Nixtla/statsforecast - CodSpeed

statsforecast

Blog Docs Changelog

Performance History

Latest Results

return float max instead of nan

jmoralez:more-arima-fixes

12 hours ago

restore expected

jmoralez:more-arima-fixes

13 hours ago

jmoralez:more-arima-fixes

14 hours ago

reduce comment verbosity and check for transform_pars when setting init

jmoralez:more-arima-fixes

1 day ago

docs: recommend --reinstall-package for re-compiling the C++ code `uv sync --reinstall-package statsforecast` rebuilds the extension from scratch on every invocation, so it always reflects the current sources. Make it the documented way to recompile, and demote `setup.py build_ext --inplace` to the faster-but-sharper alternative. Also note the two ways to end up silently testing a stale extension: a plain `uv sync` never rebuilds it at all, and `build_ext --inplace` does not track headers (pybind11's naive_recompile only compares each .cpp against its object file), so a header edit is a sub-second no-op. Verified: with an edit to include/statsforecast/distributions.h, the documented command picks the change up without deleting the .so first; `--reinstall-package` also rebuilds when nothing changed at all (~30s either way). The `--group` note matters because `uv sync` prunes packages outside the selected groups -- syncing with `--group dev` from a `--group all` environment uninstalls 37 packages. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Edar7n924DeSAMzaJzc6pr

jmoralez:more-arima-fixes

1 day ago

Revert "change efficiency test data" The noise was added to work around test_efficiency[AutoRegressive] failing with `ValueError: non-stationary AR part from CSS` on the perfectly periodic ramp, whose CSS estimate has roots on the unit circle (arCheck correctly rejects it now that it is no longer dead code). b83623b then fixed that path properly: a non-stationary CSS estimate falls back to the stationary `init` instead of raising, which is what R does. So the workaround is no longer needed, and it is expensive -- AutoARIMA on a noisy season_length=24 series runs the full stepwise search instead of exiting immediately. Measured locally, all 38 models, same harness: deterministic 6.3s (AutoARIMA 0.01s, AutoTBATS 2.0s) with noise 54.4s (AutoARIMA 43.8s, AutoTBATS 9.7s) That 8.7x is what took the CodSpeed job from ~9m to 2h29m under simulation mode. Tuning the noise down does not help -- the cost is not monotonic in sigma (totals: s=1.0 51.0s, s=0.1 35.0s, s=0.01 58.1s). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Edar7n924DeSAMzaJzc6pr

jmoralez:more-arima-fixes

1 day ago

docs: fix workflow typo in contribution guide (#1220) Co-authored-by: Mariana Menchero García <47995617+MMenchero@users.noreply.github.com>

main

2 days ago

build(deps): bump the ci-dependencies group across 1 directory with 2 updates (#1227) Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

main

2 days ago

Latest Branches

+65%

arima: model eligiblity fixes#1212

12 hours ago

1117793

jmoralez:more-arima-fixes

0%

build(deps): bump the ci-dependencies group across 1 directory with 2 updates#1227

3 days ago

7a3ebdb

dependabot/github_actions/ci-dependencies-e8b72ca420

0%

Index the per-season alpha in the seasonal ES models#1226

3 days ago

43f427b

VenishPaneliya:seasonal-es-per-season-alpha

© 2026 CodSpeed Technology

Home Terms Privacy Docs