Reinforcement learning for financial trading

The integration of reinforcement learning into financial trading is an emerging area of study, building on the success RL has seen in domains like robotics and gaming. This thesis delves into RL's principles and explores its potential for crafting trading strategies within the financial markets. The research focuses on the Proximal Policy Optimization algorithm enhanced with Long Short-Term Memory networks to test its effectiveness on historical Bitcoin price data. PPO, an advanced learning algorithm, coupled with LSTM's ability to remember long-term patterns, is particularly well-suited to the task of navigating the complexities of financial time series. The strategies developed were backtested against historical Bitcoin price data to measure their effectiveness and potential for outperforming standard market benchmarks. The assessment of three distinct reward functions provided insight into how these strategies adapt to and perform under different market conditions. The results indicated that none of the reward functions consistently outperformed the benchmark throughout the experiments. However, certain reward functions demonstrated notable adaptability to significant market shifts. This adaptability suggests that while raw performance metrics may not always surpass benchmarks, the strategic value of RL lies in its potential to adjust to changing market environments, a critical capability for managing risk and pursuing long-term profitability in financial trading.