ABSTRACT
This paper focuses on an integration of economic theory in a machine learning process for the purpose of exchange rate forecasting. Since the early 1980s literature stresses the weakness of economic theory in forecasting exchange rates. Consistent with these results, in this paper it is asked how machine learning can increase the forecasting performance of theoretical models. Therefore structural exchange rate models are implemented in a machine learning process as a framework in which and among which the possibilities of machine learning are exploited not only to identify further sources of influence but also to test alternative variables of aggregates suggested by exchange rate theory. Thus to serve as a tool for model selection. The applied approach uses a Genetic Algorithm for model selection and Neural Networks for the generation of the forecasts. It is shown that this combination of economic theory and machinery learning not only increases the “fitness” (a popular term in Genetic Algorithm literature) of theoretical exchange rate models but is also fruitful for the effectiveness and correctness of machine learning processes. As experience showed, relationships derived from machine learning techniques are often not convincing as regards their correctness and effectiveness. Most machine learning approaches do not contribute much to persuade otherwise. In this paper we tried to overcome parts of this limitation. The approach is illustrated in some detail for five exchange rates on a monthly frequency.
