Adaptive Confidence-Weighted LLM Infusion for Financial Reinforcement Learning

How can we make financial AI models more trustworthy, cautious, and accurate in their market valuations?

We’re happy to share that Aydin Javadov, PhD student at Mobiliar Lab for Analytics at ETH Zurich, coauthored a paper presented at the 2025 IEEE International Conference on Intelligent Data and Security (IDS), in the special track on Financial Reinforcement Learning and Foundation Models (FinRLFM).

In collaboration with Imperial College London, our work introduces “Adaptive Confidence-Weighted LLM Infusion for Financial Reinforcement Learning”, a method that improves AI-driven trading by using stock signals from language models, while adjusting their influence based on confidence levels and refining decision-making through entropy regularization.

Key highlights:
– LLM signals are scaled based on self-reported confidence, helping to filter out uncertain or noisy information
– Empirical results show improved stability and better risk-adjusted returns in real-world market backtests
– Findings were robust across two types of reinforcement learning agents: proximal policy optimization and risk-sensitive (CVaR) models

These results suggest that modeling confidence explicitly can support more reliable and realistic financial decision-making in AI systems.

Read the published paper: https://lnkd.in/eSfeMVra

Thanks to our collaborators at Imperial College London and to ETH Zürich, D-MTEC ETH Zurich, and die Mobiliar for their support.