DIRECTIONAL FORECASTING OF EURUSD EXCHANGE RATE USING CLASSIFICATION APPROACH

Abstract

This thesis study aims to use classification methods in forecasting EURUSD direction of change. A number of classifiers including logistic regression, knn, naïve bayes, and classification tree are used. The input variable universe is comprised of three major categories: currency pairs, interest rates and market indices (stock and commodity indices). All series are from 1.1.2004 to 8.2.2016. Two main types of models are constructed. First are models with fixed predictors that are based on ideas from literature. Second are models which select predictors at each step from a pool of predictors using an input selection algorithm. The input selection algorithms are MIM, MRMR, JMI and DISR originated from information theory field. In estimating the models two types of predictors are used: original form and discretized version. Models are estimated using both recursive and rolling window. Finally, the out-of-sample forecast is formally tested for statistical significance. Among the models built according to the combination of classifier, input selector, predictor and estimation scheme, a few models are found to be marginally significant, indicating the promising outlook of using more sophisticated methods.