bt

What it is

bt takes a pandas DataFrame of prices and a Strategy built from an ordered list of algos — RunMonthly(), SelectAll(), WeighEqually(), Rebalance() is the canonical four-liner — and walks the index one row at a time, tracking positions, transactions and costs.

The distinction that decides whether this is your tool: bt tests allocation rules, not entries and exits. The fifty-odd algos that ship are scheduling (RunDaily through RunYearly, RunIfOutOfBounds), selection (SelectMomentum, SelectWhere) and weighting (WeighInvVol, WeighERC, WeighMeanVar, TargetVol). Nothing addresses a stop, a limit order or a trade lifecycle. If the question is whether a momentum tilt beats 60/40 after costs, bt is built for it; if it is whether a five-minute breakout fills, it is the wrong library.

Strategies nest: a child is priced by running it on a notional $1,000,000 and handing its price series to the parent, so a portfolio of strategies is the same object as a portfolio of securities.

Pricing

Free, MIT-licensed, nothing to buy: no pro build, no hosted tier, no data bundle. pip install bt is the entire product.

Data & coverage

bt ships no market data. The only bundled fetcher is bt.get, re-exported from ffn, which pulls Yahoo Finance adjusted close via yfinance; everything else you load yourself, from a CSV or a market data API. Instrument modelling is specific rather than generic: Security for equities and ETFs, CouponPayingSecurity and FixedIncomeStrategy for bonds, per-security multipliers and a Margin algo for futures. Markets are whatever your price series covers.

Integrations

ffn is the companion library by the same author, and the coupling is tight: bt depends on ffn>=1.1.2, bt.Result subclasses ffn.GroupStats, and the stats table, the plots and to_csv come from there. Python 3.9 or newer, with core.py cythonized and shipped as platform wheels. Repository state on 13 September 2026: release v1.2.3 on 11 September, last commit to master on 12 September, not archived, 13 open issues.

Limitations

• No live trading and no broker. Nothing in the source routes an order.
• One price per row: a fill happens at the bar's price, with no intrabar sequencing. A bid/offer spread is modelled only if you hand the backtest a bidoffer frame alongside the prices, in which case half the spread is charged on each trade; otherwise slippage is whatever cost model you supply.
• No options — no pricing model, no greeks, no chain handling.
• Positions are integers by default, which quietly skips allocations a small account could not fill; fractional shares need integer_positions=False.
• PyPI still classifies it Development Status :: 4 - Beta after twelve years.

Alternatives

backtrader and Backtesting.py for event-driven, signal-level testing; vectorbt where speed over large parameter sweeps matters more than portfolio structure. For the statistics without the simulation, ffn alone is the smaller dependency.