Benchmarking ERCOT day-ahead against real-time at three resolutions — monthly, weekly and hourly — from raw settlement point prices. 13,847 hours per hub, North and Houston, January 2025 to July 2026. The answer changes by a factor of four and a half depending on which resolution you ask at, and in nine months of nineteen it changes sign in meaning entirely.
Third in a series with ERCOT's $500K Months and ERCOT Without the Storm.
Mean absolute basis is $2.72/MWh monthly, $4.65 weekly and $12.14 hourly. Same market, same period, same metric.
At monthly resolution the day-ahead and real-time markets look like they converge: the gap clears the SEC 10% test in 3 of 19 months. At weekly resolution it clears in 37 of 81 weeks. At hourly resolution it clears in 66% of all 13,847 hours.
The reason is arithmetic, not opinion. A monthly basis averages day-ahead across the month, averages real-time across the month, and subtracts — which nets a $600 divergence in one hour against a −$600 divergence in another and reports zero. In nine of nineteen months the monthly average basis came in under $1/MWh while the hourly basis inside those same months averaged between $5.73 and $15.00.
April 2025 is the cleanest case. Monthly average basis: −$0.04/MWh. Textbook convergence. Inside it, the hourly basis averaged $13.80 and ran from −$1,454 to +$98. The monthly figure was not an understatement of the risk. It reported the opposite of the risk.
Every section asks the same thing — how far does day-ahead sit from real-time, and how often is the gap material — and answers it at a different time resolution. The comparison between the three is the point.
Real-time is the denominator because it is what physically settled — day-ahead is the forward decision, real-time is the outcome it was judged against. A positive basis means the day-ahead buyer overpaid. The 10% threshold is the SEC's own sensitivity-analysis floor, carried over from Study II, and needs no volume assumption — so the count is identical for every customer size.
| Resolution | Observations | Mean |basis| | Std dev | Range | Material |
|---|---|---|---|---|---|
| Monthly | 19 months | $2.72 | $6.34 | −$1.78 to +$27.17 | 3 of 19 (16%) |
| Weekly | 81 complete weeks | $4.65 | $9.76 | −$17.63 to +$59.52 | 37 of 81 (46%) |
| Hourly | 13,847 hours | $12.14 | $50.87 | −$2,027 to +$1,592 | 66% of hours |
One denominator problem, handled explicitly. Real-time price goes to zero and below in ERCOT — 5.7% of North Hub hours settled at or under $5/MWh, and 3.0% went negative. Dividing by a near-zero denominator produces meaningless percentages, so the hourly test is reported two ways: against that month's mean real-time price as a stable reference (66% of hours material), and against the same hour's real-time price with hours under $5 excluded (72% material, 5.7% of hours dropped). The two agree closely; the stable-reference version is used throughout because it keeps every hour in the sample and is closer to what materiality means commercially.
Two monthly series are available and they corroborate each other. The Independent Market Monitor publishes an ERCOT-wide, load-weighted monthly average; the raw settlement files give a hub-level simple average. They agree on the finding.
| Series | Mean basis | Mean |basis| | Max premium | Max discount | Material |
|---|---|---|---|---|---|
| IMM ERCOT-wide (load-weighted) | +$2.40 | $3.70 | +$38.02 | −$2.75 | 3 / 19 |
| HB_NORTH (simple average) | +$2.05 | $2.72 | +$27.17 | −$1.78 | 3 / 19 |
| HB_HOUSTON (simple average) | +$2.32 | $2.75 | +$26.70 | −$1.52 | 3 / 19 |
Three material months out of nineteen on every series, and a mean basis between +$2.01 and +$2.40/MWh. The load-weighted ERCOT-wide figure runs a little wider because load weighting puts more mass on the expensive hours, which is exactly the effect this study is about. The conclusion from the earlier monthly-only work stands: at monthly resolution the two markets appear to converge, with a modest persistent day-ahead premium and one catastrophic month.
The monthly view cannot distinguish a month in which day-ahead sat steadily $2 above real-time from one in which it alternated between +$15 and −$11 week to week. Those are different risks. Eighty-one complete weeks answer it.
The answer is alternating, not persistent. Lag-1 autocorrelation of +0.39 is real but weak; the mean same-sign run is under two weeks and 89% of runs end within three. A term position taken on the basis would be right for a fortnight and then wrong.
Weekly averaging also destroys information, just less of it than monthly does: within these same weeks the hourly mean absolute basis was $12.21 against a weekly-average figure of $4.65 — a 2.6× compression, versus 4.5× at monthly.
13,847 hours per hub, day-ahead against the hour-averaged 15-minute real-time price. This is the resolution at which ERCOT actually settles, and the resolution at which the heat-rate dispersion found in Study II physically happens.
| Hub | 0.1% | 1% | 5% | 25% | 50% | 75% | 95% | 99% | 99.9% |
|---|
The basis is not spread evenly. It has a strong and repeatable daily shape — and the shape is the same at both hubs, which means it is a system property rather than local congestion.
This is the heat-rate finding, seen from the other side. Study II found that ERCOT price movement is driven by implied heat rate rather than fuel — grid tightness, outages and reserve scarcity. Those are hourly events, and here they are: mean absolute basis peaks at $26.45/MWh in hour ending 20 and $26.00 in HE21 against $6.89 at midday, with a second spike in the morning ramp at HE07–08. The day-ahead market is at its least able to forecast real-time precisely when the net-load ramp is steepest. Two independent methods, two datasets, the same underlying mechanism.
A small number of hours carry most of the divergence — though not quite as few as predicted.
| Share of total |basis| | North Hub | Houston Hub |
|---|
| Date | Hour ending | Day-ahead | Real-time | Basis | At 100 MW |
|---|
The two clusters tell the whole story of January 2026. On 26 January — the peak of Winter Storm Fern — day-ahead cleared between $1,556 and $1,899/MWh for hours 6 through 9 while real-time settled between $136 and $335. The day-ahead market priced a catastrophe that did not arrive: +$1,592/MWh in a single hour, or $159,000 on a 100 MW book in sixty minutes. Two days later, on 28 January, it ran the other way — day-ahead at $147, real-time at $1,382. The largest single dislocation in the whole dataset is 15 January 2025 hour ending 08: day-ahead $59.98, real-time $2,087.03, a basis of −$2,027/MWh and −$203,000 in one hour at 100 MW. None of these hours is visible in any monthly average.
| Avg load | MWh | Net cost of DA scheduling | Gross basis moved | Worst single hour |
|---|
At 100 MW the net cost of scheduling day-ahead rather than settling real-time was +$2.84M over nineteen months — close to the $3.42M the monthly ERCOT-wide series suggested, which is reassuring. But the gross basis that moved through the book was $16.82M, against $5.17M at monthly resolution. Three and a quarter times as much money was in motion as the monthly view can see, on the same book, over the same period, in the same market.
This is the finding that justifies the whole exercise. For each of the nineteen months, compare the monthly average basis — the number a monthly report would print — against the mean absolute hourly basis inside that same month.
| Month | Monthly avg basis | Hourly mean |basis| | Ratio | 5th pct hour | 95th pct hour | Worst hour |
|---|
Nine months of nineteen — almost half the sample — printed a monthly average basis under $1/MWh. In those same nine months the hourly basis averaged between $5.73 and $15.00, the 5th percentile hour ran between −$12 and −$33, and the worst hour reached −$1,477. A monthly reconciliation of any of those months would have shown a day-ahead and real-time market in near-perfect agreement.
The precise claim, stated carefully. Monthly averaging does not understate the basis — it reports a different quantity. The monthly average is the net position, and the net position genuinely was near zero in those months; a buyer who settled monthly at the average paid roughly what the average implied. What the monthly figure cannot show is the path: the gross movement, the collateral required to fund it, the hours where a contract term triggered, and whether any individual settlement was correct. For a business whose product is evidencing what was actually paid, that distinction is the entire point.
The previous version of this document recorded four expectations, so that they could be judged rather than confirmed after the fact. Two held, two did not.
| Prediction | Outcome | Verdict |
|---|---|---|
| Hourly mean |basis| between 3× and 10× the monthly figure | $12.14 hourly against $2.72 hub-monthly = 4.5×. Against the IMM ERCOT-wide ex-January figure of $1.79 it is 6.8×. | Held |
| More than half of all absolute basis carried by fewer than 5% of hours | The top 5% of hours carry 46.2%. Reaching 50% takes 6.5% of hours — 906 of 13,847. | Missed |
| Weekly mean |basis| roughly 2–3× monthly | $4.65 against $2.72 = 1.7×. Over-predicted; weekly sits closer to monthly than expected. | Missed |
| Weekly sign run-lengths rarely exceed three weeks | 89% of same-sign runs end within three weeks; mean run 1.8 weeks, longest 8. | Held |
The two misses are worth keeping. The concentration miss means the divergence is less tail-driven than expected — it is not a handful of scarcity hours, it is a broad condition affecting the majority of hours, which is a stronger argument for hourly evidencing, not a weaker one. The weekly miss means weekly resolution buys less than expected: the jump that matters is from weekly to hourly, not from monthly to weekly.
The same market, the same nineteen months and the same metric produce $2.72, $4.65 or $12.14 depending only on the window you average over — and a material-move rate of 16%, 46% or 66%. Any claim about ERCOT basis risk that does not state its resolution is unfalsifiable. That is worth saying out loud in a market where most reporting is monthly.
Weekly resolution recovers only 1.7× the monthly figure; hourly recovers 4.5×. Half-measures do not help. If a product settles or evidences at anything coarser than native market resolution, it is closer to the monthly picture than to the real one — and the monthly picture is, in nine months of nineteen, a picture of nothing happening.
Two thirds of all hours clear the SEC threshold, and the top 5% of hours carry under half the total. This is an everyday condition, not a tail event. It also has a repeatable daily shape, peaking at 4× the midday level in the evening ramp at both hubs — which means it is forecastable, structural, and the same phenomenon Study II found in the heat rate.
At 100 MW the net was +$2.84M and the gross was $16.82M. The net is what a monthly reconciliation shows. The gross is what had to be funded, collateralised and — if a contract has hourly-triggered terms — correctly settled, hour by hour, 13,847 times. Since ERCOT eliminated unsecured credit in October 2023, the gross is the number that sizes the facility.
NP4-180-ER (Historical DAM Load Zone and Hub Prices, hourly) and NP6-905-CD (Historical RTM, 15-minute), annual files for 2025 and 2026. HB_NORTH and HB_HOUSTON extracted. No intermediary, no vendor, no chart-reading.
The four 15-minute real-time intervals within each delivery hour are averaged to give an hourly real-time price, then matched to the day-ahead price for the same delivery date, hour and repeated-hour flag — so the daylight-saving duplicate hour is handled rather than dropped. 28,750 hourly hub-observations matched with zero unmatched rows on either side, which is itself a validation that the alignment is correct.
Weekly and monthly figures are computed from the same hourly rows, as simple means of the hourly prices. Nothing is interpolated or modelled. Complete weeks only — a week needs 160 of 168 hours to be included, which drops partial weeks at the boundaries.
Material when |basis| reaches 10% of real-time. At monthly and weekly resolution the denominator is that period's mean real-time price. At hourly resolution it is the month's mean real-time price, with the same-hour version reported alongside — see section 01.
The hub-level monthly figures are compared against the IMM's ERCOT-wide load-weighted series for the same nineteen months. Both give three material months out of nineteen and a mean basis within $0.40/MWh of each other.
Four limits, stated before anyone quotes this. One: hub prices here are simple hourly averages, not load-weighted — a real book pays a load-weighted price, which puts more weight on the expensive hours and would make the basis wider, not narrower. The IMM comparison in section 02 quantifies that gap at roughly $1/MWh on the monthly mean absolute figure. Two: these are hub prices. Retail load settles at load zones, which carry additional congestion; load-zone basis will differ and is a natural next cut, since the same files contain it. Three: ERCOT's real-time market was redesigned on 5 December 2025 — RTC+B replaced the ORDC price adder with per-product demand curves and split the system-wide offer cap into separate day-ahead and real-time caps. That change sits inside this window and acts directly on the quantity being measured. Comparisons across December 2025 span two market designs, and the 2025 and 2026 halves of this dataset should not be pooled for any claim about trend. Four: nineteen months is a short sample containing one extreme scarcity event, and the summary statistics move materially depending on whether January 2026 is in the window.