[quantum-like]

Owning issue

Closes #89

Outcome

7 correctness fixes, one commit each, each with a focused test:

• Return skewness and kurtosis divided 1/n central moments by the n−1 standard deviation, which understated both inputs to the deflated Sharpe ratio on short samples. They now use population moments (m3/m2^1.5, m4/m2^2). The Sharpe ratio itself is unchanged. Test: tightened tests/test_evaluation/test_moments.py::test_reference_scipy_skewness and ::test_reference_scipy_kurtosis to an exact SciPy match on 30 observations. (7237b26)
• Rolling return attribution fit betas on uncleaned factor rows after NaN rows were dropped from returns, which shifted every lagged beta. Betas now use the same cleaned rows they are applied to. Test: tests/test_evaluation/test_factor_attribution.py::test_nan_row_matches_predropped_sample. (48787d0)
• The lower Hill tail sorted absolute returns, so large gains entered the loss tail. It now uses losses only. Test: tests/test_evaluation/test_distribution.py::test_hill_lower_tail_ignores_right_tail_spikes. (1d72278)
• Follow-up: with the lower-tail fix, tail="both" raised on all-positive samples. It now falls back to the upper tail when no loss tail exists. Test: tests/test_evaluation/test_distribution.py::test_hill_both_on_positive_sample_uses_upper_tail. (8779ad5)
• String factor timestamps used cast(Date)/cast(Datetime), which Polars 2 rejects. They are now parsed with str.to_date/str.to_datetime. Test: existing TestUtf8Timestamps (both tests were failing before). (84a452c)
• Empty frames crashed outlier filtering (Polars 2 refuses abs() on a Null-dtype column). They are now returned unchanged. Test: existing test_empty_dataframe (was failing before). (947459a)
• Non-mapping YAML/JSON configs (empty, null, list) raised TypeError. They now raise the documented ValueError. Test: tests/test_config/test_io.py::test_non_mapping_documents_raise_value_error. (b3c5628)

Compatibility

• No public compatibility impact
• Compatibility impact is documented in the owning issue and migration guidance

Deflated Sharpe inputs (skewness, kurtosis) change numerically on short samples, and rolling attribution betas change when returns contain NaN rows. Note that the distribution-moments API (evaluation/distribution/moments.py) still uses unbiased sample moments; both conventions are documented.

Verification

• Tests reproduce the prior failure or verify the requested behavior (each listed test fails on main and passes with its fix)
• Ruff lint and format checks pass (ruff format --check is clean on the changed files and ruff check reports nothing new in them versus main, but I did not run the full pre-commit suite)
• ty passes (not run on this branch)
• Test suite passes: Default suite (not slow): 5261 passed, 52 failed before; 5268 passed, 49 failed after (128 skipped both times). The fixes cleared 3 failures.
• Package build passes when applicable (not run)
• Strict MkDocs build passes when documentation is affected (no documentation changes)
• Ecosystem qualification passes (not run)

The 49 remaining failures are environment limits in my run, not regressions, and were present before: SHAP is not installed (SHAP metric, interaction and end-to-end example tests fail on import shap), and one portfolio test fails inside the installed ml4t-backtest because NYSE rejects session_start_time='17:00'.

Documentation and release

No documentation changes. The behavior changes above may be worth a line in the next patch-release notes.

Amir

[stefan-jansen]

The seven repairs from this bundled PR have landed through focused replacements:

• fix: support factor timestamps and empty signal frames on Polars 2 #99: Polars timestamp parsing and empty signal frames, including the additional ISO datetime regression repair.
• fix(stats): use consistent population moments for DSR #100: consistent population moments for DSR.
• fix(factor): align rolling attribution after missing rows #101: rolling attribution alignment after missing rows.
• fix(distribution): estimate Hill loss tails without gains #102: loss-only Hill tails and one-sided combined-tail handling, including the additional negative-only regression repair.
• fix(config): reject non-mapping documents with ValueError #103: documented ValueError for non-mapping configs.

Thanks @quantum-like (Amir) for the original repairs. Contributor credit is retained in the source commits, focused PR descriptions and squash commits. Each focused head passed all 39 applicable checks against its current main base; its reviewed tree was verified equal to the resulting squash-main tree. All reported fixes are now on main at dc9768c. Package publication is a separate pending step. Closing this bundled PR as superseded by the merged replacements.

AI assistance: the original Cursor Agent authorship is retained in the source history. Codex reproduced the workflows, split and reviewed the changes, added the ISO-datetime and negative-only-tail regressions/repairs, ran verification, and prepared this maintainer closure.

[quantum-like]

Thanks for landing all of these, and for the extra ISO-datetime and negative-only-tail regressions. Much appreciated.