tsbootstrap 0.7.0__tar.gz → 0.7.2__tar.gz

{tsbootstrap-0.7.0/src/tsbootstrap.egg-info → tsbootstrap-0.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tsbootstrap
-Version: 0.7.0
+Version: 0.7.2
 Summary: A Python package to generate bootstrapped time series
 Author-email: Sankalp Gilda <sankalp.gilda@gmail.com>
 Maintainer-email: Sankalp Gilda <sankalp.gilda@gmail.com>, Franz Kiraly <franz.kiraly@sktime.net>, Benedikt Heidrich <benedikt.heidrich@sktime.net>
@@ -38,7 +38,8 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
-Requires-Python: <3.14,>=3.10
+Classifier: Programming Language :: Python :: 3.14
+Requires-Python: <3.15,>=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: numpy
@@ -51,7 +52,7 @@ Requires-Dist: statsmodels; extra == "models"
 Provides-Extra: uq
 Requires-Dist: scikit-learn; extra == "uq"
 Provides-Extra: mcp
-Requires-Dist: mcp<2,>=1.28; extra == "mcp"
+Requires-Dist: mcp<2,>=1.28.1; extra == "mcp"
 Provides-Extra: accel
 Requires-Dist: numba; extra == "accel"
 Provides-Extra: examples
@@ -108,8 +109,6 @@ Requires-Dist: pyright; extra == "dev"
 Requires-Dist: ruff; extra == "dev"
 Requires-Dist: autoflake; extra == "dev"
 Requires-Dist: typos; extra == "dev"
-Requires-Dist: tox; extra == "dev"
-Requires-Dist: tox-gh-actions; extra == "dev"
 Requires-Dist: pycobertura; extra == "dev"
 Requires-Dist: tomlkit; extra == "dev"
 Requires-Dist: memory-profiler>=0.60.0; extra == "dev"
@@ -318,7 +317,12 @@ reduce holds about 20 MB at B=50000 where materializing every replicate takes
 about 1.94 GB (roughly 96x lighter), from
 [`benchmarks/results/membench_2026-07-04.json`](benchmarks/results/membench_2026-07-04.json).
 The multivariate and ragged-panel reduce
-paths have no equivalent in `arch`. Full methodology,
+paths have no equivalent in `arch`. The panel reduce
+(`bootstrap_reduce_panel`) returns the full per-series bootstrap distribution
+of the statistic (`n_bootstraps x num_series`), so quantile and tail workflows
+on an estimator are served directly with no replicate tensor. Use the
+materializing path only when the workflow consumes the resampled paths
+themselves. Full methodology,
 single-threaded numbers, and the reproduction script are in
 [benchmarks/README.md](benchmarks/README.md).
 
@@ -436,6 +440,13 @@ To run all tests, in your developer environment, run:
 uv run pytest tests/
 ```
 
+That runs in a single process. Add the pytest-xdist flags CI uses to run the
+suite in parallel, which is several times faster on a multi-core machine:
+
+```sh
+uv run pytest tests/ -n auto --dist loadscope --max-worker-restart 3
+```
+
 The sktime adapter classes can be validated with sktime's estimator checks:
 
 ```python

{tsbootstrap-0.7.0 → tsbootstrap-0.7.2}/README.md

@@ -201,7 +201,12 @@ reduce holds about 20 MB at B=50000 where materializing every replicate takes
 about 1.94 GB (roughly 96x lighter), from
 [`benchmarks/results/membench_2026-07-04.json`](benchmarks/results/membench_2026-07-04.json).
 The multivariate and ragged-panel reduce
-paths have no equivalent in `arch`. Full methodology,
+paths have no equivalent in `arch`. The panel reduce
+(`bootstrap_reduce_panel`) returns the full per-series bootstrap distribution
+of the statistic (`n_bootstraps x num_series`), so quantile and tail workflows
+on an estimator are served directly with no replicate tensor. Use the
+materializing path only when the workflow consumes the resampled paths
+themselves. Full methodology,
 single-threaded numbers, and the reproduction script are in
 [benchmarks/README.md](benchmarks/README.md).
 
@@ -319,6 +324,13 @@ To run all tests, in your developer environment, run:
 uv run pytest tests/
 ```
 
+That runs in a single process. Add the pytest-xdist flags CI uses to run the
+suite in parallel, which is several times faster on a multi-core machine:
+
+```sh
+uv run pytest tests/ -n auto --dist loadscope --max-worker-restart 3
+```
+
 The sktime adapter classes can be validated with sktime's estimator checks:
 
 ```python

{tsbootstrap-0.7.0 → tsbootstrap-0.7.2}/pyproject.toml

@@ -1,5 +1,11 @@
 [build-system]
-requires = ["setuptools>=61", "wheel"]
+# Floor is 83, not 61: setuptools below 83.0.0 lets a MANIFEST.in exclusion be
+# bypassed in an sdist through a Unicode NFC/NFD normalization collision on the
+# case-folding macOS filesystems (CVE-2026-59890), so a file the manifest excludes
+# can still be published in the source distribution. This is the bound that governs
+# our own sdist build. 83.0.0 requires Python >= 3.10, which matches requires-python
+# below, so the floor costs no supported interpreter.
+requires = ["setuptools>=83", "wheel"]
 build-backend = "setuptools.build_meta"
 
 [tool.setuptools.packages.find]
@@ -7,7 +13,7 @@ where = ["src"]
 
 [project]
 name = "tsbootstrap"
-version = "0.7.0"
+version = "0.7.2"
 description = "A Python package to generate bootstrapped time series"
 maintainers = [
     { name = "Sankalp Gilda", email = "sankalp.gilda@gmail.com" },
@@ -19,7 +25,7 @@ authors = [
 ]
 license = { file = "LICENSE" }
 readme = "README.md"
-requires-python = ">=3.10,<3.14"
+requires-python = ">=3.10,<3.15"
 classifiers = [
     "Development Status :: 4 - Beta",
     "Intended Audience :: Science/Research",
@@ -33,6 +39,7 @@ classifiers = [
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.14",
 ]
 
 dependencies = [
@@ -60,7 +67,11 @@ uq = [
 # on top of CORE tsbootstrap only (it does NOT pull the models or uq extras): the two
 # tools are observation-resampling bootstraps, which are pure-numpy.
 mcp = [
-    "mcp>=1.28,<2",
+    # Lower bound is 1.28.1, not 1.28: the WebSocket server transport in <= 1.28.0
+    # performs no Host/Origin validation (CVE-2026-59950), so a browser page can drive
+    # a locally bound server. tsbootstrap-mcp speaks stdio, but the bound keeps a
+    # consumer that imports the SDK for its own transport off the vulnerable release.
+    "mcp>=1.28.1,<2",
 ]
 
 # Compiled acceleration: a numba kernel for the VAR recurrence (replicate-parallel).
@@ -133,8 +144,6 @@ dev = [
     "ruff",
     "autoflake",
     "typos",
-    "tox",
-    "tox-gh-actions",
     "pycobertura",
     "tomlkit",
     "memory-profiler>=0.60.0",  # For performance testing
@@ -161,7 +170,15 @@ required-version = "==0.11.28"
 [tool.pytest.ini_options]
 minversion = "6.0"
 pythonpath = ["."]  # repo root on sys.path so tests can import the shared tests._helpers package
-addopts = "-ra -q -n auto --dist loadscope --max-worker-restart 3 --durations=20"
+# Reporting only. Parallelism is deliberately NOT here: addopts applies to every pytest
+# invocation in this repository, so putting "-n auto" in it made xdist the default for
+# callers that must not have it, and each of those then had to remember to undo it. The
+# notebook job forgot, started one Jupyter kernel per worker, and the kernels raced to
+# bind their ZMQ ports. The obvious escape hatch does not work either: "-p no:xdist"
+# unregisters the -n / --dist / --max-worker-restart options that addopts still passes,
+# so pytest exits 4 on unrecognized arguments. Callers that want parallelism now ask for
+# it: see the Run tests steps in .github/workflows/CI.yml and DEVELOPER_NOTES.md.
+addopts = "-ra -q --durations=20"
 testpaths = [
     "tests",
 ]
@@ -482,8 +499,12 @@ do_not_mutate_patterns = [
 ]
 
 pytest_add_cli_args_test_selection = ["-m", "not slow", "tests/"]
-# -o addopts= drops the project's default addopts (coverage, xdist, durations). The per-test
-# timeout is mutmut's hang guard: a mutant that sends an iterative solver or a recurrence into a
-# non-terminating loop is killed at 60s (thread method, so it interrupts even C-extension calls)
-# instead of stalling the whole run.
+# -o addopts= drops the project's default addopts, which is now reporting only: -ra, -q and
+# --durations=20. mutmut runs pytest once per mutant, so a per-run durations table and a summary
+# of every non-passing outcome are pure noise across thousands of runs. It has nothing to do
+# with xdist any more:
+# parallelism has left addopts and lives on the callers that want it.
+# The per-test timeout is mutmut's hang guard: a mutant that sends an iterative solver or a
+# recurrence into a non-terminating loop is killed at 60s (thread method, so it interrupts even
+# C-extension calls) instead of stalling the whole run.
 pytest_add_cli_args = ["-o", "addopts=", "--timeout=60", "--timeout-method=thread"]

{tsbootstrap-0.7.0 → tsbootstrap-0.7.2}/src/tsbootstrap/api.py

@@ -694,7 +694,9 @@ def bootstrap_reduce_panel(
     representation is a list of per-series arrays (or a flat array plus a CSR ``indptr``),
     not the rectangular ``(n, d)`` that :func:`bootstrap_reduce` assumes. The output gains
     a series axis, and ``.values()`` is mathematically incoherent across unequal lengths,
-    so the reduce IS the panel API.
+    so the reduce IS the panel API. The compiled kernels consume the
+    flat-values-plus-offsets layout directly, so access is contiguous within each series
+    rather than per-element pointer chasing.
 
     Parameters
     ----------
@@ -737,6 +739,11 @@ def bootstrap_reduce_panel(
         ``.statistics`` of shape ``(n_bootstraps, num_series, |theta|)``, collapsed to
         ``(n_bootstraps, num_series)`` when the series are univariate and the statistic is
         scalar (mirroring the ``(B,)`` collapse of the rectangular reduce).
+
+        The reduce returns the full per-series bootstrap distribution of the statistic
+        (``n_bootstraps x num_series``), so quantile and tail workflows on an estimator
+        are served directly with no replicate tensor. Use the materializing path only
+        when the workflow consumes the resampled paths themselves.
     """
     if backend not in ("numpy", "compiled"):
         raise MethodConfigError(

{tsbootstrap-0.7.0 → tsbootstrap-0.7.2}/src/tsbootstrap/model/arima.py

@@ -8,14 +8,20 @@ back using the original initial levels. statsmodels is imported lazily.
 from __future__ import annotations
 
 from dataclasses import dataclass
+from typing import Any
 
 import numpy as np
+from numpy.linalg import LinAlgError
 from numpy.typing import NDArray
 from scipy.signal import lfilter, lfiltic
 
-from tsbootstrap.errors import Codes, MethodConfigError
+from tsbootstrap.errors import Codes, MethodConfigError, ModelStabilityError
 from tsbootstrap.model.fit import _require_statsmodels
 
+# Floor for the innovation variance in the fallback starting values: a constant (or
+# near-constant) series has zero sample variance, which is not a usable sigma2 start.
+_MIN_START_SIGMA2 = 1e-10
+
 
 @dataclass(frozen=True, slots=True)
 class ARMAFit:
@@ -66,10 +72,79 @@ def integrate_batched(w: NDArray[np.float64], levels: list[float]) -> NDArray[np
     return cur
 
 
+def _interior_start_params(
+    model: Any, demeaned: NDArray[np.float64], p: int, q: int
+) -> NDArray[np.float64] | None:
+    """Return statsmodels' starting values, or a white-noise start when they are unusable.
+
+    Returns ``None`` when the model's Hannan-Rissanen starting values survive the
+    stationarity transform, so the refit keeps statsmodels' default start. Returns an
+    explicit interior start (zero AR/MA, sigma2 at the sample variance) otherwise.
+
+    statsmodels already replaces starting AR parameters it judges non-stationary with
+    zeros, but that check accepts values sitting exactly ON the unit circle. Their
+    inverse stationarity transform divides by ``sqrt(1 - r**2) == 0`` and yields
+    ``+/-inf``, whose forward transform is NaN, so an optimization started there returns
+    NaN parameters rather than raising.
+    """
+    with np.errstate(divide="ignore", invalid="ignore"):
+        unconstrained = np.asarray(model.untransform_params(model.start_params), dtype=np.float64)
+    if np.isfinite(unconstrained).all():
+        return None
+    sigma2 = max(float(np.var(demeaned)), _MIN_START_SIGMA2)
+    return np.concatenate([np.zeros(p + q, dtype=np.float64), [sigma2]])
+
+
+def _fit_arma_mle(demeaned: NDArray[np.float64], p: int, q: int) -> Any:
+    """Maximum-likelihood ARMA(p, q) fit of ``demeaned``, robust at the unit circle.
+
+    statsmodels enforces stationarity with ``r = u / sqrt(1 + u**2)``. Once ``|u|``
+    passes roughly 1e8 that expression rounds to exactly ``+/-1.0`` in float64, so the
+    AR polynomial acquires a root exactly ON the unit circle. The stationary state-space
+    initialization must then solve a discrete Lyapunov equation whose transition matrix
+    is singular, and LAPACK fails inside the Kalman machinery -- reaching the caller as a
+    raw ``numpy.linalg.LinAlgError`` ("LU decomposition error", or "Schur decomposition
+    solver error" depending on which factorization is reached first). This happens either
+    at the Hannan-Rissanen starting values or at an intermediate optimizer step, so
+    conditioning the starting values alone does not close it.
+
+    The recovery refits the identical model under an approximate-diffuse initialization,
+    which solves no Lyapunov equation and so stays well defined for a transition matrix
+    with unit-modulus eigenvalues, starting from the interior when the default starting
+    values are themselves unusable. The recovered fit is returned as-is: it may well be
+    non-stationary, and rejecting it is the stability layer's job, not this one's. If the
+    recovery cannot produce finite parameters either, the failure is reported as a typed
+    :class:`~tsbootstrap.errors.ModelStabilityError`.
+    """
+    from statsmodels.tsa.arima.model import ARIMA as _SMARIMA
+
+    try:
+        return _SMARIMA(demeaned, order=(p, 0, q), trend="n").fit()
+    except LinAlgError as exc:
+        recovery = _SMARIMA(demeaned, order=(p, 0, q), trend="n")
+        recovery.ssm.initialize_approximate_diffuse()
+        start_params = _interior_start_params(recovery, demeaned, p, q)
+        try:
+            res = recovery.fit(start_params=start_params)
+            if np.isfinite(res.arparams).all() and np.isfinite(res.maparams).all():
+                return res
+        except LinAlgError:
+            # The diffuse re-fit reached the same singularity. Fall through to the typed
+            # error rather than let a linear-algebra internal reach the caller.
+            pass
+        raise ModelStabilityError(
+            f"ARMA(p={p}, q={q}) maximum likelihood reached an autoregressive root on the "
+            f"unit circle and could not be re-fit under a diffuse initialization; the series "
+            f"is probably under-differenced",
+            code=Codes.NEAR_UNIT_ROOT,
+            context={"p": p, "q": q, "n": int(demeaned.shape[0])},
+            hint="Increase the differencing order d, or reduce the ARMA order.",
+        ) from exc
+
+
 def fit_arma(w: NDArray[np.float64], p: int, q: int) -> ARMAFit:
     """Fit a demeaned ARMA(p, q) to the (already differenced) series ``w``."""
     _require_statsmodels()
-    from statsmodels.tsa.arima.model import ARIMA as _SMARIMA
 
     series = np.ascontiguousarray(np.asarray(w, dtype=np.float64).ravel())
     n = series.shape[0]
@@ -80,7 +155,7 @@ def fit_arma(w: NDArray[np.float64], p: int, q: int) -> ARMAFit:
             context={"p": p, "q": q, "n": n},
         )
     mean = float(series.mean())
-    res = _SMARIMA(series - mean, order=(p, 0, q), trend="n").fit()
+    res = _fit_arma_mle(series - mean, p, q)
     ar_coefs = np.ascontiguousarray(np.asarray(res.arparams, dtype=np.float64))
     ma_coefs = np.ascontiguousarray(np.asarray(res.maparams, dtype=np.float64))
     # Derive the innovations in OUR engine's convention (scipy lfilter), not statsmodels'

{tsbootstrap-0.7.0 → tsbootstrap-0.7.2/src/tsbootstrap.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tsbootstrap
-Version: 0.7.0
+Version: 0.7.2
 Summary: A Python package to generate bootstrapped time series
 Author-email: Sankalp Gilda <sankalp.gilda@gmail.com>
 Maintainer-email: Sankalp Gilda <sankalp.gilda@gmail.com>, Franz Kiraly <franz.kiraly@sktime.net>, Benedikt Heidrich <benedikt.heidrich@sktime.net>
@@ -38,7 +38,8 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
-Requires-Python: <3.14,>=3.10
+Classifier: Programming Language :: Python :: 3.14
+Requires-Python: <3.15,>=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: numpy
@@ -51,7 +52,7 @@ Requires-Dist: statsmodels; extra == "models"
 Provides-Extra: uq
 Requires-Dist: scikit-learn; extra == "uq"
 Provides-Extra: mcp
-Requires-Dist: mcp<2,>=1.28; extra == "mcp"
+Requires-Dist: mcp<2,>=1.28.1; extra == "mcp"
 Provides-Extra: accel
 Requires-Dist: numba; extra == "accel"
 Provides-Extra: examples
@@ -108,8 +109,6 @@ Requires-Dist: pyright; extra == "dev"
 Requires-Dist: ruff; extra == "dev"
 Requires-Dist: autoflake; extra == "dev"
 Requires-Dist: typos; extra == "dev"
-Requires-Dist: tox; extra == "dev"
-Requires-Dist: tox-gh-actions; extra == "dev"
 Requires-Dist: pycobertura; extra == "dev"
 Requires-Dist: tomlkit; extra == "dev"
 Requires-Dist: memory-profiler>=0.60.0; extra == "dev"
@@ -318,7 +317,12 @@ reduce holds about 20 MB at B=50000 where materializing every replicate takes
 about 1.94 GB (roughly 96x lighter), from
 [`benchmarks/results/membench_2026-07-04.json`](benchmarks/results/membench_2026-07-04.json).
 The multivariate and ragged-panel reduce
-paths have no equivalent in `arch`. Full methodology,
+paths have no equivalent in `arch`. The panel reduce
+(`bootstrap_reduce_panel`) returns the full per-series bootstrap distribution
+of the statistic (`n_bootstraps x num_series`), so quantile and tail workflows
+on an estimator are served directly with no replicate tensor. Use the
+materializing path only when the workflow consumes the resampled paths
+themselves. Full methodology,
 single-threaded numbers, and the reproduction script are in
 [benchmarks/README.md](benchmarks/README.md).
 
@@ -436,6 +440,13 @@ To run all tests, in your developer environment, run:
 uv run pytest tests/
 ```
 
+That runs in a single process. Add the pytest-xdist flags CI uses to run the
+suite in parallel, which is several times faster on a multi-core machine:
+
+```sh
+uv run pytest tests/ -n auto --dist loadscope --max-worker-restart 3
+```
+
 The sktime adapter classes can be validated with sktime's estimator checks:
 
 ```python

{tsbootstrap-0.7.0 → tsbootstrap-0.7.2}/src/tsbootstrap.egg-info/requires.txt

@@ -34,8 +34,6 @@ pyright
 ruff
 autoflake
 typos
-tox
-tox-gh-actions
 pycobertura
 tomlkit
 memory-profiler>=0.60.0
@@ -64,7 +62,7 @@ sktime
 jupyter
 
 [mcp]
-mcp<2,>=1.28
+mcp<2,>=1.28.1
 
 [models]
 statsmodels