Neural Networks for Defuzification of Fuzzy Rules: An Application in Macroeconomic Forecasting

Abstract

Macroeconomic forecasting has traditionally been performed with the use of econometric tools though these methods necessarily make many theoretical assumptions that are not valid in all circumstances. The main advantage of the use of approaches that apply machine learning algorithms to economic data is that forecasts largely free of assumptions can be made. This study presents an approach to macro economic forecasting that generates fuzzy rules from data using a fuzzy control system architecture and evolutionary programming. However, the selection of a defuzification method is typically performed subjectively in fuzzy control systems. We demonstrate that the selection of defuzification method makes a substantial impact on forecasts. In order to overcome this subjectivity and further enhance our objectives of developing forecasting systems free of any technical or theoretical assumptions we introduce a neural network to perform the defuzification. The performance of our approach compares very favourably with other data mining techniques on cross validation tests with macro economic data.