Metrics 0.3.0 d2absoluteerror

Lodestar.Metrics 0.3.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

D2AbsoluteError

R2's question asked of the absolute error: what fraction of it the model explains against the baseline of always predicting the weighted median of the truth. D2Pinball at alpha = 0.5.

Where R2 compares against the mean and is pulled by an outlier, this compares against the median and is not. That is the whole reason to prefer it: one wild value in the truth inflates R2's baseline and flatters every model measured against it, while the median baseline barely moves. On a truth of [5, 5, 5, 1] against a prediction of [1, 2, 3, 4] this scores -2 — the model is worse than the flat guess, and says so.

A truth that never varies scores 0 rather than raising, unlike D2Tweedie; below two samples the answer is nan, which zeroDivision can change.

Members

Member	What it does
D2AbsoluteError.Score	The explained fraction, one number for the whole prediction.
D2AbsoluteError.PerOutput	The same, one number per output column.