kernels.replay_conventional()
Refit fixed candidates, then measure each successive observed diagonal.
Usage
kernels.replay_conventional(
triangle,
candidates,
dates,
*,
loss_field="paid_loss",
premium_field="earned_premium",
on_error="raise"
)Dates must be strictly increasing, one development period apart. Every candidate must declare the same final development horizon. The same cohort, field, units, and previous-origin population are used throughout. on_error='record' retains interval-level failures for later selection to disqualify; it never silently scores a surviving subset of origins.