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

AI summary

Overview

Research area: Multi-agent large language model systems applied to quantitative finance and short-term equity forecasting.

Technical level: Intermediate. The paper assumes familiarity with basic portfolio metrics (Sharpe ratio, drawdown, volatility) and with LLM agent architectures, but the pipeline itself is described in plain terms.

Scope: The paper presents MASFIN, a five-stage multi-agent pipeline built on CrewAI with GPT-4.1-nano that combines structured financial metrics with news sentiment to build weekly 15–30 stock portfolios under explicit bias-mitigation protocols, evaluated over eight weeks against the S&P 500, NASDAQ-100, and Dow Jones.

What This Paper Is About

Short-term stock prediction is hard because markets are volatile and non-stationary, and because useful signals are split between quantitative data (financial ratios, prices) and qualitative data (news sentiment). Many existing AI financial systems rely on a single modality or on proprietary data such as BloombergGPT, which limits reproducibility, and traditional quantitative models remain vulnerable to survivorship bias, hindsight bias, and overfitting. MASFIN's goal is to build a transparent, reproducible, low-cost alternative that integrates both signal types while embedding safeguards against those three biases.

Key Contributions

1. The MASFIN framework: A five-stage, multi-agent pipeline (Postmortem, Screening, Analysis, Timing, Portfolio) that integrates Finnhub and Yahoo Finance data with news sentiment and statistical insights, embedding explicit safeguards against survivorship, hindsight, and overfitting bias, incorporating human-in-the-loop validation to mitigate hallucinations, and evaluated in live-market conditions against major benchmarks.
2. Design principles for multi-agent systems: Using finance as a high-stakes testbed, the paper demonstrates how task decomposition, human-in-the-loop oversight, and bias-aware modular design can improve reliability, interpretability, and affordability of multi-agent generative systems.
3. Reproducible, low-cost implementation: The system uses GPT-4.1-nano for reproducibility and cost-efficient inference, relies on openly available data, and releases Python and reproducibility files at github.com/mmontalvo9/MASFIN.
4. A live-market evaluation: An eight-week, weekly-rebalanced evaluation against the S&P 500 (SPY), NASDAQ-100 (QQQ), and Dow Jones Industrial Average (DIA).

Main Findings

• Cumulative return: MASFIN delivered a 7.33% cumulative return over eight weeks, above the NASDAQ (5.36%), S&P 500 (4.92%), and Dow Jones (4.11%).
• Win rate: MASFIN produced positive returns in six of eight weeks, described as a 75% win rate comparable to the NASDAQ and S&P 500, and it consistently outperformed in total return.
• Higher risk: MASFIN's weekly volatility was 2.61%, the highest among all benchmarks, placing it in the high-risk, high-return quadrant. The paper states the magnitude of gains suggests a favorable risk-adjusted profile.
• Market correlation: MASFIN showed strong correlation with the S&P 500 (0.97) and the NASDAQ (0.95), which the authors interpret as amplified performance within existing market trends.
• Practical constraints observed: End-to-end architectures ingesting all metrics, headlines, and analyses exceeded LLM context limits and reduced interpretability, and fully automated execution raised API costs, motivating the human-in-the-loop workflow with manual handoffs.
• Validation gaps acknowledged: MASFIN lacks a learning mechanism and statistical inference tools such as confidence intervals or hypothesis testing, and the eight-week window, while limiting overfitting to specific market regimes, is short.

Methodology in Plain English

MASFIN processes each forecasting cycle through five sequential "crews," each containing 3–5 LLM-based agents, implemented on CrewAI. Every crew except the Portfolio Crew includes a Summary Agent that condenses outputs into a structured report, which serves as the manual handoff point in a human-in-the-loop workflow.

1. Postmortem Crew — Studies Finnhub headlines for eighteen delisted or at-risk companies across sectors (EV/Auto, Biotech/Pharma, Consumer/Retail, Cannabis, Tech & Other) to extract failure signals, sentiment dynamics, and bias indicators. Analyzing delisted firms directly counters survivorship bias.
2. Screening Crew — Uses real-time Finnhub headlines plus Postmortem context to filter the market down to a shortlist of 50–100 tickers, applying sentiment evaluation, trend analysis, and rule-based screening. It uses contemporaneous data to avoid hindsight bias.
3. Analysis Crew — Evaluates screened tickers and prior-week holdings to identify 35–50 short-term outperformers, computing metrics from Yahoo Finance while using Finnhub headlines for qualitative context. Metrics cover 21D and 5D returns, momentum (21D price change), annualized volatility, maximum drawdown, Sharpe, Sortino, beta, alpha, RSI-14, 5D return z-score, volume trend, and price versus 5D moving average, contextualized through a "Global Mean" benchmarking formula against the cohort average. All metrics use fixed historical windows and contemporaneous snapshots to prevent look-ahead bias.
4. Timing Crew — Decides whether candidates are appropriately timed for near-term entry, relying strictly on historical data, and using Sortino ratio, return z-score, momentum, regression slope, and trading volume to produce buy, hold, or sell decisions. A unified decision schema flags inconsistencies or elevated risks, yielding 20–30 candidates.
5. Portfolio Crew — Consolidates outputs into a 15–30 stock portfolio with allocation weights, resolving conflicts, challenging weak selections, and ensuring diversification by balancing risk, limiting concentration, and grounding each inclusion in both statistical evidence and external signals.

Why This Matters

Impact on research: The paper positions finance as a demanding testbed for generative AI, arguing that task decomposition, human-in-the-loop oversight, and bias-aware modular design are transferable to other high-stakes domains. It also pushes back on proprietary-data pipelines by showing that an openly documented, low-cost configuration can be evaluated in live-market conditions.

Real-world applications:

• Weekly portfolio construction and rebalancing for short-term equity strategies.
• Bias auditing of financial models, particularly screening out delisted firms and preventing look-ahead leakage.
• Sentiment-aware screening tools that combine news flow with standard quantitative metrics.
• Human-in-the-loop decision support, where an analyst validates and adjusts agent output before capital is committed.

Industry relevance: The paper targets the cost and reproducibility gap between proprietary financial LLM pipelines and practical, transparent alternatives. Its use of GPT-4.1-nano and open data sources speaks to affordability, while the reported 2.61% weekly volatility signals that the return advantage came with meaningfully higher risk. The authors frame interpretability, not full automation, as the deliberate trade-off.

Future Directions

• Comparing MASFIN directly with other AI-based financial systems to better contextualize its performance.
• Adding statistical inference tools such as confidence intervals and hypothesis testing, which the system currently lacks.
• Introducing a learning mechanism, since the current system produces weekly portfolios but does not learn from outcomes.
• Extending evaluation beyond eight weeks, which the authors note would need to be balanced against the benefit of limiting overfitting to specific market regimes.

Target Audience

Researchers and practitioners working on multi-agent LLM systems, generative AI in finance, or agentic pipelines for high-stakes decision-making. It is also suited to quantitative finance professionals interested in bias mitigation and reproducibility, and to readers evaluating whether modular agent design with human-in-the-loop oversight can substitute for proprietary, single-model financial AI approaches.

Authors’ abstract

Recent advances in large language models (LLMs) are transforming data-intensive domains, with finance representing a high-stakes environment where transparent and reproducible analysis of heterogeneous signals is essential. Traditional quantitative methods remain vulnerable to survivorship bias, while many AI-driven approaches struggle with signal integration, reproducibility, and computational efficiency. We introduce MASFIN, a modular multi-agent framework that integrates LLMs with structured financial metrics and unstructured news, while embedding explicit bias-mitigation protocols. The system leverages GPT-4.1-nano for reproducability and cost-efficient inference and generates weekly portfolios of 15-30 equities with allocation weights optimized for short-term performance. In an eight-week evaluation, MASFIN delivered a 7.33% cumulative return, outperforming the S&P 500, NASDAQ-100, and Dow Jones benchmarks in six of eight weeks, albeit with higher volatility. These findings demonstrate the promise of bias-aware, generative AI frameworks for financial forecasting and highlight opportunities for modular multi-agent design to advance practical, transparent, and reproducible approaches in quantitative finance.