Agentic Trading Lab is an open-source experimental playground for LLM-powered trading agents.
Turn trading ideas into traceable experiments: prototype agents, run backtests and paper-trading simulations, inspect reasoning and decision logs, benchmark against market baselines, and study how agents behave under realistic financial constraints.
- Overview
- Key Features
- File Structure
- Architecture
- Quick Start
- Future Roadmap
- Citation
- License
- Contributing
Agentic Trading Lab is an interactive research and educational platform for exploring trading systems powered by large language models. Built alongside a systematic survey of 130+ agentic trading papers, the project aims to make agentic trading research more accessible by allowing students, researchers, and developers to customize trading agents, evaluate their performance, and study how they behave in realistic market settings.
The platform is designed to bridge the gap between alpha-seeking research and deployable trading systems. Instead of only measuring whether an agent can generate profitable signals, Agentic Trading Lab helps users understand the full trading workflow, including data processing, agent decision-making, backtesting, paper trading, execution constraints, risk control, and governance.
The current system provides three main modes. The Backtesting mode evaluates trading agents on historical market data and compares their performance against market baselines. The Paper Trading mode connects agents to an Alpaca paper trading account for simulated real-time trading. The Leaderboard mode is designed for competition-style evaluation, where different agents or teams can be compared under standardized metrics.
Agentic Trading Lab serves as both a research tool and an educational playground for studying how LLM-based trading agents perform beyond static benchmarks and under practical market constraints.
- Build and customize LLM-powered trading agents
Prototype trading agents with configurable models, prompts, asset universes, and decision logic. - Interactive backtests
Test performance on historical market data and compare results against baselines. - Live-market paper trading
Test agents beyond historical replay by deploying to paper trading accounts that interact with real-time market data. - Inspect decision logs and reasoning traces
Review each BUY, SELL, and HOLD decision with timestamps, prices, portfolio state, and reasoning. - Evaluate risk and performance
Analyze cumulative return, Sharpe ratio, volatility, maximum drawdown, win/loss behavior, and benchmark comparisons. - Compete on agent leaderboards
Join agentic trading competitions where teams and users compete against others in a live-market paper trading setting.
The main folders are backend, frontend, and orchestration. The lab uses a backtest → API → dashboard pipeline.
AgenticTrading/
├── backend/ # FastAPI app, SQLite layer, paper trading, LLM validator
├── frontend/ # Dashboard (served by backend at http://localhost:8000)
├── scripts/ # CLI backtest (backtest_hourly_agent.py, etc.)
├── config/ # Default run IDs and date range (defaults.json)
├── data/ # SQLite backtest results (backtest.db)
├── credentials/ # Local only — not in git (see alpaca.json.example)
├── backups/ # Database backups
├── docs/ # Sphinx docs (see docs/README.md for local preview)
├── readthedocs.yml # Read the Docs build config
└── orchestration/ # FinAgent multi-agent framework (separate subsystem)
┌─────────────────────────────────────────────────────────────┐
│ Backtest Engine (scripts/backtest_hourly_agent.py) │
│ ├─ Fetch Alpaca hourly bars │
│ ├─ Run agent + baseline logic │
│ ├─ Write 3 runs (agent, buy-and-hold, DJIA) │
│ └─ Store in data/backtest.db (SQLite) │
└────────────────┬────────────────────────────────────────────┘
│
┌────────────────▼─────────────────────────────────────────────┐
│ REST API (backend/app.py) │
│ ├─ GET /health │
│ ├─ GET /runs, /runs/{id}/equity, /compare │
│ ├─ POST /backtest/run, GET /backtest/status │
│ ├─ GET /ticker │
│ ├─ GET /paper/account, /paper/positions, … │
│ └─ GET /config/defaults │
│ (LLM example: backend/llm_integration_example.py only) │
└────────────────┬────────────────────────────────────────────┘
│
┌────────────────▼─────────────────────────────────────────────┐
│ Web Dashboard (frontend/) │
│ ├─ index.html, app.js, styles.css │
│ └─ images/ │
└──────────────────────────────────────────────────────────────┘
pip install -r requirements.txtUse either environment variables or a local credentials file.
Option A — .env (recommended for deploy):
cp .env.example .env
# Edit .env:
# ALPACA_API_KEY=your_key
# ALPACA_SECRET_KEY=your_secret
# ALPACA_BASE_URL=https://paper-api.alpaca.marketsOption B — local file (for backtest CLI and local API fallback):
cp credentials/alpaca.json.example credentials/alpaca.json
# Edit credentials/alpaca.json with your paper-trading keysThe credentials/ folder is not tracked in git. See credentials/README.md.
python3 backend/app.py- Open agentic-trading-lab.vercel.app or http://localhost:8000/ and stay on the Backtest tab.
- Set the date range, assets, and model in the left sidebar.
- Click ▶ Run Backtest.
- Wait for the run to finish — the UI polls
/backtest/statusand reloads the equity charts when complete.
You get interactive charts and comparison views (agent, buy-and-hold, DJIA) in the Trading Performance panel.
CLI (optional) — For headless or scripted runs only; there is little visualization in the terminal:
python3 scripts/backtest_hourly_agent.py --start 2026-03-01 --end 2026-03-31
python3 scripts/backtest_hourly_agent.py --mode buy_and_hold # validation modeUse the dashboard to inspect results after a CLI run, or call POST /backtest/run with the same parameters the UI sends.
- Leaderboard backed by real multi-agent runs (replace mock data)
- Sentiment analysis (Reddit, news APIs)
- Monte Carlo simulation baselines
- Production-ready Docker image (frontend + data included)
This repository includes the FinAgent Orchestration Framework under orchestration/, originally developed by Jifeng Li et al. at Open Finance Lab as part of the work on financial agent orchestration. The orchestration framework provides multi-agent architecture, memory systems, and DAG-based planning components. See orchestration/README.md for details.
If you use the orchestration framework in research, please cite:
@inproceedings{orchestration_finagents_2025,
title = {Orchestration Framework for Financial Agents: From Algorithmic Trading to Agentic Trading},
author = {Jifeng Li and Arnav Grover and Abraham Alpuerto and Yupeng Cao and Xiao-Yang Liu},
booktitle = {NeurIPS 2025 Workshop on Generative AI in Finance},
year = {2025},
}
Plain-text citation:
Jifeng Li, Arnav Grover, Abraham Alpuerto, Yupeng Cao, and Xiao-Yang Liu. Orchestration Framework for Financial Agents: From Algorithmic Trading to Agentic Trading. NeurIPS 2025 Workshop on Generative AI in Finance, 2025.
Documentation: finagent-orchestration.readthedocs.io (Agentic Trading Lab + Orchestration Framework). Local preview: docs/README.md
OpenMDW-1.0 — See LICENSE (Copyright Jifeng Li @ SecureFinAI Lab)
Pull requests and issues welcome!
Built with Alpaca API, FastAPI, Chart.js, and SQLite