Deep Reinforcement Learning methods and LLMs for dynamic portfolio management

Abstract

This thesis investigates the application of Deep Reinforcement Learning (DRL) for dynamic portfolio management of stocks enhanced with sentiment analysis features from Large Language Models (LLMs). The focus of the thesis is to understand if the prediction capabilities of DRL models with classical financial indicators can be improved with the addition of a sentiment score. Sentiment analysis, derived from social media posts and extracted with a BERT model, is used to add information and capture the qualitative side of financial markets. The thesis explores the performance of three RL algorithms, Proximal Policy Optimization (PPO), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Soft Actor-Critic (SAC), which proved to be effective algorithms for this task. The performance of the models is evaluated by comparing it with and without sentiment score.