One row per player per market for the coming week. "Our Proj" is the model's
median projection, with its middle-50% range (25th–75th percentile) underneath.
"Line" is the sportsbook number we priced against, with that book's over and under
prices. "P(Over)" is our calibrated probability the player clears that line, and
"Edge" is how far that sits from the book's own vig-free probability — shown both in
probability points and in stat units (our projection minus the line).
The model
Each market gets its own model, matched to the shape of the stat. Yardage markets
use a quantile ensemble that predicts the whole distribution rather than a single
number; reception and attempt counts use a count model with its own spread parameter;
rare events like passing touchdowns use a shrunk count model. Every model is trained
on nflverse player-game data from 2021 on — usage and role, team pace and pass
tendencies, opponent defensive efficiency and pressure, weather, and game script from
the market's own spread and total. Raw predictions are then pulled partway toward the
book's number (the blend weight is fit, not guessed) and passed through a calibration
curve so a stated 60% actually means 60%.
Lines and no-vig prices
Books post several lines for the same player. We price the one whose two-way
prices sit closest to an even market — nearest to −110 on both sides — because that
line carries the least built-in margin and the cleanest implied probability.
Stripping the margin out of both prices gives the "no-vig" probability: what the book
really thinks, once its cut is removed. A "fair price" is the American odds that
match our own probability with no margin at all; if the posted price is better than
our fair price, the bet has positive expected value.
Edges and official picks
A row becomes an official pick only when several things line up: the market is
approved for picks, our edge clears 4.5 probability
points and the market's stat-unit threshold, a sharp reference book does not clearly
disagree, the price isn't worse than −130, and the play survives per-player, per-team
and per-slate exposure caps plus a correlation guard. Stakes are
25%-Kelly on our calibrated probability at the posted
price, capped. During the shadow period picks are recorded internally and graded, but
not published here — nothing goes on the board until the record is real.
Closing line value
CLV is the honest short-run scoreboard. It measures whether the market moved
toward our number between the moment we priced a pick and kickoff. Prop samples are
small and variance is large, so a handful of weeks of win-loss tells you almost
nothing — but consistently beating the closing line is very hard to fake. We publish
mean and median CLV, and the share of picks with positive CLV, alongside the record.
Model accuracy
"Model Acc" is the model's season-to-date record on that player in that market:
the share of graded games where the side we leaned to (over or under, against the
line we priced) is what the player actually did, with the count underneath. Pushes
and games the player did not play are not counted. The smaller line shows the same
record across all of the player's markets. A dash means nothing has graded yet for
that player this season, so treat the first few games as noise — two or three
results say very little.
The prop card
Click a player's name to open their prop card. It shows every prop we priced for
them this week, and four tabs: Overview (projection vs the line,
P(over), edge, the model's record on this player, and hit rates against today's line
over the last 5, last 10, this season, and head-to-head), Matchup
(the player's card from our matchups board: PFF-graded player rank against the
opponent's defensive units), Hit Rate (game-by-game results against a
line you can move, over or under, with our projection as the last bar), and
Odds (every book's line and prices from the latest snapshot, best
price highlighted, Hard Rock alternates).
Past weeks and results
Use the week picker in the header to walk back through earlier boards. Once a
game is graded (the morning after it ends), the "Result" column shows what the
player actually did and whether our side was right; a pick that voided says why
(DNP, no line). Past-week boards show every game by default.
The confidence badge
High, Standard or Watch reflects how well our projections have tracked that
specific player in that specific market over time — accuracy against the line,
whether our range has covered the actual results, and whether we've been persistently
high or low. Players we've modelled poorly get no badge and no pick. New or
low-sample players sit at Standard until there is enough history to say more.
No published record yet. Official picks start after the shadow period —
until then every pick the engine makes is logged and graded privately so the first
published number is a real one. Methodology is on the
methodology page.
Past board. This is the Week 5 board as it was priced;
graded results and the season-to-date model accuracy are shown alongside each row.
Back to the current week →
