FinAgent Orchestration Framework: Shifting from Algorithmic to Agent‑Based Trading

Background

Financial markets are highly dynamic and low‑signal‑to‑noise, making them a challenging testbed for AI agents. Traditional algorithmic trading pipelines—data processing, signal extraction, portfolio management, execution, and evaluation—require years of development by specialist teams.

Problem Definition

The paper aims to build a generic orchestration framework that maps each component of a conventional algorithmic system to an autonomous agent (planner, orchestrator, alpha, risk, portfolio, back‑test, execution, audit, memory), enabling end‑to‑end trading from raw data while preserving auditability and data security.

FinAgent Orchestration Framework

Overview : The framework consists of multiple agent pools, each responsible for a stage (data, alpha, risk, portfolio, execution). An orchestrator controls pools via a Model Context Protocol (MCP) that sends small control messages describing task type, ID, input schema, flags, timeout, and retry budget, and receives confirmations, status, logs, and artifact IDs.

Data Processing : Large language models (GPT‑4o, Llama‑3, FinGPT) support agents. The data agent pulls from Polygon and yfinance, compares coverage, consistency and latency, selects the better source, aligns timestamps, cleans errors and missing values, and produces simple features.

Signal Generation : Cleaned data is fed to the alpha and risk agents. The alpha agent proposes factor structures based on literature; a tool‑based module computes numeric signals. The risk agent calculates exposure and limits, filtering signals with RankIC, rolling tests and forward‑back‑testing.

Portfolio Management : Under capital and turnover constraints, the portfolio agent back‑tests long‑only and long‑short rules, runs simulations or live trading, and logs equity curves, drawdowns and contributions. The planner and orchestrator update future plans using these logs while avoiding leakage of evaluation windows.

Message Control & Agent Communication : All pools are controlled by the orchestrator via MCP. After task publication, pools use an Agent‑to‑Agent (A2A) protocol with a memory agent to read prior context, upload logs and key results, and share progress at fixed intervals. Failed collaborators can be taken over by peers or reassigned by the orchestrator.

Context Protocol & Memory Integration : Context messages are serialized as JSON, excluding raw price series, future timestamps or direct optimization targets. Numerical arrays are stored externally and referenced by identifiers. The memory agent stores deterministic UUIDs that encode role, task description, parameters and timestamps, ensuring immutable, secure retrieval and isolation between training and evaluation.

Trading Example Implementations

Stock Back‑test (hourly data, Apr‑Dec 2024) : Data from Polygon and yfinance is de‑duplicated, aligned, and baseline features (return, momentum, volatility, volume‑ratio) are computed. The alpha agent uses published factor literature without accessing any evaluation‑period data. Risk limits are applied, and the execution agent converts weights to orders with slippage and cost models. The method achieves a 20.42 % total return, 2.63 Sharpe, and –3.59 % max drawdown, compared with the S&P 500’s 15.97 % return.

Bitcoin Back‑test (minute data, Jul 2025) : The same pipeline is reused with minute‑level data. The alpha agent employs short‑term micro‑structure factors (order‑flow imbalance, bid‑ask spread, volume spikes). Risk limits are stricter; execution occurs only when drift and threshold conditions are met. The strategy yields an 8.39 % return, 0.38 Sharpe and –2.80 % max drawdown, versus a 3.80 % buy‑and‑hold benchmark.

Experiments

Setup : Both tasks start with $100 k capital; the stock benchmark includes equal‑weight, SPY, QQQ, IWM, VTI, rebalanced weekly.

Results : The equal‑weight baseline attains the highest total return (47.46 %) and Sharpe (3.37). FinAgent’s stock method delivers 20.42 % return, the lowest volatility (11.83 %) and smallest max drawdown (‑3.59 %). Bitcoin results show a 4.6 pp excess return over buy‑and‑hold with lower volatility.

Additional Validation : Comparisons with several open‑source multi‑agent trading systems show superior scalability and auditability. The strict context and memory design prevents data leakage, confirming reliability across market conditions.