thetrading.tools

Forecast Accuracy Tracker

On July 24, 2026, the grid's own day-ahead demand forecast missed actual lower-48 demand by an average of 1.3% across the day's 24 hours, under-forecasting by 0.5% on balance. At the day's peak hour the miss was +0.8%. Over the past 12 months the typical daily miss has been 1.4%.

Latest daily miss
1.3%
avg absolute hourly error
Latest bias
-0.5%
under-forecast
12-month avg miss
1.4%
365 complete days
Worst day, past year
6.3%
April 16, 2026

Daily forecast miss

Each point is one day's average absolute hourly error; the green line is the 30-day average. Structural improvement or degradation shows up in the smooth line, weather busts in the spikes.

Jul 24, 2021Jul 24, 2026
0.5%3.4%6.3%202220232024202520261.3%1.3%
Daily miss30-day average
Daily mean absolute day-ahead forecast error for lower-48 demand, January 1, 2016 – July 24, 2026. Matched-fleet comparison of EIA-930 balancing-authority forecasts vs adjusted actual demand.

Over- or under-forecasting?

Signed 30-day average bias. Positive means balancing authorities collectively predicted more demand than materialized — the conservative direction for grid operations. Across the full history, 26% of days came in over-forecast.

Jul 24, 2021Jul 24, 2026
-1.9%-0.7%0.4%20222023202420252026-0.8%
30-day average signed forecast error. Above zero = over-forecast (demand came in below prediction); below zero = under-forecast.

Seasonal norms

Average miss by calendar month across the full history — the base rate to judge any single day against. The seasonal spread is modest; what stands out more is the slow drift in the 30-day line above.

MonthAvg daily missAvg biasDays tracked
January1.34%-0.68%341
February1.33%-0.50%311
March1.24%-0.30%341
April1.27%-0.49%330
May1.26%-0.64%341
June1.42%-0.62%330
July1.35%-0.45%334
August1.27%-0.36%310
September1.27%-0.37%300
October1.15%-0.64%310
November1.27%-0.77%300
December1.42%-0.75%310

Why traders and investors watch this

Forecast error is where physical volatility comes from: day-ahead power markets commit supply against these forecasts, and when actual demand comes in above them, the difference clears at whatever price scarcity demands. Under-forecast days are the raw material of real-time price spikes. The subtler signal is the bias — a persistent under-forecast lean means load keeps exceeding the models, which is the earliest statistical footprint of structural growth the forecasters haven't caught up to (unmodeled data-center energization looks exactly like this). It is also the honesty layer for the rest of the Energy section: it measures how predictable the system behind all these gauges actually is. No forward-return claim is made — this is measurement, not signal.

How this is computed—and what it cannot say

Every balancing authority files an hourly day-ahead demand forecast to EIA-930 alongside its actual demand. We sum forecasts and actuals over the same set of authorities each hour (a matched fleet), so authorities without a usable forecast never skew the comparison. A fixed rule excludes authorities whose forecast errors exceed 50% on average within a half-year file — those are broken data feeds, not forecasts; the excluded set is around 1% of national demand. The daily miss is the mean absolute hourly error over 24 complete hours; single-hour reporting glitches are interpolated with the same guard used on the demand page. History starts January 2016: EIA-930 collection launched in July 2015, and its first six months show roughly ten times the steady-state error — filing teething, not forecast skill.

This measures the accuracy of the operators' own filed forecasts in aggregate. It cannot rank individual utilities' forecasting skill, separate weather surprise from model error, or say anything about price or reliability outcomes. There is no "adjusted" forecast series, so forecasts are compared as filed, and EIA's ~30-day revision window applies to the actuals.

Sources, methodology & freshnessLast updated 2026-07-24 · Open ↓
Source
EIA-930 hourly balancing-authority demand forecasts and adjusted actual demand (public domain)
Methodology
Matched-fleet hourly forecast vs actual; daily mean absolute % error, signed bias and peak-hour error; non-functional forecast feeds excluded by fixed rule
Updates
Daily, after EIA publishes a complete prior dayLast: 2026-07-24
Maintained & reviewed by Yuriy Matso — methodology shown on the page.

thetrading.tools PRO

Talk to this data with your AI

PRO is an MCP server that plugs Claude, ChatGPT, Cursor or any AI agent straight into everything on this site — the live market scores, ~90 economic indicators with full histories, and every signal study. Ask a market question; your AI answers from our data, dated and sourced.

Get PRO — $39/mo

Cancel anytime

How Forecast Accuracy Tracker Works

  1. 1
    The grid forecasts itself, daily
    Every balancing authority files a day-ahead hourly demand forecast into EIA-930 alongside its actuals. We compare the two — hour by hour, summed over the same BAs on both sides (a matched fleet, so coverage changes cannot fake accuracy changes).
  2. 2
    Three error reads per day
    MAPE (mean absolute hourly error as % of actual) is the headline miss. Signed bias shows the lean — positive means the fleet over-forecast. Peak-hour error isolates the hour that matters most for scarcity, when a miss means scrambling for supply.
  3. 3
    Broken feeds are excluded by rule
    A few small BAs file junk forecasts (one printed 10x misses). A fixed rule drops any BA whose half-year MAPE exceeds 50% from both sides of the comparison — about 1% of demand, which would otherwise inflate the national miss by a third of a point.
  4. 4
    Stats start when the data stabilizes
    EIA-930's first six months (late 2015) show ~10x steady-state error — collection teething, not forecast skill. Statistics begin January 2016 and the page says why.

Who Uses Forecast Accuracy Tracker

Power Market Watchers
Day-ahead vs real-time is the fundamental spread in electricity markets, and forecast error is what drives it. Days with big under-forecasts are the days real-time prices spike above day-ahead — this page tracks that error term at fleet scale.
Grid Data Skeptics
Anyone using grid data (including ours) should know how good the system's own nowcasts are. Answer: remarkably — a ~1-2% typical daily miss — which is context for how seriously to take demand-driven readings.
Volatility Thinkers
Forecast error is where physical volatility comes from: a grid that misses by 5% needs 30+ GW of unplanned supply. The 30-day error trend rising is the physical system getting harder to predict.
Seasonality Students
The monthly norms table shows when the grid is hardest to forecast (shoulder seasons and extreme-weather months) — the calendar of when surprises, and surprise-driven pricing, concentrate.

Pro Tips

01
Bias is the tell, not MAPE
A symmetric miss is weather noise. A persistent under-forecast lean — actuals repeatedly above forecast — is the signature of load growing faster than the forecasting models assume, which is exactly what unmodeled data-center growth looks like.
02
Peak-hour error is the money number
The daily average miss can be small while the peak hour misses big — and the peak hour is where reserve margins and scarcity pricing live. Watch the two diverge.
03
Compare the 30-day mean to the monthly norm
A 2% miss is normal in April and notable in July. The rolling mean against its own month's historical norm separates "the season is hard" from "something changed".
04
Extreme misses cluster with extreme events
The record miss days are storm days. A spike here is usually the footprint of a weather event — cross-reference the Grid Stress score for the same dates.

Common Issues & Solutions

Whose forecast is this?
The balancing authorities' own operational day-ahead filings — the forecasts the grid actually runs on — not ours and not a model we built. We measure them; we do not make them.
Why do a few BAs get excluded?
A fixed, documented rule: any BA whose mean absolute error exceeds 50% within a half-year file is treated as not filing a functional forecast and removed from both sides for that file. It protects the national number from ~1%-of-demand feeds printing garbage.
Is a rising error trend bullish or bearish for anything?
We make no traded-signal claim — no forward-return study has been run against this series. What a rising trend does say, descriptively, is that the physical system is getting harder to predict, which historically accompanies the volatile power-price regimes.
Why does accuracy look worse before 2016?
EIA-930 collection launched July 2015 and its first months show ~10x the steady-state error — filing and pipeline teething at the BAs, not genuinely worse forecasting. That is why the statistics window starts January 2016.

Frequently Asked Questions

How accurate are power-grid demand forecasts?
Remarkably: the fleet's day-ahead hourly forecasts typically miss by about 1-2% on an average day. The live daily miss, the 30-day trend, the record miss and each month's historical norm are all on this page, computed from EIA-930 since 2016.
Why does forecast error matter for power prices?
Day-ahead markets commit supply against the forecast; reality clears in real time. When actual demand comes in above forecast, the difference is bought at whatever price scarcity demands — under-forecast days are the raw material of real-time price spikes. This page tracks that error term at national fleet scale.
What is a demand forecast bias, and why watch it?
Bias is the signed lean of the errors — persistently negative (under-forecasting) means actual load keeps exceeding the models. That is the earliest statistical footprint of structural load growth the forecasters haven't caught up to, e.g. faster-than-modeled data-center energization.
Whose forecasts are being graded?
The U.S. balancing authorities' own operational day-ahead filings in EIA-930 — the forecasts the grid physically schedules against. We sum forecast and actual over the identical set of BAs each hour, and exclude by fixed rule the few small feeds that file non-functional numbers.
Does this page predict anything?
No — it is a measurement of the grid's own nowcasting skill, shipped without a forward-return claim because we haven't validated one. Its value is context: how predictable the physical system is right now vs its own history.

Explore Other Tools

Last updated: 2026-07-24