SHEF 2.0 · day-ahead load forecast monitor
Every day at 10:00 Central, tomorrow's hourly electricity load for ERCOT's South Central weather zone (Austin, San Antonio) is forecast using only data available at that moment, and checked against what actually happened. Rows marked backtest are the Aug 2025 – Oct 2026 holdout, scored once on archived weather forecasts; live rows are real forecasts.
Latest forecast
Forecast vs actual, last 14 days
Live, hourly. Fallback hours (baseline served because the weather forecast was incomplete) are listed in the table.
Daily MAPE: backtest, then live
Mean absolute percentage error per Central day.
Error by lead time
Hours from the 10:00 issue to the target hour (14 = midnight, 37 = 23:00).
Against ERCOT's own forecast
ERCOT's South Central forecast as published by 09:30 CT on D-1, the same information cut-off as ours. Matched hours only.
Benchmark summary
Same hours, same actuals. "Ours better" is the share of hours where our error was smaller.
Weather forecast the model was served
Temperature, population-weighted over 4 cities, on each target day. vs ERA5: against what was observed.
vs training proxy: against the archived forecast the model was trained and backtested on.
Input drift
PSI of the last 14 live days against the same month in the 3 training years before the model's cut-off.
Thresholds are calibrated on 1,000+ historical windows: watch above their 95th percentile, drift above their 99th.
Model
What is serving, and how it got there.
Promotion log
Monthly: a retrained challenger replaces the champion only if it beats it by ≥ 0.10 MAPE points on days neither model trained on, without a worse Peak-MAPE.
Load publishing latency
Every feature cuts load off 1 hour before issue time, assuming GridStatus publishes each hour within 60 minutes.
Each daily load fetch tests that for free: the newest hour it sees was published within the first figure,
and the following hour was still missing after the second. A second figure ≥ 60 min would break the assumption.