FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting

Short summary

FinAbstain proposes a multimodal RAG framework for financial forecasting that abstains from predictions when uncertainty is too high. It uses modality-specific agents (fundamental, news, technical, risk, verification) with point-in-time retrieval and multiple calibration techniques (temperature scaling, isotonic regression, conformal prediction). Results are simulated and explicitly labeled as such, illustrating that calibrated abstention can trade coverage for lower selective error and drawdown.

• •Multimodal RAG framework with selective prediction for financial forecasting
• •Modality-specific agents aggregate probabilistic assessments with uncertainty scoring
• •Controller abstains or routes to human review when uncertainty exceeds threshold
• •Simulated results demonstrate intended hypothesis; no empirical claims yet

Generated with AI, which can make mistakes.