Eureka! Institutional Repository

This present thesis aims to study the forecasting of exchange rate EUR/USD, using machine learning techniques. Particularly with the use of SVR, a reliable machine learning model, that, using the kernel trick, transforms input data into a space with higher dimensions and thereby makes regression. In this thesis, firstly, datasets have been crafted out of raw data, which were gathered by the European Central Bank and the broker Dukascopy, following this, experiments were carried out based on the aforementioned data, by using four ε-SVR models each with a different core, whose hyper-parameters were found through the Grid Search procedure. In these experiments, three kinds of exchange rate forecasts are realised; per day, per hour and per ten minutes. In order to realise the forecasts, in addition to the historic exchange rates, data from the Stock Twits are used to uncover the predispositions of investors, towards the rise or fall of the exchange rates. The results have shown that it is difficult to construct machine learning models, that are well suited to predicting the future of exchange rates. Furthermore, it seemed unclear as to whether the sentiment of investors from the Stock Twits contributes positively in the training of models or simply constitutes noise. In some cases, the information proves useful, while in others, little more than noise.