Math, models, and baseball.
710,000+ pitches. Two seasons of Statcast data. We calculated barrel rates for every MLB pitcher and found the masters of weak contact, the batting practice machines, and the stat that explains modern pitching.
Two years off. A complete rebuild. 10,000 Monte Carlo simulations per game, 1.7 million at-bats of training data, and three seasons of backtesting. We're ready.
We simulated the Dodgers' 2025 season ten thousand times. In 934 universes they won 100+ games. In 2, they won just 67. The real season? 57.7th percentile.
The best models in the world top out around 58% in MLB. In the NBA, 70%+ is routine. A sport-by-sport look at prediction ceilings and why baseball's variance is uniquely brutal.
Our model is a 25-state Markov chain. That sounds intimidating. It's actually just baseball written as math. A visual walkthrough of one at-bat flowing through the model.
We don't predict one outcome — we simulate ten thousand of them. In 3,200 universes the Dodgers won 5-3. In 47 they lost 14-0. Here's what that looks like.
Colorado's park factor is 1.272 — 27% more runs than average. A mile-high stadium where baseballs fly like they're on the moon, and adjusting for it somehow makes predictions worse.
We calculated park factors from 7,292 MLB games and tested whether adjusting our Markov chain simulations for ballpark effects improves accuracy. The answer surprised us.
Parameter sweeps, recency weighting, park factors, bullpen modeling — we ran thousands of backtests across 7,289 games. Here's the honest truth about what moves the needle in baseball prediction.