Player props

Five markets: passing yards, rushing yards, receiving yards, receptions and anytime touchdown. The Sharp Model’s projection for each stat is the weekly projection the rest of the site runs on, set against every line the books post for that player.

The line is the consensus, taken from the books that post both sides at the same number at the newest pull: the median of their lines, and the lower of the middle two when there is an even count, so it is always a line some book actually hung and one outlier book cannot move it. The market is then the books whose line sits within 1.0 of that consensus, since a book further off is quoting a different line altogether, and it needs at least 2 of them or there is no consensus at all. The market’s probability for each side takes the median over price and the median under price across those books and removes the vig from the pair in proportion. Anytime touchdown is priced on Yes only, so there is no pair to take the vig out of; its probability is the median Yes price’s own divided by 1.07, a 7% allowance that is stated here rather than measured.

The starting point is the market’s own number, not ours, because the books price a player’s week sharper than any weekly projection does. For yardage and receptions, the consensus line and the de-vigged prices together imply an average for the stat, read off a normal curve with a fixed spread: 60 yards for passing, 24 for rushing, 24 for receiving and 1.8 receptions. A prop line sits nearer the typical game than the average one, and rushing, receiving and receptions run right-skewed, a few big games pulling the average above the typical one, so the curve’s centre sits below the average by 0.15 of the spread for rushing yards, 0.2 for receiving yards, 0.1 for receptions, and not at all for passing yards.

The Sharp Model’s projection then moves the market’s number three-tenths of the way toward itself (a weight of 0.3) and no further: enough to say where we disagree, never enough to overrule a market that knows more than a projection does. On anytime touchdown the starting point is the market’s Yes probability, and it moves the same share of the way toward the Sharp Model’s own chance of a touchdown, which takes projected rushing plus receiving touchdowns as a rate and the chance of at least one as one minus e to the minus that rate. The Sharp proj. column and the figure on each card are the projection itself, before it is weighed against the market.

The lean is where that lands against the price: the side the moved number favours, and the edge is the gap between its probability and the market’s, in percentage points. Stars are bands on the edge: under 3 points is one star, under 6 is two, under 9 is three, under 12 is four, and anything more is five. 3 stars and up are prop picks and get a card. One and two stars are leans, not picks: shown in the tables, counted nowhere. On anytime touchdown a lean on No has no price to take, so it reads “No (no price posted)” and is never a pick.

Guards. No lean for a player with no projection, or a projection of zero, for the stat; none where fewer than 2 books post both sides at the same number, or, on anytime touchdown, fewer than 2 books post Yes. A player listed Out, IR or Doubtful when the leans publish gets no lean at all, not merely no pick: his prices stay on the board with nothing set against them. A name the books spell in a way we cannot match stays on the board with no lean rather than being guessed at, and two books spelling one player two ways are one line.

Two pulls a week, three in the weeks with Saturday games, because every game costs 5 requests, one per market, out of a free plan of 500 a month shared with the game lines. Thursday at 20:10 UTC pulls the Thursday game; Saturday at 22:10 UTC pulls every game still to kick off, Sunday and Monday; and in weeks 15 to 18, Saturday at 16:00 UTC pulls the Saturday games that kick off before the evening pull. A pull that would leave fewer than 40 requests for the month is skipped whole, and no game is pulled after it kicks off.

A lean is frozen the first time a pull sees it before kickoff: the side, the line, the best price and its book, the projection, the two probabilities and the stars never change after that. A later pull adds prices to the board and nothing else. When the week’s stats land, each row is graded against its frozen line on its frozen side: right, wrong, or a push on the number. A player who did not play is void and counted nowhere.

There is no backtest. Historical prop prices are a paid product and we do not hold them, so the record starts with the first graded week and nothing is filled in behind it. The constants above are stated so they can be argued with, and they stay where they are until there are graded weeks to fit them on. This page is a record of a system, kept in public. It is not a recommendation to do anything with it.