Foreign exchange rate forecasting by using machine learning algorithms :|developed vs. emerging economies

Abstract

This thesis has twofold aim. For both of the aims, we forecast the exchange rate of currency pairs by using different models. First, we explore the exchange rate forecast performances of different linear and nonlinear algorithms tree-based and ensemble methods. Ordinary least squares, least absolute shrinkage and selection operator, decision trees, random forests, support vector machines and extreme gradient boosting are the algorithms that are used for forecasting. Second, we compare the performances between emerging and developed markets. For the period between 2002 and 2022, we use monthly data and predictions are made for the month-end high values of nine different currency pairs. We employed two different models based on uncovered interest rate parity model. We performed both regression and classification and as performance evaluation metrics we used root mean squared error and classification accuracy, respectively. Our findings imply that there exist nonlinearities in exchange rate movements. Although any algorithm does not show outstanding performance in the base model, as the model complexity increases, extreme gradient boosting and random forest stand out among others with their improved forecast performance. However, in our findings, there does not exist any difference between emerging and developed markets.