FinRL

What it is

FinRL is the AI4Finance Foundation's library for training deep reinforcement-learning agents on market data. It is built in three layers. Market environments are gym-style simulators for single stocks, a multi-stock account, portfolio allocation and crypto. Agent wrappers sit over Stable-Baselines3, ElegantRL, Ray RLlib and, since September 2026, FastTD3. Applications are worked examples for stock trading, ensembles and portfolio allocation. The loop is train, test, trade. An agent sees prices, holdings and indicators, outputs a buy or sell quantity per ticker, and is rewarded by the change in account value.

The 2026 tutorial is the shape of the whole thing. It downloads a fixed 30-ticker Dow list from Yahoo and adds MACD, RSI, VIX and a turbulence index. It trains A2C, DDPG, PPO, TD3 and SAC for 20,000 timesteps each on 2014–2025, then backtests them on 1 January to 20 March 2026 against a mean-variance portfolio and the DJIA.

The repository now describes itself as the "original" framework for education and research. Its README sends new users to FinRL-X, a separate Apache-2.0 repository where strategies return portfolio weights instead of acting inside a gym environment. MIT per the LICENSE file, about 16,600 stars, not archived, first published in July 2020.

Pricing

The library is MIT, and the site's FAQ says that commercial use is allowed. finrl.ai also advertises a Pro tier at $19.90 a month or $199 a year and an Enterprise tier on request, but the Pro button opens an email rather than a checkout (see the FAQ). The real costs are compute and data: Alpaca, WRDS and FXMacroData each need an account of their own.

Data & coverage

The README's table lists fifteen data sources. The DataProcessor class, which the pipeline goes through, accepts only four: yahoofinance, alpaca, wrds and fxmacrodata. Anything else raises "Data source input is NOT supported yet". Processor files for CCXT, EODHD, JoinQuant, QuantConnect and Sinopac sit in the tree but are not wired in. Separate downloaders cover Tushare for China, Shioaji for Taiwan and Interactive Brokers through ib_insync. Everything is OHLCV bars plus the indicators FinRL computes. It ships no data of its own.

Integrations

Python only. Agents come from Stable-Baselines3, ElegantRL (installed from its git master) and RLlib, and evaluation uses pyfolio-reloaded. The paper-trading class talks to Alpaca through alpaca_trade_api, the older of Alpaca's two Python SDKs, which has had no release since January 2024. It polls on a fixed interval and submits market orders with day time-in-force.

Limitations

• The environment is the backtest, and its fill model is simple. An order fills at the close in the agent's own state, the bar it just observed, plus a flat percentage cost (0.1% each way in the tutorial). There is no next-bar delay, no spread and no slippage. The README's comparison table calls this "custom hand-rolled evaluation loops".
• The tutorial universe is fixed. The list in config_tickers.py includes NVDA, AMZN and SHW and is applied unchanged to every day from 6 January 2014, so the backtest holds today's index members on dates when some were not in the index.
• Packaging is broken in two directions: the PyPI wheel declares no dependencies, and the repository's own pins hold yfinance and ccxt a major version behind.
• The hosted docs still build as version 0.3.1, and their install page tells you to install Anaconda first.
• "FinRL" is a trademark of FinRL LLC, licensed to the foundation. The MIT licence covers the code, not the name.

Alternatives

Qlib is the closest research pipeline. It does supervised factor models rather than RL, ships a model zoo and keeps its own data store. VectorBT and Backtesting.py are the tools to reach for once a policy has to survive a fill model you trust. FinGPT and FinRobot are the same foundation's LLM projects and solve a different problem. FinRL-X is the foundation's successor: it backtests on bt, rebalances through Alpaca and borrows only FinRL's DRL portfolio environment.