Nixtla/statsforecast - CodSpeed

statsforecast

Blog Docs Changelog

Performance History

Latest Results

arima: compute var_coef once at the reported coefficients and defer it in the search The Hessian is now taken after the fit, in the reported coefficient space (untransformed AR, unrotated xreg), so it needs no gradtrans or rotation mapping and is computed once even when maInvert re-parameterises the MA part. Steps scale with parscale, so regressors whose standard error is far below 1e-3 no longer get a biased Hessian, and _coef_var applies its rank test to the diagonal-normalised matrix, so the cutoff responds to collinearity rather than to units. The Hessian is skipped when the objective is not finite at the optimum, which was the whole cost of the AutoRegressive benchmark regression. arima() gains var_coef=True. auto_arima_f passes False for every candidate and computes the covariance once for the winner through arima_var_coef, which holds the coefficients fixed and differentiates over the estimated ones. The result dict records the method so that function can rebuild the objective. The core.py change was unrelated to this PR and is reverted. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

jmoralez:arima-coef-var

23 hours ago

tests: loosen the R reference tolerances in the xreg rotation test Pin the coefficients to 3 decimals with a 1% tolerance and the log-likelihood and AIC to 1 decimal. The fit lands within 1e-4 of R across platforms, while a mismatch between the rotation and the undo sites moves the coefficients by 20% or more, so the looser bounds keep the guard without pinning optimizer noise. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

jmoralez:fix/xreg-svd-rotation

2 days ago

tests: drop the wrong issue-649 expectation and cover the xreg rotation The rotation only runs for a multi-column xreg with all-free coefficients, and AutoARIMA on the issue-649 series never builds one, so the first candidate still scores inf; the earlier value was not produced by this change. Add a test that the design passed to the OLS initialisation is xreg %*% S$v (orthogonal columns with the singular values as norms), which fails with the previous orientation, and a coefficient/log-likelihood check against stats::arima on a correlated two-regressor fit. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LSjJy9fwzM565cFPn5FUKL

jmoralez:fix/xreg-svd-rotation

2 days ago

build(deps): bump docker/setup-qemu-action in the ci-dependencies group Bumps the ci-dependencies group with 1 update: [docker/setup-qemu-action](https://github.com/docker/setup-qemu-action). Updates `docker/setup-qemu-action` from 4.3.0 to 4.4.0 - [Release notes](https://github.com/docker/setup-qemu-action/releases) - [Commits](https://github.com/docker/setup-qemu-action/compare/1f40c72289eff860ee54a304f1438e3cff362e0a...99012661954931238ded8c8b007157a8430204e1) --- updated-dependencies: - dependency-name: docker/setup-qemu-action dependency-version: 4.4.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: ci-dependencies ... Signed-off-by: dependabot[bot] <support@github.com>

dependabot/github_actions/ci-dependencies-3b757a5d15

3 days ago

fix: preserve AutoTheta configuration

DresdenGman:fix/autotheta-preserve-configuration

3 days ago

feat: expose fixed theta parameter

DresdenGman:feature/theta-parameter

3 days ago

fix(mfles): preserve naive fallback scale

DresdenGman:fix/mfles-naive-original-scale

4 days ago

perf: reuse the ETS and Theta optimizer scratch buffers (#1232) Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

main

6 days ago

Latest Branches

-14%

arima: fix coef stderr#1206

23 hours ago

df8e21d

jmoralez:arima-coef-var

0%

fix(arima): rotate multi-column xreg onto the singular directions as R does#1234

2 days ago

7098986

jmoralez:fix/xreg-svd-rotation

0%

build(deps): bump docker/setup-qemu-action from 4.3.0 to 4.4.0 in the ci-dependencies group#1242

3 days ago

e4c6b37

dependabot/github_actions/ci-dependencies-3b757a5d15

© 2026 CodSpeed Technology

Home Terms Privacy Docs