mktlib-scan

Compiled entry/exit chain resolver for mktlib's backtest engine.

pip install mktlib[fast]

That is the whole user-facing story. This package has no API you are meant to call — mktlib discovers it and dispatches to it automatically. Installing it changes performance and nothing else.

What it does

mktlib resolves which entry signals actually open a position by walking the entry/exit chain forward. In pure Python that walk is fast, but getting the data to it is not: Series.to_list() boxes one Python object per element.

Measured in mktlib at 500k bars, before this package existed:

Two-thirds of the cost was marshalling, not computing. So this package takes polars.Series directly and never builds a Python list of floats. An FFI that accepted lists would re-pay that and capture almost nothing — which is the single most important thing to know if you ever change this boundary.

Correctness

mktlib/backtest/_scan.py remains the reference implementation and the fallback. This crate must agree with it bit for bit, and that is enforced rather than asserted:

• mktlib runs its full equivalence corpus and all of its frozen golden baselines against both backends. The baselines were frozen before this package existed, so any divergence fails.
• The kernel is a separate crate with no pyo3 and no polars, unit-tested against the rules that are easy to get wrong — entry-wins-on-a-shared-bar, session forcing, NaN-as-null, strictness folding, and the pair/general tie-break.
• nextafter is pinned bit-for-bit against a table generated from CPython, including the signed zeros, where a sign-propagating implementation diverges.

Bit-identity is achievable here rather than aspirational: the outputs are indices and booleans, and the only float operations are comparisons and nextafter. No accumulation, no reassociation, nothing for an optimizer to reorder.

The semantic contract both implementations satisfy is in CONTRACT.md.

Layout

License

Apache-2.0