Skip to content

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

Conference: NeurIPS 2025 arXiv: 2512.21878
Code: None
Area: LLM Agent
Keywords: multi-agent system, financial reasoning, decomposed forecasting, LLM agent, time-series analysis

TL;DR

This paper proposes MASFIN, a multi-agent system that decomposes financial forecasting into multiple sub-tasks (macroeconomic analysis, industry analysis, technical analysis, sentiment analysis, etc.), with specialized LLM agents collaborating to produce more accurate and interpretable financial predictions than single-model approaches.

Background & Motivation

Background: Financial forecasting requires simultaneous consideration of multiple dimensions, including macroeconomic conditions, industry trends, technical indicators, and market sentiment, which is difficult for any single model to cover comprehensively.

Limitations of Prior Work: (1) Traditional time-series models lack reasoning capabilities; (2) Single LLMs tend to overlook critical factors in multi-dimensional analysis; (3) The heterogeneous, multi-source nature of financial data makes it difficult for a single model to handle effectively.

Key Challenge: A fundamental tension exists between comprehensiveness and specialization โ€” a single model is either broad but shallow, or narrow but deep.

Key Insight: Multi-agent division of labor โ€” each agent focuses on one analytical dimension, with a coordinator agent synthesizing the results.

Method

Overall Architecture

A Coordinator Agent receives the forecasting task โ†’ assigns it to multiple expert agents (Macro Agent, Industry Agent, Technical Agent, Sentiment Agent) โ†’ each agent conducts independent analysis โ†’ the Coordinator performs integrative reasoning โ†’ final prediction is produced.

Key Designs

  1. Task Decomposition Strategy

    • Function: Decomposes financial forecasting into 4โ€“5 independent sub-tasks
    • Mechanism: Macroeconomic analysis (GDP, interest rates, inflation), industry analysis (competitive landscape, supply and demand), technical analysis (price trends, indicators), and sentiment analysis (news, social media)
    • Design Motivation: Each agent contributes signals from a distinct dimension; aggregation yields more robust predictions
  2. Integrative Reasoning Mechanism

    • Function: The Coordinator synthesizes analytical reports from all agents
    • Mechanism: Signals from each dimension are weighted and fused, with consistency and conflicts taken into account, to generate a final prediction along with a reasoning chain
    • Design Motivation: Interpretability โ€” users can trace predictions back to specific analytical dimensions

Loss & Training

No training is required. Each agent performs in-context reasoning using pre-trained LLMs (e.g., GPT-4).

Key Experimental Results

Main Results

Method Directional Accuracyโ†‘ MSEโ†“ Interpretability
ARIMA 52.3% 0.045 Low
LSTM 56.7% 0.038 Low
GPT-4 (single model) 58.2% 0.035 Medium
MASFIN 63.5% 0.029 High

Ablation Study

Configuration Directional Accuracy Note
Technical only 55.1% Price trends
Macro + Technical 59.3% Two dimensions
Sentiment + Technical 58.8% Sentiment supplement
All agents 63.5% Best

Key Findings

  • The multi-agent system improves directional forecasting accuracy by 5.3 pp over single-model GPT-4
  • Each additional analytical dimension contributes 1โ€“3 pp improvement, with diminishing marginal returns
  • Interpretability is significantly enhanced โ€” users can inspect the analytical report from each dimension

Highlights & Insights

  • Divide-and-conquer paradigm: The inherently multi-dimensional nature of financial forecasting is well-suited to multi-agent specialization, allowing each agent to employ the most appropriate prompts and tools.
  • Interpretable forecasting: Final predictions are accompanied by a complete reasoning chain, enabling users to understand and verify the basis of each forecast.

Limitations & Future Work

  • High API call costs due to multiple agents across multiple dialogue turns
  • Limited iterative interaction among agents (currently single-round analysis followed by synthesis)
  • Evaluation covers a restricted time period; long-horizon forecasting performance remains unverified
  • No comparison against professional financial analysts
  • vs. FinGPT: FinGPT is a single fine-tuned model, whereas MASFIN achieves more comprehensive analysis through multi-agent specialization
  • vs. MetaGPT: MetaGPT targets software engineering; MASFIN introduces the multi-agent paradigm to the financial domain

Rating

  • Novelty: โญโญโญ Multi-agent framework applied to finance, though the approach is relatively straightforward
  • Experimental Thoroughness: โญโญโญ Limited datasets; large-scale backtesting is absent
  • Writing Quality: โญโญโญโญ Clear and well-structured
  • Value: โญโญโญโญ Interpretable financial forecasting addresses a genuine practical need