What the next official numbers are likely to be, how wide the honest range is, and how often that range has been right.
+8.2%
our estimate for September 2026
80% range +6.5% to +10.0%
· published 14 October
EIA's weekly surveys put the average US pump price at $4.355 a gallon over the month, +7.3% on the month before. Last official figure: +3.9% for August 2026.
| Period | Forecasts | Inside the 80% range | Error vs no-change | Direction right | Typical miss |
|---|---|---|---|---|---|
| Since 2010 | 200 | 84% | 69% smaller | 87% | 1.2 pts |
| Last 3 years | 36 | 78% | 72% smaller | 86% | 1.2 pts |
| Last 12 months | 12 | 67% | 80% smaller | 100% | 1.4 pts |
+18.0%
our estimate for September 2026
80% range +5.4% to +30.5%
· published 16 October
Brent averaged $114.16 a barrel over the month, +25.3% on the month before. Last official figure: -3.4% for August 2026.
| Period | Forecasts | Inside the 80% range | Error vs no-change | Direction right | Typical miss |
|---|---|---|---|---|---|
| Since 2010 | 200 | 82% | 57% smaller | 85% | 2.3 pts |
| Last 3 years | 36 | 83% | 16% smaller | 81% | 3.2 pts |
| Last 12 months | 12 | 58% | 1% smaller | 75% | 6.0 pts |
$4.35
our estimate for 19 October 2026
80% range $4.11 to $4.58
· EIA survey for 19 October
Starts from $4.354 on 5 October; New York Harbor wholesale moved +0.167 dollars a gallon over the previous four weeks. Last official figure: $4.35 for 5 October 2026.
| Period | Forecasts | Inside the 80% range | Error vs no-change | Direction right | Typical miss |
|---|---|---|---|---|---|
| Since 2012 | 769 | 85% | 10% smaller | 67% | 6.8¢ |
| Last 3 years | 157 | 89% | 5% smaller | 65% | 7.9¢ |
| Last 52 weeks | 52 | 77% | 5% smaller | 68% | 12.9¢ |
Withheld. This model failed today's accuracy check: improves on the no-change forecast by 4.8%, below the 5% bar; its 80% range held only 29 of the last 52 outcomes. It comes back on its own once its record clears the bar again.
| Period | Forecasts | Inside the 80% range | Error vs no-change | Direction right | Typical miss |
|---|---|---|---|---|---|
| Since 2012 | 769 | 82% | 5% smaller | 68% | 6.7¢ |
| Last 3 years | 157 | 80% | 6% smaller | 59% | 11.6¢ |
| Last 52 weeks | 52 | 56% | 7% smaller | 65% | 21.6¢ |
12.8%
chance that September 2027 is a month of US recession, from the New York Fed's yield-curve model
Built from September 2026: 10-year Treasury 4.99%, 3-month bill 4.04% on a bond-equivalent basis, a spread of +0.95 points. The curve has to invert — short rates above long — before the probability passes 30%.
| Warning for | Peak | What followed |
|---|---|---|
| Jan 1967 | 32% | No recession — false alarm |
| Sep 1967 – Feb 1968 | 41% | No recession — false alarm |
| Dec 1969 – Feb 1971 | 42% | Recession Jan 1970–Nov 1970 |
| Jun 1974 – Nov 1975 | 68% | Recession Dec 1973–Mar 1975 |
| Nov 1979 – Sep 1982 | 95% | Recession Feb 1980–Jul 1980 and Aug 1981–Nov 1982 |
| Jun 1990 – Dec 1990 | 33% | Recession Aug 1990–Mar 1991 |
| Jul 2001 – Jan 2002 | 46% | Recession Apr 2001–Nov 2001 |
| Aug 2007 – May 2008 | 42% | Recession Jan 2008–Jun 2009 |
| Jun 2020 – Feb 2021 | 38% | No recession — false alarm |
| Nov 2023 – Apr 2026 | 71% | No recession so far |
| When the model said | Months | Share that were recession months |
|---|---|---|
| 0–10% | 425 | 2% |
| 10–30% | 256 | 14% |
| 30–50% | 72 | 47% |
| 50–100% | 48 | 33% |
16cm
measured 11 Oct, 06:30 +0200
· +10 cm in 7 days
30-day range -6 to 36 cm
Official forecast issued 10 Oct, 07:00 +0200: 14 cm by 14 October. Lowest point in the forecast 9 cm on 11 October. BfG labels the first days a forecast and later days an estimate, which is less certain. Both are shown as published.
Reference line: GlW, the equivalent low-water level WSV publishes for this gauge, 77 cm. The forecast is at or below it. The lowest level WSV lists for the gauge is 25 cm, on 22 October 2018; the latest reading is below it. PEGELONLINE levels are raw readings, published before WSV checks them.
9.6ft
measured 11 Oct, 04:00 UTC
· +2.7 ft in 7 days
30-day range -4.8 to 11.3 ft
Official forecast issued 10 Oct, 15:13 UTC: -1.3 ft by 25 October.
Reference line: the NWS low-water threshold for this gauge, -8.0 ft. The forecast stays above it. Lowest stages NWS lists: -12.1 ft (Oct 2023, preliminary); -10.8 ft (Oct 2022); -10.7 ft (Jul 1988).
| For | Measure | Forecast (80% range) | Official | Result |
|---|---|---|---|---|
| 19 October 2026 | US regular gasoline, two weeks out | 4.35 (4.11 to 4.58) | — | Due 19 Oct |
| September 2026 | US CPI gasoline, next print | +8.2 (+6.5 to +10.0) | — | Due 14 Oct |
| September 2026 | US import prices for fuel, next print | +18.0 (+5.4 to +30.5) | — | Due 16 Oct |
Nothing on this page is an opinion about the future. Each figure is one of three things: an estimate of a number that is already settled but not yet published, a short-range estimate whose error we have measured over years, or a forecast made by the official body responsible for it, reproduced unchanged.
Every model of ours meets four conditions before anything is shown.
The check runs every day. A model that fails it is withheld, the page says which condition failed, and it returns only when its record clears the bar again.
The range widens fast and narrows slowly. It is the larger of two measures: the 80th percentile of the last 120 misses, and of the last year’s (26 weeks for the weekly pump-price model). After oil prices jumped in March 2026, ranges built only from the long history were too narrow for months: in the 52 weeks to 21 September 2026 the two-week gasoline range caught 28 outcomes. With the short window the same model caught 40. We changed the rule after seeing that failure, and say so here because a rule changed after the fact deserves suspicion. The test that rules out a model on its last year was added at the same time, for the same reason.
CPI gasoline. The Bureau of Labor Statistics prices gasoline from its own sample and publishes the index seasonally adjusted; EIA’s weekly survey, which drives the estimate, is a different sample and is not adjusted. The model learns the gap between the two, calendar month by calendar month. It cannot see a revision to the BLS seasonal factors, and its backtest is scored against today’s revised figures rather than the first prints; the ledger uses first prints.
Fuel import prices. Part of a move in Brent reaches the index a month late; the model’s second term measures how much. It has been weakest exactly when it matters most: over the twelve months to August 2026 its typical miss was more than twice its long-run average, and it only just beat “no change”. That is why its range is wide. Read the range, not the midpoint.
Pump prices two weeks out. Retail prices follow wholesale prices with a lag, and the model measures that lag. It cannot see a refinery outage, a hurricane or a tax change coming. The diesel version is withheld: over the last year its range missed too often.
Recession probability. This is the New York Fed’s model, not ours. It turns the gap between the 10-year Treasury yield and the 3-month bill into a probability of recession twelve months later. It flagged the recessions of 1970, 1973–75, 1980, 1981–82, 1990–91, 2001 and 2008–09, missed 1960–61 and the pandemic recession of 2020, and gave false alarms in 1967–68 and, most recently, for every month from November 2023 to April 2026, peaking at 71%; no recession has been declared. Its record is on the page; quote it with the number.
Rivers. The Rhine level at Kaub and the forecast are published by the German waterways administration (WSV) and its Federal Institute of Hydrology (BfG); the Mississippi at Memphis by the US National Weather Service. We add no thresholds of our own. Reference lines are drawn only where the agency publishes one: the equivalent low-water level (GlW) for Kaub, the NWS low-water threshold for Memphis.
Scored on 11 October 2026 with the same real-time test as the published
models. The code is in research/outlook/rejected.py in the site’s
repository; anyone can rerun it.
| Idea | Result | Why it is not here |
|---|---|---|
| US goods imports from PortWatch port tonnage, next monthly print | Error 5.9% larger than “no change” | Tonnage does not track import values month to month |
| Core goods CPI over the next 6 and 12 months, adding the NY Fed supply-chain pressure index | Error 3.6% smaller (6 months) and 1.5% larger (12 months) than the same model without it | Too small to matter, and an earlier test of ours overstated it (see below) |
| US goods imports over the next six months from the OECD leading indicator | Error 0.4% smaller than without it | No gain; FRED holds the revised indicator, so even this flatters it |
| 5-year breakeven inflation as a forecast of the next five years’ CPI | Error 1.37 points against 1.40 for “the next five years repeat the last five”; on average 0.65 points too low | No better than the naive guess, and biased |
| US diesel, one, two and four weeks out | Error 2.5–6.4% smaller than “no change” | Its range caught too few outcomes over the last year |
| US gasoline, one and four weeks out | Error 14% and 6% smaller | Both pass. One week out the figure arrives days before the official one; four weeks out the gain is small. We publish two weeks |
| Probabilities from the Stress Index’s own history (“how long do Severe spells last?”) | 7 completed runs at 80 or above since 2008, the longest 24 business days | Too few episodes to estimate a probability |
| EIA’s Short-Term Energy Outlook price path | Not scored | We do not republish a forecast whose record we cannot check |
A correction. Our first test of the supply-chain index let the model learn from inflation figures that would not yet have been published on the forecast date. That put its gain at 11–12%. Tested properly it falls to 3.6% and −1.5%. Every figure on this page now comes from the corrected test.
The data is free to reuse under CC BY 4.0: /data/outlook.json holds every figure on this page, the full backtests and the whole ledger. Quote a forecast with its range, the release it is for and its record — never the midpoint alone.