In the theoretical literature on exchange rates we find many results with little empirical support. This paper seeks to contribute to a better understanding on one of the main issues that were not fully answered: the forecasting puzzle on exchange rates. By analyzing in-sample and out-of-sample forecasts for nominal and real exchange rates and the commodity price index of Brazil, we find empirical evidence on some of the explanations for this puzzle. The results insample and out-of-sample indicate that the Brazilian nominal exchange rate has an important forward looking component, which could explain the failure of many models to predict this variable. Past values of nominal exchange rate forecasts commodity prices substantially better than a random walk, both in short and long term horizons, while the reverse relation does not hold. An additional analysis in which we compare our model with a univariate autoregressive one suggests that the evidences may not be as robust as we thought at first. For the real exchange rate we extend the analysis made in the literature on commodity currencies to measure the predictive power of these models. Considering the in-sample exercise, commodity prices play an important role in predicting the Brazilian real exchange rate. However, there is also Granger causality in the opposite direction. Regarding the out-of-sample forecasts, our error correction model could not overcome the random walk model. Our results are robust to structural breaks.