The 2026 MLB Season by the Numbers: How Our Picks Held Up

The Brewers closed out the Rays in five games, and our MLB model goes quiet until spring. It went live mid-season, on July 30, so this is a two-and-a-half-month record, not a full season: every pick we graded from that day through the last out of the World Series. That covers wins, losses, and a couple of results we'd rather not print.

One ground rule first. The model picks from past performance only: results, scores, opponents, team and pitcher stats. It never sees a betting line. Sportsbook totals show up below only to grade the over/under calls after the fact.

The headline

Moneyline: 451–347 across 798 games, 56.5%.
First five innings: 154–113, 57.7%.
Postseason: 22–14, 61.1%.
World Series: picked Milwaukee in all five games and was right in four. The only miss was Game 4, a 4–3 Rays win.
Totals: 265–275 against the posted line, 49.1%. More on that below.

Is 56.5% any good?

Baseball is the hardest of the major sports to pick game by game. The best team in the league loses about 40 times a year. So the useful question is how the model did against simple ways of picking, not whether 56.5% sounds high.

Always take the home team: 54.8% over the same stretch. The model beat that by 1.7 points.
Always take the team with the better record: 57.5%. The model finished one point behind that.

That second line is the honest receipt. "Pick the better team" is a tough baseline, and by August the standings already know a lot. The model's pitcher matchups and recent form added plenty of opinions, but they didn't add up to more correct picks than the standings alone. That's the first thing to work on before next season.

Month by month

August: 55.4% (417 games)
September: 58.6% (326 games), its best month
October: 53.6% (28 games), mostly the postseason; far too few games to read much into

It was right on more than half the games on 43 of the 76 slates it picked. The best day was August 6 (10 of 11). The worst was August 14 (4 of 14).

Did its confidence mean anything?

A model that says "57%" should be right about 57% of the time on those picks. On the bulk of its picks, it was:

50–55% picks (472 games): said 52.2%, won 54.2%
55–60% picks (237 games): said 57.1%, won 59.5%
60–65% picks (76 games): said 61.9%, won 57.9%
65% and up (11 games): won 9

It was slightly modest on close games and slightly overconfident in the 60–65% band, with nothing wildly off. It also rarely got loud: only 11 of 796 picks reached 65%. That's appropriate for baseball.

Most confident calls

The model's top three picks of the year came in a single series: the Rays at Oakland, August 10–12, at 68%, 69% and 72%. Tampa won all three, 10–6, 12–4 and 8–4.

The two most confident misses came in a single series too: the Red Sox at home against the Athletics on August 8 and 9, at 66% and 68%. Oakland won 7–3 and 4–3.

Teams it read best, and worst

Pick accuracy in games involving each team:

Best read: Dodgers 45 of 59 (76%), Rockies 38 of 50 (76%; it picked against Colorado in 46 of 51 games), Athletics 36 of 51 (71%)
Worst read: Mets 22 of 50 (44%), Orioles 23 of 50 (46%), Tigers 24 of 51 (47%)

Totals: the part that didn't work

Over/under calls went 265–275 against the posted line. That's a coin flip, and a coin flip loses once you pay the juice. Run projections missed the final total by an average of 3.4 runs.

The pattern behind it: the model leaned hard to the over. About six of every seven over/under calls were overs. Those overs hit 50.4%. The few unders it called hit 45.3%. A model that almost always says "over" isn't telling you much about any one game. It also passed on 242 games where its projection sat too close to the line to make a call.

What changes for 2027

Live on Opening Day. This year's record starts July 30. Next year we grade every game from the first pitch, when the standings know the least and a model has the most room to help.
Beat the standings. The moneyline model has to clear "pick the better record" before its extra inputs are earning their keep.
Fix the over lean. The totals model needs to be re-centered so it calls overs and unders at roughly the rate they actually happen.

See you in March.

For entertainment only. 21+. Gambling problem? Call 1-800-GAMBLER.