FinRL-X

What it is

FinRL-X is the AI4Finance Foundation's portfolio-trading framework, and the project FinRL's README now sends new users to. It is not FinRL renamed. FinRL still takes commits in its own repository; FinRL-X lives in FinRL-Trading, a repository begun in May 2020 for a DRL stock-trading paper, rewritten from September 2025 and retitled FinRL-X in March 2026 alongside an arXiv paper (2603.21330).

What changed is the interface. FinRL gives an agent a gym environment and takes back trade quantities. In FinRL-X every strategy returns a vector of target portfolio weights. The vector passes through four optional stages (stock selection, allocation, timing and a risk overlay) and then goes to either a backtest or a broker. The shipped modules are:

• an ML stock selector, which scores quarterly fundamentals with random forest, LightGBM and XGBoost over point-in-time S&P 500 membership;
• equal-weight and minimum-variance allocation inside the selector, and a separate module with mean-variance and DRL allocators;
• a rules-based weekly rotation across 21 stocks and ETFs, which deploy.sh runs end to end.

"AI-native" and "the LLM era" describe less code than they suggest. The only LLM call in the tree sends FMP news articles to OpenAI (gpt-4o-mini by default) for a sentiment label and stores the label in SQLite. No strategy, backtest or trading module reads it.

The DRL module is where FinRL-X is still FinRL. It imports FinRL's StockPortfolioEnv and its Stable-Baselines3 DRLAgent wrapper, plus PyPortfolioOpt for mean-variance. Neither FinRL nor PyPortfolioOpt is listed in requirements.txt or setup.py, so that path means installing both yourself.

Apache-2.0 per the LICENSE file since 3 October 2025; the repository was MIT before that. About 3,800 stars, not archived.

Pricing

Nothing is for sale. The code is Apache-2.0 and there is no hosted version. finrl.ai, where the foundation advertises a paid FinRL tier, does not mention FinRL-X. What you pay for are the accounts the code calls. The ML selector needs a Financial Modeling Prep key for its fundamentals and news, trading needs an Alpaca account, and sentiment labels need an OpenAI key.

Data & coverage

• The unified data manager registers one source, Financial Modeling Prep. Without a key it runs "offline" from its SQLite cache. It pulls quarterly statements and ratios, S&P 500 constituents, daily prices and news.
• Yahoo Finance is not behind that manager, although the README calls it the free default. deploy.sh downloads daily bars through yfinance into data/fmp_daily/, keeping open, high, low, close and volume and dropping the adjusted close.
• WRDS appears in the README's architecture table and in the settings class. No fetcher reads it.
• The repository ships two datasets. One is a 24 MB file of quarterly fundamentals for 715 tickers, June 2015 to March 2026, which its documentation says was built from FMP's API. The other is a set of daily S&P 500 membership snapshots from 2 January 1996 to 17 April 2026. Neither file states terms of use.
• Quarterly figures are mapped to a trade date about two months after quarter end (31 March to 1 June), so the selector does not trade on a report before it could have been filed.
• US stocks and ETFs, one broker (Alpaca), daily bars.

Integrations

Python 3.11 or later. Backtests go through bt, at a flat 0.1% per trade unless you pass a bt cost model such as Almgren-Chriss (supported since April 2026). Execution calls Alpaca's REST API with requests directly, not through the alpaca-py SDK that requirements.txt installs. Several accounts can be configured in .env. A rebalance cancels open orders, sells before it buys, and sends market orders with day time-in-force. When the market is closed it either skips, the default, or queues at-the-open orders. A Streamlit dashboard and a Dockerfile are included.

Limitations

• Installing it is the first obstacle. The PyPI wheel (2.0.2, November 2025) predates the release and lacks the src package its modules import, as well as the base-strategy and rotation modules. It also installs top-level packages named config, data and utils into your environment. From a clone, pip install -r requirements.txt fails on the finnhub line, and the Dockerfile runs the same command.
• The bt backtest fills at the close of the weight's own date. Each weight date rebalances at that day's adjusted close, with no next-bar delay and no spread or slippage unless you configure a cost model. Weights are rescaled to sum to one, so cash left in a weight vector gets invested.
• The rotation backtest that deploy.sh runs does not use bt. It compounds weekly close-to-close returns on the weekly weights, with no transaction costs and on closes that are not adjusted for dividends. Its daily stop-loss and fast risk-off triggers are printed but never applied to the equity curve; the code's own comment says position sizes "would be reduced in live trading". The rotation's 21 symbols are fixed in its YAML file and used unchanged on every backtest date.
• The README's results tables are the authors' own runs, and the README does not say which backtest path produced them. This card does not reprint them.
• The pre-trade risk checks cover one path. The $100,000 per-order cap and 50% turnover cap run only in TradeExecutor.execute_strategy. The rebalance call that the README demonstrates, and that deploy.sh uses, skips them.
• It has broken quietly before. Until a fix merged on 18 September 2026, the ML selector's generate_weights() raised AttributeError on every call, according to the fix's commit message. The repository has no tests and no CI.
• The README is out of step with the code. Its "start here" notebook was deleted in October 2025, and its Python example imports MLStockSelectorStrategy, a class that does not exist.

Alternatives

FinRL is the same foundation's gym-environment library, still the one for reproducing deep-RL trading papers. bt is the engine under FinRL-X's backtest; with a weight vector already in hand, use it directly. Lumibot and LEAN run one strategy class in backtest and live, with more brokers. Qlib covers ML stock selection with its own data store and no broker.