Since the underlying economic structure is likely to be affected by changes in preferences, technologies, policies, crises, etc., data in the previous time period may be irrelevant to the present data-generation process. Thus, econometric forecasts are often based on rolling estimation. However, it is far from clear how to choose an optimal sample to estimate a predictive model. We propose a novel approach to selecting the optimal window in a predictive linear regression model with time-varying parameters, by minimizing suitable criteria based on forecast errors, including unconditional/conditional/global mean square forecast errors. A practically feasible cross-validation procedure is developed to choose an optimal window, which is asymptotically equivalent to the infeasible optimal window based on unconditional mean square forecast errors. Simulation studies are conducted to evaluate the accuracy of forecasts using our methods under various types of structural changes. Empirical applications, to forecasting the US GDP growth rates, inflation rates and stock returns, highlight the merits of the proposed methods relative to other popular methods available in the literature.