Using 18 different specifications of the GARCH-M model and high frequency data for the Colombian exchange market index (IGBC), we evaluate the out-of-sample performance of the models. The models considered take in account the leverage effect, the day-of-the-week effect, and the hour-of-the-day effect. We evaluate 1000 one-step-ahead rolling forecasts foreach of the 18 models. Using different descriptive statistics and the Granger and Newbold (1976) test and the Diebold and Mariano (1995) test, we found that the best model would bethe GARCH-M without the leverage effect, the day-of-the-week effect, and the hour-of-the-day effect.