Research
The studies behind the engine.
Most changes to the model are tested first, against a pass rule written down before the data is read. We publish studies here, including ones that failed. Not every study is written up, and some engine changes went live before their test finished; each market's model card shows the evidence behind it now.
Contact from balls in play onlyAlmost half of what the engine counted as a hitter's contact was foul balls. Dropping them fixes about 40% of the gap between our hitter projections and what actually happened.Strikeouts at the book's own lineThe model is good at counting strikeouts. It is not yet good at a harder question: will this pitcher go over the line the book set for him?What each engine channel earnsWe switched off, or halved, every hand-set piece of the engine and asked whether the numbers got better or worse on games the engine was never tuned on.Defense and bullpen rest don't fix game totalsTeam defense is real and bullpen fatigue sounds obvious. Neither made our game totals or win probabilities better, so neither shipped.Last season is the biggest untapped signalHow much should one month of at-bats move a projection? We tested the engine's pull-to-average amounts, its handedness splits, time decay and last season. Last season won by a distance.