Reproduce “FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Dorothy Torres, Wei Cheng, Henan Huang

FinAbstain proposes a financial forecasting system that checks whether time-valid, multimodal evidence supports a directional call and abstains or seeks help when uncertainty is high. Specialist agents assess filings, news, market data, and risks, while calibration and a selective controller govern decisions. In simulated results, five-day selective accuracy reached .688 at .72 coverage, but no real-market payoff was demonstrated.

Key insights

  • Abstention is treated as an explicit decision governed by risk–coverage or cost-sensitive objectives, rather than as a failure to forecast.
  • Temporal validity requires tracking both publication and ingestion times, versioning amendments, and recomputing market features only from data available at the forecast timestamp.
  • Calibration should be assessed across time and market regimes because temporal dependence and regime shifts weaken the assumptions behind standard conformal guarantees.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.