Inflation forecasting using hybrid ARX and neural network model

Abstract

Getting more accurate predicted values was always in the focus of the forecasting literature. In my research, this topic is elaborated, as well: how can price indexes be forecasted in a more accurate way using time-series analysis. Studying inflation is a highly relevant question since the decision of all members of the economy is influenced by it. In my work, American price index data were analysed with macroeconomic rates. From a methodological view, four different techniques were compared: two single and two complex models. In the case of the first category, a linear approach, such as the ARX model, moreover a nonlinear machine learning method, such as the artificial neural network (ANN) was applied. In the case of the complex techniques, the averaging method was calculated as the average of the predicted values of ARX and ANN models, whilst these methods were hierarchically combined in the case of the hybrid method. The results of the forecasting were evaluated by the mean squared error (MSE) rate and by the Diebold-Mariano test. As the main results of the research, it can be claimed that the complex methods, applied on price index time-series, can always perform at least as accurate as single ones. Furthermore, the linear approach, such as the ARX model, was generally outperformed by these complex models significantly, while ANN model was not. It is also shown that my results are not robust, the length of the in- and out-of-sample time-series influence whether a complex or the ANN model has the lowest MSE value. It is also observed that after the financial crisis in 2008-2009, the residuals are more volatile.