SEMINAR: LLM News Sentiment and PPO for Stock Portfolio Optimization
Guest: Kemal Kırtaç
Title: LLM News Sentiment and PPO for Stock Portfolio Optimization
Date/Time: February 2026, 13:40
Location: https://sabanciuniv.zoom.us/my/balcisoy
Abstract: Large Language Models (LLMs) are transforming financial sentiment analysis by capturing contextual, emotional, and semantic cues that traditional dictionary-based methods often miss. In this seminar, I will present two published papers on LLM-based financial news sentiment and trading performance, and one ongoing paper that integrates sentiment into PPO-based reinforcement learning (SAPPO) for stock portfolio optimization. This presentation introduces an end-to-end framework for sentiment-driven trading using advanced transformer models, including BERT, FinBERT, OPT, and LLaMA-3, fine-tuned on nearly one million firm-specific financial news articles from Refinitiv. After fine-tuning the models on 3-day abnormal returns, I evaluate their predictive accuracy, econometric significance, and trading performance across long, short, and long–short portfolios. Models such as OPT and BERT consistently outperform dictionary baselines, delivering the highest classification accuracy, strongest predictive coefficients for next-day returns, and robust long–short Sharpe ratios under realistic market frictions. I further integrate LLM-derived sentiment into a reinforcement-learning portfolio strategy—Sentiment-Augmented PPO (SAPPO)—which improves risk-adjusted performance relative to standard PPO and benchmark equity indices. The results demonstrate that LLM-based sentiment signals can enhance market prediction, support systematic trading, and strengthen reinforcement-learning policy optimization. The presentation concludes with implications for quant research, model interpretability, and cross-market generalization, highlighting LLMs as a scalable and powerful component of modern investment pipelines.
Bio:Kemal Kırtac is a PhD Candidate in Computer Science at University College London, working at the intersection of reinforcement learning, large language models, and quantitative finance. His research develops LLM-based sentiment signals from large-scale financial news and integrates human-derived information into sequential decision-making and portfolio optimization. He has published on financial sentiment and trading using LLMs, with work appearing at ACL (2024, 2025) as well as ICLR (2025) and ICML (2025), and has presented his research at various international conferences. He also studies policy-gradient and actor–critic methods such as PPO for decision-making under uncertainty.