MLB · model accountability

How the model
has held up

Every prediction we've made on MLB games, with every result. Pre-game probabilities are frozen at the moment of prediction — the backtest never reads from a future rating.

Hit rate

55.1%

1650 / 2996 games

Home-always baseline

53.5%

Naive: pick home every time

Edge over naive

+1.6pp

What the model adds

Calibration gap

0pp

Avg pred 53.5% vs actual 53.5%

By confidence bucket

Higher confidence → higher hit rate?

Buckets defined by edge from 50%: tossup < 4pp, lean 4–8pp, edge 8–15pp, lock ≥ 15pp.

BucketGamesWinsHit rate
tossup115260852.8%
lean89948353.7%
edge74142056.7%
★ Lock20413968.1%

How we report this

Every row above carries the probability we'd have shown you before first pitch — never a backfilled rating. A game counted as "model right" if the team we favored won, regardless of margin. Hit rate is total wins divided by total finals. No selection bias, no excluded results.

← Tonight's slate