A Bad Start Is Forever. A Good One Isn't.

Every July the same argument starts: which first-half teams are real, and which are about to be found out? We profiled all 30 clubs at the 2026 break over at the team explorer, then went after the question underneath it — how much does a first half actually tell you about the second half to come?

We pooled five full seasons, 2021–2025, 150 team-seasons, and tested it out of sample. The headline number is modest: the first half explains only about a third of the second. But the interesting part is where that third hides and when it shows up. It isn't spread across the league — it's concentrated almost entirely in the basement, it's locked in by Memorial Day, and the “cleverer” stats everyone reaches for don't help at all.

A team's quality is real

Start with the foundation. Take a team's whole season, scramble its games, and split them into two random halves: run differential in one half predicts the other at r = 0.74, pooled across all five seasons. Teams are not coin flips — they have a true talent level, and run differential measures it cleanly. The profiles on the explorer are a real read on how good a club is right now.

But the first half sees only about a third of the second

Now the test that matters for a July prediction: use the chronological first half (games 1–81) to forecast the second (82+). Real, but modest.

Each dot is one team-season, 2021–2025 (n = 150). Pooled r = 0.56, but trained on four seasons and tested on the fifth, the first half explains only ~30% of the second — R² ≈ 0.30. The dashed diagonal is perfect persistence; the cloud falls well short. Roughly two-thirds of the second half is genuinely new.

Almost all of that third is in the basement

Here's the wrinkle that makes the number useful. Split each season into thirds by first-half run differential and the predictability is wildly lopsided: a bad first half is a far surer thing than a good one.

How much of a team's second half its first half explains (R²), by first-half tier, pooled 2021–2025. The cellar's first half explains a third of its second; the penthouse's, almost nothing; the .500 muddle is pure noise. The cellar effect holds in all five seasons; the penthouse correlation is erratic and went negative in two of them.

A team buried in the first half stays buried: 82% of clearly-bad starts stayed under water. A good first half is much shakier — only 67% of clearly-good starts stayed above it. And the regression is asymmetric: good teams gave back -0.57 runs per game on average, while bad teams clawed back only +0.36. Gravity pulls down harder than it lifts up. If your team is in the dead-even muddle, the first half tells you essentially nothing — that's where the real races live.

And you knew it by Memorial Day

The All-Star break is a lagging checkpoint. We measured how much of the (fixed) second half you can predict from just the first N games, and the signal saturates fast: out-of-sample R² runs 0.14 at 10 games, 0.23 at 20, and 0.27 by game 40 — about 90% of everything the full 81-game first half will ever tell you, reached by late May. Watching another two months of baseball barely sharpens the picture. Whatever the first half is going to say, it has mostly said it before the weather turns.

The biggest frauds — and the great escapes

The averages have faces. The most spectacular collapse in our five years was a genuinely good team:

The 2023 Angels — Shohei Ohtani's last half in Anaheim — were five games over .500 and outscoring opponents at the break, then went 29–52. Surprises run the other way but are rarer, exactly as the asymmetry predicts.

What the rule says about 2026

We're now ~75 games into 2026 — well past the Memorial Day lock-in — so the five-season rule already has opinions. The basement is, by the rule, close to decided: Colorado (-1.17 run diff/game), Cincinnati, the Giants, Athletics, Angels, and Royals are all clearly-bad, and four in five such teams stay buried. At the top, only the Dodgers (+1.91), Yankees (+1.66), and Brewers (+1.58) clear the +1.5 line where collapses essentially stop. The Braves (+1.36) are excellent — and, by the rule, the contender most exposed to a 2023-Angels swoon. Everyone else, from the Padres to the Mets to the Red Sox, sits in the muddle, where the first half is silent and the races are genuinely still open.

The simplest numbers win

If a third is the ceiling, which first-half number gets you there? We checked them all against the second half, pooled across the five seasons.

How strongly each first-half number tracks a team's second-half run differential, pooled 2021–2025. Run differential and the raw win–loss record tie at the top; the “cleverer” expected stats carry less; bat speed is near-noise (2024–25 only).

The punchline is deflating in the best way: the two best predictors are the two simplest. Run differential (0.57) and the plain win–loss record (0.58) tie at the top and are interchangeable. Every fancier number carries less — K−BB% (0.48), offense xwOBA (0.39), strike-getting CSW% (0.36) — and bolting any of them onto run differential improves an out-of-sample forecast by 0.000. They're a fine way to explain a team on the explorer; none is a separate window into October.

Bat speed is the emptiest tell

The Statcast era gave us a thrilling new toy — bat tracking — and a tempting new way to be wrong. Team bat speed, the number that lights up broadcasts, tracks the second half at r = 0.14, the weakest of everything we tested. Bat speed is an input, not an output. It measures how hard a team swings, not how well it hits; at the team level that conversion is swamped by approach, contact, and luck. An ability you can admire, not a result you can bank.

The honest part

We earned this the hard way, getting it wrong twice on thin data — first chasing a buy-low board that only predicted the bounce, then, from a single season, nearly publishing that the second half was a coin flip. Both were small-sample mirages; pooling five seasons dissolved them. We also checked the tempting wrinkles and killed the ones that didn't hold: pythag-lucky teams do not regress harder than their run differential implies, and no model we built could call which specific team will flip beyond the base rates. The aggregate is knowable; the individual identity isn't.

We run these as dual-agent studies — an interpretability-first analysis and an independent machine-learning one, plus from-scratch recomputes — precisely so the exciting-but-fragile version doesn't sneak through. So here is the whole method for reading a team at the break: glance at the run differential — or honestly, just the record. If it's deeply negative, believe it; that team is cooked. If it's good, enjoy it but hold it loosely — a third of good first halves don't last. And if it's near even, stop trying to read tea leaves; the season hasn't decided yet. The full half-season profile for all 30 of this year's teams is at the team explorer.