2026 · Week 1 · Player props
New England Patriots at Seattle Seahawks
Thursday 10 September at 01:15 UK·66 lines, 19 priced·captured 20 Aug, 11:57
The board
The bookmaker’s number, then ours, then how often his own games have cleared that line. Open a row for the working and the twenty games behind it. Nothing here is a tip, and the rows are ordered by how far apart the two numbers are rather than by anything resembling attractiveness.
Published, not priced by us — 47
A line we hold no comparable per-game column for, or a player with too few games to say anything with. Anytime touchdown is the bulk of it: it has no line to clear, only a price on an event, so the machinery on this page does not fit it and it is left alone rather than given a number that would look like the others.
Last season’s defences
What each side allowed per game by position in 2025, against the league average. This is not in the projections above — measured over 16,523 held-out games, last season’s defensive table is worth −0.01% this early in a season and actively harmful later. It is here because it is true, not because it improves the number.
NE defence
SEA defence
Your own lines
The board above is limited to lines we have captured. Type any player, market and number your bookmaker is showing and it goes through the same arithmetic — recent form, regressed toward the middle of its market, counted against your line with the sample size shown. These stay in your browser and never reach the server.
Nothing on the board yet. Add a player, pick the stat and type the line your bookmaker is offering — 48.5 receiving yards, say — and this shows how often he has cleared it, what the defence he faces next allows, and where the two disagree. Prices are optional; enter both sides and the margin is stripped before the comparison. Everything stays in this browser.
What the number is
A player’s recent games, weighted towards the last five, then pulled back towards the middle of his market. That last step is the most useful thing in the model and the easiest to get wrong: a ten-game average over-predicts, because form regresses and because a player only reaches a ten-game sample by having been worth starting. For receiving yards the line is 3.786 + 0.848 × his average, fitted on seasons before the one being predicted.
It is a line rather than a single multiplier because the pull is towards the middle, not towards zero. A player averaging 15 yards does slightly more than that next week; one averaging 90 does noticeably less. An earlier version of this page multiplied everyone by one constant, which is right on average and wrong at both ends — it ran about five yards high on players averaging 60 or more, which is to say on most of the lines anyone actually bets. The line drops that to under a tenth of a yard.
Nine factors were tested that way, and almost none survived. Home or away, rest, divisional games, wind, snap share and the market’s own implied team total were each worth between −0.02% and +0.17% — noise. Target share made the forecast worse in all four seasons, by 1.21% on the games where it says most, because a player’s role is already inside his own recent average and adding it again counts it twice.
The opponent adjustment is real, but only from games already played this season. Built that way it helps in all four test seasons. Built from last season’s table it makes things worse in every season it covers. So it is applied from week 5 onwards and not before, and a week-1 board like this one carries none at all.
Re-measure rather than trusting this: npx tsx scripts/fit-prop-factors.ts. Two runs are byte-identical.
What this cannot see
Stated because the gaps decide how much weight a number deserves, and a page that lists only its strengths is not research.
Who is playing
No injury or inactive feed is imported. A line that has moved on news will look like a disagreement here, and it is the most likely explanation for a large one.
Whether we are any good
Every projection is written down at capture time with the model version on it, so calibration can be settled later. It has not been settled yet, and nothing on this page claims an edge.
Why the price is where it is
Bookmakers price for balanced liability, not for accuracy. A number that disagrees with ours may be absorbing money rather than making a forecast — and closing lines show no exploitable bias across 6,967 games.
Prefer to type your own lines? The betting page carries a board you fill in yourself, kept in your browser.