AgenticTrading

AgenticTrading is a lab for building and comparing LLM trading agents rather than an agent itself: it wraps agent construction, backtesting, live-market paper trading, drawdown and risk metrics, an open leaderboard against market baselines, and Alpaca and Robinhood broker integrations. Its companion paper maps the classical algorithmic-trading pipeline onto planner, orchestrator, alpha, risk, portfolio, backtest, execution, audit and memory agents. The repository vendors the FinAgent orchestration framework under orchestration/FinAgents/, including a qlib_local alpha-agent pool, ships an adapter for the external TradingAgents framework, and pins ai-hedge-fund at a fixed commit in an isolated virtualenv. Live trading against a real Robinhood account is supported but off by default, behind a review-only mode that shows what an agent would do before it can act and per-order risk caps.

The paper is “Orchestration Framework for Financial Agents: From Algorithmic Trading to Agentic Trading” (arXiv:2512.02227), a poster at the NeurIPS 2025 Workshop on Generative AI in Finance — a workshop track, not the main conference. The return and Sharpe figures in its abstract are author-reported and unaudited; they have not been independently verified.

The licence is OpenMDW-1.0, a model-materials licence rather than a familiar OSI-style code licence: permissive in substance, but it carries a notice-retention duty on redistribution and a patent-litigation termination clause, so read it rather than assuming MIT or Apache equivalence. GitHub reports NOASSERTION only because OpenMDW is absent from its licence database, and the copyright holder in the LICENSE file is SecureFinAI Lab while the publishing org is Open-Finance-Lab. Two practical notes: adoption is modest — a few hundred stars, with provenance and cross-project integration the stronger signal — and the PyPI SDK (agentictrading 0.1.0, its only release) points its Source and Issues links at a fork, so link the canonical org repository.