Skip to content

Orchestration Framework for Financial Agents: From Algorithmic Trading to Agentic Trading

🕒 Published (v1): 2025-12-01 21:50 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

FinAgent is an orchestration framework that maps every component of a traditional algorithmic trading system (data, alpha, risk, portfolio, execution, backtest, audit, memory) to specialized LLM agents coordinated via MCP and A2A protocols. On a 7-stock equity backtest it achieves 20.42% return (Sharpe 2.63) versus SPY's 16.60% (Sharpe 1.86), and on BTC minute-data it achieves 8.39% versus buy-and-hold 3.80%, with lower drawdown in both cases.

Problem

Traditional algorithmic trading systems require professional teams and years of development; they are monolithic, hard to adapt, and inaccessible to non-experts. Existing LLM-trading work lacks a systematic, production-grade multi-agent architecture with clean data-leakage prevention and auditable inter-agent communication.

Method

The framework decomposes the AT pipeline into named agent pools: planner, orchestrator, data agents, alpha agents, risk agents, portfolio agents, backtest agents, execution agents, audit agents, and a memory agent. The orchestrator sends structured control messages to agent pools via MCP (Model Context Protocol); agents communicate peer-to-peer via A2A (Agent-to-Agent protocol). Critically, LLM agents never touch evaluation-window returns or prices—all numerical computation (signal construction, optimization, backtesting) is delegated to deterministic tool-based modules. The memory agent stores states indexed by SHA-256 UUIDs of (role, task, params, timestamp) to enable reproducibility and prevent label leakage. For BTC, an XGBoost regressor predicts 1-minute forward returns from 100+ microstructure features under rolling walk-forward retraining; the trading signal blends model output with regime-conditioned price-action rules.

Key Contributions

  • Full mapping of algorithmic trading pipeline components to a multi-agent architecture with defined roles and communication protocols.
  • Use of MCP for orchestrator-to-pool control and A2A for intra-pool and peer agent communication in a financial context.
  • Leakage-prevention architecture: LLMs propose factor structures; tool modules execute all numerical computation; only the backtest agent sees evaluation-window data.
  • UUID-based memory system providing immutable, reproducible, isolation-enforced state across orchestration cycles.
  • Two end-to-end trading examples (hourly equities, minute-level BTC) on a single unified pipeline with concrete performance numbers.

Results

  • Stock trading (04/2024–12/2024, 7-stock universe, hourly): Return 20.42%, Sharpe 2.63, MaxDD −3.59% vs. SPY: 16.60%/1.86/−8.89%, QQQ: 21.59%/1.79/−14.13%. Equally-weighted benchmark dominates on return (47.46%, Sharpe 3.37) but with far larger drawdown (−16.21%).
  • BTC trading (07/27–08/13/2025, minute data, 17 days): Return 8.39%, Sharpe 0.38, MaxDD −2.80% vs. Buy-and-Hold: 3.80%/0.17/−5.26%. Calmar ratio 166.06 vs. 23.30 for buy-and-hold.
  • Strategy exhibits lower volatility and smaller drawdown than benchmarks in both tasks, at the cost of some upside return.

Limitations

  • Evaluation windows are very short (9 months for stocks, 17 days for BTC), making statistical significance of Sharpe ratios and returns impossible to establish.
  • The equally-weighted benchmark outperforms on total return for equities, raising questions about whether the added complexity is justified for return maximization.
  • The 7-stock universe is hand-picked (FAANG-adjacent), which may inflate results due to survivorship/selection bias.
  • No ablation studies isolating the contribution of individual agent types, MCP/A2A protocols, or memory mechanisms.
  • Live trading is not demonstrated; results are entirely backtested, and slippage/transaction cost models may not reflect real execution conditions.
  • BTC strategy details (XGBoost + price-action rules) are highly engineered and may overfit the narrow 17-day window.

Relevance to Agentic AI / LLM Agents

This paper is a concrete engineering blueprint for deploying multi-agent LLM systems in a high-stakes, data-intensive domain where correctness guarantees matter—specifically addressing the challenge of keeping LLM reasoning decoupled from numerical computation to prevent data leakage. The MCP+A2A protocol stack it adopts aligns with emerging standards for inter-agent communication, making it a relevant reference architecture for production agentic systems beyond finance. It also demonstrates a working approach to memory persistence with reproducibility guarantees (UUID-hashed entries), which is a recurring open problem in long-horizon agentic workflows. The explicit agent-role decomposition (planner → orchestrator → specialized pools) mirrors patterns seen in general-purpose agentic frameworks and provides a domain-grounded case study of hierarchical multi-agent orchestration.