We use machine learning (ML) models, namely Random Forest (RF), Complete Subset Regression (CSR), LASSO, adaLASSO, Elatic Net and Ridge to forecast Brazilian yearly CPI inflation. In particular, our goal is to determine whether accumulating predictor variables enhances the accuracy of inflation forecasts. We compare these models with the random walk (RW) and, mainly, with a survey of expectations from the Central Bank of Brazil called Focus. We show that the Random Forest beats the Focus consensus in all possible datasets with gains up to 58 percent in terms of RMSE. On the other hand, the performance of the shrinkage methods exhibits significant heterogeneity across different datasets. We show that machine learning models consistently outperform the benchmarks when predictor variables are not accumulated or accumulated in 12 months. Finally, we show that the models (especially RF) consistently outperform the benchmarks during periods when inflation is more volatile.