Financial data analysis by means of Reinforcement Learning techniques

These state-of-the-art Reinforcement Learning techniques have already demonstrated to be profitable, being efficient in different market conditions and beating the benchmarks. Despite the results, the classic rewarding methodology brings to model instability, overfitting and ineffectiveness in taking into account medium and long-term trends, and difficulties in diversifying the portfolio exposure to the market, while not being interpretable enough to be inspected before any deployment. Solely providing the agent with historical price data information is often not enough to enable it to learn comprehensive and reliable strategies. Especially when dealing with multistock trading problems the complexity of the process increases, due to the fluctuating nature of the market and the demand to manage a structured portfolio, as it is necessary for the agent to take into account several factors such as distinguishing how much to allocate on each asset, which positions to open in what direction, and what is the best condition to close them in order to accumulate the maximum reward in an unknown environment. In these circumstances, the choice of the reward function turns out to be a fundamental aspect, and must be tailored to the trading context in order to allow the model to learn how to act efficiently. For these reasons, in this thesis we empower the RL algorithms with a large series of data containing, for each trading day, information on prices’ moving averages, historical volumes, trend momentum indicators, volatility-related information and a correlation matrix between the stocks considered. In particular, the thesis research aims to take advantage of the implementation of the metrics associated to the historical returns and the asset allocation into the reward function in order to give to the model a clear view about the trader risk-aversion profile and to assist a potential user in his trading operations. The models studied in this thesis are trained on price and volume data from January 1st, 2001 to October 31th, 2021, for each stock in the Dow Jones provided by Yahoo Finance. The analysis of the results shows that our approach can accurately interfere with the behaviour of the agent with a high interpretability level and can be applicable in a real-world working application, with an integration in the workflow of the risk-management office, empowering the operators with a tool tailored on their risk-aversion profile.