Machine learning methods for inflation forecasting in Brazil: New contenders versus classical models

We conduct an extensive out-of-sample forecasting exercise, across a variety of machine learning techniques and traditional econometric models, with the objective of building accurate forecasts of the Brazilian consumer prices inflation at multiple horizons. A large database of macroeconomic and financial variables is employed as input to the competing methods. The results corroborate recent findings in favor of the nonlinear automated procedures, indicating that machine learning algorithms (in particular, random forest) can outperform traditional forecasting methods in terms of mean-squared error. The main reason is that some machine learning methods can yield a sizeable reduction in the forecast bias, while keeping the forecast variance under control. As result, forecast accuracy can be improved over traditional inflation forecasting models. These findings offer a valuable contribution to the field of macroeconomic forecasting, and provide alternative methods to the usual statistical models often based on linear statistical relationships.