Leveraging LLM-based sentiment analysis for portfolio allocation with proximal policy optimization

Kemal Kirtac, Guido Germano

Track: New scientific result

Keywords: Deep reinforcement learning, stock return forecasting, large language models, portfolio allocation

Abstract:

Portfolio optimization requires adaptive strategies to maximize returns while managing risk. Reinforcement learning (RL) has gained traction in financial decision-making, and proximal policy optimization (PPO) has demonstrated strong performance in dynamic asset allocation. However, traditional PPO relies solely on historical price data, ignoring market sentiment, which plays a crucial role in asset movements. We propose a sentiment-augmented PPO (SAPPO) model that integrates daily financial news sentiment extracted from Refinitiv using LLaMA 3.3, a large language model optimized for financial text analysis. The sentiment layer refines portfolio allocation by incorporating real-time market sentiment alongside price movements. We evaluate both PPO and SAPPO on a three-stock portfolio consisting of Google, Microsoft and Meta, and we compare performance against standard market benchmarks. Results show that SAPPO improves risk-adjusted returns with a superior Sharpe ratio and reduced drawdowns. Our findings highlight the value of integrating sentiment analysis into RL-driven portfolio management.

Submission Number: 28