Every skill-position player's Week 9. Click any row for the game environment underneath the number.
The bands are the part worth reading. They are measured, not guessed, taken from the real distribution of 28,450 player-weeks around their own expectation. That is why a receiver's tenth percentile is a sixth of his projection while a quarterback's is a third: quarterbacks are a far safer weekly bet than the position rankings alone suggest, and that spread is the whole reason start/sit is hard.
Each projection starts from a per-game rate and is adjusted twice. First for the game environment: passing volume scales with the game total and rushing volume with the spread, both fitted within team-season across 7,051 team-games so the adjustment measures game script rather than team quality. A seven-point favorite runs about 3% more; the total moves passing far more than the spread does.
Second for the opponent. Catch rate and yards per catch are blended between what the player has done and what the defense allows to his position, weighted by sample size, so a player with 100 prior targets is trusted more than one with 25.
An honest note on that second adjustment. We measured how much the opponent actually moves a projection here: about 0.11 points on average, roughly 1.4%, and 3.4% at the ninetieth percentile. Those weights are the ones that minimize error, so this is not a shortcoming to be tuned away. It is what the data says. A player's own role and the game environment dominate a weekly outcome; the matchup is small beside them. The defensive numbers are shown in full anyway, because a person deciding between two close options should see them even where the model barely moves.
Early in the season the model leans on last year deliberately. Usage corrects fastest, and by Week 3 half the weight sits on what a player is actually being asked to do, while efficiency and defensive rates move much more slowly, because one game of a catch rate is mostly noise.
The market is in the number. Every stat with a posted prop line is blended toward the de-vigged line at 50%, one stat at a time. A player with a touchdown price but no rushing line gets a market-informed touchdown and a pure-model rushing number. A missing line never counts as a zero. Two corrections matter more than they sound: a posted line is a median while a projection is a mean, and anytime-touchdown prices carry about 23% margin even at the best price available, because there is no "no" side to price against.
Still to come: snap and usage trend from Week 2.
| Rk | Player | Pos | Opp | Proj |
|---|
Sort the board by Proj and start the higher number, once you have set your scoring to match your league. That last part matters more than people expect; half PPR and full PPR do not order receivers and backs the same way, and a six-point passing touchdown moves quarterbacks several spots on its own.
For two players you are genuinely torn between, the start/sit tool puts them side by side with the matchup, the prop lines and the injury status underneath. A gap of a full point is a clear call. Inside two tenths, you are choosing between the same bet twice, and that is when the tiebreakers earn their keep.
All three, and you pick. The receptions toggle above the table recomputes every projection, every position rank and the sort order under whatever you choose; the passing touchdown toggle does the same for quarterbacks.
Nothing is relabelled. The board stores each player as components rather than a points total, so a half PPR number is genuinely built under half PPR rules instead of a full PPR figure with a different name on it.
Volume, efficiency, and the opponent. How often a player touches the ball comes from his role and his team's projected pass and run split; what he does per touch is blended against what the defense he is facing actually gives up to his position, weighted by how much of a sample each side has.
Then the betting market gets a say. The published number is half our model and half the de-vigged prop lines, blended one stat at a time, so a player with a touchdown price but no rushing line keeps a pure-model rushing number rather than being penalised for not being widely bet. The full method is written out if you want the constants.
One thing it will not do is pretend about availability. Anybody ruled out is zeroed rather than left in the rankings at fourteen points because the model has not heard.