Kendall Tau

Kendall Tau is a way to describe how similarly two things are ordered. If you rank houses by price and also rank them by size, Kendall Tau tells you whether those two rankings tend to agree.

What it measures

Kendall Tau (often written as τ) is a rank correlation: it measures association based on the relative ordering of values, not their exact numeric distances. It looks at all pairs of observations and counts how often the two variables “move together” in rank.

• A pair is concordant if the order matches (if A > B on X, then A > B on Y).
• A pair is discordant if the order disagrees.

In the simplest case (no ties), τ = (C − D) / (C + D), where C is the number of concordant pairs and D is the number of discordant pairs. Values range from −1 (perfectly reversed ranking) to +1 (perfect agreement), with 0 meaning no consistent ordering relationship.

Why it’s useful (especially with messy data)

Because it uses ranks, Kendall Tau is more robust when:

• Relationships are monotonic but not linear (consistently increasing, but not at a constant rate).
• Data have outliers (one extreme sale price won’t dominate the result).
• You care about ordering more than exact values (recommendation rankings, search results).

Where it shows up in AI/ML

Kendall Tau is common in learning-to-rank and evaluation of ranking models: compare predicted rankings vs. true rankings (e.g., which products a user prefers). It’s also used in feature analysis when variables are ordinal (survey ratings, risk categories). In practice you’ll see it in SciPy as scipy.stats.kendalltau, with variants (like τ-b) that handle ties.