FactorLasso 1.0.2

FactorLasso 1.0.2 adds general weighted inference and opt-in residual-alpha uncertainty, calibration, and bootstrap diagnostics. Existing estimator defaults and production factor-model estimates are unchanged.

Install or upgrade with pip install --upgrade factorlasso==1.0.2.

Added

• Add conditional recursive EWMA residual-alpha uncertainty, with exact influence
  weights, joint calendar Bartlett-HAC covariance, overlap and effective-history
  diagnostics, and explicit initialization targets. Existing alpha values and
  production defaults are unchanged.
• Add the inference capability for general WLS coefficients, weighted means,
  optional full HAC covariance, fixed coefficient contrasts and Gaussian/AR1
  interval calibration with explicit mean and target validation.
• Add exact unnormalized weighted-mean HAC geometry and an opt-in recursive-alpha
  adapter. Retain calendar, residual fingerprints and units; unsupported AR1
  calendars return unavailable bounds.
• Add generic fixed-linear intervals, a joint residual dependent-wild bootstrap
  with re-estimated standard errors and optional residual-panel draws for downstream
  refits, and central Gaussian quadratic quantiles. Bootstrap coverage remains
  explicitly unvalidated.
• Add joint Gaussian sensitivity scenarios, conservative quadratic-region bounds
  with signed noise corrections, and exact joint weighted-mean/HAC covariance
  expectations under a declared temporal shape and fixed ragged masks.
• Add offline coverage, joint covariance-bias and production-mask calibration
  benchmarks. These diagnose declared-model sensitivity and finite-history bias;
  they do not establish empirical or post-selection coverage.
• Add explicit common-calendar zero-innovation residual correlation sensitivity;
  keep existing pairwise-FFILL defaults and reject indefinite matrices without repair.

Changed

• Move prior numerical inference into reusable implementations while preserving
  existing public signatures, result fields, class module paths and prior policy.
  Production prior statistics retain lag-based aggregation without dense geometry.
• Reuse a shared Bartlett kernel in prior HAC geometry and residual-alpha inference.
• Document the separation between generic weighted inference, selected-factor
  prior applications and the existing recursive residual-alpha estimator.
  Production EWMA, HAC covariance correction and factor-model estimates are unchanged.

Validation

• Local Python 3.12 suite: 1,083 passed, 10 skipped.
• Staged-source preflight, Ruff, documentation ownership, and strict Sphinx HTML/doctest builds passed.
• Docstring coverage: 94.5% against the required 89.8% floor.
• Tagged-source sdist and wheel passed Twine, metadata, publication-boundary, and independent installation checks.
• All 29 required GitHub checks passed, including Python 3.10–3.14 on Linux, Windows and macOS, coverage gates, dependency floors, and downstream integrations.
• PyPI Trusted Publishing succeeded; the attached wheel and sdist match PyPI SHA-256 digests recorded in release-manifest.json.