This dissertation consists of three essays concerning inflation forecasting, taking the Brazilian case as an application. In the first essay, we examine the effectiveness of several forecasting methods for predicting inflation, focusing on aggregating disaggregated forecasts. We consider different disaggregation levels for inflation and employ a range of traditional time series techniques, as well as linear and nonlinear machine learning (ML) models that deal with a larger number of predictors. For many forecast horizons, aggregation of disaggregated forecasts performs just as well as survey-based expectations and models generating forecasts directly from the aggregate. Overall, ML methods outperform traditional time series models in predictive accuracy, with outstanding performance in forecasting disaggregates. In our second essay, we investigate the potential benefits of combining individual inflation forecasts by proposing a time-varying bias correction for the average forecast. Our analysis includes estimations using both rolling windows and state-space models that use the recursiveness of the Kalman filter. We achieve good forecast performance for models based on small rolling windows for shorter and intermediate forecast horizons, while a state-space model performs slightly worse than procedures based on rolling windows. In the third essay, we use supervised learning to generate forward-looking indexes based on tweets and news articles for accumulated inflation and investigate whether these indexes can improve inflation forecasting performance. Our results indicate that news-based indexes provide significant predictive gains, particularly for 3- and 12-month-ahead horizons. These findings suggest that incorporating more information sources than just expectations based on experts opinions can lead to more accurate forecasts.