McNemar's test for comparing two classifiers

The problem, from first principles

Two classifiers evaluated on the same test set will each get some samples right and wrong. Samples both get right, or both get wrong, say nothing about which is better; only the samples where they disagree matter. McNemar's test asks whether the disagreements lean one way more than chance would explain.

From theory to code

Implement mcnemar_statistic(y_true, pred_a, pred_b) and return the continuity-corrected chi-square statistic computed from the two disagreement counts.

Constraints

• All three inputs are one-dimensional arrays of equal length.
• Let b be the number of samples only A got right and c the number only B got right.
• Return max(|b - c| - 1, 0)**2 / (b + c) as a Python float, or 0.0 when b + c == 0.

Hints

Open one at a time. Each gives away a little more than the last.