Reinforcement Learning for Dynamic Portfolio Optimization in Fintech

Abstract

This study assesses the performance of reinforcement learning (RL) algorithms for dynamic portfolio optimization, with a special emphasis on how they compare with conventional portfolio management techniques, like mean-variance optimization, under the fintech sector. The study applies different RL methods, like Q-learning, Deep Q Networks (DQN), and Proximal Policy Optimization (PPO), to maximize portfolio return and risk management under varying market scenarios. Utilizing a large-scale dataset of historical financial metrics and a variety of simulated market scenarios, this work evaluates performance on a cumulative return, Sharpe ratio, maximum drawdown, and Sortino ratio. Results indicate RL-based models, especially PPO, far exceed traditional methods in terms of returns generated and risk-adjusted performance. Despite the decreased performance on the models when faced with heightened volatility and backtesting financial crisis conditions, implications suggest the enhancement needs to stabilize robustness to extreme market behaviors. Generally speaking, the paper demonstrates the possible enhancement of portfolio optimization by the application of RL techniques but cautions further enhancement for better resistance to market turmoil.