Changelog

Notable changes to OpenQuant, newest first within each section. The Rust crate is published on crates.io as openquant and the Python package on PyPI as pyopenquant (import name openquant). Changes not yet in a tagged release sit under Unreleased; when a version is tagged, its entries move under a heading with that version and date.

Each entry links the pull request that made the change; the linked issue, where there is one, is in parentheses. Entries were seeded from the titles of the merged pull requests. From now on, a pull request with a user-visible effect adds one line to the right section below, in the same style as its title. The docs site renders this file as its Changelog page (python3 scripts/docs/generate_site_pages.py --write; CI fails if the page and this file disagree).

Unreleased

0.1.0 - 2026-10-01

First release.

Added

• Release: the PyPI distribution name is pyopenquant (#58)
• cross_validation: walk-forward splits over the purged k-fold folds, a no-signal null through any splits, and the purged k-fold bias toward features with memory documented; runbooks 11 to 14 audited, no conclusion changed (#217)
• Runbook 14, fracdiff features against returns in a purged-CV classifier, net of costs (#209) (#216)
• Docs: a research gallery generated from the executed runbooks, a performance page generated from benchmarks/, and this changelog on the site; the governance stubs folded into one page (#55) (#207)
• hcaa: linkage option (single, complete, average, ward), Ward by default (#197) (#203)
• Runbooks 11 and 13: a sequential-bootstrap bagging comparison, and the workarounds for #187 removed (#199)
• hcaa: distance option shared with HRP, pairwise default kept (#188) (#195)
• Runbook 13, bet sizing from predicted probabilities (#50) (#191)
• Runbook 12, CPCV backtest with PSR and the deflated Sharpe ratio (#48) (#179)
• Runbook 11, triple-barrier labeling and meta-labeling (#47) (#181)
• Runbook 10, HRP against inverse-variance and CLA out of sample (#51) (#149)
• Runbook 09, fracdiff stationarity versus memory (#49) (#148)
• Notebooks executed with nbclient, committed with outputs and figures, and run in CI (#45) (#146)
• Python bindings for purged CV, CPCV backtesting, feature importance and tuning (#42) (#143)
• openquant.data.fetch with a Parquet cache, the dataset hash in run manifests, and DATA_SOURCES.md (#43) (#141)
• openquant.evaluation: PSR, DSR, MinTRL and a trial registry (#44) (#139)
• Recovered stranded work: purged-CV diagnostics, CPCV and pipeline plots re-ported, outcomes recorded (#33) (#137)
• AFML Chapter 21 pigeonhole-partition trajectory search (#112) (#131)
• The Critical Line Algorithm; the cla module was a stub (#76) (#83)
• streaming_hpc analytics utilities, AFML Chapter 22 (#12)
• Combinatorial optimization adapters, AFML Chapter 21 (#11)
• hpc_parallel utilities, AFML Chapter 20 (#10)
• Typed error APIs for bet_sizing and filters (#9)
• Strategy risk diagnostics, AFML Chapter 15 (#8)
• Synthetic backtesting, AFML Chapter 13 (#7)
• Backtesting engine and CPCV, AFML Chapters 11–12 (#6)
• hyperparameter_tuning, AFML Chapter 9 (#5)
• ensemble_methods, AFML Chapter 6 (#4)
• AFML notebook platform and a real-data end-to-end research workflow (#3)
• Notebook-first AFML research platform: bindings, pipeline, notebooks, CI (#2)

Fixed

• onc: k-means++ initialisation, as in MLAM Snippet 4.1; results were fragile to the random stream; get_onc_clusters_with_seed and a Python seed argument (#218) (#224)
• data_processing: keep_last keeps the last input row of a duplicated key (stable sort), and the calendar join keeps its order explicitly (#220) (#222)
• A binding default that always failed, ignored parameters and silent drops (#194) (#202)
• Input validation and minor cleanups from the rustdoc audit (#186) (#204)
• Docs/code mismatches in pipeline, hcaa, onc, cross_validation, combinatorial_optimization, microstructural_features and streaming_hpc (#185) (#201)
• sb_bagging rejects a wrong column count and wraps the warm-start seed (#184) (#196)
• Typed errors instead of reachable panics, overflows and infinite loops (#184) (#193)
• Python API gaps found by the meta-labeling runbook (#187) (#190)
• cla error variants and panic guard, data_processing off-grid report and gap rule, strategy_risk unreachable targets (#168) (#189)
• hrp clusters on AFML’s distance of distances by default (#167) (#183)
• Chu-Stinchcombe-White: σ²ₜ is the mean of the squared differences (#173) (#182)
• get_sadf quadratic and sub/super-martingale models follow AFML §17.4 (#166) (#178)
• get_yang_zhang_vol implements Yang and Zhang (2000) (#165) (#176)
• An event on the last bar with no vertical barrier keeps t1 = None (#162) (#175)
• ETF-trick holdings are sized from the previous bar, per AFML §2.4.1 (#164) (#174)
• bet_sizing limit price defined for reducing, short and sign-flipping moves (#163) (#172)
• get_weights_ffd honours lim for every value, so thresh <= 0 cannot hang (#169) (#171)
• PurgedKFold embargoes only after each test block, from the end of the purge (#134) (#170)
• Python binding ergonomics before the API is published (#77) (#145)
• bias_variance_noise reports real noise only given the noiseless target (#128) (#140)
• ef3m five-moment variant fits about the mean, so μ₂ can be negative (#115) (#135)
• streaming_hpc alerts on the CDF of VPIN and weights HHI by volume (#113) (#133)
• feature_importance and backtesting_engine match their AFML references (#94) (#132)
• Python bindings keep sub-second timestamps (#87) (#129)
• MDA shuffles the test-fold column instead of rotating it (#98) (#127)
• portfolio_optimization uses simple returns and reports annual risk and Sharpe ratio (#110) (#126)
• sb_bagging draws each estimator’s sample with the sequential bootstrap (#90) (#125)
• feature_diagnostics fits a real logistic regression (#99) (#124)
• Time-decay weights keep one weight per event when starts coincide (#91) (#122)
• Conditional drawdown at risk averages the drawdown tail (#102) (#121)
• Chu-Stinchcombe-White statistic divides by the standard deviation (#104) (#120)
• Hasbrouck’s lambda regresses the signed return (#105) (#119)
• HCAA allocates down the tree and honours optimal_num_clusters (#108) (#118)
• ONC keeps the re-clustered partition when it scores higher (#107) (#117)
• ef3m fit rows report their own error; fit_m2n honours n_runs (#116)
• Nested partitions and the trajectory path cap in the Chapter 20–22 modules (#114)
• efficient_risk failed on feasible targets (#111)
• Weekly and monthly resampling scrambled the price matrix in four modules (#93) (#96)
• Invalid arguments return Err instead of panicking (#36) (#84)
• The constrained mean-variance problems are solved instead of clamping the unconstrained ones (#82)
• Daily volatility, event resolution and return attribution match AFML (#81)
• Hard-coded HRP, HCAA and ONC answers removed, and the ONC bugs they hid fixed (#75) (#80)
• Python matrices are read as rows of observations, not column-major (#74) (#79)
• PurgedKFold::split leaked: the test window now starts at the first test label’s end time (#65) (#66)
• The first tick threshold is consumed, so features align with bars (#67) (#68)
• Unpurged feature-importance CV can no longer happen silently (#27)
• get_parkinson_vol spelled correctly in the public API (#24)

Changed

• One util::stats module for mean, std_dev, covariance and quantiles (bit-identical results); one shared dollar_bar_sample.csv fixture (#38) (#210)
• The docs site and README use the Monograph identity (#57) (#86)
• Typed errors across the Rust core; the bindings raise the core’s message (#35) (#85)
• One visual identity: palette, type, logo (#56) (#72)
• The Python distribution is named pyopenquant, with accurate metadata (#70)
• AFML book-scraping tooling removed (#32) (#69)
• LazyFrame-native openquant.data path and a benchmark scaffold (#15)

Documentation

• Package metadata points at the real docs (rustdoc for the crate, a Changelog URL for Python); README and docs no longer present the PyPI name as decided (#58)
• Rustdoc under /api/rust/, a generated Python reference, docstrings and .pyi stubs (#54) (#192)
• Rustdoc and doctests for the remaining openquant modules (#37) (#180)
• Rustdoc and doctests for 15 openquant modules (#37) (#144)
• Contributor and community files; test fixture provenance recorded (#61) (#136)
• The vendored pyo3-polars patch documented and guarded with a test (#59) (#142)
• Hand-written hrp, hcaa and onc pages; a claim on the risk_metrics page corrected (#109)
• Hand-written structural_breaks and microstructural_features pages (#106)
• risk_metrics, strategy_risk and codependence pages; a codependence panic fixed (#103)
• Hand-written backtesting pages (engine, statistics, synthetic) and a phone-width math fix (#101)
• Hand-written Chapter 8–9 module pages (feature importance, diagnostics, fingerprint, tuning) (#100)
• Hand-written Chapter 5–7 module pages (fracdiff, ensemble_methods, cross_validation) (#97)
• Hand-written Chapter 4 module pages (sampling, sample_weights, sb_bagging) (#92)
• Hand-written labeling and bet_sizing pages (#89)
• Hand-written Chapter 2 module pages, the authored status and the citation gate (#88)
• The two serious known defects named on the landing page (#78)
• The landing page describes the library, not the documentation (#52) (#71)
• The coverage dashboard recomputes itself (#30)
• check:examples wired into CI, the status badge derived, the information architecture consolidated (#26)
• The get_parkinson_vol rename propagated into the docs (#25)
• The eight setup and workflow pages deliver on their titles (#22)
• Every module page says something true and compile-checked (#21)
• The API-drift gate sees the whole Rust surface (#19)
• The site no longer claims a review that never happened (#18)
• Docs site: 13 production redirects fixed, doc gates wired into CI, sitemap emitted (#17)
• Notebook-first workflow page on the docs site (#14)

Tests

• Fixtures and inline values regenerated without mlfinlab provenance (#138) (#147)
• Measured test-sensitivity audit and reference tests for eight modules (#40) (#95)
• Every Python submodule covered and exposed from the package (#39) (#73)

CI, build and tooling

• Release workflow: abi3 wheels for Linux, macOS and Windows plus an sdist, each smoke-tested on Python 3.11 to 3.13; PyPI trusted publishing, cargo publish and a GitHub Release with notes from this file, on a v* tag only; a dry run from the Actions tab or a packaging pull request (#60)
• The workspace is migrated to Rust edition 2024 (#200)
• Rust caching, parallel lint and test jobs, release-mode parallel notebooks, a docs-only skip (#198)
• CI gates what it claims to gate (#41) (#130)
• Remaining local tooling ignored and the last stale doc lines fixed (#123)
• The workspace is clippy-clean under -D warnings (#34) (#64)
• Repository root cleaned, local tooling ignored, merged branches pruned (#31) (#63)
• AI-DLC adopted and the production-readiness backlog recorded (#62)
• The documented Python examples run in CI, and the three that were broken fixed (#29)
• ethnum bumped from 1.5.2 to 1.5.3 so the crate compiles on stable (#28)
• cargo fmt applied to pyopenquant (#20)
• Per-function data-processing benchmark metrics (#16)
• Beads closure history reconciled with the delivered work (#13)

Dependencies

• rand 0.8 → 0.10 and rand_distr 0.4 → 0.6, superseding #160. Seeded results change. seq_bootstrap_with_rng and sample_log_uniform now take a rand 0.10 generator (R: rand::Rng, the former RngCore), a public API change (#219) (#225)
• polars 0.55, PyO3 0.29 and pyo3-polars 0.28 from crates.io; the vendored pyo3-polars patch is gone and Python polars 1.28.1 or later is required (#220) (#222)
• The cargo minor/patch group, 4 updates (#154)
• The Python minor/patch group, 2 updates (#152)