Accuracy
Every forward call we have published, against what the week turned out to be. Scored on the edition as it first went out — not on a correction written afterwards — and updated with each new edition.
16 of 18 scored weeks landed between the low and high figures we published for them.
Our typical miss against the typical miss you would have made by assuming this week matches the same week last season. We are closer.
Both figures are medians, not averages. One freak week — a near-empty year-ago comparison against a normal one — moves an average far enough to describe a season nobody lived through. The freak weeks are not hidden by that: they are listed individually below.
By how far ahead
A call made one week out and a call made four weeks out are different claims, and averaging them together flatters the near one. Each edition names four weeks ahead, so most weeks are graded four times over — once at each distance.
| Made | Weeks scored | Inside the range | Typical miss | Last-year baseline |
|---|---|---|---|---|
| 1 week ahead | 6 | 5 of 6 | — | — |
| 2 weeks ahead | 5 | 5 of 5 | — | — |
| 3 weeks ahead | 4 | 4 of 4 | — | — |
| 4 weeks ahead | 3 | 2 of 3 | — | — |
A dash is a horizon that has not reached 13 scored weeks yet. The count beside it is real and is what we have.
Where we were most wrong
The five worst calls on the record, largest miss first, and they stay here permanently. A record page where the bad weeks are hard to find is an advertisement.
| Week ending | Called | We said | It was | Outside by |
|---|---|---|---|---|
| Aug 22, 2026 | 1 wk ahead, in the Aug 15 edition | 1,477,500 – 1,811,700 | 1,356,400 | 121,100 cwt |
| Sep 12, 2026 | 4 wk ahead, in the Aug 15 edition | 1,516,900 – 1,860,000 | 1,398,000 | 118,900 cwt |
The ledger
Every scored call, newest first. The bar is the range we published; the marker is what shipped. Each row is drawn to its own range, so a wide band and a narrow one look the same width — the question a row answers is whether we landed inside our own stated uncertainty, and how far out we were when we did not.
| Week ending | Called | Expected | Actual | Range published, and where it landed | Miss |
|---|---|---|---|---|---|
| Sep 26, 2026 | 1 wk ahead · Sep 19 edition | 1,548,800 | 1,588,500 |
1,386,500 – 1,688,400
· band ±10%
|
inside
+2.5% vs expected
|
| Sep 26, 2026 | 2 wk ahead · Sep 12 edition | 1,558,000 | 1,588,500 |
1,386,500 – 1,688,400
· band ±10%
|
inside
+1.9% vs expected
|
| Sep 26, 2026 | 3 wk ahead · Sep 5 edition | 1,557,700 | 1,588,500 |
1,381,100 – 1,688,400
· band ±10%
|
inside
+1.9% vs expected
|
| Sep 26, 2026 | 4 wk ahead · Aug 29 edition | 1,584,900 | 1,588,500 |
1,381,100 – 1,844,700
· band ±15%
|
inside
+0.2% vs expected
|
| Sep 19, 2026 | 1 wk ahead · Sep 12 edition | 1,578,700 | 1,598,300 |
1,404,900 – 1,710,900
· band ±10%
|
inside
+1.2% vs expected
|
| Sep 19, 2026 | 2 wk ahead · Sep 5 edition | 1,578,500 | 1,598,300 |
1,399,500 – 1,710,900
· band ±10%
|
inside
+1.2% vs expected
|
| Sep 19, 2026 | 3 wk ahead · Aug 29 edition | 1,606,000 | 1,598,300 |
1,399,500 – 1,869,200
· band ±15%
|
inside
-0.5% vs expected
|
| Sep 19, 2026 | 4 wk ahead · Aug 22 edition | 1,673,500 | 1,598,300 |
1,399,500 – 1,869,200
· band ±14%
|
inside
-4.7% vs expected
|
| Sep 12, 2026 | 1 wk ahead · Sep 5 edition | 1,570,700 | 1,398,000 |
1,392,600 – 1,702,400
· band ±10%
|
inside
-12.4% vs expected
|
| Sep 12, 2026 | 2 wk ahead · Aug 29 edition | 1,598,100 | 1,398,000 |
1,392,600 – 1,860,000
· band ±15%
|
inside
-14.3% vs expected
|
| Sep 12, 2026 | 3 wk ahead · Aug 22 edition | 1,665,300 | 1,398,000 |
1,392,600 – 1,860,000
· band ±14%
|
inside
-19.1% vs expected
|
| Sep 12, 2026 | 4 wk ahead · Aug 15 edition | 1,717,700 | 1,398,000 |
1,516,900 – 1,860,000
· band ±10%
|
118,900
-22.9% vs expected
|
| Sep 5, 2026 | 1 wk ahead · Aug 29 edition | 1,531,200 | 1,631,200 |
1,334,300 – 1,782,200
· band ±15%
|
inside
+6.1% vs expected
|
| Sep 5, 2026 | 2 wk ahead · Aug 22 edition | 1,595,600 | 1,631,200 |
1,334,300 – 1,782,200
· band ±14%
|
inside
+2.2% vs expected
|
| Sep 5, 2026 | 3 wk ahead · Aug 15 edition | 1,645,800 | 1,631,200 |
1,453,400 – 1,782,200
· band ±10%
|
inside
-0.9% vs expected
|
| Aug 29, 2026 | 1 wk ahead · Aug 22 edition | 1,437,500 | 1,403,100 |
1,202,100 – 1,605,600
· band ±14%
|
inside
-2.5% vs expected
|
| Aug 29, 2026 | 2 wk ahead · Aug 15 edition | 1,482,700 | 1,403,100 |
1,309,400 – 1,605,600
· band ±10%
|
inside
-5.7% vs expected
|
| Aug 22, 2026 | 1 wk ahead · Aug 15 edition | 1,673,000 | 1,356,400 |
1,477,500 – 1,811,700
· band ±10%
|
121,100
-23.3% vs expected
|
On the record, not yet gradeable
Calls already published for weeks that have not been reported yet. They are listed here before the answer is known, which is the only time listing them proves anything — every one of them will appear in the ledger above, whichever way it goes.
| Week ending | Called | Expected | Range published | Same week last season |
|---|---|---|---|---|
| Oct 3, 2026 | 1 wk ahead · Sep 26 edition | 1,604,900 | 1,466,500 – 1,680,200 | 1,614,300 |
| Oct 3, 2026 | 2 wk ahead · Sep 19 edition | 1,638,200 | 1,466,500 – 1,785,900 | 1,614,300 |
| Oct 3, 2026 | 3 wk ahead · Sep 12 edition | 1,647,900 | 1,466,500 – 1,785,900 | 1,614,300 |
| Oct 3, 2026 | 4 wk ahead · Sep 5 edition | 1,647,700 | 1,460,800 – 1,785,900 | 1,614,300 |
| Oct 10, 2026 | 2 wk ahead · Sep 26 edition | 1,458,700 | 1,332,900 – 1,527,100 | 1,467,200 |
| Oct 10, 2026 | 3 wk ahead · Sep 19 edition | 1,489,000 | 1,332,900 – 1,623,100 | 1,467,200 |
| Oct 10, 2026 | 4 wk ahead · Sep 12 edition | 1,497,800 | 1,332,900 – 1,623,100 | 1,467,200 |
| Oct 17, 2026 | 3 wk ahead · Sep 26 edition | 1,587,200 | 1,450,300 – 1,661,700 | 1,596,500 |
| Oct 17, 2026 | 4 wk ahead · Sep 19 edition | 1,620,200 | 1,450,300 – 1,766,200 | 1,596,500 |
| Oct 24, 2026 | 4 wk ahead · Sep 26 edition | 1,586,400 | 1,449,600 – 1,660,800 | 1,595,700 |
How this is scored
Fixed before there was a record to report, and versioned. If any of it changes, this section says so and the old rule stays readable — a scoring rule chosen after seeing the results is not a scoring rule.
Scoring rules — version 1, in force since the first edition
- The forecast graded is the one first published. Corrections republish a week at the next revision and never overwrite the original. We grade the first edition that went out for each week, so a number written once part of the answer was visible can never improve the record. Rows whose week was corrected later are marked as such above.
- The actual is the figure first reported, not the latest. Movement gets revised. We take the number printed in the first edition that reported that week and never restate it, so the record cannot drift and you can re-derive it from the archive and get what we got.
- A hit is the actual landing inside the published low-to-high range. The range is a promise about our own uncertainty, kept or broken. We do not score "close enough" against a threshold, because we would be choosing that threshold after seeing the answer. A range wide enough to always be right shows up as a wide band on every row.
- Every horizon counts separately. Each edition names four weeks ahead, so a week is usually graded four times — once from each distance. Nothing is dropped for being an awkward result.
- Typical miss is a median, and it is shown against the naive alternative — assuming the week matches the same week last season. A forecast that does not beat that is not worth paying for, and the comparison is published whether or not it flatters us.
- No rate is stated below 13 scored weeks. A quarter. Until then the count is published and the percentage is withheld, for the same reason the letter withholds a confidence line off a thin base.
- Drafts and unreleased editions are never scored. A rehearsal shown to nobody, or an edition still under embargo, is not a forecast we have made.
What is graded here is the national four-week movement outlook, in cwt, exactly as each edition printed it. Every figure on this page was published before its week happened, in an edition stored as the bytes that were signed — so this is the archive scored, not a record kept alongside it. The internal back-test that shapes the model is a separate thing and stays where it is; nothing on this page depends on it.