LLM Trading Research Landscape
Agentic and LLM-driven trading research has gone from a handful of 2023 papers to a fast-moving field spanning multi-agent debate frameworks, live-market benchmarks, and open-source reinforcement-learning toolkits. This page is a cited, dated map of that field — not a tutorial and not a recommendation. Every reference below was re-fetched and verified live the same session this page shipped; see Multi-Agent Trading Architectures for the diagrammed deep dive on the three frameworks below, and Do LLM Trading Agents Actually Work? for what the live benchmarks actually found.
Part of the Agentic AI module: Build an LLM Trading Agent (the from-scratch architecture guide) · Multi-Agent Trading Architectures · Do LLM Trading Agents Actually Work? · Commercial Agentic Trading Products.
Multi-Agent Trading Frameworks
Three research frameworks anchor the current field, each dividing the trading decision across agents differently. All three are diagrammed in depth on Multi-Agent Trading Architectures; this table is the citation summary.
| Framework | Org / Date | Architecture | Status |
|---|---|---|---|
| TradingAgents | UCLA / MIT, Dec 2024 | Analyst Team → bull/bear Researcher debate → Trader → Risk Mgmt Team → Fund Manager | arXiv 2412.20138 + project site |
| FinMem | Stevens Institute et al., Nov 2023 | Profiling + layered Memory (adjustable “cognitive span”) + Decision-making | arXiv 2311.13743; reference GitHub stale since 2024-08-18 |
| ContestTrade | Aug 2025, revised Jul 2026 | Data Team + Research Team, “Quantify-Predict-Allocate” capital allocator | arXiv 2508.00554 — still being actively revised |
Open-Source Ecosystem & Tooling
Below the research papers sits an active open-source layer — some of it reinforcement-learning infrastructure that predates the LLM-agent wave, some of it agent-native from the start.
| Project | What It Is | Maintenance (as of Jul 2026) |
|---|---|---|
| FinRL | Deep reinforcement-learning trading framework (AI4Finance Foundation); ships a documented Alpaca paper-trading notebook | 15,759 stars, last push 2026-07-13, MIT, actively maintained |
| FinGPT | Open financial LLMs, same AI4Finance Foundation as FinRL | 20,904 stars, last push 2026-07-18, MIT, actively maintained |
| HKUDS AI-Trader | Open-source project describing itself as “100% Fully-Automated Agent-Native Trading”; companion benchmark paper covered on the benchmarks page | 20,896 stars, last push 2026-06-11 |
| QuantConnect LEAN + Alpaca | Brokerage-adapter repo connecting QuantConnect's LEAN engine to Alpaca for live/paper execution — a platform-side path for agent-driven strategies | Last push 2026-07-17; 16 stars is normal for an adapter submodule of a much larger core engine |
For hands-on next steps rather than research reading, AlgoDrill's own First Trading Bot and Build an LLM Trading Agent guides use a smaller, from-scratch stack (yfinance/vectorbt/Alpaca paper API) rather than any of the frameworks above.
Beyond Execution: LLMs for Alpha Research
Not every multi-agent trading paper is about deciding when to trade. A newer cluster applies the same debate/contest patterns to finding tradeable signals in the first place — a research-assistant role rather than a decision-maker role:
- R&D-Agent-Quant (arXiv 2505.15155, Microsoft Research Asia co-authorship) — a multi-agent framework that jointly optimizes data-centric factors and models; reported at NeurIPS 2025 per secondary sources, not independently confirmed on a NeurIPS-owned page this session.
- FactorMAD (ICAIF '25, Tsinghua University + Microsoft Research) — a multi-agent debate framework specifically for interpretable stock alpha-factor mining, confirmed on the official ICAIF '25 accepted-papers listing.
- AlphaCrafter (arXiv 2605.05580, Nanjing University, May 2026) — a full-stack multi-agent framework for cross-sectional quantitative trading; a recent preprint, roughly two months old as of this page's verification.
The distinction matters for how you read any of these papers: a trade-execution agent (TradingAgents, ContestTrade) is evaluated on realized returns and risk metrics, while a factor-mining agent (FactorMAD, R&D-Agent-Quant) is evaluated on the quality and interpretability of the signals it surfaces for a human or a downstream system to use — a different, and generally easier to validate, claim.
Why Evaluation Is Still Immature
A 2025 survey on evaluating and benchmarking LLM agents generally (arXiv 2507.21504, not finance-specific) proposes a two-dimensional taxonomy — evaluation objectives and evaluation process — and flags challenges that apply directly to trading agents even though the survey itself does not single out finance: role-based access to data, the need for reliability guarantees, dynamic and long-horizon interactions, and compliance. Those four challenges describe exactly what a trading-agent benchmark has to solve. See Do LLM Trading Agents Actually Work? for how three recent live-market benchmarks attempted it, and what they actually found.
What this page does not cover: product recommendations, investment advice, or any first-party live-trading results from AlgoDrill. A first-party engineering case study is planned as separate, clearly-labeled content and is not part of this survey. For commercial products that put an agent in front of a real brokerage account today, see Commercial Agentic Trading Products.
Frequently Asked Questions
- What is the TradingAgents framework?
- TradingAgents (arXiv 2412.20138, UCLA/MIT, December 2024) is a multi-agent LLM trading framework that mirrors the org chart of a trading desk: an Analyst Team (fundamental, sentiment, news, and technical analysts) gathers data, a Researcher Team debates it as a bull researcher versus a bear researcher, a Trader agent synthesizes that debate into a decision, a Risk Management Team checks exposure, and a Fund Manager gives final approval. The paper and project site report improved backtested return, Sharpe ratio, and drawdown control versus baseline strategies; AlgoDrill has not independently reproduced those figures.
- Is FinMem still actively maintained?
- The paper (arXiv 2311.13743, November 2023, accepted at the ICLR 2024 Workshop on LLM Agents) is a fixed, citable research artifact, so 'actively maintained' does not really apply to it. Its reference GitHub implementation is a different story: pipiku915/finmem-llm-stocktrading has 927 stars but its last commit was 2024-08-18 -- roughly two years stale as of this page's most recent verification. Treat the architecture (Profiling, layered Memory, Decision-making) as the citable contribution, not the reference code as a maintained tool.
- What is ContestTrade's contest mechanism?
- ContestTrade (arXiv 2508.00554) names its own mechanism 'Quantify-Predict-Allocate': a Data Team turns market data into textual factors, a Research Team produces multiple parallel trading decisions, and each agent's output is scored only after the market outcome becomes observable. Future utility is then predicted from those historical scores, and capital is allocated toward agents with positive predicted utility -- the paper describes this as inspired by institutional investment workflows, where managers with the strongest track record get more capital to run.
- Where can I find more LLM trading papers than this page covers?
- Awesome-LLM-Quantitative-Trading-Papers (github.com/Tom-roujiang/Awesome-LLM-Quantitative-Trading-Papers) is a curated, actively updated list -- 203 stars, last pushed 2026-06-17 -- organized into eight sections: Trading Agents, Financial Benchmarks, Arenas, LLM Post-Training, Stock Prediction, Factor Mining, Forecasting, and Surveys. This page cites a deliberately small, verified subset; that list is the place to go deeper.
- Do LLM trading agents use reinforcement learning or just prompting?
- Both, and they are different lineages worth keeping straight. FinRL (AI4Finance Foundation, 15,759 GitHub stars, actively maintained) is a deep reinforcement learning framework -- the agent's policy is trained on historical reward signals, no LLM involved. TradingAgents, FinMem, and ContestTrade are a newer lineage: the 'agent' is one or more large language models reasoning over data via prompting and structured output, with no gradient-based training on trading reward. FinGPT, from the same AI4Finance Foundation as FinRL, sits at the intersection -- open financial LLMs intended as building blocks for either style of system.
See the three frameworks above diagrammed side by side, or find out what live benchmarks actually measured.
Architectures, Diagrammed → Do They Actually Work? →