How often do utilities push back their restoration times?
A living analysis, recomputed continuously as outages complete. The figures below are the latest verified aggregates — the "data updated" stamp under them is the dataset's own time. Methodology last reviewed 2026-09-26 (v2026-09-26.v3).
By state
The same ledger, grouped the way the rest of the site is — by where the outage happened. Only states with at least 5 graded utility-posted events show; smaller samples add too much noise to compare fairly.
| State | Events | Slip % | Avg revisions | Utility err | Ours err |
|---|---|---|---|---|---|
| New York | 24 | 52% of 21 | 1.4× | ~1.4 h | ~1.5 h |
| California | 9 | 100% of 6 | 2.8× | ~2.5 h | ~4 h |
| Massachusetts | 9 | 56% of 9 | 5.2× | ~1.3 h | ~1.4 h |
| Michigan | 7 | 43% of 7 | 1.0× | ~3.7 h | ~0.8 h |
| Utah | 5 | 80% of 5 | 5.4× | ~3.3 h | ~1.1 h |
By utility (named)
Per-utility numbers publish only when a named utility has at least 10 graded events in the current window — a pre-registered threshold, fixed here in advance. Utilities close to the threshold are shown below without names, so the sample is honest.
| Utility | Events | Slip % | Avg revisions | Utility err | Ours err |
|---|---|---|---|---|---|
| National Grid | 22 | 53% of 19 | 3.3× | ~1.6 h | ~1.4 h |
What we found
This report grades the restoration times that reach us through ODIN, the U.S. Department of Energy's free county outage data — not the times a utility posts on its own outage map, which this dataset never sees. Read every figure below as “of the posted times we recorded”. When a power outage starts, the utility usually posts an estimated time of restoration (ETR) on its outage map. Customers plan around it: whether to buy ice, book a hotel, move a medical device. This report tracks a simple question — once posted, does that time hold? The live figures above say how often, across the completed outages in the current window whose posted time we could follow, that time moved later at least once before the lights actually came back, and how many times it did.
Correction, 2026-09-26 (methodology v3). A county's posted time is the earliest of the times posted for its outages. Until this date, every move of that earliest time by more than 30 minutes counted as a revision, and every later one as a pushback — including the time passing from one outage to the next when the earlier outage was restored, which is nobody revising anything. That overstated both figures: on 2026-09-26 the report read 82% of outages with the time moved later, and 5.7 later revisions on average. From v3 a pushback is counted only when an outage whose time the county showed, still out, moves its own time later. The figures rebuild on that rule from outages followed from their start, and each shows how many outages it covers; until enough have completed, the tiles show a dash.
Two honest framings belong next to these figures. First, a revision is not deception — early ETRs are posted before crews have assessed damage, and utilities update them as reality comes in; a moving estimate can still be a good-faith one. Second, we grade our own independent estimates by the same ruler, on the same events, and publish both numbers above. Which of the two is the closer one more often has changed from one window to the next, so the live figures, not this paragraph, say which is ahead right now. The practical takeaway for customers is not "distrust the utility"; it is that the first posted time is an early draft, and planning should carry a margin.
Methodology
Ground truth. We track each county outage in our data from first report to restoration (the data sources are ODIN, the U.S. Department of Energy's county outage data, plus ~540 utilities' own public outage maps; not every utility reports into either). A posted time is graded against when the county's count next reached 10% of the outage's peak or less (the "90% restored" mark) — the same yardstick our own estimates are graded against publicly. The event ends once its count has stayed at or below the restored level (5% of the peak, never below 20 customers) for 3 hours; the start of that stretch is the county's restoration moment.
What counts as a pushback. While an outage is active we sample the utility's posted ETR on every collection pass. A county's posted time is the earliest of the times posted for its outages, so we also note which outage it belongs to, and from then on follow that outage until it is restored. A pushback is that outage moving its own posted time later by more than 30 minutes while it is still out, measured from the time it had when we began following it or at its last counted pushback, so small steps add up; a time withdrawn and posted again later counts too.
What does not. Only moves before the county's restoration moment count: when its count fell to the restored level for the last time, at the start of the 3 hours that end the event. (Errors are graded at the 90% mark, above.) The county's time moving from one outage to another, when the earlier outage is restored or a new outage posts, is not a pushback. Moving earlier, or posting no ETR at all, is not counted against the utility.
Error. For each completed event, every posted ETR is compared against the actual restoration: the median absolute error in hours is the event's score. Our own county-wide estimate, re-run at each of the same event's readings after its peak where its time was not withheld, is scored identically — same events, same ruler.
Scope and limits. The sample is a rolling window, not a running total. The published record keeps the 600 most recently graded completed county-level outages, and the figures above come from those among them where the serving utility posted at least one ETR; the first two only from those whose posted time we could follow outage by outage, and the sample line says how many. Each newly graded outage pushes the oldest one out, so the figures describe the last few days of outages and move with whatever storms those days held; the live count and the number of days it covers are shown above.
Coverage and thresholds. Not every utility reports into these sources, and the level of detail varies. Named per-utility reliability tables publish only when a utility has at least 10 graded events in the current window — a pre-registered threshold, fixed here in advance. State breakdowns use a lower threshold (n≥5) because they aggregate across utilities.
How to cite this analysis
Cite as “When's the Power Back? — utility restoration-time reliability (living report)”, with the
date you accessed it, the methodology version shown above the tiles, and a link to this page. The
aggregate figures — sample sizes, headline rates, and the state/utility tables — are also published as machine-readable
JSON at accuracy-summary.json (field definitions ride in its
metrics section; the license block states these same terms). Quoting and citing the published
aggregates with attribution is welcome.
This report publishes aggregate findings, methodology, sample sizes, and update dates. Event-level records and programmatic bulk access are not part of the public report — they are reserved for a possible future API. Contact us if you are interested in future API access (registering interest promises nothing about availability, pricing, or service levels). Questions or corrections: the contact page.
