Season review
One forecast in this project was written before the games it predicts. That one is the evidence. Everything else on this page is reported next to it, including what lost.
The champion was the number one pick, and the pick was time-stamped.
The postseason forecast was committed at 2026-09-28T18:26:40+00:00, 23.5 hours before the first Wild Card pitch, at commit e3f719d00119. The Milwaukee Brewers were ranked #1 of 12 at 26.4% and won the World Series. Both pennant winners were the top-ranked team in their league: AL Tampa Bay Rays #1 at 31.2% NL Milwaukee Brewers #1 at 38.6%.
Nothing in that file can have been tuned to the outcome, because the outcome did not exist. That is the only claim here the clock guarantees.
Eleven series, all of them forecast in advance
| Round | Winner | Loser | Games |
|---|---|---|---|
| Wild Card | San Diego Padres | Chicago Cubs | 2-0 |
| Wild Card | Atlanta Braves | Philadelphia Phillies | 2-1 |
| Wild Card | Chicago White Sox | Houston Astros | 2-0 |
| Wild Card | New York Yankees | Boston Red Sox | 2-0 |
| Division | Los Angeles Dodgers | Atlanta Braves | 3-1 |
| Division | Cleveland Guardians | Chicago White Sox | 3-2 |
| Division | Milwaukee Brewers | San Diego Padres | 3-1 |
| Division | Tampa Bay Rays | New York Yankees | 3-0 |
| Championship | Milwaukee Brewers | Los Angeles Dodgers | 4-0 |
| Championship | Tampa Bay Rays | Cleveland Guardians | 4-2 |
| World Series | Milwaukee Brewers | Tampa Bay Rays | 4-1 |
Brier score against two baselines
| Event | n | Brier | vs uniform | vs chalk | Calls |
|---|---|---|---|---|---|
| Title | 12 | 0.0514 | +32.8% | +24.4% | 11/12 |
| Pennant | 12 | 0.0894 | +35.6% | +14.9% | 10/12 |
| Reached LCS | 12 | 0.0935 | +57.9% | +2.1% | 12/12 |
| Wild Card series | 4 | 0.2209 | +11.6% | +3.4% | 3/4 |
Read the chalk column, not the uniform column. Uniform treats every team as equally likely, which is a strawman for reaching the LCS: bye teams are obviously more likely to get there than teams playing an extra round, and a model that says so is not showing skill. Chalk encodes exactly that much knowledge, byes advancing at their historical 56.3% and better seeds at 54.7%, and nothing more.
Against chalk, reaching the LCS scores +2.1%, not +57.9%. The 12 of 12 is the obvious call landing. The title row is the real result.
The two internal strength models disagreed on two of the four Wild Card series and split 1-1. A disagreement that resolves to a coin flip is not evidence for either, and pre-registration is what stops it becoming a story about whichever one happened to win.
The season win-total model, reported at the same size as the win
The game-level model held up prospectively. The season win-total model did not. On 2026 it posted an MAE of 8.07 against a baseline of 6.64, a 21.4% loss, and its 90% intervals covered 63% of outcomes.
Extending the backtest to seven seasons produced the more uncomfortable finding: the ranking of model variants flips depending on which seasons you score. On all seven, the hierarchical variant looks best at +1.69%. Drop 2019, the one season with no component data, and that advantage disappears. Score only 2023 onward and the same variant is the worst of four. No pairwise difference anywhere reaches significance; the best p-value is 0.116.
The honest reading is that these variants are indistinguishable and the apparent winner is an artifact of the window. The season model is not a result this project claims.
Milwaukee
The case study on this site is about the model systematically under-projecting the Brewers. In 2026 it projected them for 89 wins. They won 103 and the World Series. That is the largest miss in the seven-season series, in the year it mattered most, on a bias that had already been written up.
The bigger number is Colorado, over-projected by 15.3 wins a year, roughly 70% worse than the Milwaukee effect and pointing at a specific mechanism: the park adjustment at altitude. That is the next thing being fixed.