In this paper we develop a trading strategy in which the difference in observed returns of value and growth stocks in the US stock market is exploited. In the literature this return spread is often called the "value premium". In our modeling process we use a procedure similar to the recursive modeling approach as proposed by Pesaran and Timmerman (1995). We first estimate a universe of parsimonious models in an in-sample setting using a base set of technical and economic forecasting variables. Subsequently, we generate out-of-sample forecasts in a rolling window framework for all models and evaluate the performance in a so-called model training period. This adjustment directly relates to the critique of Bossaerts and Hillion (1999), who showed the insufficiency of in-sample criteria to forecast out-of-sample information ratios. Model selection is based on three well-known selection criteria: hit ratio (% correct sign), the mean return of the strategy and the realized information ratio. Finally, we start implementing our investment strategy in a second stage out-of-sample period: the trading period. This model estimation and selection procedure enables us to address the issue whether we could have historically exploited the value/growth rotation strategy in a practical context. In the empirical section we show that it is possible to successfully forecast the time-varying value premium based on this strategy. Moreover, we observe that the set of relevant forecasting variables varies considerably through time.