Backtesting.py

What it is

A Python library of about 3,900 lines — 1,800 of them the engine itself — that takes one OHLC(V) DataFrame and one Strategy subclass, walks the bars, and returns a pandas Series of statistics — return, CAGR, Sharpe, Sortino, Calmar, max drawdown, win rate, profit factor, SQN — plus the trade list and an interactive Bokeh chart of the equity curve and fills. Two classes and about a dozen keyword arguments is the whole surface, and that is the point.

The engine is event-driven per bar. Market orders fill at the next bar's open unless you set trade_on_close=True; limit, stop and stop-limit orders fill when the bar's range touches them. spread, commission (a rate, a (fixed, relative) tuple, or a callable on size and price), margin, hedging and exclusive_orders are the entire cost and position model.

Indicators are declared in init() and computed once across the whole series — self.I() refuses anything that is not an array as long as the data — so an indicator never sees the state the simulation is in, and anything that depends on an open position is written by hand in next(). The run also does not begin until every declared indicator is non-NaN, which quietly costs a 200-bar moving average the first 200 bars of the sample.

Pricing

Free, with nothing to buy — no pro build, no hosting, no data bundle, just a GitHub Sponsors link. The cost is the licence: AGPL-3.0-or-later is stronger copyleft than most Python quant libraries carry, and it is the thing to check before this goes near a product.

Data & coverage

Ships no market data beyond three sample sets used in the documentation. You supply the DataFrame — Open, High, Low, Close, optionally Volume, on a datetime or plain range index, extra columns readable from the strategy. Any instrument, any market, any bar size works, because the engine never learns what the instrument is.

Integrations

Indicator-library-agnostic: self.I() wraps any callable returning an array, so TA-Lib, pandas-ta or your own NumPy function work unchanged. Optimisation is grid search or the SAMBO model-based optimiser, run across processes, with a heatmap of the parameter space. NumPy, pandas and Bokeh; Python 3.9 and up.

Limitations

• No live trading and no paper trading. Not a broker interface nobody implemented — there is no interface.
• One instrument per backtest. No portfolio-level sizing, no rebalancing, no shared cash across symbols.
• Bar resolution caps the realism. When a contingent stop-loss or take-profit would fire in the same bar its parent stop or limit order became a trade, the library warns that it cannot assert the intra-candle path, defers to the next matching bar, and calls the resulting trade "somewhat dubious" in its own message.
• Whole units only, unless you wrap the run in FractionalBacktest.
• Volume is decoration. The column is optional, filled with NaN when absent, and the broker never reads it: an order fills in full at the price the bar touched whatever its size. No partial fills, no volume cap, no market impact.
• No borrow cost, no financing, no dividends. Beyond the spread and the commission, none of the three appears anywhere in the engine.
• An order the account cannot afford is cancelled, not reduced. Short of margin, the broker drops it with a UserWarning; a relative size (0 < size < 1) is floored to whole units of the margin available, and when that floors to zero the order simply goes away. On the default cash=10_000 the library also warns at construction if any close exceeds the account.
• No options and no multi-leg positions.
• Slow release cadence. 0.6.5 in July 2025, 0.6.6 in July 2026.

Alternatives

VectorBT for parameter sweeps at a scale this will not reach, bt for portfolio-level strategy composition, NautilusTrader when the strategy has to run live without a rewrite. Backtrader is the other name everyone offers; Backtesting.py's own alternatives list files it under "Obsolete / Unmaintained".