This paper investigates a common approach to forecast stock returns. The forecasts are obtained in three steps. First a base set of potential forecasting variables is determined. Then a subset of forecasting variables is selected at each time period. Finally, a regression is run on the selected subset and the estimated regression parameters are used to forecast the return of the next time period. While this approach appears to have high forecasting power, a closer look reveals that none of the three steps contributes significantly to its performance. Moreover, we show that its high forecasting power is simply due to the fact that it mimics a very primitive technical trading strategy, which is based only on the signs of past returns.