Read this first: everything on this page is a statistical best guess from public data — a free heads-up, not professional advice. The model has been wrong before (we keep score below) and it will be wrong again. Check it against your own curtailment economics before acting on it.
Checking today's odds of setting this month's 4CP peak…
Every June–September since 2015 has exactly one 4CP day. Each square below is one month, colored by where this model ranked the real peak day that very morning, scored leave-one-year-out (the model never saw the year it was graded on).
Curtailment costs an afternoon; missing a peak costs a year of transmission charges. The trade-off, measured across all 43 months:
The 4CP question isn't "how hot is today?" — it's "will today beat every other day this month?" ERCOT already forecasts the load. The hard part is the rest of the month, so this model races today against it: it predicts today's peak from temperature, Gulf humidity, the last three days' heat, weekends and holidays — then simulates the remaining days of the month a few thousand times from ten years of Texas climatology, and reports how often today wins.
It ranks the day against the month, not just the thermometer. That's the whole trick.
Trained on 1,310 summer days, 2015–2025 (EIA hourly demand + archived weather). Two-stage cadence: a morning call from forecasts, then a 2 PM verdict recomputed from ERCOT's actual midday load — which can upgrade the day or publicly stand the morning call down. The month ledger updates every 10 minutes from ERCOT's live feed.
Version 1 was deliberately weather-blind about the future: it raced today against pure climatology. The obvious upgrade — let it read the 7-day forecast — turned out to be a story worth telling straight.
First attempt: failure. Fed four years of archived forecasts, the model got worse. The autopsy found the forecast archive running a systematic +4.5°F warm bias against the reanalysis data the model was trained on — at ~260 MW per degree, every tomorrow looked like a phantom scorcher and real peaks got buried. Standard meteorology (a per-lead bias correction) fixed the inputs; the rematch was rerun honestly, graded only on 2022–2025 — the four hardest years on record.
The verdict: an ensemble. The climatology model and the forecast-aware model fail in different months — climatology lets scorchers stand out; forecasts correctly humble them when a hotter week looms. Their geometric mean, which only scores a day high when both agree, beat each parent where it counts:
| Model (2022–25, 16 peaks) | Alert days / mo | Peaks caught |
|---|---|---|
| v1 climatology | ~3.5 | 9 of 16 |
| v2 forecast-aware | 4.4 | 13 of 16 |
| v2.0 ensemble (live now) | 2.6–3.6 | 11–12 of 16 |
The ensemble also arrived better calibrated: its HIGH days claimed 39% and won 28%, versus v1's 43%-claimed, 20%-won. Demanding agreement makes confidence honest.
And one discovery we didn't go looking for. A midday-load model trained on the pre-2022 grid systematically over-predicts the evening peak on true 4CP days — by roughly 900 MW on 14 of the 16 peak days. That missing load is the fingerprint of large flexible consumers — crypto miners, data centers — curtailing into the very peaks everyone forecasts. To our knowledge nobody publishes a direct measurement of this; consider it measured. It is also why no honest model of this market — ours or anyone's — gets to claim certainty: the act of predicting the peak now moves the peak.
The raw model used to claim ~43% on its confident days; reality paid ~20%. Two honest reasons: Texas heat is fatter-tailed than any simulation, and since ~2020 large flexible loads — crypto miners and now data centers — deliberately shut down on predicted peak days, shaving the very peaks everyone forecasts. So every probability shown here is calibrated down to what days like it have actually paid, and capped at 65%. Nobody gets to be sure in a game with adversaries.
The misses we can't fix: July 2024 — Hurricane Beryl deleted Houston's load and the month peaked on July 1st (we had it at ~0%). June 2025 — a quiet dry-heat day won a month the model read completely wrong (ranked it 15th). Hurricanes and freak verdicts stay unforecastable; a model that claims otherwise is lying to you.
And the benchmark, since you should ask: paid forecast desks with analyst teams and same-day cancel calls have published records around 90%+ peak capture at ~2.5 alert days a month (Amperon, 2017–2022). This model isn't there — it's a free morning call that grades itself with no knowledge of the weather forecast, by design, to establish an honest baseline. The forecast-aware rematch has now been run — results below, losses included, as promised.
On ELEVATED and HIGH days the model speaks once — a morning email and a post — and otherwise stays quiet, so an alert from us means something. Sign up free →
4CP follows grid demand, not price. The four peak intervals are only confirmed after each month ends. This is a heads-up, never a promise — on the order of $50+ per kW-year rides on these four 15-minute intervals, so verify against your own curtailment economics.
No ads · no tracking · no commissions · gridglancer.com