This study evaluates the efficiency gains in forecasting three commodity prices (live cattle, coffee and cotton) using time series models. Different leves of temporal aggregations are tested (weekly, monthly, quarterly and annually). The objective is to test whether models based on disaggregated data can produce better price forecasting than the corresponding model using a higher level of aggregation. For example, we test if weekly models can predict better monthly prices than monthly models. Because the high volatility in real prices, we evaluate the possible non-stationarity behavior and heteroskedasticity of each commodity. Then, we use time series methods to model the prices and select the best estimators at each aggregation level and commodity. For the three commodity prices, models based on disaggregated levels effectively provided an efficiency gain in forecasting. Among these levels, the best models were the weekly models. The same behavior was consistent across all possible levels of aggregations.