Modelos de previsão da inflação brasileira e combinações de previsões: uma análise de curto prazo

Abstract

Inflation is a key variable to the economic environment, being its control a crucial element to the agents’ decision making and to create favourable growth conditions. Therefore, it is of fundamental importance to foresee its trajectory in order to use mechanisms which can keep it under control. Thus the objective of this paper is to make forecasts of the Brazilian monthly inflation in 2017, measured by the Brazilian price index, called IPCA. To achieve them, three statistical models were used: that is, Autorregressive Integrated Moving Average (ARIMA), Autoregressive Distributed Lag (ADL) taking into account the Phillips Curve approach, and the Error Correction Model (VECM). Trying to obtain improved forecasts, the combination of forecasts methodology was implemented using the following methods: simple mean, weighted average and linear regression. The criterions used to rate the forecasts were the Root Mean Squared Error (RMSE), and the Mean Absolute Percent Error (MAPE), where the analysis emphasis is given to this last one, because it shows, in percentage terms, how distant is every forecast series from the real inflation values in the period. The forecasts made were still compared to two benchmarks: the Naive model, which has the supposition of inflation pass-through, and to the market forecasts, available in a report made by the Brazilian Central Bank called Focus. Among the results achieved, one can mention is, considering the group of forecasts made by the three models, the best forecast series was the one from the Error Correction Model, with an error percentage of 0.1274%. Now considering the group of combinations, the method of combining using linear regression was the one which had the best accuracy in its group, with an error percentage of 0.1221%, and it was also the series, among all, which has reached closer the real monthly inflation in 2017. Concluding, combine forecasts, in general, provides more accurate forecasts than the individual models.