The Last Start Tells You Almost Nothing About the Next One
We measured the correlation between a pitcher's earned runs in consecutive starts across 13,604 pitcher-starts over three seasons. The answer is essentially zero.
One of the most common patterns in baseball conversation goes something like: "He got shelled last time out, I'm fading him tonight" or "He's been dealing lately, ride the hot hand." The assumption underneath both takes is the same -- that what happened in a pitcher's most recent start carries useful information about what will happen next.
We tested it. Rigorously.
That's not a typo. The correlation between earned runs allowed in start N and earned runs allowed in start N+1, measured across every qualified start in three full seasons, is 0.0008. For context, a correlation of 1.0 would mean perfect predictability. A correlation of 0.0 would mean pure randomness. Our measured value is so close to zero that it's statistically indistinguishable from noise.
A pitcher who gave up 7 earned runs last start is no more likely to get shelled again than a pitcher who threw a shutout. And a pitcher coming off a dominant outing has no statistically meaningful advantage in his next one -- at least not from the fact that the previous start went well.
What We Actually Tested
We didn't stop at one correlation. We built this test because we were looking for any usable "form" signal -- anything in a pitcher's recent history that could improve predictions beyond his season-long baseline stats.
We tested 12 rolling box-score features across multiple window lengths. Every feature was evaluated for its ability to predict the next start's outcome after controlling for the pitcher's established baseline rates.
| Feature Tested | Windows | Predictive Value |
|---|---|---|
| Earned runs (rolling) | 3, 5, 7, 10 starts | None |
| Innings pitched | 3, 5, 7, 10 starts | None |
| Strikeouts | 3, 5, 7, 10 starts | None |
| Walks | 3, 5, 7, 10 starts | None |
| Hits allowed | 3, 5, 7, 10 starts | None |
| Home runs allowed | 3, 5, 7, 10 starts | None |
| K rate (rolling) | 3, 5, 7, 10 starts | None |
| BB rate (rolling) | 3, 5, 7, 10 starts | None |
| K-BB rate | 3, 5, 7, 10 starts | None |
| WHIP (rolling) | 3, 5, 7, 10 starts | None |
| Pitch count | 3, 5, 7, 10 starts | None |
| Game score | 3, 5, 7, 10 starts | None |
Twelve features. Four window lengths each. Forty-eight tests total. Zero passed validation.
None of these rolling features -- at any window length -- improved prediction of the next start's outcome beyond what the pitcher's season-long rates already captured. The 3-start window didn't work. The 10-start window didn't work. Nothing in between worked either.
The Physical Metrics Story Is the Same
Maybe box-score stats are too noisy. Maybe the signal is hiding in the physical data -- velocity, movement, spin. If a pitcher's stuff is changing start to start, that should show up in the underlying pitch characteristics, right?
We tested that too. We built a combined model using velocity, command, spin, movement, and zone rate -- every measurable physical dimension of a pitcher's "stuff" on a given day.
Every measurable dimension of a pitcher's stuff on a given day -- velocity, command, spin, movement, zone rate -- combined, they explain 3.2% of the variance in strikeout outcomes from one start to the next. The remaining 97% is matchup context, sequencing, count leverage, umpire tendencies, and all the other moving parts that make baseball chaotic at the single-game level.
Velocity deviation alone is even weaker: r = 0.017, explaining essentially 0% of the variance. A pitcher who lost 1.5 mph on his fastball might get shelled, or he might throw 7 shutout innings. The measurable "stuff" changes barely move the needle.
Why This Matters
This finding has a direct implication for anyone trying to project pitcher performance: season-long baseline rates are the signal. Recent form is noise.
A pitcher's K%, BB%, HR/9, and other established rates -- built over hundreds of batters faced -- tell you far more about his next start than anything from his last two weeks. The sample size in 3-5 starts is simply too small to distinguish a real change from random variation.
This doesn't mean pitchers never change. Injuries happen. Mechanics get adjusted. New pitches get developed. But those changes show up in the season-long rates as the sample grows. They don't reliably appear as short-term "hot" or "cold" streaks that you can identify and trade on in real time.
We went looking for a recent form signal that could sharpen our projections. We found that the most honest answer is: there isn't one.
The Market Angle
Here's the part that gets interesting for anyone betting pitcher props. If recent form is noise, but the market reacts to recent form, that creates a specific kind of opportunity.
Watch what happens to a pitcher's ER prop line after he gives up 8 runs. Watch what happens after he throws a complete game. The lines move. Bettors have short memories and strong recency bias. The question is whether the books adjust enough to offset that -- or whether some of the recency reaction leaks through into the posted odds.
We're not going to answer that question here. But the autocorrelation finding is the foundation: if you know the signal is zero, you can identify when the market is pricing in something that isn't there.
Methodology
Data source: Walk-forward dataset covering all regular season pitcher starts from 2023-2025. Total: 13,604 pitcher-starts with complete box-score data (earned runs, innings pitched, strikeouts, walks, hits, home runs).
Autocorrelation measured as Pearson correlation between ER in start N and ER in start N+1 for the same pitcher, excluding season boundaries (a pitcher's last start in 2023 is not paired with his first start in 2024). Only starts by the same pitcher in the same season are paired.
Rolling features computed as trailing means over 3, 5, 7, and 10 start windows. Predictive value assessed by whether adding the rolling feature to a model containing season-long baseline rates improved next-start prediction accuracy in out-of-sample testing. Physical metrics analysis used Statcast pitch-level data (velocity, command, spin, movement, zone rate) aggregated to the start level, with a combined regression model explaining 3.17% of game-to-game K% variance (R-squared). Velocity deviation alone: r = 0.017.