Qlib

What it is

Qlib is a research pipeline for machine-learning stock selection. You describe a dataset, a model and a strategy in one YAML file, and qrun builds features, trains, predicts, backtests and writes an evaluation report. Its unit is the cross-section rather than one strategy on one ticker. A model scores every stock in a universe such as the CSI 300 or the S&P 500 each day. A top-k strategy then rebalances into the best-scored names. The report shows information coefficient, grouped returns and excess return over a benchmark, with and without costs.

What it has that most backtesters lack is the layer before the strategy. It has its own binary data store and an expression language for features (Ref($close, 1), Mean($close, 3)). It ships two ready-made factor sets, Alpha158 and Alpha360, and benchmark configs for 26 models, from LightGBM and XGBoost to LSTM, Transformer, TabNet and Microsoft's own research models such as HIST and DoubleEnsemble. It also has a point-in-time store for quarterly fundamentals and a reinforcement-learning framework for order execution.

Microsoft Research publishes it under the MIT licence. RD-Agent, its LLM tool for proposing factors and tuning models, is a separate repository and is not part of this package.

Pricing

Free and MIT-licensed, with nothing to buy. The cost is data and compute: the deep-learning models in the benchmark zoo want a GPU, while LightGBM on Alpha158 runs on a laptop.

Data & coverage

The engine knows three regions: China, with 100-share lots and a ±9.5% limit-move check; the US, with single shares and no limit check; and Taiwan, with 1,000-share lots. Yahoo collector scripts also build daily and one-minute bars for India and Brazil. Each region's universes are index-constituent lists: CSI 100, 300 and 500 for China, and the S&P 500 and Nasdaq-100 for the US.

The datasets the CLI downloads are frozen in 2020 (see the FAQ). The README says that the "official dataset is disabled temporarily". Instead it points to chenditc/investment_data, a community project that publishes a fresh China A-share bundle in Qlib's format every day. Its latest release was dated 7 October 2026. There is no equivalent for the US. For that market you run the Yahoo collector yourself, and Yahoo's history leaves out delisted names, which builds survivorship bias into any backtest on it. Qlib's own crowd-source page says the same.

A crypto collector exists, but its README says that the dataset supports data retrieval only and not backtesting, because it lacks OHLC bars. A mutual-fund collector reads Chinese funds.

Limitations

• The bundled data ends in 2020, and the official download has been "temporarily" disabled since at least May 2024, when the archives were last refreshed.
• There is no live trading, no paper trading and no broker adapter. The exchange is a simulator.
• The defaults are China-shaped. All 56 benchmark configs target CSI 300 or CSI 500 and carry A-share costs: 5 bps to open, 15 bps to close and a minimum fee of 5. A US run means rewriting most of the config, not just changing the region flag.
• The release cadence has stalled: 0.9.7 in August 2025, while the hosted docs already describe a 0.9.8 development build that has not been released.
• No wheels for Python 3.13 or later, and LightGBM needs brew install libomp on Apple Silicon.
• The MIT licence covers the code, not the bars the Yahoo collector fetches, and training a model on them is a question for Yahoo's terms. Can you train a model on licensed market data? maps it.

Alternatives

Zipline-reloaded's Pipeline API is the nearest thing to Qlib's factor layer, but it has no model zoo. VectorBT is the fast way to sweep rules you already have. QuantConnect solves the US data problem and connects to a broker, at the price of running on its platform. NautilusTrader is the choice when the goal is execution rather than research. FinRL is the reinforcement-learning counterpart from the AI4Finance Foundation, with gym-style environments and RL agents in place of supervised factor models. PyBroker backtests a model you trained yourself on each symbol's data, retrained walk-forward, rather than a cross-sectional ranker. Kronos, an open-weights model pre-trained on OHLCV bars, runs its fine-tuning demo and its only backtest on Qlib's engine, a top-k backtest over the fine-tuned model's output.