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

AI summary

Overview

Research area: Multi-agent systems applied to quantitative finance, specifically the transition from rule-based algorithmic trading to LLM-driven "agentic trading."

Technical level: Advanced. The paper assumes familiarity with large language models, multi-agent orchestration protocols (MCP, A2A), quantitative portfolio construction, and walk-forward backtesting.

Scope: The paper proposes a full agent-orchestration framework (FinAgent) that maps every stage of a traditional algorithmic trading pipeline onto cooperating LLM agents, and validates it on two in-house backtests — a seven-stock hourly equity strategy and a BTC/USDT minute-level strategy.

What This Paper Is About

Building an effective algorithmic trading (AT) system typically requires a professional team working over years, spanning data engineering, signal research, risk management, portfolio construction, and execution. The authors ask whether this pipeline can be decomposed into specialized LLM agents — each responsible for one stage — coordinated by an orchestrator, so that sophisticated financial tooling becomes accessible to the general public rather than only to institutional desks. The core challenge is that financial markets have low signal-to-noise ratios and strict temporal ordering, meaning naive LLM use risks data leakage and hallucinated signals; the framework is designed around that constraint.

Key Contributions

1. A one-to-one mapping from AT components to agents. Traditional pipeline stages (data, alpha, risk, portfolio, execution, evaluation) are each realized as an agent pool, supplemented by new roles specific to agentic trading: planner, orchestrator, audit agents, and a memory agent. The design is presented as a paradigm shift from algorithmic to agentic trading.

2. A dual-protocol communication architecture. The orchestrator dispatches tasks to agent pools via Model Context Protocol (MCP) using small, schema-typed control messages (node type, task id, declared inputs, policy flags, timeout, retry budget) and receives acknowledgements, logs, and artifact IDs. Peer agents inside a pool coordinate via Agent-to-Agent (A2A) using lightweight speech-act messages (ask, tell, propose, confirm) with role tags and context IDs.

3. A leakage-resistant design principle: LLMs reason, tools compute. Agents propose structures — factor formulas, risk gates, optimization problems — but never receive raw prices, returns, or labels from the evaluation window. All numerical work is performed by deterministic tool modules behind the orchestration layer. Only the Backtest Agent touches evaluation-window data, and it returns only aggregated metrics.

4. Deterministic, hash-indexed memory for auditability and reproducibility. The memory agent stores records keyed by UUID = SHA256(role ∥ task ∥ params ∥ time). Training and evaluation memories live in separate UUID namespaces, entries are immutable, and stored payloads contain only structural summaries (e.g., feature hashes, average IC by bucket) that cannot be inverted to recover raw test-period data.

Main Findings

• Strong risk-adjusted equity performance, but not the highest total return. On a static seven-stock universe (AAPL, MSFT, GOOGL, JPM, TSLA, NVDA, META) with hourly bars, the agentic strategy delivered 20.42% total return, 11.83% volatility, Sharpe 2.63, and −3.59% maximum drawdown over the 04/2024–12/2024 test window. This beat SPY (16.60%), IWM (11.45%), and VTI (16.29%) with substantially lower volatility, but trailed QQQ (21.59%) and, notably, an equal-weighted weekly-rebalanced portfolio (47.46% return, Sharpe 3.37).

• The strategy is risk-controlled rather than return-maximizing. The agentic portfolio produced the lowest volatility and smallest drawdown of all equity variants tested, consistent with the pipeline's explicit gating by risk and portfolio agents.

• BTC outperformed Buy-and-Hold on both return and risk. Over a 17-day minute-level test window (23,500 one-minute bars, 27/07/2025–13/08/2025), the strategy returned +8.39% versus +3.80% for Buy-and-Hold, with a Sharpe of 0.380 vs 0.168, a Calmar ratio of 166.06 vs 23.30, lower annualized volatility (24.23% vs 25.82%), and a shallower max drawdown (−2.80% vs −5.26%).

• Moderate trading activity despite the minute-data setup. The BTC strategy executed only 17 trades (~1.04/day), with a 64.7% win rate vs 58.8% for Buy-and-Hold, an average daily return of 0.48% vs 0.23%, and an average holding time of 16.07 hours (median 0.65 hours) — reflecting regime gating and minimum-holding rules rather than continuous churn.

• The BTC system is an XGBoost regressor, not an LLM predictor. The model predicts next-minute returns from 100+ engineered features (momentum across 1/3/5/10/15/30/60/240-minute horizons, RSI, MACD, Bollinger features, volatility clustering proxies, order-flow imbalance) using 300 trees, depth 6, learning rate 0.08, and L1/L2 regularization. It retrains every 1,440 minutes on a rolling 7-day minimum window and keeps only the top 70% of features by importance; all features are computed from data at or before t−1 with at least a two-minute gap to the label.

• Explicit regime-aware sizing and layered drawdown control. Position size factors vary by regime (2.5 in breakout, 1.8 in strong trend, 0.8 in high volatility, 0.7 in sideways), individual positions are capped at 3–5% of capital with leverage ≤ 4.0, and drawdown thresholds at 1%/2%/3% trim positions by 20%/30%/50%, with a hard −3.0% max-drawdown liquidation.

Methodology in Plain English

The authors take the familiar stages of a trading operation and turn each into a specialist agent. A planner sets the agenda for a run; an orchestrator hands out tasks to pools of agents and monitors their health; data agents reconcile feeds from multiple providers; alpha agents propose factor ideas grounded only in published literature; risk agents compute exposures and set limits; portfolio agents translate signals into weights under capital and turnover constraints; execution agents simulate or place orders with slippage and cost modeling; backtest/evaluation agents produce metrics and attribution; and a memory agent records state so subsequent cycles can build on prior runs.

Two design rules distinguish this from simply "having GPT-4o trade." First, the LLM is kept away from numbers. Agents describe what to compute — a momentum factor, a volatility gate, a convex optimization objective — and deterministic tool modules do the arithmetic. Second, the LLM is kept away from the future. Every prompt is restricted to training-window summaries and published priors; agents are explicitly forbidden from tuning on Sharpe ratios, IC, or any evaluation-window statistic. The Backtest Agent is the single component permitted to load test-period data, and it returns only aggregates like volatility and max drawdown, never timestamp-level P&L.

Communication follows two protocols: MCP for orchestrator-to-pool control messages, and A2A for peer-to-peer and agent-to-memory exchanges. Each context message is a JSON object with fields like task_id, agent_role, run_mode (train/test/live), time_window, universe, inputs, tool_outputs, and diagnostics; raw arrays are never embedded, only referenced by path or dataset ID. Memory entries are keyed by a SHA-256 hash of role, task, parameters, and timestamp, which makes runs replayable and keeps training and evaluation records in separate namespaces.

For the BTC experiment specifically, the pipeline identifies market regimes (strong trend, breakout, sideways, high volatility) and blends a model prediction with a price-action signal, weighting the model more heavily when its confidence proxy is high and falling back on rules when it is weak. Final signals are smoothed with two exponential moving averages, a dead-band prevents tiny position changes, and an eight-minute minimum holding time suppresses noise-driven reversals.

Why This Matters

Research impact. The paper reframes agentic trading as an orchestration and information-isolation problem rather than a prompting problem. Its leakage-prevention architecture — LLMs propose structure, tools compute numbers, only one agent sees test data, and memory stores only non-invertible summaries — offers a concrete template for running LLM agents in any temporally ordered, high-stakes domain where lookahead bias is fatal. The framework is also directly comparable to existing open-source agent-trading projects (the paper tabulates TradingAgents at ~24,800 GitHub stars, AI Hedge Fund at ~42,300, and others), making it a candidate baseline for replication studies.

Real-world applications:

• Retail and semi-professional trading: A single user could run a pipeline that would otherwise require a quant research team, with risk gates enforced automatically.
• Systematic strategy prototyping: Portfolio managers can test whether an LLM-proposed factor structure survives walk-forward validation before committing engineering resources.
• Regulatory and compliance tooling: The immutable UUID-indexed memory and full audit logs of prompts, tool calls, and decisions map naturally onto audit-trail requirements in regulated finance.
• AI safety in high-stakes agents: The structural separation between reasoning and computation is a reusable pattern for any agentic system where agents must not see ground-truth labels.

Industry relevance. The results are honestly framed: the agentic stock strategy beats broad market ETFs on a risk-adjusted basis but loses decisively to a naive equal-weighted portfolio on total return. That candor matters — it positions agentic trading as a risk-control technology at this stage, not a return-generation silver bullet, which is a more defensible claim for institutional adoption.

Future Directions

• Longer horizons and broader markets. The BTC test covers only 17 days and the equity test a fixed seven-stock universe; both are too short and narrow to establish robustness across regimes or asset classes.
• Ablations on the core mechanisms. The authors explicitly call for isolating the contributions of risk gating, the memory agent, and the MCP/A2A messaging layer to determine which design choices actually drive performance.
• Adaptive planning under regime shifts. The planner currently updates from logs between cycles; making it responsive to intra-run regime changes is an open problem.
• Broader information sources. Extending beyond price-derived features to news, filings, and alternative data would test whether the agentic architecture adds value where LLM reasoning has a comparative advantage.
• Benchmark and log release. Publishing the full benchmark suite and orchestration logs would let the community verify the leakage-prevention claims independently.

Target Audience

This paper is best suited for: quantitative researchers and ML engineers building LLM-agent systems for finance; practitioners at systematic funds evaluating whether agentic architectures add value over conventional pipelines; AI safety and evaluation researchers interested in information-isolation patterns for agents in high-stakes settings; and graduate students in financial engineering or multi-agent systems who want a concrete, end-to-end reference implementation with published prompts, context schemas, and code. Readers seeking a deployable alpha source should temper expectations — the paper's most transferable contributions are its orchestration and anti-leakage design patterns, not its specific return numbers.

Authors’ abstract

The financial market is a mission-critical playground for AI agents due to its temporal dynamics and low signal-to-noise ratio. Building an effective algorithmic trading system may require a professional team to develop and test over the years. In this paper, we propose an orchestration framework for financial agents, which aims to democratize financial intelligence to the general public. We map each component of the traditional algorithmic trading system to agents, including planner, orchestrator, alpha agents, risk agents, portfolio agents, backtest agents, execution agents, audit agents, and memory agent. We present two in-house trading examples. For the stock trading task (hourly data from 04/2024 to 12/2024), our approach achieved a return of $20.42\%$, a Sharpe ratio of 2.63, and a maximum drawdown of $-3.59\%$, while the S&P 500 index yielded a return of $15.97\%$. For the BTC trading task (minute data from 27/07/2025 to 13/08/2025), our approach achieved a return of $8.39\%$, a Sharpe ratio of $0.38$, and a maximum drawdown of $-2.80\%$, whereas the BTC price increased by $3.80\%$. Our code is available on \href{https://github.com/Open-Finance-Lab/AgenticTrading}{GitHub}.