Metrics 0.3.0 d2tweedie

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.

D2Tweedie

R2's question asked of a deviance: what fraction of it the model explains, against the baseline of predicting one constant. One minus the model's mean TweedieDeviance over the mean deviance of predicting the weighted average of the truth.

1 is a perfect prediction, 0 is one no better than that constant, and negative is worse than doing nothing. Unlike the deviance itself, it is unitless and comparable across models on the same data.

At power 0 it is R2 exactly, since that regime's deviance is the squared error and its constant baseline is the mean. Measured on the worked case, both give 0.65.

A truth that never varies is refused here and answered elsewhere

If every yTrue is the same value, the constant baseline is already perfect, the denominator is zero, and there is no deviance to explain. This throws UndefinedMetricException; D2AbsoluteError and D2Pinball answer 0 on the same input.

That split is the reference's, not this library's: d2_absolute_error_score masks the zero denominator and returns 0, while d2_tweedie_score divides by it and raises ZeroDivisionError. Reproducing the asymmetry copies a divergence rather than inventing one.

Multioutput is not offered, because the reference refuses it — d2_tweedie_score on a 2-D input raises "Multioutput not supported in d2_tweedie_score". The two pinball scores do take it.

Members

Member	What it does
D2Tweedie.Score	The fraction of Tweedie deviance the model explains.