The Anatomy of a Slump: What Eleven Straight Losses Actually Tell You

Every July slump gets the same autopsy on television: the clubhouse is tight, the lineup is pressing, they’ve forgotten how to win close games. We wanted the boring version instead. Take every bad three-week stretch a team has had since 2021 — there are 491 of them — measure everything about each one (was it the bats or the arms? real decline or bad sequencing? close losses or blowouts?), and then check what actually happened over the next thirty games. If slumps have anatomy, the anatomy should predict something.

It mostly doesn’t. Across 491 slumps, no detail of the slump itself — its depth, its luck, its close-loss count, its bullpen meltdowns — added any out-of-sample forecasting power beyond two numbers the team already owned: its season record and its season-long expected-contact differential. We tried to break that null two independent ways, including turning a gradient-boosted model with 33 engineered features loose on it. The honest measured lift from everything slump-shaped: +0.000, interval [−.019, +.021]. The slump is the noise. The season is the signal.

Everything this piece measures now runs as a living tool: The Form Table re-computes the 15-game form-vs-luck decomposition for all 30 teams every night, with forecasts for whichever teams are in the slump window that morning.

Four slumps, three diseases

Start with July 2026. The standings said four teams were collapsing at once. The process data says they were doing four different things.

Each dot is a team’s last 15 games through July 24, measured two ways: did the underlying play change (horizontal), and did the results run ahead of or behind that play (vertical). The interesting teams are the ones far from where the standings put them.

Miami’s streak is the strangest object in baseball right now: eleven straight losses through July 25, the first ten all by three runs or fewer, and an 0-4 record in streak games they led or were tied after six innings. The bats genuinely froze — expected wOBA fell .313 → .284 during the streak, bottom-five in baseball over that window — but the pitching never broke: the staff’s expected contact allowed barely moved (.310 → .316) while its results ran 22 points worse than that. Half real cold spell, half coin-flips landing tails eleven times.

Seattle is the ordinary version: hitters running 22 points of expected wOBA below their own established level, run prevention fine. (At our data freeze the Mariners had, absurdly, outscored opponents while going 6-9; Saturday’s 1-7 loss in Texas finally pushed the window run differential negative. That fact was never stable enough to build on — which is rather the point of this piece.) Toronto is the counterintuitive one: the Jays’ blowout losses look like a pitching collapse, but the staff’s underlying contact quality improved during the slide — results simply ran 24 points hot against it while the offense supplied nothing. And the Athletics are what an actual collapse looks like: the largest process decline in baseball (−.065), driven by a bullpen whose expected wOBA allowed went .315 → .395.

The control case matters too: Boston’s 14-1 was the league’s best process differential over the window (+.067), both rotation and bullpen improving, with almost no luck in it (+.010) — a whole-roster leap, not a heater. (Baseball being baseball, they were promptly shut out by Toronto the night we froze the data.)

What predicts the next 30 games

So which slumps do teams escape? We anchored on every 15-game window since 2021 in which a team won six or fewer (491 events, all 30 teams represented), measured each one the same way we measured July’s, and raced models on the next 30 games:

That last row deserves a sentence, because a null is only as good as the effort to kill it. Two frontier models built this analysis independently — separate code, separate methods, forbidden from reading each other until done — and one of them existed purely to attack the null with engineered features (window volatility, opponent strength, momentum, home/away mix) under nested cross-validation. It found nothing transportable. When both the structure-first model and the kitchen-sink model land on the same shrug, we believe the shrug.

Forecast honesty, up front: even the winning model explains about 17% of the variance in a team’s next 30 games, with a typical miss of ±.086 of win percentage. That’s the ceiling. A month of baseball is mostly noise, and anyone selling you a sharper slump forecast is selling.

The things that predict nothing

The useful product of a strong null is the list of stories it retires.

Left: slumps sorted by how “unlucky” they were (results running behind underlying play). Bounce-back rates are flat — and non-monotonic, with the third quartile worst, which is what noise looks like. Right: the famous “good teams bounce back” gap, shrinking as you control for the fact that good teams are simply good.

“They’re in every game” is worth nothing. Teams whose slump losses were mostly close bounced back at 48.4%; teams getting blown out, 48.7%. Identical. “They’re due” is worth nothing. The most process-unlucky quartile of slumps outbounced the most “earned” quartile by four points with an interval spanning zero. And the one we half-expected to survive — contenders bounce back — turns out to be an accounting trick: the raw gap is 24 points, but a hierarchical model that stops counting the 2024 White Sox six separate times cuts it to 10.9 [−0.6, 22.9], and a within/between decomposition shows the information lives entirely at the level of which team you are, not that you’re slumping. Good teams bounce because they’re good. The slump added nothing. One sobering corollary survives everything: slumping contenders return to winning, but not to their prior pace — next-30 records run .038 below their pre-slump season mark. Regression, not restoration.

The lone survivor: cold bats revert. Slumps driven by an offense running below its own established level — Seattle’s kind, Miami’s kind — predicted better next-30s in every specification we and the attacking model ran (direction negative in 14 of 14, roughly 1 point of win percentage per standard deviation of bat-chill). It’s small, and its magnitude wobbles across eras, so we’ll say it the careful way: if you must buy a slumping team, buy the one whose hitters stopped hitting. Broken pitching carries no such promise.

The streak cliff, read honestly

Here is where we tell on ourselves. Our first pass at the losing-streak question produced a terrifying table: teams whose streaks reached nine-plus games went on to play .366 ball with a 13% bounce rate. We nearly led with it. It’s also survivorship arithmetic — that table sorts streaks by how long they eventually got, which no one standing inside a streak gets to know. Both of our independent analyses killed it within a day of each other.

Red: the raw chance of losing your next game, by how many you’ve already lost — it climbs from .48 to .66 and looks like doom. Dashed: the same question after controlling for who’s playing whom, and where. The cliff is mostly who has long streaks, not what streaks do.

Conditioned on team quality, opponent, and venue, streak length adds no statistically resolvable change to the odds of losing the next game (p .29–.43 across control sets; the data can’t rule out effects bigger than ~7 points at nine losses, so we say “unresolved,” not “disproven”). What does survive, in every clean specification, is a small drag on the month: about −0.46 points of next-30 win percentage per consecutive loss. Eleven losses in, that’s a real but modest tax — on the order of one extra loss over the next thirty games, not a death sentence. Beyond twelve losses the historical record simply runs out: 39 team-games from seven team-seasons, most of them the 2024 White Sox. Nobody knows what L13 does, including us.

The forecast, on the record

Which brings us to the payoff. Here’s the identity model’s next-30 forecast for every team in July’s slump conversation, with uncertainty bands sized so that 80% of historical forecasts landed inside them — and a public promise to re-grade this chart in thirty days.

Forecasts frozen on July 24 data. Hollow dots sit outside the model’s slump definition (Cleveland was 7-8; Boston is a surge, scored for fun) — context, not forecasts. The ordering is soft: the model says PHI over SEA at only 57/43.

Philadelphia gets the best forecast (.545) for reasons that are almost embarrassingly simple: it’s a good team, its season-long process is intact, its staff’s ugly three weeks were the most results-behind-process stretch in baseball, and its bats are the cold kind that revert. Seattle (.522) is the same story at a lower altitude. The Athletics (.474) get the worst read because theirs is the one slump the process data confirms.

And Miami? The model that can’t see the streak says .489 — a coin-flip team. The streak-aware analyses, once we fixed both of our first attempts (one simulator flunked its own history exam; one tree model was extrapolating past its training data — details in the methodology), converge on .48, honest band .44–.52, with a 35–50% chance of playing .500-or-better ball the rest of the month. The closest historical comparables — 24 team-seasons that hit seven-plus straight losses while holding a solid record — played .482 afterward and bounced 11 times out of 24. The eleven-game streak is real, agonizing, franchise-record-tying — and worth about one win of forecast. What it’s actually done is what results always do: moved the deadline calculus. That’s a decision about this month’s evidence, and this month’s evidence is, on the record above, the least informative kind there is.

(The nightly-updated version of this scoreboard lives on The Form Table.)

One clear takeaway for the fan watching tonight: when your team drops its fourth straight, check two numbers — the season record and the season expected-wOBA differential. If those are intact, you are watching variance wearing a story. If the bats are the problem, history is mildly on your side. Everything else in the postgame autopsy is, as far as five seasons of data can tell, narration.