncertainty about which variables to include as regressors and the exact form of the disturbance process. This paper uses the Monte Carlo method to investigate the predictive performance of nine different pretesting strategies for misspecification. Data generating processes used in the study include first-order autoregressive (AR(1)) disturbances, first-order moving average disturbances, an extra exogenous regressor and the lagged dependent variable as an extra regressor. We find that remarkably robust predictions for a range of misspecified models result from applying the Durbin-Watson test for autocorrelation and correcting for AR(1) disturbances when the test is significant.