In my master thesis, I examined financial time series, combining novel types of machine learning approaches. I chose one of the branches of machine learning discipline, the neural networks based deep learning, which gained a lot of attention among machine learning scientists recently, to build a complex system that is capable modelling time series, taking into account the effect of economical events (fundamental analysis) beyond using the regular historical data (technological analysis).
In my research I examined and applied most popular architectures of deep neural networks. To begin with, I examined and used a network made of fully connected layers, then continued with exploring the potentiality of the recurrent neural networks. Last but not least, I applied the architecture of convolutional neural networks, which is one of the most popular architecture nowadays.
I used historical price data from both foreign exchange market and Budapest Stock Exchange. During my research, I examined the movements of the EUR-USD currency pair and stock prices of one the of companies, OTP, listed on the BSE, depending on time. In order to use the full potential hidden in the deep neural networks, I seek methods to improve the results of the previous classification problem. That is why, I analysed economic news based on a method, called sentiment analysis, then implement their effects to the forecasting model.
It is a mandatory task to measure the goodness of the forecasted results, so I used a testing platform to see how much return the model can achieve on historical data.