[PatrickHenkel1]

Final review of 1.0.0 before the merge into main.

Merge of main (0.3.0, #86)

main is merged into this branch, so the PR refactoring_v1.0.0 → main has no conflicts. All conflicts are resolved in favour of 1.0: the old workflows, requirements.txt and .github/copilot-instructions.md stay removed, and pyproject.toml keeps the static version and the bounded dependencies. These are at or above the security fixes of 0.3.0 (constraints.txt checked against the pins of main). The tree after the merge is identical to refactoring_v1.0.0.

Fixes

• Leakage in the rollout: lags (Q_lag1) and changes (Change(Q)) of derived columns that depend on the recursive feature were taken from the data, so the model saw the measured T over the horizon. This affected the horizon metrics, RecursiveTraining and the feature selection criterion. They are now written in the predicted history (Q_lag1 → expression of Q one step back); lagged raw columns are the data shifted by their lag.
• shift=0: Change(Q) with Q derived from the output, and nested derived columns, are rejected (holds_current_value).
• CMNN / PINN: without monotonicities, activation_split was [1, 0, 0]. Only convex neurons plus the non-negative weights of the following layers made the prediction convex in all inputs (affected PINNModel by default, e.g. example 08). The default is now [1, 1, 1].
• Split settings (test_size, val_size, block_length, gap_steps) passed next to a BlockSplit raise a ValueError instead of being ignored.
• The output rescaling keeps the scale 1 for a constant target (it was 0).

Methodology

• ModelComparison.paired: the confidence interval and the p-value now come from the corrected resampled t-test (Nadeau & Bengio 2003): var(d) · (1/n + n_test/n_train), with the ratio taken from the split summary of the runs (new column resampling). There is no correction if all runs share one split (vary=('init',)). This replaces the plain t interval and the Wilcoxon test, which assume independent runs. A note on multiple testing is added to the docs.
• The embargo of multi step data no longer adds the warmup, which lies inside the split of its sequence.
• Feature selection: the docs and the docstring note that the score of the selected set is optimistic (validation data).

Documentation

• CHANGELOG: 1.0.0 is marked unreleased (the checklist says to set the date before tagging), new entry for 0.3.0, wording of requirements.txt, entries for the fixes above.
• README: the constrained regression is guaranteed on the box of the training data or on bounds, with a certificate per constraint (not "certified on the whole operating range").
• No runtime claims ("about 1 s for 1e5 samples").
• branch_review.md: branches already deleted are no longer listed.
• Release checklist: merge main first.

Checks

• pytest: 266 passed, 1 skipped; ruff: clean; smoke run of all 13 examples: OK; sphinx-build -W: clean.
• New regression tests: rollout with Q_lag1 / Change(Q) against a recursion by hand, shift=0 cases, split settings next to a BlockSplit, CMNN neurons, output scale, multi step embargo, corrected t-test.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.