[arXiv]score: 0.13
Bayesian Source Memory Improves Financial RAG Macro-F1 from 0.438 to 0.471
June 1, 2026
A frozen-LLM RAG system for event-driven financial prediction adapts its retrieval layer via a Bayesian source memory updated from realized market return feedback, rather than retraining the reader. On an 89-stock Nasdaq universe using FNSPID news and EDGAR filings, macro-F1 improves from 0.438 to 0.471 with downstream Sharpe ratio gains.
cs.CL
HOW THIS AFFECTS YOU
●
builderYou can improve financial RAG accuracy by routing retrieval through a feedback-updated source trust layer rather than relying solely on embedding similarity.
●
researcherBayesian source memory provides a tractable mechanism for adapting retrieval trust without touching the LLM, with measurable financial signal improvement.