SlateSims

NFL Game Stack Scenarios

A tournament is not won by nine players each having a nice day. It is won when one game goes off and you own several pieces of it, because players in the same game rise together. So this does not rank players. It plays the slate out thousands of times, solves the lineup that would have won each of those slates, and then asks what those winners had in common. Almost always the answer is a game rather than a player. Deep is how many of your nine come from that one game.

Cheap correlation is the piece of that game nobody is taking. A lineup that stacks a game has to choose which players to stack it with, and the obvious ones are obvious to everybody. The column shows whoever carries the most ceiling per point of ownership, among players with a real enough role to be worth it: he moves with the game almost as much as the popular piece, at a fraction of the ownership, and the salary he saves is what pays for somebody expensive elsewhere. Ownership is never divided by less than one percent, because the ownership model carries about four points of error and the difference between a tenth of a percent and eight tenths is noise rather than an edge. The lineup that won the 1 November Millionaire did exactly this, taking a $2,500 tight end owned by 3% of the field as the fourth piece of its stack rather than the $5,800 running back everyone else used.

There are two win columns, and they disagree. Optimal% is the share of simulated slates whose best possible lineup had that shape. Contest% is the share whose best lineup inside a simulated field had it. A tournament is won by the best entry somebody actually submitted, not by the best entry that could exist, and those are far apart: four-deep stacks are 10% of optimal wins and 41% of contest wins, and the two definitions agree on which game to stack only 40% of the time. Contest% is nearer what a contest pays. Optimal% is the only column here that owes the ownership model nothing. Both are shown rather than one replacing the other, because where they disagree is itself the information.

One column is measured and the rest lean on a model of the crowd. Optimal% comes from the simulation alone. Field%, Contest% and both leverage columns rest on projected ownership, which carries about four points of average error against the two real contests it has been checked on. Scored against the 8 November Millionaire, the leverage ordering had no measurable skill on that slate: the rank correlation between what it predicted and what happened was −0.01 across twelve games, which is noise. It ranked the game that went on to win 5th of 12, and it rated the game that drew the most entries, Detroit against Miami, slightly positive at 1.15×, where it returned 0.29×. That is one slate and twelve games, far too little to call the method good or bad, but it is what there is. Rank with these columns if you like. Do not read 3.0× as "three times the edge".

Free with your email: the field, leverage and cheap correlation columns. Win rates are free to everyone.
Seen in at least

Shapes, with the game names removed

The same winners counted by form rather than by which game it was. 4-2 means four players from one game and two from another. This is the more stable half of the page, because it does not depend on naming the right game, but it still divides by the same modelled field.

How this is built

It keeps the joint structure instead of averaging it away

The projection build already simulates the whole slate thousands of times with team mates correlated, so that a quarterback and his receiver rise and fall together. It then throws that structure away and reports an average for each player, which is what every projection source does.

This page keeps it. For each simulated slate it solves for the lineup that would have won that particular world, then groups those winners by where their points came from. A game only needs a fat enough right tail to show up here. It does not need a high projected total, which is exactly why a board ranked by mean projection cannot see it.

Why the field column is the weak half

To say that nobody else is on a game, you have to know what everybody else is doing. That means simulating a field, and a simulated field is only as good as the ownership it is built from. Ownership here is a model rather than a measurement, and when it was finally checked against a real contest it was systematically over-dispersed: the players it called heavy chalk came in at less than half its number, and the ones it buried came in at nearly three times.

The field is also built to stack the way real entries do, with a quarterback deliberately paired with his own receivers and often somebody from the other side. That matters more than it sounds: a field that only stacks by accident makes deep stacks look rare and precious when they are neither. The rates it stacks at are reasonable assumptions, not measurements, and until they are measured from a real contest the shape ranking should be read as provisional.

What to do with it

Read the win column as which games have the fattest tails on this slate, and the leverage column as a rough ordering of which of those the field is least likely to be crowding. Neither column tells you which game will actually go off. Nothing can. What it does tell you is where the payoff is concentrated if one of them does, which is the part that is knowable before kickoff.