Exploiting deep learning techniques for stock price prediction

Over the last decade, the analysis of financial time series has undergone an important development, both in terms of basic research and direct market applications. In particular, the computerization of data has made detailed information on price trends and traded volumes easily available, thus creating new fields of investigation. At the same time, the introduction of electronic trading systems has made large financial institutions interested in automating trading processes through algorithms. In this thesis, the problem of predicting closing stock prices over multivariate time series using Deep Learning models was addressed. Due to the volatility and mutability of the time series, it is necessary to introduce models that allow predicting future values based on past ones, with the objective that the investor faces a lower risk of loss. The use case consists of an analysis of data coming from the site Yahoo! Finance related to the stock index ISP.MI. In the beginning, the theory on which the neural networks are based will be deepened, the mathematics behind it will be explained, and it will be shown how they can adapt to the characteristics of the problem to be faced. We will then move on to the collection and preparation of the necessary data, and the implementation of the different Networks in Python code, exploiting an open-source library called Keras, the most widely used to use Deep Learning models. Starting from a database of 4 years, used to train several models, we want to predict the closing values of the days immediately following the input days. In particular, four models have been analyzed which, although based on the same principle of pattern detection through data encoding and decoding, present different properties. Finally, the performance of the models will be evaluated and, based on these, conclusions will be drawn on the actual effectiveness of the model, its limitations, and possible future developments. The objective of this thesis is to find a model that performs well in the approximation and prediction of time series in the financial domain.