PyBroker

What it is

PyBroker is a Python backtesting framework built around models rather than rules. You register a training function, and PyBroker calls it with each symbol's training data. It then backtests an execution function that reads the model's predictions bar by bar, through ctx.preds(), alongside indicators, positions and pending orders. The engine is event-driven per bar and per symbol, with NumPy and Numba under it. It handles long and short positions, stops, limit orders, margin, rotation and ranking across a universe.

Two features set it apart from the rest of this category. walkforward() splits the history into a fixed number of windows, each with its own train/test split. It retrains the model per window and moves forward in time, adding each window's test data to the next window's training set, with a lookahead setting to stop labels leaking across the boundary. calc_bootstrap=True then reports confidence bounds for the Sharpe ratio and profit factor, plus drawdown probabilities, rather than a single number from one path.

The repository is not archived. The last release is v2.0.1 (28 August 2026), after v2.0.0 on 17 August 2026, and the last commit to master was the same day. It has about 3,570 stars and no open issues against 143 closed ones. The ten open pull requests are all Dependabot's. One author wrote about 90% of the commits.

Pricing

Free, with no account, tier or paid edition. The licence is the thing to read instead: Apache 2.0 with the Commons Clause v1.0, the same arrangement as VectorBT, which makes it source-available rather than OSI open source.

Data & coverage

Built-in sources are YFinance, YQuery (yahooquery), Alpaca for US stocks and AlpacaCrypto, with AKShare for China in an extension module. Since 2.0.0, AKShare is no longer installed with the package, so you install it yourself. Any pandas DataFrame with symbol, date and OHLC columns works as a source, and custom columns can be registered for use inside the strategy. Downloads, indicators and trained models can be cached to disk. The engine models shares, not contracts: there is no options chain, futures multiplier or roll.

Integrations

Optuna for parameter search (hyperparam and Strategy.optimize), joblib for parallel indicator and model work with Ray as an optional backend, and any model library you can call from a training function. scikit-learn, gradient-boosting and time-series models appear in the docs. Results are pandas DataFrames, and 2.0.0 added to_json(). The repo also ships six Agent Skills, which are SKILL.md instruction files that teach a coding agent such as Claude Code, Codex or Cursor to write PyBroker strategies. They are not an AI feature of the library itself.

Limitations

• The defaults are kind to a backtest. An order is placed one bar after its signal and fills at that bar's midpoint, (high + low) / 2, with fees at zero until you set fee_mode. Set the fill price and fees explicitly before you believe a number.
• The bootstrap draws bars independently, with no blocks. That shuffles away serial correlation and volatility clustering, so the drawdown bounds assume the path does not matter.
• There is no live trading, no paper account and no broker adapter.
• The licence is not OSI open source, and the PyPI classifier misstates it (see the FAQ).
• Python 3.11 or newer for 2.x. The 1.2 line, last released as 1.2.14, supported 3.10.

Alternatives

VectorBT is much faster for sweeping rules but has no walk-forward model harness. Backtesting.py is smaller and single-asset. Qlib is the choice when the model is a cross-sectional factor ranker over an index. NautilusTrader and Lumibot are the ones to look at when the same code has to place real orders. Kronos is the opposite starting point: pre-trained open weights that generate bars, with no backtester of its own, rather than a model you train per window.