The List / Agents and Multi-Agent Systems / paper
P1GPT: a multi-agent LLM workflow module for multi-modal financial information analysis
Abstract
Recent advances in large language models (LLMs) have enabled multi-agent reasoning systems capable of collaborative decision-making. However, in financial analysis, most frameworks remain narrowly focused on either isolated single-agent predictors or loosely connected analyst ensembles, and they lack a coherent reasoning workflow that unifies diverse data modalities. We introduce P1GPT, a layered multi-agent LLM framework for multi-modal financial information analysis and interpretable trading decision support. Unlike prior systems that emulate trading teams through role simulation, P1GPT implements a structured reasoning pipeline that systematically fuses technical, fundamental, and news-based insights through coordinated agent communication and integration-time synthesis. Backtesting on multi-modal datasets across major U.S. equities demonstrates that P1GPT achieves superior cumulative and risk-adjusted returns, maintains low drawdowns, and provides transparent causal rationales. These findings suggest that structured reasoning workflows, rather than agent role imitation, offer a scalable path toward explainable and trustworthy financial AI systems.
In one sentence
Summary
P1GPT addresses the lack of coherent, multi-modal reasoning in financial LLMs by proposing a layered multi-agent architecture that replaces role-playing with a structured pipeline. The system comprises an Input Layer for data parsing, a Planning Layer for task decomposition, an Analysis Layer with domain-specific Intelligent Specialized Agents (ISAs) for fundamentals, technicals, news, and sectoral insights, and an Integration Layer that fuses these outputs via standardized reports and conflict resolution. This design aims to provide transparent, auditable decision-making suitable for regulated environments, contrasting with single-agent predictors or loosely connected ensembles that lack unified reasoning workflows. The framework is evaluated through backtesting on three major U.S. equities—Apple, Google, and Tesla—spanning February to September 2025, a period marked by post-election uncertainty and tariff volatility. Data sources include real-time news, social media sentiment, technical indicators (MACD, RSI), and fundamental metrics from EDGAR and Yahoo Finance. The system is compared against Buy & Hold, MACD, KDJ+RSI, Zero-Mean Reversion, and SMA baselines. The experimental design ensures no lookahead bias, with agents accessing only information available prior to the decision timestamp, and outputs are generated as Buy/Sell/Hold signals accompanied by natural language rationales. Results indicate that P1GPT achieves superior cumulative and risk-adjusted returns with lower maximum drawdowns than all baselines, demonstrating the efficacy of structured multi-modal fusion. The system exhibits stable behavioral patterns, such as disciplined exits during regime shifts and re-entry after drawdowns, driven by aligned multi-modal signals. However, the study abstracts away transaction costs, leverage, and portfolio interactions, limiting external validity. Future work must address realistic execution frictions, multi-asset allocation, and formal causal tracing to assess scalability and robustness in live trading environments.
Research problem
Current financial LLM frameworks often rely on isolated single-agent predictors or loosely connected analyst ensembles that lack a coherent reasoning workflow to unify diverse data modalities. Existing systems also struggle with unstructured inter-agent communication, limited explainability, and scalability in complex markets. This paper addresses the need for a structured, multi-agent reasoning pipeline that systematically fuses technical, fundamental, and news-based insights to provide interpretable and trustworthy trading decisions.
Core contributions
- Proposes P1GPT, the first end-to-end layered multi-agent LLM workflow tailored for multi-modal financial analysis.
- Designs a structured reasoning and communication protocol using standardized agent reports to improve interpretability and reliability.
- Introduces a five-layer architecture (Input, Planning, Analysis, Integration, Decision) that mirrors professional trading firm workflows.
- Demonstrates through backtesting that structured reasoning workflows outperform rule-based and single-agent LLM baselines in cumulative and risk-adjusted returns.
- Provides a mechanism for dynamic conflict resolution and cross-agent synthesis to handle discrepancies in multi-modal data.
Method
- Input Layer: Parses user queries and multi-modal data using entity extraction and NLP preprocessing.
- Planning Layer: Decomposes tasks, assigns them to Intelligent Specialized Agents (ISAs), and maps dependencies.
- Analysis Layer: Executes domain-specific analysis using Fundamental, Technical, News, and Semiconductor ISAs.
- Integration Layer: Aggregates outputs, performs cross-agent reasoning, and synthesizes insights using rule-based and LLM-driven fusion.
- Decision Layer: Generates actionable Buy/Sell/Hold signals with risk assessment and natural language explanations.
- Structured communication: Agents exchange standardized reports rather than free-form text to reduce ambiguity.
- Conflict resolution: The system triggers iterative feedback loops or external search when agent assessments conflict.
- Dynamic weighting: Confidence scores and dynamic weights are assigned to agent inputs based on context and performance.
Results & findings
- P1GPT achieves superior cumulative and risk-adjusted returns compared to all five baselines.
- The system maintains low maximum drawdowns, indicating better risk control than momentum or reversion strategies.
- Structured reasoning allows the system to adapt to regime shifts, such as tariff policy changes, by aligning multi-modal signals.
- The system provides transparent causal rationales for each Buy/Sell/Hold decision, enhancing auditability.
- Conflict resolution mechanisms trigger additional analysis when agent assessments disagree, improving decision robustness.
- Case studies show the system correctly identifying sell opportunities during overvaluation and buy opportunities during technical rebounds.
- The layered architecture enables modular extensibility, allowing new agents or data sources to be integrated without redesigning the entire pipeline.
Data & experiments
- Backtesting on three major U.S. equities: Apple (AAPL), Google (GOOGL), and Tesla (TSLA).
- Evaluation period from February 1, 2025, to September 30, 2025, covering post-election market uncertainty and tariff volatility.
- Data sources include real-time news (Bloomberg, Reuters), social media (X, Reddit), technical indicators (MACD, RSI), and fundamental data (EDGAR, Yahoo Finance).
- Baselines include Buy & Hold, MACD, KDJ+RSI, Zero-Mean Reversion, and Simple Moving Average strategies.
- Metrics used: Cumulative Return, Annualized Return, Sharpe Ratio, and Maximum Drawdown.
- Strict no-lookahead bias protocol ensuring agents only access information available prior to the decision timestamp.
Limitations & caveats
- Backtesting abstracts away transaction costs, leverage, and slippage, limiting real-world applicability.
- Evaluation is limited to single-stock positions, ignoring portfolio-level interactions and diversification benefits.
- The study covers a single market window (Feb-Sep 2025), which may not represent all market regimes.
- No formal causal tracing is provided to link specific agent outputs to final trading actions.
- The system's performance on less liquid or non-U.S. equities is not evaluated.
- Reliance on LLM-generated rationales may introduce hallucination risks, though mitigated by structured reporting.
Curator notes
Discovery status: round3_promoted. Source links: Automate Strategy Finding with LLM in Quant investment; Can Large Language Models beat wall street? Evaluating GPT-4’s impact on financial decision-making with MarketSenseAI; FLAG-Trader: Fusion LLM-Agent with Gradient-based Reinforcement Learning for Financial Trading; Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning; FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making; FinGPT: Democratizin