Finrankgrpo improves financial asset ranking with large language models
FinRankGRPO: Optimizing LLMs for Listwise Financial Asset Ranking via Group Relative Policy Optimization
Computational Engineering, Finance, and Science
Summary
Financial experts use lists of assets to decide how to build portfolios, but Large Language Models (LLMs) usually predict one word at a time and struggle with precise numbers. The authors present FinRankGRPO, a method that teaches LLMs to rank assets instead of just guessing numbers. They train the model to think in steps and then reward it for matching real market rankings. Their approach outperforms other methods in balancing risk and returns and better matches actual financial asset orders.
What this means in practice
- •For portfolio managers: Use FinRankGRPO to generate more reliable rankings of assets for better portfolio selections.
- •For financial software developers: Incorporate FinRankGRPO’s ranking method to improve AI-driven asset allocation tools.$Commercial implications: Enables development of commercial AI-based portfolio management products with improved ranking accuracy.
Authors
Ningyuan Deng, Jinyuan Wang, Qi Li, Jia Zhang, Yi Yang
Abstract
While Large Language Models (LLMs) excel at understanding unstructured financial contexts, their direct use in portfolio optimization is limited by a mismatch between next-token prediction and the listwise ranking objectives required for asset allocation. They also struggle with precise numerical forecasting, leading to instability and arithmetic hallucinations. To bridge this gap, we propose FinRankGRPO, a framework that shifts LLM based portfolio construction from direct numerical prediction to listwise ranking of financial assets. We introduce a two-stage training process, supervised finetuning on Chain-of-Thought reasoning data, followed by our Financial Asset Ranking via Group Relative Policy Optimization with a Spearman rank correlation reward that aligns generated asset rankings with ground truth market orderings. The second stage uses a novel Spearman rank correlation reward to explicitly align the model's generative preferences with ground truth market orderings. Experimental results show that FinRankGRPO outperforms traditional quantitative and state-of-the-art commercial models, achieving a Sharpe ratio of 0.636 and a Spearman correlation of 0.023.