VectorBT

What it is

A Python library that treats a backtest as array arithmetic rather than as a loop over bars. Prices, signals and parameter grids are broadcast into NumPy arrays, the hot path is compiled by Numba, and thousands of configurations are simulated in one call. The README's own example runs 10,000 dual-SMA window combinations across three pairs and plots the result as a heatmap; an optional Rust engine installs as vectorbt[rust] for the paths where Numba's JIT overhead is the cost.

That shape is the whole trade-off, and it is the one architectural choice on a backtester you do not get to revise afterwards. A vectorised engine is fast because it computes over the entire series at once, and for the same reason it struggles to express a rule that depends on what just happened — a position size derived from the previous trade's result, or an exit conditional on the drawdown of the position currently open. Stops and from_order_func recover some of it at the cost of the speed you came for. If your strategy is path-dependent in that sense, an event-driven engine is the other answer, not a slower version of this one.

It is also the free half of a commercial product. The vendor's own sentence is that VectorBT is the open-source community edition of VectorBT PRO, and PRO is a separately installed package (vectorbtpro) behind a paid membership and a private repository. What that split leaves out of this edition is in Limitations, and it is not trivia.

Liveness, as of 2026-09-19: the repository is not archived, the last commit to master was 2026-09-17, and v1.1.0 shipped on 2026-07-05, the second release after v0.28.5 in March. 9,124 stars and 121 open issues — the REST API's open_issues_count reads 139 because it counts pull requests as issues.

Pricing

The community edition is free, for individuals and organisations alike, and there is no pricing page for it: the price is stated in the README and in LICENSE.md. What the licence withholds is the right to sell it, not the right to use it — see Limitations.

The upgrade is a different purchase. On 2026-09-19 VectorBT PRO membership was advertised at $25 a month, $240 for twelve months (shown as reduced from $300, i.e. $20 a month) or $500 once for lifetime access, payable through GitHub Sponsors, Ko-fi, Stripe, Patreon, Paddle or Gumroad, with a note that individual memberships cover personal, non-commercial use only and that the figures are a limited-time rate for early contributors. Those are another product's prices, on a page this catalogue does not have a card for; check them before planning against them.

Data & coverage

VectorBT is not a data source and has no feed of its own, which is why this card records no data latency: whatever you hand it is what you get, at whatever age you fetched it.

It does ship downloaders, and the two crypto ones are candle-only. BinanceData calls python-binance's get_klines, which is the Binance spot klines endpoint. CCXTData calls fetch_ohlcv on whichever CCXT exchange you pass it, so its venue list is CCXT's venue list and its maintenance is CCXT's maintenance, not this library's. Neither fetches trades or order-book depth, and neither is installed by default — ccxt, python-binance, yfinance and alpaca-py all live in the full extra. There is also YFData, AlpacaData and a synthetic GBM generator for tests.

DataUpdater schedules repeated fetches so a dataset can be kept current. That is data plumbing; it does not place orders and it is not why the card's automation flag is false.

Integrations

The library defines no exporter of its own — no to_csv, no to_json; whole objects pickle through save() and load() with dill, and anything else comes from pandas, because every result it returns is a pandas object. Indicators come from TA-Lib, ta and pandas-ta-classic through the indicator factory, or from your own function wrapped by it. Plots are Plotly and Jupyter widgets. Performance reporting can be handed to QuantStats. vectorbt.messaging.telegram wraps python-telegram-bot with send_message and send_to_all so a script can push its own output to you — a sending helper, with no condition engine behind it, which is why this card sets alerts to false.

Version 1.1.0 requires Python 3.11 to 3.14, numpy 2.4.6 or newer and pandas 3.0.3 or newer, which is a more aggressive floor than most libraries in this category.

Limitations

No leverage, no limit orders, no contract multipliers. All three are PRO features and the vendor says so in the README and in its own comparison table. Searching the published 1.1.0 package for leverage and contract_multiplier returns nothing at all.

Which means no perpetuals, in any useful sense. funding and liquidat are likewise absent from the entire package. Nothing models an hourly funding payment, a mark price, maintenance margin or a liquidation. You can feed it a perp's price series and it will backtest it as spot — arriving at a number that omits the two largest cash flows the position actually has. For a crypto reader this is the line that decides it, and it is not fixed by a parameter.

The default fee is zero and the default slippage is zero. _settings.py ships fees=0.0, fixed_fees=0.0 and slippage=0.0. Every README example that does not pass them is a frictionless backtest, and on a strategy that turns over daily that is the difference between an edge and a mirage.

"Open source" is the vendor's word, not the licence. LICENSE.md is Apache 2.0 with the Commons Clause: the source is public, use is free, and the right to sell a product or service whose value derives substantially from the software is explicitly withheld. That is source-available, which is why this card sets open_source to false and writes the condition into the licence string. The README calls it fair-code, the badge reads "Fair Code", and GitHub's own licence detector gives up and returns NOASSERTION. If you intend to run a paid service on top of it, read the file first; the author invites licence-exception requests by email.

No execution of any kind. There is no order placement in the package — no broker adapter, no create_order. Going live means writing that yourself against the signals this produces, and the vendor points paying members at PRO's streaming simulation for it.

No portfolio optimiser. The vendor's comparison table states it plainly: allocations are yours to compute, and PortfolioOptimizer is PRO.

No MCP server, despite the README's talk of AI agents. The documentation chat, the CLI and the MCP server are all PRO.

Alternatives

The first fork in this category is architectural, not commercial. If your rule depends on what happened at the last fill, an event-driven engine will express it and this one will fight you; if you are sweeping parameter grids over long series, nothing event-driven will keep up. Decide that before comparing feature lists.

The second is the licence. If you are building something you intend to sell, an MIT- or BSD-licensed engine removes the Commons Clause question entirely rather than answering it by email.

The third is perpetuals, and it is the one with the shortest shortlist: most general-purpose Python backtesters, this one included, model spot and stop there. The rest of the section is at Backtesting & Research Libraries.