The issue of forecast aggregation is to determine whether it is better to forecast a series directly or instead construct forecasts of its components and then sum these component forecasts. Notwithstanding some underlying theoretical results, it is generally accepted that forecast aggregation is an empirical issue. Empirical results in the literature often go unexplained. This leaves forecasters in the dark when confronted with the option of forecast aggregation. We take our empirical exercise a step further by considering the underlying issues in more detail. We analyse two price datasets, one for the United States and one for the Euro Area, which have distinctive dynamics and provide a guide to model choice. We also consider multiple levels of aggregation for each dataset. The models include an autoregressive model, a factor augmented autoregressive model, a large Bayesian VAR and a time-varying model with stochastic volatility. We find that once the appropriate model has been found, forecast aggregation can significantly improve forecast performance. These results are robust to the choice of data transformation.