// the find
AI4Finance-Foundation/FinRL
FinRL®: Financial Reinforcement Learning. 🔥
FinRL is the original open-source deep reinforcement learning framework for quantitative trading, built around a train-test-trade pipeline with pluggable market environments, DRL agents (A2C, DDPG, PPO, SAC, TD3 via Stable Baselines3), and financial applications like stock and crypto trading. It's aimed at researchers, students, and anyone prototyping RL-for-trading ideas rather than running real money through it.
The three-layer separation (environments / agents / applications) is a genuinely useful mental model and matches several NeurIPS/ACM papers behind it, so the research claims aren't just marketing. Data layer supports 15+ sources (Yahoo Finance, Alpaca, Binance, CCXT, WRDS, Tushare, etc.) through a common processor interface, which saves real integration work. It composes cleanly with the rest of the AI4Finance ecosystem — FinRL-Meta for environments, ElegantRL for lighter-weight algorithm implementations — if you need pieces rather than the whole stack.
The README spends more space telling you to go use a different repo (FinRL-X/FinRL-Trading) than documenting this one, and by its own comparison table this version's risk management is 'Gym environment constraints only' with no order- or portfolio-level controls — don't mistake the backtest numbers for anything production-safe. There are binary model artifacts (actor.pth, recorder.npy) committed directly into finrl/applications/, which bloats the repo and gives you no way to see how those weights were produced. Config is still hardcoded Python (config.py, config_tickers.py) with no validation layer, and the tooling cited in tutorials (Yahoo Finance scraping, DOW 30 CSVs) is the kind of brittle data pipeline that breaks silently when an upstream API changes.