Part of The 7 OraclesThe Pump Line
Where the national average gas price is going in 1, 7 and 14 days — and a permanent record of how wrong we have been, misses included. We are the only people in America who publish a gas price forecast and then tell you how it did.
AAA national average $4.042 as of 2026-08-08 · RBOB $2.715
Quick Answer
Over 1,124 graded seven-day forecasts we have been 37.9% closer than assuming no change, and right on direction 80.2% of the time.
This tool was recalibrated on 29 July 2026. Version 1 was a Kalshi trading signal built on a different model, and it lost — 2 wins from 25 settled positions. It was retired rather than tuned: the model had no wholesale pass-through structure, and the market it traded has about $50 of depth, so no amount of accuracy would have made it a business.
Version 2 is a different product measured a different way — a price forecast scored on how far off it lands, not a position scored in profit and loss. The two records share no units and cannot be added together, so the record below is V2's only. V1's is retained in full and is not going anywhere; it just is not this tool's record.
The forecast — from the 2026-08-08 print
We do not publish a 30-day number. At a month out the model beats a coin flip by about 6% — close enough to nothing that quoting it would be pretending. Anyone giving you a confident gas price for next month is guessing.
The record — every graded forecast
| Horizon | Graded | Our avg miss | “No change” miss | Better by | Right direction |
|---|---|---|---|---|---|
| Tomorrow | 1091 | 0.57¢ | 0.91¢ | 37.5% | 78.9% |
| 7 days out | 1124 | 3.32¢ | 5.35¢ | 37.9% | 80.2% |
| 14 days out | 1097 | 7.40¢ | 9.89¢ | 25.2% | 75.2% |
“No change” is the benchmark everyone actually competes against: assume tomorrow’s price equals today’s. Beating it is the only thing that makes a forecast worth reading.
Last 90 days only
| Horizon | Graded | Our avg miss | “No change” miss | Better by | Right direction |
|---|---|---|---|---|---|
| Tomorrow | 50 | 0.90¢ | 1.39¢ | 35.3% | 80.0% |
| 7 days out | 51 | 5.09¢ | 8.43¢ | 39.5% | 86.3% |
| 14 days out | 46 | 12.43¢ | 17.78¢ | 30.1% | 69.6% |
Same table, recent window only — so a strong past cannot hide a weak present.
The last 24 calls, unfiltered
| For | Out | We said | It was | Miss | Dir |
|---|---|---|---|---|---|
| 2026-08-08 | 7d | $4.082 | $4.042 | 4.0¢ | ✓ |
| 2026-08-08 | 1d | $4.042 | $4.042 | 0.0¢ | ✓ |
| 2026-08-07 | 1d | $4.059 | $4.063 | 0.4¢ | ✓ |
| 2026-08-07 | 7d | $4.032 | $4.063 | 3.1¢ | ✓ |
| 2026-08-06 | 7d | $4.047 | $4.080 | 3.3¢ | ✓ |
| 2026-08-06 | 1d | $4.075 | $4.080 | 0.5¢ | ✓ |
| 2026-08-05 | 1d | $4.083 | $4.089 | 0.6¢ | ✓ |
| 2026-08-05 | 7d | $4.058 | $4.089 | 3.1¢ | ✓ |
| 2026-08-04 | 7d | $4.045 | $4.095 | 5.0¢ | ✓ |
| 2026-08-04 | 1d | $4.086 | $4.095 | 0.9¢ | ✓ |
| 2026-08-03 | 1d | $4.093 | $4.096 | 0.3¢ | ✓ |
| 2026-08-02 | 1d | $4.099 | $4.101 | 0.2¢ | ✓ |
| 2026-08-01 | 1d | $4.090 | $4.106 | 1.6¢ | ✗ |
| 2026-07-31 | 1d | $4.083 | $4.098 | 1.5¢ | ✗ |
| 2026-07-30 | 1d | $4.090 | $4.091 | 0.1¢ | ✓ |
| 2026-07-29 | 1d | $4.108 | $4.099 | 0.9¢ | ✓ |
| 2026-07-28 | 14d | $3.955 | $4.099 | 14.4¢ | ✓ |
| 2026-07-28 | 7d | $4.140 | $4.099 | 4.1¢ | ✓ |
| 2026-07-28 | 1d | $4.115 | $4.099 | 1.6¢ | ✗ |
| 2026-07-27 | 1d | $4.119 | $4.110 | 0.9¢ | ✗ |
| 2026-07-26 | 14d | $3.887 | $4.110 | 22.3¢ | ✓ |
| 2026-07-26 | 7d | $4.096 | $4.110 | 1.4¢ | ✓ |
| 2026-07-24 | 14d | $4.029 | $4.105 | 7.6¢ | ✓ |
| 2026-07-24 | 7d | $4.163 | $4.105 | 5.8¢ | ✓ |
Our five worst misses, ever
| 2022-03-10 | 14d | $3.605 | $4.318 | 71.3¢ | ✓ |
| 2022-03-09 | 14d | $3.573 | $4.252 | 67.9¢ | ✓ |
| 2022-03-12 | 14d | $3.676 | $4.326 | 65.0¢ | ✓ |
| 2022-03-14 | 14d | $3.703 | $4.325 | 62.2¢ | ✓ |
| 2022-03-07 | 14d | $3.551 | $4.065 | 51.4¢ | ✓ |
These stay on the page permanently. A forecast record you can only see the good half of is marketing, not a record.
How it works, and how it was tested
Retail gas follows wholesale gas with a lag. When RBOB futures move, the pump takes one to three weeks to finish reacting — faster upward than downward. The model is built on that lag directly: today's forecast is driven by where wholesale has already been over the past three weeks, plus retail's own recent momentum.
It was tested walk-forward on 2,121 archived AAA daily observations going back to June 2016. A forecast only counts if both the day it was made and the day it targets are real observations — no interpolating across gaps, which would leak the answer into the question. Coefficients were refit as the window expanded, and everything reported is out-of-sample from 2021.
- 1 day: 36.5% better than a no-change assumption, 82.4% directional, over 1,080 out-of-sample days.
- 7 days: 34% better than a no-change assumption, 78.6% directional, over 1,119 out-of-sample days.
- 14 days: 21.2% better than a no-change assumption, 74.4% directional, over 1,092 out-of-sample days.
A tell that this is economics rather than trend-following: wholesale terms alone beat retail momentum alone (+30.2% vs +20.9% at one week). If we had merely rediscovered momentum, that would be the other way round.
The honest caveat. Most of the graded rows on this page are reconstructed: we ran the model over history at launch rather than waiting five years to accumulate them, and those rows are flagged is_backfill in the data. Because the coefficients were fit on that same history, backfilled rows read slightly better than live ones will. The accuracy figures quoted in this section are the out-of-sample walk-forward numbers, which do not have that problem. Forecasts written from tonight forward are live calls, graded after the fact and never revised.
What about the gas market on Kalshi?
Kalshi lists contracts on the AAA national average, and this page used to be built around trading them. It is not any more, for a reason worth stating plainly: at the money there is roughly $50 of depth. A model that was perfect would make a few hundred dollars a week there before it started moving the price against itself. That is not a business, and pretending otherwise is what produced a losing record on a public page.
There is also a trap in the comparison. When our forecast and that market disagree loudly, the overwhelmingly likely explanation is that our inputs went stale, not that we found an edge — the market watches live futures continuously. So we treat a large divergence as a fault alarm to investigate, never as a signal to act on. That single inversion is the difference between this rebuild and what it replaced.
No position sizing is published here, and none returns until at least 26 weekly settles are on the board under this model.
Frequently Asked Questions
Will gas prices go up or down next week?
Our model publishes a specific number every day for where the AAA national average lands 1, 7 and 14 days out, along with a 95% range. The forecast at the top of this page is the current one. Over 1,119 graded seven-day forecasts it has been about 38% closer than assuming no change, and has called the direction correctly 80% of the time — which also means it has been wrong on direction about one week in five.
How accurate is this gas price forecast?
At one day out the average miss is under a cent. At seven days it is about 3.3¢, against 5.4¢ for assuming no change. At fourteen days it is about 7.4¢ against 9.9¢. Every graded forecast is listed on this page, including the five worst misses, and the accuracy table always shows the no-change benchmark next to our number so you can check the comparison yourself.
Why do you not publish a 30-day gas price forecast?
Because we tested it and it barely works. At a 30-day horizon the model beats a no-change assumption by roughly 6%, with a fitted coefficient that flips sign — the signature of fitting noise rather than signal. Publishing that number with a confident face would be dishonest. Anyone quoting you a precise national average for next month is guessing, whether or not they say so.
What actually drives gas prices?
Wholesale gasoline, with a lag. Retail pump prices follow RBOB futures over roughly one to three weeks — the "rockets and feathers" pattern, where prices climb faster than they fall. Our model is built directly on that lag structure: the wholesale terms alone beat retail momentum alone, which is the evidence that it is tracking economics rather than just extrapolating a trend.
Where does the data come from?
Retail prices are the AAA national average, which surveys up to 120,000 stations daily. Wholesale is the RB=F front-month gasoline futures close. Our AAA history runs to 2,121 daily observations back to June 2016, reconstructed from public archive snapshots and stamped with AAA’s own as-of date rather than the crawl date, then appended nightly.
Is this a trading tool?
No, and it used to be, which was the mistake. Kalshi lists gas contracts and we show them below for comparison, but that market has roughly $50 of depth at the money — a perfect model there earns a few hundred dollars a week before it starts trading against itself. We do not publish position sizing for it. The forecast is the product; the market comparison is context.
Forecasts are model output for general information, not financial advice. The model is wrong regularly and the record above shows exactly how often. Data: AAA national average (retail) and RB=F front-month futures (wholesale). Model pumpline-v1.