[CBravoR]

Summary

This pull request fixes how EMP-PY 2.0.4 fails on degenerate or invalid input, and it prepares version 2.0.5. Results on valid input do not change. It also adds a workflow that runs the tests on pushes and pull requests, and a workflow that publishes to PyPI with Trusted Publishing.

Changes

EMP/metrics.py:

• Replace scipy.spatial.ConvexHull with a monotone-chain upper hull from (0,0) to (1,1). Qhull raised QhullError when the ROC curve had fewer than three points or all its points on one line, for example with constant scores or empCreditScoring([0.3, 0.3, 0.7, 0.7], [0, 1, 0, 1]). The hull uses the integer counts of false and true positives, so the test for collinear points is exact. empChurn uses the same helper.
• Raise ValueError instead of calling sys.exit, which ended the Python process.
• Raise a clear ValueError when the labels contain only one class. Version 2.0.4 failed with ValueError: Points cannot contain NaN.
• Drop the check that scores lie in [0,1]. EMP depends only on the order of the scores, so logits and scorecard points work. Scores that are NaN, infinite or not numeric raise ValueError.
• Raise ValueError for input that is not one-dimensional.
• Suppress the divide-by-zero RuntimeWarning from a horizontal last hull segment. Its lambda (or gamma) is infinite, and the existing < 1 filter drops it.
• Replace the Python loops in the input checks with numpy operations.
• Explain in the comments of empCreditScoring how EMPC_fraction is chosen when two cutoffs tie at lambda = 1.

Other files:

• README.md: the functions return the output and do not print it by default. The README now states the label and score conventions.
• tests/: 54 pytest tests. tests/brute_force.py computes EMP from its definition on a grid, without the hull.
• setup.py, README.md, NEWS: version 2.0.5 and a changelog.
• setup.py: require Python 3.10 or later (was 3.8). Give the license as the SPDX expression GPL-3.0-or-later instead of the deprecated classifier, list pandas once, and drop setup_requires.
• .github/workflows/tests.yml: runs the tests on each push to main and on each pull request. The matrix is Python 3.10 to 3.15 on Linux, Python 3.15 on Windows and macOS, and numpy 1.x on Python 3.12.
• .github/workflows/publish.yml: a published GitHub release runs the tests (through tests.yml), builds the package, checks that the tag matches the version, and uploads to PyPI after a reviewer approves the pypi environment.

Results on valid input

The PyPI 2.0.4 wheel and this branch ran the same 6,030 calls: 1,005 data sets with 6 parameter settings. Both used numpy 2.5.4, scipy 1.18.1, scikit-learn 1.9.1 and pandas 3.1.0 on Python 3.14.6.

	Count
Calls that 2.0.4 computes	6,018
Of those, equal bit for bit in this branch	6,018
Calls that fail in 2.0.4 (QhullError) and work in this branch	12
Calls that work in 2.0.4 and fail in this branch	0
RuntimeWarning: divide by zero, 2.0.4 / this branch	2,598 / 0

Against brute force on 4,850 calls, EMPC and EMP agree within 2.3e-7, which is the grid error.

Behavior changes

• Invalid input raises ValueError instead of SystemExit.
• Scores outside [0,1] are accepted.
• Input that is not a vector raises ValueError. Version 2.0.4 flattened nested lists of any shape, so [[0.1, 0.2], [0.3, 0.4]] was read as four scores.
• A single-column DataFrame works. In 2.0.4 it raised KeyError.

Known limitation, not changed

Sometimes two ROC points tie exactly for the maximum profit at lambda = 1. Then the fraction at the p1 point mass is not unique. Version 2.0.4 and this branch both choose by how lambda rounds near 1. In 15 small-sample cases with ROI = 0.5, the fraction differs from brute force by up to 0.06. EMPC is not affected. This pull request keeps the 2.0.4 behavior and documents it in the comments of empCreditScoring. In all 165 test calls with an exact tie, the choice follows the documented rule.

Tests

pip install pytest
python -m pytest

All 54 tests pass on this branch and on the built wheel in a clean environment. On 2.0.4, 19 fail, which are the edge cases and the new behavior. The Tests workflow runs them on this pull request.

Before the first release with the workflow

1. On PyPI, open the EMP-PY project, then Manage > Publishing. Add a GitHub publisher with owner Banking-Analytics-Lab, repository EMP-Py, workflow publish.yml and environment pypi.
2. On GitHub, open Settings > Environments and create the environment pypi with at least one required reviewer.
3. After the merge, create a GitHub release from main with the tag v2.0.5.

Not in this pull request

• metrics.py still imports pandas but does not use it, so pandas stays a requirement.
• DESCRIPTION, man/ and inst/CITATION are files from the R package.