Artificial neural networks in inflation prediction: application like analysis tool for financial decisions at small organizations

Abstract

Estimation of inflation rates is crucial for managers, because investment decisions are closely linked to them. However, the behavior of inflation tends to be nonlinear and even chaotic, making it difficult to be estimated. This characteristic may become simplest models, accessible for small organizations, inaccurate to forecasting the phenomenon, since many of these require large data manipulations and / or specialized software. This article aims to evaluate, through formal statistical analysis, the effectiveness of artificial neural networks in inflation forecasting at small organizations reality. ANNs are appropriated tools to measure the phenomena of inflation, as they are approximations of polynomial functions, capable of dealing with nonlinear phenomena. This article selected 3 basic models of Multi Layer Perceptron artificial neural networks, simple enough to be implemented whit open source spreadsheets. These three models were tested from a set of independent variables suggested by Bresser-Pereira and Nakano (1984), with a lag of one, six and twelve months. For that were used Wilcoxon test, coefficient of determination R² and the average percent error of tested models. Data set was divided into two, one group used for artificial neural networks training and another group used to verify models predictive ability and their ability to generalize. This work concluded that certain models of artificial neural networks have a reasonable ability to predict inflation in the short run and constitute a reasonable alternative for this type of measurement.