How it works
We are not trying to be more accurate. We are trying to be earlier.
A Saturday call that every beat writer already agrees with is worth nothing. The same call on Wednesday is worth something even if it is no more correct. So the thing we maximise is lead time on a call we do not later regret.
What a call has to survive to count
We call a forecast bankable only when three things are true at once: it is confident, it is stable — no later day in the week flips it — and it turns out correct. The middle one is the point. A Wednesday call the model abandons on Friday is not an edge; it is luck we would have talked ourselves out of.
The model
Every case starts at the measured play rate for its position and injury site — a quarterback with a back problem and a lineman with a knee are not the same question — and the model learns only the departure from that baseline. It is fitted by Newton-Raphson on fifteen seasons of injury reports joined to who actually took a snap, and validated leave-one-season-out, because a sweep pooled across seasons reversed its own conclusion twice when we held a season back.
Concussion is modelled separately. It is not a club decision — the protocol decides, with an independent consultant — and it measurably behaves differently from every other injury.
What we cannot do
The book is roughly forty will-play calls for every will-sit call, and the sit calls are the least reliable thing we publish. At this base rate almost nothing moves a case far enough to call a scratch with confidence. So the product is this player will be available, and we knew on Wednesday — never this player will be scratched. Roughly two cases in five never become bankable at all, on any day of the week. We would rather say so than manufacture a call.
Where the numbers on this site come from
Backtest is not the live record
Figures describing how the method behaves — how often a Wednesday call still agrees with Saturday, how much of the book is callable early — are backtest results, measured on three seasons of dated forecasts scored against real outcomes. They are not our live record, they are labelled backtest wherever they appear, and the two must never be added together. The live record is on the record page, and it says not yet measured until a game has actually been played.
The record cannot be edited
A forecast that can be revised after the fact proves the opposite of what it
exists to prove. Ours is append-only — not by convention but by database grant: the
account this system writes with holds no UPDATE and no
DELETE on the forecast table, and the database refuses a call timestamped
after kickoff. If we change our mind during the week, both calls stay on the
player's page.
Where the data comes from
Official club injury reports and participation data via nflverse, used under CC-BY-4.0. Every capture is stored with its licence and the raw bytes are kept so any call can be replayed exactly as it was made.