NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #3139 most downloaded on PyPI
Fast implementations of common forecasting routines
Latest stable
0.0.20
Last release 18 days ago
03 Nov 2026
Release timing varies
gaps range from 1 weeks to 11 months
Nearly every release is documented
notes for 21 of 21 stable releases
Nothing withdrawn
no release was ever pulled
3 years old
22 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
feat: run without the GIL on free-threaded Python @jmoralez
AutoDifferences.update and AutoSeasonalDifferences.update raised
ValueError: Last element of indptr must be equal to the size of data
whenever the series had been assigned different numbers of differences.
fit_transform only keeps the tails of the series that are differenced in
each round, but update read them as if every series had them. It now uses
the same layout (#172).AutoDifferences.stack and AutoSeasonalDifferences.stack took the number
of differencing rounds from the first scaler, so stacking scalers whose
largest number of differences wasn't the same raised an IndexError or
silently dropped the later rounds, and the stacked scaler's
inverse_transform and update were wrong for those series. They now keep
every round.cannot open file 'python314.lib': CMake looked up Python through pybind11's
legacy module, which ignores the free-threaded ABI. It now uses CMake's
FindPython. Wheels are published for 3.14t and 3.15t as well. Source builds
need CMake 3.31 or later, the first version whose FindPython builds for
free-threaded Python on Windows; scikit-build-core downloads it if the
installed one is older.Your coding agent can read these notes before it upgrades. Set up the MCP server →
build: add Python 3.15, drop Python 3.10 @nasaul
coreforecast.rolling, coreforecast.expanding and
coreforecast.exponentially_weighted now preserve a leading run of NaNs and
compute the statistic over what follows it. They used to return all NaN as
soon as the series started with one. With skipna=True, min_samples counts
positions after that run, which is what pandas does. The seasonal variants
split what follows the run into seasons, so a leading run whose length isn't a
multiple of season_length now groups different positions together than it
used to.2**24 on were silently rounded
before; the result now takes twice the memory. Pass a float32 array to get
float32 back, which was already the only way to be sure of it.ValueError or an IndexError instead: non-positive window_size,
min_samples and season_length, negative lag, d and k, quantile
levels outside [0, 1], indptr entries that are negative or decreasing,
out-of-range group indices, and ds, stats,
periods and tails arrays whose size doesn't match the group count.
Calling update on a lag transform built with lag=0 also raises now: an
update reads the value that isn't in the array yet, so there is nothing for
it to consume. A negative max_season_length is a ValueError like the
other counts rather than a TypeError from the binding.GroupedArray.indptr is stored and returned as int64 instead of int32,
so arrays are no longer limited to 2**31 elements. Any integer dtype is
still accepted as input; the values that used to be rejected as "not
representable with 32-bit integers" are now checked like every other offset.
It is signed rather than unsigned so that arithmetic mixing it with numpy's
default integer stays integral instead of promoting to float64.ExpandingMean, ExpandingStd, ExpandingMin, ExpandingMax and
ExponentiallyWeightedMean ignored skipna in update(). Every other lag
transform forwards it to a kernel, but these five carry their accumulator in
Python and incorporated a NaN unconditionally, so one arriving through the
incremental path made the statistic NaN for that group permanently even with
skipna=True. transform() was always correct, so training features and
the features produced while predicting recursively disagreed with nothing
raised. With skipna=True the mean and std now leave their state untouched
by a NaN, the min and max ignore it and the EWM forward-fills its mean, and
a group whose whole lagged history was NaN, which the driver skips and whose
stats row it fills with NaN, seeds from the first value it sees, as the
transform does after a leading run of NaNs. Without skipna the updates are
unchanged.ExpandingMin, ExpandingMax and ExponentiallyWeightedMean seeded the
state of an empty group from the output position before its end, which
belongs to another group, so a value the group got later through update()
was folded into that group's statistic. ExpandingMean read the same
position, and on an array with no elements the read raised an IndexError
from all four. An empty group now starts from a NaN state, as with the
other accumulators.LocalBoxCoxScaler.stats_ had an uninitialised second column: the lambda
kernels write one value per group and the array holding them was never
cleared. Transforms were unaffected since that column is not read, but the
attribute differed between identical fits. It is now deterministic padding:
zero for a group that produced a lambda, NaN for an empty or all-NaN group,
whose whole row the driver fills.rand, which glibc
serialises on a process-wide lock. They now come from a per-thread generator,
so quantiles scale like the other kernels (4x on four threads, from 0.5x) and
are slightly faster on a single thread. Results are unchanged: the heights
only affect how the list is searched, not the order it keeps.num_diffs computed the number of KPSS lags once, from the length of the
original series, and reused it after differencing. R's forecast::ndiffs
uses the length of the series being tested, which is one shorter after each
difference, so the two disagreed for lengths where trunc(3 * sqrt(n) / 13)
changes (19, 76, 169, ...). The lags are now recomputed for each difference.tests/cpp, built with
-DCOREFORECAST_BUILD_TESTS=ON against the external_libs/doctest
submodule; existing checkouts need git submodule update --init for it.
They are not part of the sdist or the wheels._lib.rolling.rolling_quantile and seasonal_rolling_quantile
take p right after data, the order the public functions already used.skipna documentation said an interior NaN "may result in
NaN statistics". It doesn't for min-max, where the result is whichever value
wins an unordered comparison. With skipna=False only a leading run of NaNs
is supported; pass skipna=True for anything else.[CHORE] Updating packages settings @nasaul
[FEAT] Implements NaN handling @nasaul
docs.json @deven367 (#108)Bumps to dev version in order to test release
Bumps to dev version in order to test release
fix: remove _greatest_autocovariance by @jmoralez in https://github.com/Nixtla/coreforecast/pull/65
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.15...v0.0.16
fix(rolling): return NaN when n < min_samples by @jmoralez in https://github.com/Nixtla/coreforecast/pull/62
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.14...v0.0.15
feat: add GroupedArray._greatest_autocovariance to estimate season length by @jmoralez in https://github.com/Nixtla/coreforecast/pull/59
GroupedArray._greatest_autocovariance to estimate season length by @jmoralez in https://github.com/Nixtla/coreforecast/pull/59Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.13.1...v0.0.14
fix: do not edit library output path by @jmoralez in https://github.com/Nixtla/coreforecast/pull/56
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.13...v0.0.13.1
add ruff formatting, pre-commit and codspeed by @jmoralez in https://github.com/Nixtla/coreforecast/pull/43
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.12...v0.0.13
return NaN for empty series in reduce by @jmoralez in https://github.com/Nixtla/coreforecast/pull/41
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.11...v0.0.12
use standard thread instead of openmp by @jmoralez in https://github.com/Nixtla/coreforecast/pull/37
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.10...v0.0.11
stack allocate GroupedArray by @jmoralez in https://github.com/Nixtla/coreforecast/pull/35
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.9...v0.0.10
Create CODE_OF_CONDUCT.md by @tracykteal in https://github.com/Nixtla/coreforecast/pull/29
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.8...v0.0.9
ensure dtypes match in scalers by @jmoralez in https://github.com/Nixtla/coreforecast/pull/28
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.7...v0.0.8
fix boxcox inequality check by @jmoralez in https://github.com/Nixtla/coreforecast/pull/25
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.6...v0.0.7
add loglik method to boxcox by @jmoralez in https://github.com/Nixtla/coreforecast/pull/23
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.5...v0.0.6
add take method to scalers and lag_transforms by @jmoralez in https://github.com/Nixtla/coreforecast/pull/21
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.4...v0.0.5
add docs by @jmoralez in https://github.com/Nixtla/coreforecast/pull/11
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.3...v0.0.4
add quantiles by @jmoralez in https://github.com/Nixtla/coreforecast/pull/8
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.2...v0.0.3
add templates to support float64 by @jmoralez in https://github.com/Nixtla/coreforecast/pull/4
Full Changelog: https://github.com/Nixtla/coreforecast/compare/v0.0.1...v0.0.2
initial packaging by @jmoralez in https://github.com/Nixtla/coreforecast/pull/1
Full Changelog: https://github.com/Nixtla/coreforecast/commits/v0.0.1
One column per quarter.