The record
How we have actually done
Every call we publish is scored against what happened, and the result is shown here whether it flatters us or not. A results page that can only display good news is marketing.
Every week of the season
One row per man per week — his current call, which is the one we published. Not one row per call: a man re-called five times is still one man.
This is our own model only. Excluded: 13 restated, 245 from a previous model. Repeating a team's announcement scores near 100% and is not a forecast; a model we have retired is not this one. Both are kept in the record and neither is counted here.
A rate needs a denominator. A week marked too few has under 100 scored calls of our own — the figure is shown because hiding it would be worse, and it is not yet a rate anyone should quote.
| Week | Men | Accuracy | When confident | Brier | State |
|---|---|---|---|---|---|
| Week 1 | 0 | we were not live. We hold the outcome for 182 men that week, and no call — a week that did not happen is not a week we got wrong | not live | ||
| Week 2 | 0 | we were not live. We hold the outcome for 251 men that week, and no call — a week that did not happen is not a week we got wrong | not live | ||
| Week 3 | 0 | no calls of our own were scored this week — 245 from a model we have retired and 13 restating a team's own announcement. Kept in the record, not counted as ours | none of ours | ||
| Week 4 | 10 | 60.0% | — | 0.2106 | too few |
By game day
A Thursday-night call has two days of lead on it and a Monday-night call has six. They are not the same product, so they are not averaged together here. Our own calls only — excluded: 13 restating a team's own announcement and 245 from a model we have retired. Kept in the record, not counted here. This is 10 scored calls of our own across the whole season — shown because hiding it would be worse, and not yet a rate anyone should quote.
| Game day | Men | Accuracy | Brier |
|---|---|---|---|
| Thursday night | 10 | 60.0% | 0.2106 |
Called by champion/sides v1: 10/9 families (off/def), regime per week, threshold per side; trained on 16,686 cases 2016-2026 <wk4.
By band
The band is the product, not the decimal. Only the two ends are the ones we stand behind — the middle three are published with their measured rates so you can see why.
| Band | Calls | Accuracy | Brier |
|---|---|---|---|
| COIN FLIP | 3 | 0.0% | 0.3228 |
| LIKELY | 4 | 75.0% | 0.2333 |
| UNLIKELY | 3 | 100.0% | 0.0680 |
By call source
A call that reads the team's own announcement is not model skill, and it is counted separately so it cannot flatter the rest.
| Source | Calls | Accuracy | Brier |
|---|---|---|---|
| before his team said anything | 10 | 60.0% | 0.2106 |
By model
The retired model's rows sit beside the champion's, graded the same way on the same cases. Every n is shown: a rate without a denominator is not a measurement.
| Model | Calls | Accuracy | Brier |
|---|---|---|---|
| champion/sides v1 | 10 | 60.0% | 0.2106 |
Every call in week 4, including the misses
| Player | Team | Band | Call | P(plays) | What happened | Result |
|---|---|---|---|---|---|---|
| Brandin Echols | PIT | COIN FLIP | SIT | 42.5% | played | wrong |
| Elgton Jenkins | CLE | UNLIKELY | SIT | 22.8% | did not play | right |
| Grant Delpit | CLE | LIKELY | PLAY | 65.8% | played | right |
| Jalen Ramsey | PIT | COIN FLIP | SIT | 48.9% | played | wrong |
| Joey Porter Jr. | PIT | COIN FLIP | PLAY | 61.3% | did not play | wrong |
| Rico Dowdle | PIT | UNLIKELY | SIT | 32.9% | did not play | right |
| Teven Jenkins | CLE | LIKELY | PLAY | 78.8% | did not play | wrong |
| Tylan Wallace | CLE | UNLIKELY | SIT | 21.0% | did not play | right |
| Tyson Campbell | CLE | LIKELY | PLAY | 68.7% | played | right |
| Tytus Howard | CLE | LIKELY | PLAY | 68.8% | played | right |