Portfolio Management Application based on Ranking and Reinforcement Learning

This thesis presents an integrated system for constructing “intelligent” portfolios by combining multimodal machine learning techniques with reinforcement learning for autonomous investment. In the first stage, the ranking algorithm incorporated technical and fundamental indicators to decide which stocks to select on a monthly basis, signaling the best-performing stocks. The ranking process relies on the XGBoost ranking algorithm, which arranges the monthly future returns of S&P 500 stocks from worst to best and divides them into ten quantiles, with the lowest-ranked stocks assigned to rank 1 and the highest return to rank 10. After testing the algorithm’s stability, the optimal portfolio is selected based on performance and utility (with a maximum of 20 stocks). An additional filtered portfolio is then constructed by selecting the top-performing stocks across business sectors (no more than two per sector), ensuring both diversification and limited exposure to specific industries. The final stage involves the automatic allocation of capital among the selected stocks through the use of reinforcement learning. In this framework, the agent dynamically adjusts the portfolio composition to optimize returns, leveraging both technical indicators, stock characteristics and real-time news (e.g. general and stock-specific news). A comparative evaluation of state-of-the-art reinforcement learning algorithms (A2C, DDPG, PPO, TD3, and SAC) is conducted to assess their ability to learn effective capital allocation strategies. Performance metrics and backtesting demonstrate that the proposed hybrid approach, which combines ranking methods and reinforcement learning with financial news integration, leads to an intelligent, adaptive, and strategically balanced investment system. Our study demonstrates that the agent learns robust and adaptive investment policies, and the proposed hybrid framework outperforms both the SP500 and the ranking model, offering a data-driven and self-adaptive approach to portfolio management that is responsive to both market trends and external information.