Skip to content

MASFIN: A Multi-Agent System for Decomposed Financial Reasoning and Forecasting

🕒 Published (v1): 2025-12-26 06:01 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MASFIN is a five-stage multi-agent pipeline (built on CrewAI) for short-term equity portfolio construction that integrates structured financial metrics (Finnhub, Yahoo Finance) with unstructured news sentiment under explicit survivorship-, hindsight-, and overfitting-bias mitigation protocols. It uses GPT-4.1-nano for cost efficiency and human-in-the-loop (HITL) handoffs between stages. Over an 8-week live evaluation it returned 7.33% cumulative, beating S&P 500, NASDAQ-100, and DJIA.

Problem

Existing AI-driven financial forecasting systems suffer from: (1) single-modality inputs (metrics or sentiment, not both); (2) reliance on proprietary data pipelines (e.g., BloombergGPT) limiting reproducibility; (3) inadequate safeguards against survivorship bias (ignoring delisted firms), hindsight bias (data leakage), and overfitting; and (4) monolithic architectures that exceed LLM context limits and reduce interpretability.

Method

MASFIN decomposes financial forecasting into five sequential "crews," each staffed by 3–5 LLM agents:

  1. Postmortem Crew — analyzes 18 delisted/at-risk firms via Finnhub headlines to extract failure signals, directly countering survivorship bias.
  2. Screening Crew — filters the market to 50–100 tickers using real-time sentiment and Postmortem signals.
  3. Analysis Crew — evaluates 35–50 candidates with a fixed-window quantitative framework (21D/5D returns, Sharpe, Sortino, max drawdown, RSI-14, beta/alpha, volume trend, price vs. 5D MA) from Yahoo Finance; Finnhub headlines provide qualitative context without real-time leakage. Global Mean Benchmarking normalizes metrics across the cohort.
  4. Timing Crew — issues buy/hold/sell decisions using ex-ante-only data (Sortino, momentum, regression slope, volume) to prevent look-ahead bias.
  5. Portfolio Crew — resolves conflicts and allocates weights to a 15–30 stock portfolio under diversification and risk-adjusted constraints.

All inter-crew handoffs pass through a Summary Agent producing structured reports, and a human validates each handoff (HITL). The framework runs on CrewAI with GPT-4.1-nano.

Key Contributions

  • Five-stage modular multi-agent pipeline with explicit, per-stage bias mitigation (survivorship, hindsight, overfitting).
  • Open-data reproducibility: Finnhub + Yahoo Finance only; full code released.
  • HITL workflow as a practical solution to LLM context-length limits and hallucination risk in end-to-end pipelines.
  • Design principles for multi-agent systems in high-stakes domains: task decomposition, HITL oversight, and bias-aware modularity as generalizeable patterns.
  • Live-market evaluation (not backtest) against major index benchmarks over 8 weeks.

Results

  • Cumulative return: 7.33% vs. NASDAQ-100 5.36%, S&P 500 4.92%, DJIA 4.11% over 8 weeks.
  • Positive weekly return in 6 of 8 weeks (75% win rate), matching NASDAQ and S&P 500 win rates.
  • Weekly volatility: 2.61%, highest among all compared instruments (higher risk, higher return quadrant).
  • Correlation with S&P 500: 0.97; with NASDAQ: 0.95 — amplified market-beta behavior.

Limitations

  • 8-week evaluation window is too short for statistical inference; no confidence intervals or hypothesis testing reported.
  • No learning/adaptation mechanism; portfolio logic does not update from prior-week outcomes.
  • HITL manual handoffs reduce automation and scalability; full automation exceeded LLM context limits and API cost budgets.
  • No direct comparison against other AI-based financial systems; outperformance relative to index benchmarks only.
  • Higher volatility (2.61% weekly) than all benchmarks; risk-adjusted superiority is asserted but not formally quantified.

Relevance to Harnesses / Meta-Harnesses

MASFIN is a concrete domain-specific meta-harness: it orchestrates multiple specialized LLM agent crews in a fixed sequential topology, with structured inter-stage data contracts (Summary Agent reports) and HITL validation gates functioning as reliability checkpoints between pipeline stages. The paper's candid discussion of why a monolithic end-to-end architecture failed (context limits, cost, interpretability loss) and how decomposition into sequential crews resolved it is directly instructive for harness designers. The explicit separation of concerns across crews — each with its own data sources, bias-mitigation mandate, and output schema — exemplifies the modular pipeline pattern central to harness design. The HITL handoff mechanism is a practical pattern for injecting human oversight into otherwise automated multi-agent chains without requiring full human-in-the-loop at every agent step.