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.

TL;DR¶

MASFIN is a five-stage multi-agent pipeline built on CrewAI that integrates GPT-4.1-nano with structured financial metrics (Finnhub, Yahoo Finance) and unstructured news sentiment to construct weekly equity portfolios with explicit survivorship, hindsight, and overfitting bias mitigations. Over an eight-week live evaluation it achieved 7.33% cumulative return, outperforming S&P 500 (4.92%), NASDAQ-100 (5.36%), and Dow Jones (4.11%) in six of eight weeks, at the cost of higher volatility (2.61% weekly).

Problem¶

Existing financial AI systems suffer from: (1) single-modality inputs that ignore either quantitative metrics or sentiment; (2) reliance on proprietary data (e.g., BloombergGPT) that undermines reproducibility; and (3) failure to guard against classic financial ML pitfalls—survivorship bias (excluding delisted firms), hindsight bias (look-ahead data leakage), and overfitting. Traditional quantitative models share these structural weaknesses.

Method¶

MASFIN is a sequential five-crew pipeline where each crew contains 3–5 LLM agents and produces a structured report as input to the next stage:

Postmortem Crew: Analyzes 18 delisted/at-risk firms across sectors using Finnhub headlines to extract failure signals, explicitly countering survivorship bias.
Screening Crew: Filters the full market to 50–100 candidate tickers via sentiment evaluation, market trend analysis, and rule-based screening using real-time Finnhub data.
Analysis Crew: Scores 35–50 tickers on a multi-metric quantitative framework—21D/5D returns, Sharpe/Sortino ratios, momentum, beta/alpha, RSI-14, volume trends, and a Global Mean Benchmarking method (Gm = (1/N)Σ M_i,m)—using strictly historical, contemporaneous Yahoo Finance snapshots to prevent look-ahead.
Timing Crew: Issues buy/hold/sell decisions for 20–30 candidates using Sortino ratio, return z-score, regression slope, and volume; all decisions are ex-ante only.
Portfolio Crew: Allocates weights across a 15–30 stock portfolio balancing risk-adjusted returns, diversification, and conflict resolution.

Human-in-the-loop (HITL) manual handoffs between crews validate outputs and mitigate hallucinations. The backbone LLM is GPT-4.1-nano for cost and reproducibility.

Key Contributions¶

Five-stage CrewAI multi-agent pipeline with explicit bias-mitigation at each stage (survivorship, hindsight, overfitting)
HITL validation at crew boundaries as a practical safeguard against LLM hallucinations and context-window overload
Fully reproducible system using only open data sources (Yahoo Finance, Finnhub) with public code
Live-market evaluation against major index benchmarks over eight weeks
Design principles for bias-aware, modular multi-agent systems transferable to other high-stakes domains

Results¶

MASFIN cumulative return: 7.33% over 8 weeks vs. NASDAQ-100 (5.36%), S&P 500 (4.92%), Dow Jones (4.11%)
75% win rate (positive return in 6/8 weeks), comparable to NASDAQ and S&P 500 weekly win rates
Weekly volatility: 2.61%—highest among all compared benchmarks (higher risk, higher return quadrant)
Pearson correlation with S&P 500: 0.97; with NASDAQ: 0.95 (amplified market beta, not alpha generation)

Limitations¶

Evaluation window is only eight weeks—insufficient to distinguish skill from luck or to test across market regimes
No statistical inference (confidence intervals, hypothesis testing) to assess significance of outperformance
No learning mechanism; portfolio construction does not improve from prior weeks
HITL manual handoffs were required due to LLM context-length limits on end-to-end processing, reducing full automation
High API costs with fully automated execution; HITL was partly adopted as a cost mitigation
Not compared against other AI-based financial systems, limiting contextualization of results
High market correlation (r≈0.97) suggests returns may reflect leveraged market exposure rather than genuine alpha

Relevance to Agentic AI / LLM Agents¶

MASFIN is a concrete demonstration of task decomposition via sequential multi-agent pipelines in a high-stakes domain, making it a useful reference architecture for agent system designers. The HITL integration at crew boundaries addresses a practical pain point in agentic systems—context-length limits and hallucination risk—by treating human review as a first-class component of the workflow rather than a fallback. The explicit bias-mitigation protocols per agent crew illustrate how domain-specific constraints (temporal data integrity, survivorship accounting) can be encoded structurally rather than through prompt engineering alone. The paper also surfaces the tension between full automation and reliability that is central to the agentic AI field.