How accurate are these estimates?
The public ledger of our independent restoration estimates, scored against what actually happened.
How estimates are evaluated
Every county outage we track is followed from first report to restoration. While the outage is active we publish an independent, county-wide restoration estimate computed from the live recovery rate (how fast the reported outage count is falling), bounded deceleration, current weather, outage size, and — where enough history exists — how long similar past outages in that area took. When the outage closes, the estimates we published along the way are compared against the actual restoration time.
What counts as accurate
For each completed event we record the mean absolute error (how far our predicted hours-to-restore were from the real hours, averaged across the estimates we published during the event) and the bias (whether we ran optimistic or pessimistic). Smaller error is better; bias near zero means we don't systematically over- or under-promise. An event only counts once it is fully closed — we never grade an estimate against another estimate.
Not every completed outage is scored: when the recovery pattern never stabilized, the site showed "no clear recovery yet" instead of a time. That's a deliberate non-answer, not a prediction — so those events are tracked to completion but have nothing to grade. That's why "scored" can be smaller than "tracked to completion" above.
How restoration completion is determined
A county event is considered restored when its reported outage count falls below a small fraction of its peak and stays there across consecutive data refreshes (hysteresis, so a flickering feed doesn't count as restored). The event archive then stores the real duration alongside the estimates that were published.
Current limitations
- Results shown here are preliminary until a large sample of completed events accumulates; early numbers can swing.
- Estimates and scoring are county-grain — an individual street or circuit can be restored much earlier or later.
- Coverage depends on utilities reporting into the public county feed; missing data is not the same as zero outages.
- Very small events (a handful of customers) are noisy and can dominate early averages.
- Estimator: recovery-rate model with bounded deceleration and an optional historical prior (v2, 2026-07). Scoring is automatic; nothing here is hand-picked.
Estimates are unofficial and independent — not affiliated with any utility. When a utility posts an official restoration time, the site shows both, labeled separately.