kernels.point_summary()
Point errors of a PredictiveDistribution’s draw means against outcomes.
Usage
kernels.point_summary(
pred,
observed,
)errors has one row per target: the target metadata plus estimate (the draw mean), outcome, error and pct_error (missing where the outcome is missing or zero).
metrics is point_metrics() over the scorable targets - those whose label is not "total" (the total is the sum of the others and would count every error twice) and whose outcome is a finite number. It is None when the cohort cannot be scored at all, which happens three ways: no scorable target, a scorable target whose draw mean is not finite (a fit with missing draws; scoring the rest would put this cohort on fewer targets than its neighbours), and scorable outcomes summing to zero or less (pool_ape divides by that sum). excluded counts each reason, so a caller reading None can tell which one it was without re-deriving it.