Resolved wildfires — how good was the burn estimate?

Read the horizon before the error. Every fire graded here ended within about 20 days of the forecast being made, so these numbers say nothing about the 14-to-21 day range. A median error under a day is a statement about short-horizon calls only.

582 of 3,509 fire-days were discarded. NIFC publishes in batches, so a fire is recorded as gone at the next pull that omits it, not the day it actually stopped. Where that window is wider than a day and a half the end date is not known well enough to grade against, and the forecast is dropped rather than scored against a guess. Grading everything instead would report a -5.6 day bias that is an artifact of the publishing schedule, not the model.