The statistical record
Every claim this project makes, with the interval around it. Two of these results are the reason to take the model seriously. One is the reason not to oversell it.
And the game-level claim does, decisively
An interval crossing the dashed line is indistinguishable from no effect
The season projection is the impressive-looking artifact and the weaker evidence. Its edge over a regressed run-differential baseline is +0.142 wins per team-season, and the interval runs from −0.47 to +0.76. It wins in 76 of 150 team-seasons, which is a coin flip.
Reaching 80% power at that effect size would take roughly 5,800 team-seasons, or 193 years of baseball. This is not a problem more work fixes. The honest phrasing is "indistinguishable from the baseline over five seasons", and the game model is the claim to lead with.
Which is why the run model uses a negative binomial
18,180 team-games, 2023 to 2026
Which is how three different values coexisted in the codebase
7,289 games, 2023 to 2025. Lower is better.
A 55% call should win 55% of the time
Points on the diagonal are perfectly calibrated. Dot size is the number of games.
Predictions timestamped before first pitch, and only those
The shaded band is where home-field advantage alone would land