What is FinRL?
FinRL trains deep reinforcement-learning agents to trade. It wraps market data in Gym-style environments, trains agents such as A2C, DDPG, PPO, TD3 and SAC through Stable Baselines 3 or ElegantRL, and backtests them against simple baselines. The pipeline runs train, test and trade in three steps, with example applications for stock trading, portfolio allocation, cryptocurrency and high-frequency trading.
The AI4Finance Foundation first released it in 2020 and published it at a NeurIPS 2020 workshop. Its README now describes this repository as the original framework for education, benchmarking and research prototyping, and sends anyone building production or live trading systems to FinRL-X in a separate repository, FinRL-Trading.
Data loaders cover 15 sources, including Yahoo Finance, Alpaca, Binance, CCXT, WRDS and several Chinese market feeds. The code is MIT-licensed and the README states that nothing in it is financial advice.
Is FinRL worth it?
FinRL is a teaching framework that happens to have 17,100 GitHub stars. The maintainers say so directly: the README points new users who want production or live trading to FinRL-X, a rewrite in the FinRL-Trading repository with about 3,950 stars, and keeps this repository for the classic train-test-trade workflow.
As a learning tool it is well judged. The 2026 tutorial, added in v0.3.8 on 20 March 2026, downloads the Dow 30 from Yahoo Finance, trains on 2014 to 2025, trades from 1 January to 20 March 2026, and compares five agents with a mean-variance portfolio and the DJIA. Three short scripts take you from data to a backtest chart, which makes the gap between a learned policy and a dull baseline easy to see.
The weaknesses are the ones that matter in finance. Until October 2026 the tutorial used today's Dow members for past years, a survivorship bias that flatters results; a merged fix now uses point-in-time constituents, and that it took this long says something about how carefully earlier results were checked. Live trading is basic Alpaca support with no order-level risk controls, according to the FinRL-X comparison table, and Yahoo Finance, the default source, is limited to 2,000 requests an hour. 327 issues and pull requests are open.
The repository is still maintained, with merged fixes as recently as 25 October 2026, but the README is clear that new development belongs in FinRL-X. FinRL is worth reading for the ideas and the papers; money belongs in something with proper risk controls.
Where it earns its place
A master's student in quantitative finance
Reproduce a reinforcement-learning trading result for a thesis chapter
Run the three 2026 tutorial scripts on a laptop, swap in a different ticker list or date range, and compare the five agents against the mean-variance and DJIA baselines. Software is free; Yahoo Finance data costs nothing within its 2,000 requests an hour.
A university lecturer teaching a machine-learning-for-finance module
Give students a codebase where they can change the reward and see what happens
Use FinRL's Gym-style stock-trading environment in notebooks, have students change the reward function or features, and discuss why the October 2026 survivorship-bias fix changes the backtest.
A quant developer at a small fund evaluating reinforcement learning
Decide whether RL timing signals are worth a proper research project
Prototype with FinRL against the fund's own data through a custom loader, then move anything promising to FinRL-X or in-house infrastructure, since FinRL lacks order and portfolio risk controls.
Before you commit
- Setup
- Clone the repository, create a Python virtual environment and run pip install -e .; the three tutorial scripts then download data, train and backtest on a laptop.
- Integrations
- Stable Baselines 3 and ElegantRL for agents; data from Yahoo Finance, Alpaca, Binance, CCXT, WRDS, QuantConnect, EOD Historical Data, Tushare, Akshare and others; Alpaca for paper or live trading.
- Data and privacy
- Runs entirely on your machine; market data comes from the providers you configure under their own terms. The README states it is shared for academic purposes and is not financial advice.
- Limits
- Yahoo Finance allows 2,000 requests an hour; live trading is basic Alpaca support without order-level risk management; the maintainers' production work has moved to FinRL-X; 327 open issues and pull requests.
What’s good
- Complete data-to-backtest pipeline in three scripts
- Five standard agents compared against simple baselines
- Loaders for 15 market-data sources
- MIT licence and a long publication record
What’s not
- Maintainers steer production users to FinRL-X instead
- Survivorship bias in the tutorial was fixed only in October 2026
- Only basic Alpaca support for live trading
- 327 open issues and pull requests
Best for
Students and researchers who want a working reference for applying deep reinforcement learning to trading, not a system to run money through.
FinRL alternatives
FinRL: common questions
- Is FinRL worth it?
- FinRL is a good way to learn how reinforcement learning is applied to markets and a poor foundation for trading real money, which its own README now says in so many words. We rate it 3.5 out of 5.
- How much does FinRL cost?
- Free under the MIT licence. There is no hosted or paid tier; costs are your compute and any paid market-data subscriptions.
- What are the downsides of FinRL?
- Maintainers steer production users to FinRL-X instead; survivorship bias in the tutorial was fixed only in October 2026; only basic Alpaca support for live trading; 327 open issues and pull requests.
- Who is FinRL best for?
- Students and researchers who want a working reference for applying deep reinforcement learning to trading, not a system to run money through.
- Is FinRL safe for company data?
- Runs entirely on your machine; market data comes from the providers you configure under their own terms. The README states it is shared for academic purposes and is not financial advice.
- What are the best FinRL alternatives?
- The closest alternatives we have reviewed are Kronos and TimesFM.
Answers checked Nov 16, 2026.