Omaha
The simulation engine behind every number we publish, and how it has changed.
Omaha runs the game forward play by play, thousands of times, and records what each player did in every one. Versions are numbered so that a figure you read last month can be traced to the engine that produced it.
Each release below is measured the same way: the model is rebuilt for every week of a season using only what was known at the time, then scored against what actually happened. We publish where it beats a player’s own season average and where it does not. The second list is not shorter than the first.
These are historical measurements, not a forecast of future accuracy. Every figure comes from rebuilding the model for a season that has already been played and scoring it against what happened, so a different season, a rule change, or a change to the engine can move it. And accuracy here is measured against a player’s own season average, never against a market price. Nothing on this page is a claim about betting returns, which we have not measured.
The simulation now knows the difference between a throw to the first receiver a quarterback looks at and a checkdown to his running back. Receptions and targets, the two weakest figures on this page, improve the most.
Questions with a whole-number answer are answered in whole numbers again. The chance of a receiver catching two or more passes was reading 23.5% where it should have read 31.3%, against an actual 32.7%.
The engine can now tell a tied game from a three-point lead. None of the eleven figures below moved: the average is +9.7% before and after, and what improved is confined to close games late.
Simulations come back about nine times faster. This is the same model rather than a better one, and three of the eleven figures below moved the wrong way.
Fourth and one and fourth and twelve were being treated as the same situation. So were first and goal from the one and first and goal from the five.
On second and twelve, teams throw. The simulation was calling it the same way it calls second and eight, because it could not see how far there was to go.
When a team's starting tight end is out, real teams line up differently. The simulation kept calling the same formation and sent the backup out instead.
Near the goal line, real teams hand the ball to different people. The simulation was choosing a carrier as though it were standing at midfield.
The same team does not call the same game every week. The simulation was treating each team's tendencies as settled, so it produced a narrower range of games than football actually produces.
When a game stops being competitive, real teams start resting their starters. The simulation was playing its best players to the final whistle.
Before kickoff, the simulation was too sure of itself. It now leans harder on what a player has actually done this season, and every figure improved or held.
Who catches the ball now depends on how far it traveled. Every figure we publish is now better than a player's own season average, targets, the last holdout, has turned.
Rushing figures now count kneel-downs, the way official statistics do. Our quarterback rushing numbers were understated by about a quarter and now match what you would look up elsewhere.
The simulation now runs a two-minute drill at two-minute-drill speed, kneels out a win, and keeps track of timeouts. Ten of the eleven figures we publish improved.
The offensive line now counts. Overall accuracy is unchanged; late-game accuracy improved by half a point.
Play calling now reads the clock as well as the scoreboard. The gain is small overall and much larger late in games, which is exactly where it should be.
Ten of the eleven figures we publish are now more accurate than a player's own season average. Rush attempts, the weakest of them, has improved by forty-three points across six releases.
Players are now simulated for the team they play for, and a player who changed clubs in the off-season brings his role with him. Rush attempts improved by nineteen points.
The simulation now knows who dressed. Every figure we publish improved, and the ones about how often a specific player touches the ball improved a great deal.
Play calling now responds to the scoreboard. A team two scores behind throws far more often than one two scores ahead, and until this release the simulation called the same game either way.
Running back workloads are noticeably closer to reality, because the simulation now decides whether a play is a run or a pass before it decides who is on the field for it.
The first version measured end to end by walking forward through a full season, one week at a time, against what a reader could work out from a player's own season average.