Avatar for the Nixtla user
Nixtla
mlforecast
BlogDocsChangelog

Performance History

Latest Results

docs: fix lag transform guide links (#709) Co-authored-by: Eduardo Lopez <5743857+loama@users.noreply.github.com>
main
1 day ago
docs: fix lag transform guide links
fix/lag-transform-doc-links
3 days ago
docs: fix broken documentation links
fix/broken-documentation-links
3 days ago
Merge branch 'main' into feature/nan-target-support
simonez-tuidi:feature/nan-target-support
6 days ago
[FEAT] Make `skipna` tri-state so `allow_null_target` implies it Requiring both `allow_null_target=True` and `skipna=True` on every transform was two flags for one intent, and the warning about propagating transforms amounted to the library saying "you probably wanted skipna=True". If it knows the answer it should apply it. `skipna` is now `Optional[bool] = None`, where None means "follow `allow_null_target`", resolved in `_fit` once the flag is known. True/False remain explicit overrides, so opting out of skipping is still possible -- a null-containing window yielding a null feature is a legitimate way to signal insufficient data to the model. A blanket flip wasn't possible: the five transforms whose coreforecast `update` ignores `skipna` reject it outright, so inferring True would have turned working `ExpandingMean`/`ExponentiallyWeightedMean` configurations into hard errors just for opting into null targets. An inferred True now downgrades to False for those and warns; an explicit True still raises. The warning that used to fire for any propagating transform is now limited to this case, which makes it both quieter and actionable. Resolution mutates `_core_tfm.skipna` rather than the declared `skipna`, which matters twice: coreforecast reads the attribute at call time so the post-construction flip takes effect, and re-fitting with a different `allow_null_target` re-resolves instead of inheriting the previous answer. Feature names are built during `_parse_transforms`, before any of this runs, so inference can't rename features -- a name reflects the declared spec, not the resolved behaviour. `Offset` and `Combine` take no `skipna` of their own, so they delegate resolution to the transforms they wrap; `_accepts_skipna` is derived from the constructor signature so it can't drift. `Lag`/`LookupLag` are untouched, since the lag of an unobserved value genuinely is unobserved. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
simonez-tuidi:feature/nan-target-support
11 days ago

Latest Branches

CodSpeed Performance Gauge
0%
docs: fix lag transform guide links#709
3 days ago
06799d6
fix/lag-transform-doc-links
CodSpeed Performance Gauge
0%
3 days ago
88fdf14
fix/broken-documentation-links
CodSpeed Performance Gauge
0%
[FEAT] Allow NaN target values at fit/preprocess via `allow_null_target`#706
6 days ago
89cbcf9
simonez-tuidi:feature/nan-target-support
© 2026 CodSpeed Technology
Home Terms Privacy Docs