ERCOT
ERCOT demand forecasts. 1.56% MAPE.
69.4% lower error than the ISO baseline. Built for Texas energy traders, IPPs, and competitive retailers.
Why ERCOT is different
An isolated grid where forecast errors cost you directly.
ERCOT is the only major U.S. grid with no meaningful AC interconnection to the rest of the country. When demand surges or generation drops, there is no import cushion — scarcity pricing kicks in at the $5,000/MWh cap, ORDC adders layer on top, and load-serving entities either hedged correctly or watch a month of margin evaporate in six hours. Forecast accuracy matters more here than in any other market, because there is nowhere for the error to hide.
The competitive retail structure adds another layer. Eighty-plus REPs serve 29M customers in the deregulated footprint — Oncor, CenterPoint, AEP Texas, and TNMP deliver the electrons, but the commercial risk sits with the retailers. Every REP with any scale has a forecasting function, either in-house or vendor-supplied. The question we consistently hear isn't whether better forecasts are valuable; it's whether a new provider can actually be measured against what the desk is already using.
That's what the benchmark on this page is for. ERCOT publishes their own day-ahead forecast. We publish ours next to it, with methodology, over a full calendar year. The Aug 2023 heat dome and Jan 2024 cold snap are in the evaluation window — those hours are where baseline errors compound and where a 3-4 percentage-point MAPE gap translates to real imbalance exposure.
If you already model nodal price, think of Gramm as a sharper load input to whatever LMP or congestion model you run — not a replacement for Yes Energy, Genscape, or your in-house stack. ERCOT's own day-ahead load forecast is what most price models carry today; swapping in a 1.56% MAPE input where the ISO baseline is 5.1% measurably tightens the downstream price estimate without changing anything else in the pipeline.
Market impact
ERCOT by the numbers.
A 3.54 percentage point MAPE reduction over the ISO baseline, measured across 8,760 hours. Lower error means tighter DA bids and less imbalance exposure.
Evaluated against ERCOT’s own published day-ahead forecast over the full 2025 calendar year. Every hour, every season — no cherry-picking.
Hourly generation data by fuel type (gas, wind, solar, nuclear, coal) across ERCOT. Know when wind dies and gas must ramp — before the market clears.
Benchmark
ERCOT benchmark details.
Full-year evaluation against the published ERCOT day-ahead forecast. Every number is reproducible.
Worked example
Translating MAPE to imbalance exposure on ERCOT
A Houston-area REP serving 10 GW of peak load. ERCOT's published DA baseline runs at ~5.1% MAPE; our evaluation shows Gramm at 1.56% MAPE — a 3.54 percentage-point reduction in load-weighted forecast error. Multiplied across a year of operations:
Illustrative calculation for a 10 GW peak REP at ERCOT's average imbalance penalty. Your exposure depends on your hedging profile, scarcity-hour participation, and position size. The reduction scales roughly linearly with load — doubling the footprint doubles the exposure benefit.
Use cases
ERCOT workflows.
Day-ahead scheduling
Submit optimal DA bids with forecasts that beat ERCOT’s own projection. Reduce over-commitment and imbalance exposure.
Wind integration
ERCOT leads the nation in wind capacity. Our generation mix API shows real-time wind output, so you can anticipate ramp events hours ahead.
Gas asset dispatch
If you operate gas peakers or CCGT units, knowing when demand spikes — and when wind drops — determines your dispatch schedule and profit margin.
Quick start
One request. Full day-ahead forecast.
Most teams integrate in under an hour. Bearer token auth, JSON response, prediction intervals included.
View full API docsFAQ
Questions ERCOT desks actually ask.
Do you forecast by load zone (Houston, North, West, South, Coast)?
System-wide ERCOT forecast is available on all plans. Zonal forecasts (Houston, North, South, West, and the LZs) are available on Enterprise — the model is trained per-zone because the load shapes diverge significantly, especially during the Houston summer peak versus North Texas industrial load.
How does the forecast behave during scarcity pricing events?
This is where we spent the most tuning effort. The second of our two founding papers (arXiv:2601.01410) was specifically about tail behavior — state space architectures don't collapse the way transformer variants do under extreme weather. During August 2023 heat-dome hours when ERCOT's baseline error spiked to 8-12%, our backtested MAPE stayed in the 2-3% range.
Do you provide 15-minute settlement-period forecasts?
Yes, on Developer, Team, and Enterprise plans. The day-ahead benchmark published on this site is hourly to match ERCOT's own DAM reporting, but the API also returns 15-minute granularity for RT operations.
Can you integrate with ERCOT API feeds or SCED outputs?
Delivery is REST API (api.gramm.ai), or Snowflake / SFTP on Enterprise. Teams consume the output as an input to their own scheduling or bidding systems.
What about Uri-type events — do you claim accuracy during those?
No. Uri was outside the bounds of any demand forecast's training distribution, and we don't think anyone forecasting MW should claim otherwise. What we do claim is that during normal-but-stressed conditions (heat domes, cold snaps), the state space model degrades gracefully where transformer-based forecasts tend to fail catastrophically.
How do you handle load from the ADER pilot and behind-the-meter solar?
BTM visibility is an ongoing problem across all grids and we don't claim perfect attribution. What we do is explicitly train the model to treat BTM-adjusted net load as the target, using interval data from AMI where available and WRF solar irradiance as a proxy elsewhere. ERCOT's ADER pilot is still small enough that it's not a material forecast driver yet, but we're tracking it.
Get started
Start evaluating ERCOT forecasts.
Free plan includes 100 req/hr across all 7 grids. No credit card required.