The NFL Forecasting Challenge. Pup
Put a probability on every game. So do four models, a sealed crowd line, and eventually the market. Everyone is scored with the same formula on the same slate, and every model number was written down and dated before kickoff — so when the season ends there is an argument about skill rather than about who remembered what.
Drag the football between the goal lines to set P(home). Untouched is an absence, not 50.
Leaderboard
Loading entrants and game results…
Tracked entrants / total points
| Entrant | Type | Next game ready? | Submitted | Graded n | Coverage | Points | Brier ↓ | Log loss ↓ |
|---|
Total points rewards participation. Zero graded games means no performance ranking yet. “Submitted” counts real receipts, not a promise that a registered bot will play. The linear-mean control is tracked but excluded from the main comparison.
Common-sample comparison
| Entrant | Type | Next game ready? | Submitted | Shared n | Coverage | Points | Brier ↓ | Log loss ↓ |
|---|
Game-by-game audit
The lines you are playing against
Live forecasts refresh before kickoff, then lock separately for each game. Receipts show source and capture times. The original preseason ledger remains unchanged. Toto’s public input packet contains site data, never other players’ unlocked probabilities.
| Line | Version | Games | Captured | Independent? |
|---|
⚠️ Two of these are not independent of each other, and the correlation above is measured rather than asserted. Averaging four numbers that are mostly the same number is not an ensemble — it is one model with extra steps, and it would look far more confident than it has earned.
The crowd line
At least three touched human forecasts are needed. The crowd uses robust log-odds aggregation and excludes bots. Eligibility closes at kickoff; the next processing tick seals the receipt. People can see model hints, so independence from the models is not guaranteed.
How this is scored
⚠️ Everything below is rendered from /data/model-contracts.json, not typed into this page. If the contract and the page ever disagree, the page is wrong — and it cannot, because there is only one copy.