Changelog¶

0.6.0 — 2026-09-24¶

Joint-column scalar metrics, RatioOfSums, and design-aware BCa.
Optional row-unit ID checks and compact, copy-safe reporting metadata.
Offline assigned-user/order example preserving non-buyers and explicitly distinguishing assigned-user metrics from buyer-only descriptions.
Cluster identifiers retain types/large-integer precision; incompatible labels and duplicate matrix-DataFrame column labels raise explicit errors.
Known-truth benchmarks now include dependent activity/price, changes in their relationship between arms, and clustered dependence at two unit-count regimes.
Release evidence corrected the practical guidance: BCa is not presented as an automatic ratio choice, and small-cluster/skew limitations are explicit.
A supplemental 300-dataset run covers correlated ratios at 100/120 clusters; all methods' Monte Carlo intervals include the nominal 95% target.
See Composite metrics and the source-matched statistical/performance evidence.

BootstrapCV rejects unsupported groups rather than silently allowing entity overlap between train and test; OOB and .632 terminology is clarified.
Numba kernel typing supports both compiled and fallback configurations, and both configurations are checked by CI.
Offline notebooks now execute in CI and verify documented results.
A separate manual workflow executes the real-data Hillstrom analysis with source checksum verification, independent of ordinary PR checks.
Core dependencies, the product-example seed, and benchmark instructions are corrected. The statistical algorithms and 0.5.0 benchmark evidence are unchanged.

A native two-sample API estimates treatment-minus-control differences, ratios, relative lift, and custom scalar effects.
Independent, paired, and separately clustered experiment designs have explicit resampling behavior and result metadata.
Percentile, basic, and BCa intervals are supported; clustered BCa deletes complete clusters in its jackknife step.
New coverage and runtime suites compare applicable workflows with SciPy, record invalid/failing trials, checkpoint long runs, and preserve metadata.
The audited release evidence includes 9,900 valid interval trials, matched runtime/allocation measurements, exact environment metadata, and plots.
A controlled product A/B reference fixes one primary metric, effect, interval, and decision threshold before generating data with known truth.
A real-world Hillstrom notebook retrieves a checksum-verified public experiment and demonstrates the extra boundaries required for archived data, sparse conversion, and concentrated spend.
Experiment documentation now explains analysis units, user aggregation, ratio instability, paired data, and repeated events.
The clustered A/B notebook now demonstrates the native two-sample cluster interface with independent experiment arms.

Results export to compact dictionaries and one-row pandas DataFrames.
The documented numba install extra now exists and has its own CI check.
Numba guidance now states exactly which methods it accelerates, its JIT startup tradeoff, and how to reproduce the fallback comparison.
A limitations guide makes one-sample, scalar-output, dependent-data, and pre-1.0 boundaries explicit.
DataFrame documentation distinguishes separate column intervals from an actual treatment-effect comparison.
The guide now starts from data shape and separates supported paired/grouped workflows from unsupported unpaired A/B effects.
Versioned 0.4.4 runtime, memory, Numba, and 160-cell coverage evidence is published with exact metadata and generated plots.
Coverage benchmark random streams, failure accounting, metadata, and CI smoke validation have been corrected.
Release benchmark profiles now preserve separate outputs and support safe checkpoint/resume for long coverage studies.
Optional-dependency errors, issue forms, and contribution guidance are more actionable for practitioners.

Strict mypy validation now covers the complete source tree and runs in CI.
Time-series tests cover all six methods across batch sizes, optional Numba fallback, taper variants, wild multipliers, and sieve edge cases.
Invalid taper-window errors are stable across supported SciPy versions, and strict typing uses NumPy stubs compatible with the Python 3.10 target.
GitHub Actions use immutable commit pins maintained by Dependabot.
Publishing can be rerun safely after a partial release: an existing complete PyPI version is skipped and the GitHub Release is created or updated.
Documentation is automatically deployed to GitHub Pages when relevant changes reach main, and can also be deployed manually.
Release checks include an 85% coverage floor, dependency consistency, byte-compilation, and core import smoke tests.

Tapered block bootstrap now preserves constant series, location shifts, and variance scale by tapering centered observations with energy normalization.
NumPy fast paths honor batch_size instead of allocating all resamples at once.
BootstrapCV always produces the requested number of nonempty OOB splits and validates small samples and constructor parameters.
Constant time series produce a degenerate sieve-bootstrap distribution instead of failing with a singular matrix.
Release publication now runs tests, lint, documentation, and coverage gates before uploading to PyPI and creates a GitHub Release.

Studentized estimates and nested standard errors now come from the same outer resample.
Bayesian bootstrap evaluates functionals directly under Dirichlet weights; custom statistics provide a weighted callable.
Subsampling intervals use centered, rate-scaled roots.
Bernoulli subsets include a finite-population correction.
Statistical regression tests cover scaling, weighting, cluster uncertainty, and nested-resample pairing.

BootstrapCV for scikit-learn.
pandas accessor for Series and DataFrame.
Expanded README examples and optional dependency model.
New notebook/documentation directions for ML metrics, A/B testing, and time series.