Repository navigation
Releases
EconHDFE v0.7.0
EconHDFE v0.7.0 — Alpha research release
中文
本次发布整合 Post/Prediction、IV 正确性修复以及中英文文档。项目仍为 Alpha 阶段研究软件。
- 标准 OLS/线性 IV 支持冻结设计状态、分块
predict()、原始样本行位置恢复,以及显式save_fe=True保存的分类固定效应预测。未知类别、违反嵌套关系和未识别组合明确拒绝。 - 线性组合与 Wald 检验保留非有限值、负方差、秩亏和零方差保护;
stdp仅包含 beta 协方差,不包含固定效应不确定性或完整预测区间。 - IV 增加矩条件识别与数值秩保护;敏感 2SLS 使用压缩工具变量空间 QR/SVD;修正加权 first-stage fitted 值的坐标及 fixed-k 协方差,并使用调用内有界诊断缓存。
- 整合此前 0.6.4/0.6.5 开发线的资源边界守卫与测试体系整理。测试入口为
python scripts/run_tests.py --suite full;普通 pytest 只运行日常核心集。 - 中英文首页、技术文档、历史目录和发布打包入口统一。旧持久化结果缓存按 ABI 安全失效;已有 within 数值缓存 ABI 不变。
支持 Python 3.10–3.13。升级命令:pip install --upgrade econhdfe==0.7.0。
发布流程使用成功的 16 环境 CI(Linux/macOS/Windows 与可行最低依赖环境)及 bundle 验证产出的冻结制品,经独立验收后发布;GitHub Release 与 PyPI wheel/sdist 字节相同,哈希见 SHA256SUMS.txt。
边界:不包含 PPML/IV-PPML、varying-slope/group+individual FE 样本外预测、margins/AME;不新增 licensed-Stata/upstream 外部认证、普遍提速或全流程内存预算保证。Lean 证明覆盖明确假设下的数学命题,不等于 Python/Numba 浮点程序的形式验证或历史原创性认证。
English
This alpha research release brings together Post/Prediction, IV correctness fixes, and bilingual documentation.
- Standard OLS and linear IV gain frozen prediction state, chunked prediction, original-row restoration, and explicitly saved categorical fixed effects. Unknown categories and unidentified or inconsistent FE combinations fail explicitly.
- Linear contrasts and Wald tests retain finite-value, variance and rank safeguards. Prediction
stdpis beta-only; it is not a full prediction interval or an estimate of FE uncertainty. - IV estimation adds cross-moment identification/rank guards and a compressed instrument-coordinate QR/SVD solve for sensitive 2SLS systems. Weighted first-stage coordinates and fixed-k covariance are corrected, with bounded call-local diagnostic reuse.
- Includes resource guards and the reorganized full test suite from the unreleased 0.6.4/0.6.5 development lines, together with English/Chinese documentation and packaging cleanup.
Python 3.10–3.13 is supported. Install with pip install --upgrade econhdfe==0.7.0.
Publication consumes the frozen candidate from successful matrix and bundle validation after detached acceptance; it does not rebuild after tagging. GitHub and PyPI share the exact wheel/sdist bytes.
PPML/IV-PPML prediction, varying-slope/group+individual FE prediction, margins/AME, complete FE uncertainty and full external Stata/upstream certification remain outside this release's claims. No universal performance, end-to-end memory-bound or implementation-formal-verification claim is added.
EconHDFE v0.6.3 — Alpha Research Release
EconHDFE v0.6.3 — Alpha Research Release
v0.6.3 is the first release candidate to complete the package's full cross-platform release gate.
Highlights
- Keeps the shared OLS / linear-IV / PPML / IV-PPML HDFE architecture and public estimator contracts unchanged.
- Fixes Python 3.10/3.11 compatibility-gate reflection differences.
- Fixes Windows newline-sensitive release checks.
- Hardens managed Numba
workqueueconcurrency and bootstrap BLAS thread control. - Uses an isolated sdist-to-wheel build path and verifies clean installations.
- Separates original and public-redacted benchmark provenance without weakening hash checks.
Validation
- Linux / Windows / macOS × Python 3.10–3.13: 12/12 environments passed.
- Four feasible minimum-runtime-dependency environments passed.
- Each CI environment: 928 passed, 1 skipped, 0 failed.
- Local full suite with the optional exact backend: 929 passed.
- Frozen wheel numerical recheck: 350 passed.
- Final bundle verification and detached formal release gate passed.
The single CI skip is an optional exact-backend test; it was executed successfully in the local full suite.
Release identity
This release is built from and tagged at commit:
2782d93f901ec1eed28ac0f963834701e0a44b09
The uploaded artifacts are the already-validated frozen artifacts; they are not rebuilt after tagging.
Status and external validation boundary
EconHDFE remains an alpha research release. Licensed-Stata/upstream external certification remains a separate boundary. This release does not add a universal performance claim or an independent mathematical-novelty certification.
See SHA256SUMS.txt in the release assets for exact artifact hashes.