When's the Power Back?

How accurate are these estimates?

Our independent, county-wide restoration estimates, graded against what actually happened. Unofficial — always follow your utility's guidance.

At a glance

Across 600 scored outages since Nov 14, 2026, median absolute error 1.6 h (57% within 2 h). For 64 time-matched cases, our median error was 1.7 h versus the utility's 1.5 h; ours was lower in 53% of them.

23,946
outages scored all-time
of 48,384 tracked (49%)
1.6 h
median absolute error
across the 600 most recent scored
2.3 h
error at first graded reading
the earliest re-run after an outage's peak (n=600)
0.9 h
error at last graded reading
the last re-run before 90% were back
1.5 h
utility's own estimate
median error at the same readings (n=64); ours: 1.7 h

Every error figure is a median absolute error across outages — each outage is scored by the times graded during it (the utility's posted times as recorded; our county-wide estimate re-run at each reading after the outage's peak where the time was not withheld), and half of outages scored at or under the figure shown. The utility comparison is time-matched — both graded on the same readings against the same restoration.

How close: of the 600 most recent scored outages, 38% landed within 1 hour of when 90% of customers were back, 57% within 2 hours, 72% within 4. Typical direction of the miss: 0.1 h late — under one reading's interval, so no clear lean either way.

Utilities pushed their posted time back at least once in 44% of the 61 outages whose posted time we could follow outage by outage (median 3× when they did).

Median error by time into the outage: 2.2 h at 1 h in (n=535) · 2.2 h at 3 h in (n=365) · 1.9 h at 6 h in (n=207). Each figure is a different set of outages — only those that lasted at least that long — so read them as separate cohorts, not one outage sharpening over time (a survivor effect: longer outages dominate the later horizons).

Median error by outage size: 1.5 h for under 500 customers out (n=257) · 1.7 h for 500–2,000 customers out (n=245) · 1.7 h for 2,000–10,000 customers out (n=91) · 1.6 h for 10,000–50,000 customers out (n=7).

Data updated Nov 19, 12:45 AM EST · methodology 2026-10-26.v5 · new events land ~3 h after restoration holds. Every statistic above except the all-time count is computed over the 600 most recently scored outages (about 4.7 days of outages) — the window the published ledger retains — not all 23,946. Each newly scored outage pushes the oldest one out, so these figures move with whatever storms those days held, and one large storm can dominate them. The gap between tracked and scored is outages with no gradable reading: fewer than 6 readings, the 90%-restored mark never reached in the record, or no reading after the peak where the estimator gave a time (a withheld or holding pace). Those are not scored.

Related: how often utilities push back their posted restoration times — same ledger, utilities graded.

Method

While an outage is active, the model watches how many customers are out and how fast the number is falling, projects the rest of the recovery forward with the pace easing off as it drags on, since crews clear the easy faults first — and leans on weather and past outages in that area. Recalculated ~every 15 min; independent of the utility. When a utility also posts a time, both are shown, labeled separately.

How we grade

  • Typical error — each outage is scored by the median hours off across the times graded during it (the utility's posted times as recorded; our county-wide estimate re-run at each reading after the outage's peak where the time was not withheld, as the methodology sets out), and the figure shown is the median of those scores. ±3 h means half of outages scored within 3 h of truth. Smaller is better.
  • "Restored" (what every grade is measured against) — the "90% restored" mark: the county's reported outage count at 10% of its peak or less. Each graded time (ours and the utility's) is measured to the next reading at that mark — "time until most reported outages are restored", not the last single customer. Until the count first reaches the mark, that is the same moment for every reading; a reading taken after the count rose back above it is measured to the next time it fell to it.
  • Bias — near zero means we don't systematically over- or under-promise.
  • First vs final — the earliest graded re-run vs the last one before 90% of customers were back; shows whether estimates sharpen.
  • Effective restoration — the target we grade against: the next reading at or below 10% of the peak (~90% recovered), not the literal last customer. The time on our live label runs until the count is gone ("to restore"); the graded re-run is aimed at this mark instead, so what is scored is our live pace's projection to the mark, not the label's own time — stated here so the numbers can't quietly flatter us.

Only outages that have fully ended are graded, and we never score an estimate against another estimate. What is graded is a re-run of our estimator after the outage ended, at each reading after its peak, not each time exactly as the site showed it. The re-run measures its pace from the outage's own first reading on, while the live page looks back about 2.5 hours and can count the readings from just before the outage began, so in an outage's first 2.5 hours the page can have shown the count still climbing, and no time, at a reading the re-run grades. An outage is not scored when it has fewer than 6 readings, never reaches the 90%-restored mark in the record, or has no reading after its peak where the estimator gave a time (that is why "scored" is smaller than "tracked").

How the error moved across the window

Error here is normalized by each outage's length — a 5-hour miss on a two-day outage isn't the same as on a five-hour one. Raw hours-of-error mostly tracks how severe that week's storms were, so this is the cleaner way to watch the estimator itself improve; the typical outage column shows whether a cohort was simply harder. Outages that lasted under an hour are left out of this table — dividing a miss by a very short outage would turn minutes into hundreds of percent and swamp the trend.

Prediction error by cohort of close date, normalized by outage length.
Cohort (by close date)OutagesError ÷ lengthMedian errorTypical outage
Earliest19225%±1.7 h~6 h
Middle19230%±1.6 h~5 h
Latest19427%±1.5 h~7 h

Roughly flat across the window. The window holds only the 600 most recently scored outages — about 4.7 days of them — so this compares the storms of those days, not the estimator over time; it cannot show a long-term trend.

What each number measures

What each number measures — the published definitions for this methodology version (2026-10-26.v5), quoted from the metric contract itself.
our median absolute error (whole event)
Median absolute error, in hours, of our county-wide estimate re-run at each graded reading of a completed outage (each reading after its peak where the time was not withheld; the live pace alone), against its observed effective restoration (the next reading at or below 10% of peak: the 90%-restored mark). Median across events.
posted-time median absolute error (as recorded)
Median absolute error, in hours, of the posted restoration times we recorded — a restoration time carried by the national county feed (the only posted-time source the event series records) — over the readings where one was present while meaningfully out, against the same observed restoration. Median across those events. Events where the utility posted a time that never reached that feed are NOT counted.
our median absolute error on the same posted-time events
independentWholeEventMae restricted to the events counted by utilityPostedMae — same event population, whole-event alignment (not the same readings). It inherits that metric's posted-time source limitation exactly.
time-matched head-to-head (same readings)
Both graded at the identical readings where a recorded posted time — a restoration time carried by the national county feed (the only posted-time source the event series records) — AND our estimator both produced a value: the strictest like-for-like. Reported as a pair (posted vs ours) with its own n.
posted time moved later at least once
Share of events carrying a recorded posted time — a restoration time carried by the national county feed (the only posted-time source the event series records) — where an outage whose time had been the county's posted time then moved its own time later by >30 minutes at least once before the county's restoration (the start of its last stretch at or below the restored level, which ends the outage), while still out, measured from the time it had when first followed or at its last counted move, so small steps add up (a time withdrawn and posted again counts). A county's posted time is the earliest of its outages' times, so it also moves when the outage holding it is restored or another outage posts; those moves are NOT counted, and neither are earlier moves. The denominator is the events followed outage by outage from their start; events whose posted time never reached that feed, or whose moves could not be traced to one outage, are not in it.
average later revisions per posted-time event
Mean count, per event carrying a recorded posted time — a restoration time carried by the national county feed (the only posted-time source the event series records) — of >30-minute later moves an outage made to its own time after that time had been the county's posted time, counting 0 when none did, over the same events as utilityPushbackRate (moves measured the same way). Moves of the county's time from one outage to another are NOT counted.

Recent examples from the scored sample

The newest 10 scored events (most recent first — not a curated or representative sample); the aggregate statistics above cover the whole window. Records of individual outages and downloadable datasets aren't part of the public site; contact us about future API access.

The newest scored outage events, with our median prediction error and the utility's where one was posted.
CountyClosedPeak outRestore tookOur errorUtility's error
Marin, CA2026-11-192,10711.5 h±2.1 h—
San Mateo, CA2026-11-197,66811 h±0.4 h—
Hinds, MS2026-11-1961412 h±1.6 h—
Calaveras, CA2026-11-191,9009.2 h±3 h—
Douglas, MN2026-11-192106.7 h±3.3 h—
Division No. 5, AB2026-11-1955022.5 h±5.5 h—
Sangamon, IL2026-11-198131.2 h±0.2 h—
Palm Beach, FL2026-11-191,5380.2 h±0.5 h—
Lee, FL2026-11-192,25714.5 h±1.4 h—
Milwaukee, WI2026-11-189077.5 h±3.6 h—

Newest 10 of 600 scored. Error = median difference between predicted and actual hours-to-restore across the times graded during that outage (the utility's posted times as recorded; our county-wide estimate re-run at each reading after the outage's peak where the time was not withheld). "—" in the utility column means no official restoration time was posted.