Application of advanced Machine Learning algorithms in Finance

This thesis explores the application of advanced machine learning techniques in finance, specifically focusing on portfolio management using deep reinforcement learning. We propose and implement a hybrid PPO-LSTM model, combining the stability of Proximal Policy Optimization with the memory capacity of a recurrent LSTM network. The trained agent learns to optimize daily portfolio weights across a diversified basket of stocks, explicitly accounting for transaction costs. We evaluate three architectural variants along with a simple ensemble created by averaging their outputs. Using walk-forward validation, we assess model performance on data from 2010 to 2023. The results show that the agents tend to mitigate extreme drawdowns and achieve Sharpe ratios comparable to a passive benchmark, though often at the cost of lower exposure and higher turnover. The compact model performs best, offering a favorable balance between network complexity, gradient stability, and generalization. Additionally, we analyze turnover dynamics and sector exposure, which helps us to interpret the agent's behavior over time.