We Tried to Build a Pressure Grade. Most of It Was Skill in a Costume.

There’s a pitcher on your team you don’t trust with a one-run lead. There’s another you’d hand the ball to in any inning of any game. The belief that some arms handle pressure and others come apart is one of the most durable ideas in baseball — and one of the least measured. So we tried to measure it: build a single “Pressure Grade,” league-wide, that says who actually pitches well when the game is on the line.

We ran it as a two-method bake-off — a Bayesian pipeline and a gradient-boosted ML pipeline, each working the same six seasons of Statcast independently, with the win conditions written down in advance. The point of two methods is simple: when both land on the same answer, you can trust it; when they split, the split tells you what’s fragile. They converged. And the answer to “can you grade who handles pressure” turned out to be a careful, three-part no — and here’s the part that’s yes.

None of that is the article we expected to write. It’s the one the data supports, and — because this is the rare myth-bust that ships a working tool — every claim below is something you can now check yourself on any pitcher’s page.

1. The number everyone wants — and why one number lies

Start with the read that does work, because it’s the reason the whole project was worth doing. Instead of one grade, score each pitcher on separate axes — command, stuff, and contact suppression in high-leverage spots, role-relative. Now a profile tells a story a single number can’t.

Click any card for the live profile. The point is Andrés Muñoz: the league’s nastiest reliever (A+ stuff) and one of its leakiest strike-throwers (D command), the same arm. A single “pressure grade” would file him at the 67th percentile and bury exactly the thing that makes him volatile. Chapman has no tax to pay; Castillo is eroding on every axis at once.

That Muñoz split is the entire case for going multi-dimensional. It’s legible, it matches what you see when you watch him walk the leadoff man and then strike out the side, and — crucially — it’s not an artifact of one method’s choices. We’ll prove the two pipelines agree on it in a minute.

The same split sorts the whole league. The pressure-proof — elite stuff and command, no weakness to exploit — are arms like Aroldis Chapman, Josh Hader, Garrett Whitlock, and Gabe Speier. The volatile pair an ace-level out-pitch with shaky control: Muñoz, Bryan Abreu, Jesús Luzardo — filthy, and a walk from trouble. A single grade calls both groups “good”; the profile tells you which kind — the difference between trusting a lead and holding your breath.

So what do those letters actually measure? Here are three contrasting arms with the real high-leverage numbers behind every grade — the same inputs the model scores, each shown next to the average for the role:

Three cards are three dots in a much bigger cloud — and 450-odd dots is just a blob. The clearer way to read the whole population is by the shape of the grades: the recognizable pitcher types, each with its own command/stuff/contact profile.

Bars are each type’s average percentile per axis (high = better); click any arm for its profile. “Power, leaky command” is the volatile bucket — ace stuff over thin control, where Muñoz, Bryan Abreu, and Jesús Luzardo live. These are descriptive shapes of the current grades, not a stable typology (we tested that — it didn’t hold).

Prefer the raw cloud? Here is every graded pitcher on all three axes at once — drag to rotate the cube, search any arm to pin and label it, and click a dot for the profile:

Command, stuff, and contact are the three spatial axes; color marks role (blue starters, orange relievers) and nearer dots are larger. Spin it to read the structure — the complete arms cluster in the high-on-all-three corner, the volatile arms hug the stuff axis with nothing on command.

2. What persists, and what’s a costume

A “skill” has to repeat. If a number is real, a pitcher who posts it one year should tend to post it the next; if it doesn’t carry over, it was luck wearing a label. So we took every pitcher with a real workload in back-to-back seasons and correlated each metric with itself, year over year.

Skill metrics (green) repeat strongly. The leverage-specific ones — a FanGraphs-style Clutch score, and the gap between a pitcher’s high- and low-leverage results — sit at 0.08 and 0.03: noise. WPA-per-leverage lands in between, because most of it is just skill leaking through.

This is the sabermetric consensus on hitter clutch, now confirmed on pitchers over six seasons: the residual — how much better or worse a pitcher does in big spots than his own overall level — does not stick. The pitcher who “came up huge” in September wasn’t accessing a stable trait. He was good, and the sample was small, and small samples in high leverage are where narratives are born.

But notice the callout under the chart, because it’s the part that rescues the question from pure nihilism: a pitcher’s overall skill this year predicts his high-leverage results next year (r = 0.29) better than his actual high-leverage line does (r = 0.13). Pressure performance is forecastable. You just forecast it with talent, not with a clutch stat — because talent is measured on thousands of pitches and the clutch stat on a few dozen.

3. Two methods, one honest disagreement

Here’s where dual-agent earns its keep. We computed the per-pitcher grades two completely different ways — a hierarchical Bayesian model that shrinks small samples toward the role mean, and a gradient-boosted model that learns its weights from what actually forecasts next-season results — and asked how much they agree.

The descriptive sub-grades replicate across methods: two engines with nothing in common grade stuff at r = 0.95, contact at 0.80, command at 0.72. The one thing they don’t reproduce is the single composite ranking (0.46) — the same null from a third angle.

Read that carefully, because it’s the cleanest result in the study. When you ask the two methods “how good is this pitcher’s stuff under pressure,” they answer the same. When you ask them to collapse everything into one rank-ordered Pressure Grade, they diverge — because there’s no stable signal in the middle of the pack to rank by. The multi-dimensional description is robust. The one-number verdict is method-dependent fiction.

4. We pre-registered the kill shots

The first pass produced one overconfident sentence — “there is no pressure trait at all” — that the cross-review caught: our test had quietly over-controlled. So we ran a convergence round with the corrections baked in and the thresholds fixed in advance, and made both methods answer the same three questions.

Both methods, all three questions, same side of the line. The corrected residual is weak — about 0.08–0.12 — not the zero our first draft claimed, and not large enough to bet on either.

Note the second row, because it’s a finding we wanted to be true and killed anyway. The intuitive idea that pitchers change what they throw under pressure — abandon the slider, lean on the heater — showed up at a tempting strength in one configuration. Vary the thresholds even a little and it falls apart. A real effect survives reasonable changes to how you measure it. This one didn’t.

5. So what do you do with it

You ship the part that’s real, labeled honestly, and you don’t ship the part that isn’t. Every pitcher’s page now carries the Pressure Profile — the command/stuff/contact read, role-relative, with the uncertainty stated — right next to a Scouting Report of league percentiles and year-over-year trajectory, and a Match Tracker that follows velocity and breaking-ball movement start by start. What you won’t find anywhere is a single “Pressure Grade” leaderboard, because we couldn’t build one that survives its own holdout.

That’s the honest shape of it. “Who handles pressure” is a real question with a real answer, and the answer is mostly: the pitchers who are good, and especially the ones who throw strikes. The multi-dimensional profile is worth having because it tells you how a given arm gets there — whether he’s a Chapman with no weakness or a Muñoz one leaked walk from disaster. It just won’t tell you whether he’ll come through tonight any better than knowing he’s nasty already does. Anyone selling you that number is selling you a costume.