It describes what has happened. It does not tell you what to bet.
Every number comes from NFL play-by-play — the same public record the league produces — aggregated over completed regular-season games. Team pace, pass rate, expected points added, explosive-play rate; what each defense has allowed to running backs, receivers and tight ends; and each player’s share of his own offense’s targets and air yards.
Nothing is smoothed, weighted toward a prior, or tuned. If a number looks strange, it’s because the season was strange.
Our baseball model beats the closing line by about half a percentage point. That’s real, and it takes two to three years of betting to tell it apart from luck — in a sport with 2,430 games a year.
The NFL plays 272, across 18 weeks. Evidence about whether a football forecast is any good arrives roughly ten times slower. We could publish picks tomorrow; we could not honestly tell you whether they worked for about a decade. So we don’t publish them.
Confidence here describes a statement, never a wager. “Strong” means the rate is far from the league average and there’s most of a season behind it. “Slight” means it’s close to average, or the sample is thin.
Anything under six games is downgraded automatically, no matter how extreme it looks. A league-leading rate over three games is a rumour.
You will never see “87% confidence on the over” here. That is a pick wearing a lab coat, and it’s the exact thing this model was built to avoid.
Where we do estimate a player’s output, every term of the arithmetic is printed beside it — his target share, his catch rate, his yards per catch, and what the defence he’s facing has allowed. You can disagree with any one of them, which is the point. A number you can’t take apart is a black box, and we don’t ship those.
Where a market number exists, ours never appears without it. Two facts, side by side. We don’t say which to take, and we don’t rank them. NFL player props aren’t posted this far out, so today these estimates sit on the stats and fantasy pages — where there is no price in sight — and nowhere near a bet slip.
Those estimates are explicitly unvalidated, and labelled that way wherever they appear. We’ve written down, in advance, the sample and the bar they must clear before we’d claim anything more — and what we’ll do if they miss.
No, and the difference is worth being precise about, because it is the whole reason one is allowed and the other is not.
Everything on this page is about betting markets, where a bookmaker prices your opinion and takes the other side. To recommend a bet we would have to beat that price. We have tested whether we can, twice, and published both answers: we cannot.
Fantasy has no counterparty. A draft recommendation ranks players against a market price anyone can look up and against a roster you own — and the value figure behind it is a claim about a draft slot (“the market takes him as the RB8; the RB8 returned 243 points last season”), not a forecast of what any player will score. Different claim, different evidence, different answer.
The one probability the Draft Room does quote — whether a player lasts to your next pick — is scored against what actually happened, and you can read the scorecard.
Every skill player on The Week carries a 0–100 index and a letter grade. It is a weighted average of percentiles within his position group, over games already played — where he stands, not where he is going.
The weights are not a judgement call. Each component is weighted by how reliably it has been measured to describe a player, read from a stability study of 35 metrics. That is why snap share and target share carry most of a receiver's grade and yards per target carries very little: one of them stabilizes in a single game and the other needs more than twenty.
What it is not.We tested it against the obvious baseline — a player's own points per game — across 135 season-weeks. It does not meaningfully beat that, and for quarterbacks it loses. So the grade describes a player's standing and is never a projection, a ranking of next week, or a start/sit call. It does beat WOPR, a well-known two-term public formula, by 0.023 — and WOPR is named on the board because of how close it runs.
One number inside it is modelled rather than counted: expected touchdowns, the chance each carry or target had of scoring given where and when it happened, with the player removed. It replaces raw touchdowns, which are the least stable thing in football. Its calibration is measured on five seasons the model never saw, and the places it failed are published too.
Start with this week’s matchups, or read where the lines have moved.