compare

SuperLSS.compare(

other: SuperLSS | DenseDistributionalModel,

X: object | EagerFrame,

y: NDArray,

*,

which: Literal['log', 'crps'] = 'log',

by: str | Sequence[Any] | NDArray | None = None,

sample_weight: NDArray | None = None,

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

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

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

**kwargs: Any,

) → Comparison

Return the paired score difference against another fitted candidate.

The comparison is paired row by row, so its standard error is that of the mean difference and not of two independent means; by splits it into segments, and murphy_quantile adds the Murphy diagram of Ehm, Gneiting, Jordan and Krüger (2016), which shows at which thresholds one candidate’s advantage actually comes from. sample_weight follows the candidates’ declared likelihood-weight semantics; incompatible non-unit semantics are refused. A negative mean difference favours this model. Use a_offsets and b_offsets for different predictor offsets; offsets is shared shorthand and cannot be mixed with either.