APPLYING VARIOUS MACHINE LEARNING METHODOLOGIES INTO THE FINANCIAL MARKET

The modernization of the financial market, with the introduction of the internet, made it easier for the average, everyday people, around the world to invest in the plentiful trading assets in the market. This created a revolution, propelling the foreign exchange market to be the most valuable and tradeable financial asset on the planet, with a daily turnover that surpasses $6 Trillion. As a result, predicting the future price can be very profitable, causing analysts and hedge funds to start a race toward searching for the best tools or algorithm that allows them to be ahead of the competition. With the introduction of faster and more powerful computers, the dream of automated, lightning-fast trading became a reality. Studies believe that more than 60% of the total traded volume in developed nations is performed by automated systems and algorithms. This thesis will investigate the claims by different studies that machine learning algorithms can be used to accurately predict the future prices of the market. The thesis chose the EUR/USD exchange rate, to study, as it is the most volatile asset and it has the highest trading volume. Based on this, ten years of daily closing prices that included many trading assets, such as currencies, indices, and commodities were collected to study the effect of different trading assets on each other and understand the correlation effect. The investigation starts with the use of linear regression techniques, including mean least-squares estimations and multiple linear regression, which failed to provide sustainable results or achieve an accuracy above 60%. In addition to that, a support vector machines model was built using a linear and a radial basis function kernel, where the linear kernel model recorded an accuracy of 60% when predicting the future price trend of the EUR/USD. Finally, the thesis dives into the use of artificial neural networks, in the form of multi-layer perceptrons (MLPs), and long-short term memory (LSTMs). All forms of artificial neural networks have failed to predict the future price trend when using one day of previous closing prices as input. This has changed when an MLP regressor was trained to use the previous closing price of 30 days to predict one day into the future. This allowed the network to achieve accuracies that exceeded 80% when predicting the future price trend