Graded in public.
Misses included.
Every version of the model has to beat the easy guess ("he'll do what he did last year") on seasons it has never seen. Here's how it's doing.
Does it beat the easy guess?
Each season was projected using only the three seasons before it, then checked against what actually happened. Lower error is better.
Hitters need 300+ PA and pitchers 60+ IP in both the test season and the year before. Error is the average miss, in the stat's own units.
The fancy stats had to earn it
Version 2 adds Baseball Savant data: expected stats, and whiff rate for pitchers. It learned on two seasons and got one shot at a third it had never seen. Whichever version wins stays in, stat by stat, and the loser sits. The results flip from year to year, which is exactly why the model re-runs this test on its own every time it updates.
Get the model's calls in your inbox.
One free email a week in the offseason, a quick one every morning once games start. Trades graded before your group chat finishes arguing about them.
Free. No spam. Unsubscribe whenever the Cubs break your heart.