Backtesting.py review (2026)

Backtesting.py is the closest thing Python has to a "just works" backtesting library. You feed it a pandas DataFrame of OHLCV candles, subclass Strategy with an init() and a next() method, and get a full statistics table plus an interactive Bokeh chart in a handful of lines. That economy of API is the product. Nothing else in the Python ecosystem gets a newcomer from idea to first backtest faster, and at $0 under an open-source license the cost of trying it is zero.

The trade-off is scope. The library simulates one instrument per backtest, offers no live trading, ships no data, and models orders at a basic level. Serious multi-asset portfolio research belongs in VectorBT, Zipline Reloaded, or QuantConnect; production execution belongs elsewhere entirely. For years the bigger worry was abandonment, with no releases between December 2021 and January 2025. The project has since revived convincingly: nine releases through v0.6.6 (July 2026) added fractional trading, a multi-dataset runner, spread and commission modeling, and the SAMBO optimizer.

We score it 7.1/10: elite ease of use and cost-value, mid-pack power. It sits above the now-dormant Backtrader (6.8) on maintenance and approachability, and below VectorBT (7.8), which trades simplicity for far more analytical firepower.

Backtesting.py is a free, open-source Python library (AGPL-3.0) created by Slovenian developer kernc, hosted on GitHub with roughly 8.8k stars and 1.5k forks as of August 2026. Installed with pip install backtesting, it backtests trading strategies on any financial instrument for which you have OHLC(V) candlestick data: stocks, forex, crypto, or futures. It requires Python 3.9+ and builds on pandas, NumPy, and Bokeh.

The design philosophy is minimalism. A strategy is a small class with two methods; indicators are declared in init() (vectorized over the whole series) and trading logic runs bar-by-bar in next(). Results come back as a pandas Series of ~40 statistics (Sharpe, CAGR, max drawdown, Kelly criterion, alpha/beta versus the market) plus per-trade DataFrames.

After a three-year dormancy that pushed many users toward forks like lucit-backtesting, the original project resumed active development in January 2025. Releases 0.4.0 through 0.6.6 (July 2026) added the SAMBO optimizer, FractionalBacktest for sub-unit crypto position sizes, MultiBacktest for running one strategy across many datasets, percentage trailing stops, bid-ask spread modeling, and commission tracking in trade statistics. There is no company behind it and nothing to buy. Funding is voluntary GitHub Sponsors donations.

The engine is a hybrid by the project's own description. Indicators are computed vectorized up front, then the simulation steps through bars event-style, which keeps execution fast while letting logic react to state (open positions, equity) the way a purely vectorized engine cannot. On the daily-bar tutorial datasets a run completes in milliseconds. Grid optimization uses multiprocessing with shared memory, and the SAMBO model-based optimizer (added in 0.5.0) searches large parameter spaces more sample-efficiently than brute force.

Accuracy is respectable for bar-level research but bounded by bar-level assumptions. Orders fill at the next bar's open by default (avoiding the most common look-ahead trap), and the engine supports commissions, bid-ask spread (added in 0.6.0), stop-loss/take-profit brackets, and margin/leverage. It does not model queue position, partial fills, order-book depth, or slippage beyond a fixed spread/commission. A GitHub discussion notes users doubling TradingView-level commissions to reconcile results, which shows how cost assumptions dominate at this fidelity. Intra-bar SL/TP sequencing is inferred from OHLC, so it can be optimistic on volatile bars unless you feed finer-grained candles.

The optimizer returns heatmaps that make overfitting visible, but there is no built-in walk-forward or combinatorial cross-validation: you must script train/test splits yourself. For single-instrument, bar-resolution strategy research, the results are as trustworthy as your data and cost assumptions.

Backtesting.py ships no market data. The package bundles only small tutorial datasets (daily Google stock and intraday EURUSD samples) so the documentation examples run out of the box. Everything else is bring-your-own. Any pandas DataFrame with Open/High/Low/Close (and optionally Volume) columns works, at any bar interval: daily, hourly, minute, or finer. Tick data must be resampled into candles first.

In practice, users pull free data from yfinance for equities, ccxt or exchange APIs for crypto, and broker exports or paid vendors (Polygon, Databento, Dukascopy) for intraday and forex data. Because the input is just a DataFrame, no adapter layer is required. That is a genuine advantage over frameworks with rigid data-feed abstractions, but it also means data cleaning (gaps, splits, dividends, timezone alignment) is entirely your responsibility, and dividend-adjusted total-return backtests require you to pre-adjust the series yourself.

There are no integrated data downloaders of the kind VectorBT provides, no point-in-time fundamentals, and no corporate-action handling. For indicators, the library computes anything you can express as a NumPy/pandas function and interoperates cleanly with TA-Lib, Tulip, and pandas-ta. Our data score reflects this: maximal flexibility, minimal batteries included.

There is nothing to price-compare: Backtesting.py costs $0, full stop. No free-tier limits, no pro edition, no cloud service, no license keys. We verified that on the official documentation site, PyPI, and the GitHub repository in August 2026. The maintainer accepts optional GitHub Sponsors donations.

The one "cost" worth understanding is the license. AGPL-3.0 is a strong copyleft. You own your strategies outright and can trade them commercially without obligation, but if you build a network-accessible service (for example, a hosted backtesting SaaS) on top of the library, the AGPL requires you to offer that service's source code to its users. Individual quants and prop-desk researchers are unaffected. Startups embedding it in a product should read the license carefully or architect around it.

Total cost of ownership is therefore dominated by data and time. A realistic free stack (yfinance daily data plus Backtesting.py) costs nothing; a serious intraday stack might add $30-$200/month in vendor data fees, and those costs would apply to any BYO-data library. Against subscription platforms charging $30-$120/month for comparable single-instrument backtesting, the cost-value here is close to unbeatable, which our 9.3 cost-value score reflects.

Choose Backtesting.py if you are a Python-literate retail trader testing directional strategies on one instrument at a time: a moving-average system on BTC, a breakout strategy on EURUSD, a mean-reversion idea on SPY. It is the best first backtesting library in Python. The API teaches correct habits (next-bar fills, explicit costs, optimization heatmaps that expose overfitting) without framework ceremony, and the interactive charts make debugging a strategy's behavior unusually concrete.

It also serves experienced quants as a rapid-prototyping scratchpad: many use it to sanity-check an idea in ten minutes before committing to a heavier research stack.

Look elsewhere if you need portfolio-level logic (position sizing across correlated assets, sector rotation, pairs trading). That is VectorBT or Zipline Reloaded territory. If you want an integrated path from backtest to live execution, QuantConnect or NautilusTrader are built for exactly that. Options traders are out of scope entirely: there is no chain, greeks, or expiration modeling. And if you cannot or will not code Python, this library offers nothing; TradingView is the usual no-code alternative. Finally, commercial SaaS builders should weigh the AGPL license before embedding it.