This paper addresses the issue of model selection risk by examining whether a model's past performance in forecasting expected returns provides an indication of its future forecasting performance. For this purpose, we implement a wide range of different forecasting models and then apply the Aiolfi-Timmermann methodology for relative performance persistence measurement. We find no evidence of performance persistence in forecasting models at the top end of the historical forecasting performance rankings. Economic consequences of this purely statistical study are subsequently quantified by an out-of-sample asset allocation exercise. Simulating an asset allocator, who selects ex ante return forecasting models based on their ex post performance, we show that investors should make portfolio decisions based on forecasting models from the middle of the historical forecasting performance rankings.