Point-in-Time Financial RAG with Frozen LLMs and Market-Feedback Adaptive Retrieval

The paper presents a novel approach to financial retrieval-augmented generation (RAG) systems that utilizes a frozen language model (LLM) and an adaptive retrieval mechanism based on Bayesian source memory. The method enhances predictive performance on a fixed dataset of 89 Nasdaq stocks, achieving a macro-F1 score improvement from 0.438 to 0.471 and a portfolio Sharpe ratio increase from 0.52 to 0.84 by incorporating market-context cards and feedback from residual-return signals. This work emphasizes the significance of optimizing retrieval strategies in financial applications, suggesting that effective evidence selection can be as critical as the reading capabilities of the model itself.