[antonsoo]

Continuous verification scores count finite forecast–observation pairs, but currently compute each variable's mean and variance using its own non-missing values. Unpaired observations can therefore change the correlation and regression scores:

import numpy as np
from pysteps.verification import det_cont_fct

pred = np.array([1., 2., 3., np.nan])
obs = np.array([1., 2., 3., 100.])
print(det_cont_fct(pred, obs, scores="corr_p"))
# currently {'corr_p': 0.019238330156554226}; expected {'corr_p': 1.0}

This masks both arrays with their joint finite mask before computing the moments. Counts, means, variances and covariance then describe the same samples. Tests cover missing values in either input, conditioning, axis reductions, streaming updates and input preservation.

I checked six lead times from each of the project's Finnish and Swiss radar examples. One Swiss forecast currently reports correlation 1.015110; the fix returns 0.853784, matching np.corrcoef on the valid pairs. All 48 failing direct/streaming moment checks now pass, while the finite-data and double-conditioned controls remain unchanged.

Validation: 813 tests pass in the full suite; both plugin tests pass when rerun after installing the missing pip dependency. One optional GDAL module is skipped. The focused tests and radar checks also pass on Python 3.11 / NumPy 1.26, and Black passes.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	🔄 Running since 2026-10-24T03:04:29.141827Z	e1e9423	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.