scores

SuperLSS.scores(

X: object | EagerFrame,

y: NDArray,

*,

which: Sequence[str] = ('log', 'crps'),

thresholds: Sequence[float] = (),

sample_weight: NDArray | None = None,

offsets: Mapping[str, NDArray] | None = None,

**kwargs: Any,

) → DataFrame

Return one column of proper scores per requested rule, one row per row.

The log score and the continuous ranked probability score of Gneiting and Raftery (2007) are both proper, so a lower mean is evidence for the model that produced it; thresholds adds one threshold-weighted CRPS column per value, which scores the tail above that point alone. Every column reads sample_weight under the fitted model’s contract: prior weights select the row’s weighted predictive law, while frequency weights return the compressed contribution of the repeated unit law.