Nixtla/mlforecast - CodSpeed

mlforecast

Blog Docs Changelog

Performance History

Latest Results

test: build the benchmark statics without the 1k-series panel `statics` derived the column names from the `series` fixture, so the many-series benchmarks also built the unrelated 1.25M-row panel. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01C9961UnEfaPugEEdjHLAoB

perf/vectorize-grouped-array

3 days ago

fix: train LightGBMCV boosters on features_order_, harden categorical_feature carry-over - build the lgb.Dataset from ts.features_order_ so the CV boosters see the same features that predict and MLForecast.fit use (id column as a static feature, groupby lag transforms, positional categorical_feature) - warn when params carries a categorical_feature alias, since LightGBM ignores it once the Dataset is constructed - from_cv tolerates CV objects pickled before categorical_feature existed and copies the list instead of sharing it - _LGBMRegressor.fit takes categorical_feature as an explicit keyword Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01K3e3v5YVqjboL2Cw4C14py

lgbcv-categorical

3 days ago

feat: categorical_feature in LightGBMCV, carried over by MLForecast.from_cv LightGBMCV builds the lgb.Dataset before the boosters see params, so categorical_feature in params was silently ignored. Add it as an argument of setup/fit forwarded to lgb.Dataset, and have from_cv return an LGBMRegressor subclass that keeps it as an attribute (outside get_params, preserved through __sklearn_clone__) and applies it in fit. Requires scikit-learn>=1.3. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UTU4n15Z3py2kzx61nKZ9i

lgbcv-categorical

3 days ago

build(deps): bump anyio from 4.12.1 to 4.14.2 Bumps [anyio](https://github.com/agronholm/anyio) from 4.12.1 to 4.14.2. - [Release notes](https://github.com/agronholm/anyio/releases) - [Commits](https://github.com/agronholm/anyio/compare/4.12.1...4.14.2) --- updated-dependencies: - dependency-name: anyio dependency-version: 4.14.2 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com>

dependabot/uv/anyio-4.14.2

3 days ago

perf: cheaper future ids and date feature positions _get_future_ids went through numpy for pandas ids, which re-factorizes a CategoricalIndex; repeating the index keeps the codes. The date features built a dict of every unique date to map the time column, which get_indexer does in one pass. At 50k ids x h=14: _get_future_ids 38 -> 0.5 ms (object ids stay on the previous path). Date features: 17.5 -> 12.5 ms per 1.25M rows. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

perf/vectorize-grouped-array

3 days ago

Merge branch 'main' into feat/model_report

janrth:feat/model_report

6 days ago

refactor: one writer for TimeSeries fit settings, copies for sub-steps (#739) Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

main

6 days ago

fix: address review on clone paths added by this PR `_fit_settings` dereferences `horizon_features_`, which `__init__` sets but pre-v1.1.0 pickles lack, so `_clone_warm` raised `AttributeError` on them. Give it a class-level default next to the other fit settings. `_clone_cold` now clones target transforms where it used to deep-copy them, so the `_clone`/`clone` overrides that built a hardcoded class were silently downgrading user subclasses. Build from `type(self)` instead, via the `_BaseLocalScaler._clone` pattern. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VZ8NAmp6gycFtM6XTpvhSL

refactor/declared-state

7 days ago

Latest Branches

+2%

perf: vectorize the GroupedArray group loops#740

3 days ago

047d63e

perf/vectorize-grouped-array

0%

feat: categorical_feature in LightGBMCV#742

3 days ago

4832898

lgbcv-categorical

0%

build(deps): bump anyio from 4.12.1 to 4.14.2#741

3 days ago

c5f4322

dependabot/uv/anyio-4.14.2

© 2026 CodSpeed Technology

Home Terms Privacy Docs