The Park in Your Swing: Baseball's Buildings Bend the Swings Taken Inside Them. Fenway Bends Them Backwards.

Broadcasters say it every road trip: he’s changed his swing for this park. It’s the kind of claim that used to be unfalsifiable. But Statcast has now tracked the physical bat path — attack angle, swing-path tilt, bat speed, swing length — on every competitive swing for two and a half seasons, which means the question finally has a testable form: does the same hitter, on the same kind of pitch, swing differently depending on the building he’s standing in?

Collectively, unmistakably: yes. Across 649 qualifying hitter-seasons in 2024–25, park-linked swing measurements differ by venue far beyond chance, the park-by-park pattern repeats across seasons at r = .66–.74, and — the part we care most about — predictions frozen from 2024–25 data called the 2026 park pattern for swing-path tilt at r = .73 in one of our two independent designs and .42 in the other, scored against a cryptographically sealed file. (Individual parks come with wide intervals; it’s the league-wide pattern that’s ironclad.) Coors Field flattens the swings measured in it — the one named-park story that passed both designs’ strictest joint test. Yankee Stadium appears to steepen its residents, though only one design resolves it. And the signature follows the building, not the franchise: when the Rays spent 2025 in a different park after the hurricane, their tilt signature jumped with the move.

We’ll show you the fingerprint, the sealed test, and Fenway’s beautiful wrong answer. Then we’ll show you the test that failed, because it’s the one the trade-season takes depend on.

The fingerprint

The measurement is deliberately narrow. Take only tracked swings against ordinary fastballs in the heart of the zone — the most standardized swing situation baseball offers — and control for who is swinging, so the question is how swings measured in a park differ from the same hitters’ swings elsewhere. That hitter-controlled design removes the roster explanation (Rockies hitters aren’t just built differently), and the pitch restriction blunts the other easy objection, that you see different pitching in different places. One of our two analysis pipelines went further and matched every home swing to a road swing by the same hitter on a nearly identical pitch — same type, same hand, median velocity mismatch 0.8 mph — 23,068 one-to-one pairs in all. The fingerprint survives both designs.

Attack-angle venue effects for all 32 physical buildings that hosted MLB baseball in 2024–25 (the Rays and A’s each used two). Whiskers are 95% intervals from the hierarchical model; many individual parks are not resolved on their own — the heterogeneity claim is about the pattern, not every dot. The three named parks in our pre-registration are colored.

Coors Field is the flagship — and the only named park whose claims passed both pipelines’ strictest joint test (its attack-angle and swing-length effects). Swings measured at Coors run 1.46° flatter [−1.77, −1.15] than the same hitters’ swings elsewhere, with the resident-specific estimate larger still; the hierarchical design also sees them about 0.4 mph slower, a single-design result the matched-pairs test leaves unresolved. Yankee Stadium runs the other way: +0.82° steeper [+0.51, +1.13] in the hierarchical design — single-design as well, so we hold it a full shade looser than Coors.

The cleanest evidence that this is about buildings came from baseball’s accidental experiment. The 2025 Rays — same franchise code, different physical park — show a tilt signature of +0.41° at Steinbrenner Field where Tropicana had been −0.17°; the matched-pairs estimate of that gap is over a full degree. (The A’s parallel relocation points the same direction, though in per-hand aim rather than tilt.) The signature followed the building — which, we note carefully, is exactly what both behavioral adaptation and per-stadium measurement quirks would predict. It rules out the franchise. It does not, by itself, pick the mechanism.

We froze the predictions before 2026 could grade them

Park-level baseball claims have a way of evaporating when the next season arrives, so we treated this one the way we’d want anyone to: the park effects were fitted on 2024–25 only, written to a file, hashed, and only then scored against 2026 — data no model was allowed to touch. Adversarial cross-review forced three revision generations of this analysis; each generation was re-frozen before its single permitted scoring, the revisions were mandated by review rather than tuned on results, and re-running the code returns the sealed artifact instead of re-opening the data. All three generations are archived. The gate passed in all three.

Each dot is a park: the frozen 2024–25 tilt effect (x) against what 2026 actually did (y). Both independent pipelines’ correlations shown; the scatter is the matched-pairs series.

Swing-path tilt — the geometry of the swing plane — passes the sealed test in both pipelines (r = .73 [.59, .84] hierarchical, .42 [.19, .61] matched pairs), and it leads the hierarchical design’s rankings; the matched-pairs design ranks bat speed first (.51) — the two estimator families order the survivors differently, but attack angle, tilt, and bat speed all clear the pre-registered gate in both. Swing length is the honest casualty: it replicated beautifully in-sample and then failed the sealed test in both pipelines (r = .06 and .30, both intervals spanning zero). Two pipelines, one verdict. In-sample replication is not a prophecy, which is exactly why we seal things.

Fenway, backwards

Before running the directional analysis we wrote down what the lore predicted: right-handed hitters at Fenway should get pull-happy — everyone knows the Monster turns righties into dead-pull hackers. That prediction didn’t just fail. It failed in the opposite direction.

Attack direction — where the swing is aimed at contact — for right-handed hitters, by park. Positive means toward the opposite field, verified against actual spray on balls in play — and for a righty, the opposite field is away from the Monster. Fenway is seam-red; Coors and Yankee Stadium for contrast.

Right-handed swings measured at Fenway aim +0.93° toward the opposite field [+0.28, +1.59] in our hierarchical design, relative to the same hitters’ swings elsewhere — away from the wall, not at it. Our candidate reading — a hypothesis, not a finding — is that the Monster doesn’t need to be aimed at: a wall that close pays off ordinary contact to left all on its own, so what stops earning is the pull-side sellout, and the swing drifts the other way. The comparison hand fits: left-handed swings at Fenway, facing an ordinary-depth right field, show no resolved effect at all (+0.00° [−0.70, +0.71]) — though we’ll note handedness groups aren’t randomized, so that’s a negative-control comparison, not a controlled experiment. Left-handed swings at Yankee Stadium aim 1.2° more pull-ward — the short-porch story working exactly the way the lore says — and Coors bends everyone oppo by startling amounts (+2.6° righty, +3.5° lefty), consistent with a park whose outfield is a prairie and whose win condition is line drives into space.

We’re publishing the Fenway result as what it is: a failed pre-registration with a coherent replacement hypothesis (the per-hand park pattern repeats across seasons at r = .62–.75), from one of our two designs. File it as strong-but-single-sourced — and file the lore as backwards.

The test that failed

Everything so far is about swings in parks. The trade-season version of the claim is stronger: move a hitter to a new building and his swing will re-bend to fit it. This winter and deadline gave us 36 transaction-verified hitters who changed home parks for 2026, and we knew exactly what each new building “should” do to each swing. So we scored it.

Each dot is a mover: predicted tilt change from the park swap (x) vs the observed change in his home swing (y). The red dashed line is the fitted relationship. The grey wedge is the killer: slopes inside it are what non-movers generate from ordinary year-to-year drift at these sample sizes.

The slope is positive — +0.73 — and it means nothing yet, because the 95% interval [−0.64, +2.03] spans zero and the whole estimate sits inside the drift wedge. We know the wedge is real because we built it the humbling way: hand non-moving hitters fake “predictions” and their matched-volume drift produces a slope of +0.85 — bigger than the real movers’ — from pure noise. Our second pipeline’s independent estimate (+1.24, interval also spanning zero) lands the same verdict. One season of movers cannot distinguish “the swing adapts” from “swings wander.” The estimand is pre-registered and sealed; we’ll re-score it when the 2027 movers arrive, and we’re on record either way.

Why would the building win?

Everything above is measurement. This section is interpretation — hypotheses the data suggested, not findings it proved — but three candidate forces fit the pattern too well to leave unsaid.

First: the park defines what a good swing is worth. A swing is a solution to an optimization problem, and the building writes the reward function — if hitters are optimizers, the swing should follow the money. Read Fenway that way and the inverted result starts to make sense: a wall that close pays ordinary contact toward left without being aimed at, so the expected-value case for selling out to pull collapses, and the measured swing drifts the other way. The lefty null fits the same story: no wall, ordinary reward function, no bend. We stress again — this is the hypothesis that best survives our data, not one the data has confirmed.

Second: physics re-prices contact. Coors is the clean case. Thin air means less drag and flatter pitch movement, and the outfield is a prairie — fly balls need no help at altitude, while line drives into all that space are premium contact. The direction of the residents’ change is what an optimizer would choose: flatter, a touch slower, aimed at the gaps, with both hands. That’s not lore compliance; that’s what a park physically changing the price of contact would look like.

Third: repetition as the delivery mechanism. The fingerprint shows up in visitors too, and some response could be immediate — you can decide to use a wall in one weekend series. Residents show a significant additional bend, which is consistent with a full home schedule and daily batting practice grooving an adjustment in — though that interaction is observational, and it cannot prove the mechanism. If bends accrue with repetition, a hitter half a season into a new building would show exactly the faint, drift-sized mover signal we found. But so would a league where swings never adapt at all. The movers test cannot yet tell slow adaptation from none — that’s not a shrug, it’s a schedule: the estimand is sealed, and the 2027 re-score arbitrates.

One mechanical footnote, also offered as hypothesis: attack direction and contact depth tend to travel together (a relationship we did not measure here), so some of Fenway’s oppo shift may be hitters committing a beat later rather than re-aiming the bat path. Whether it’s aim or timing, our data can’t say.

And a callback for longtime readers: The Adjustable Swing found that hitters own a real, repeatable dial on their swing path — and then couldn’t find anything the dial was wired to. This piece supplies our best current candidate: the dial is wired to the building. That sentence is a hypothesis, not a finding — and Round 2’s job is to prove it or kill it.

What this isn’t

Three claims came out of review dead or capped, and one caveat survives everything. Teams show no support for positive park-fit selection — six roster-construction tests found nothing in the hypothesized direction, and the one association that did resolve leaned weakly negative — so whatever the fingerprint is, front offices don’t appear to be selecting for it. Swing length is dead as a forecastable headline signature, as the sealed test showed (it does replicate in-sample, and its Coors effect passed the joint gate — it’s the out-of-sample stability that failed). And the “adaptation transfer” story is unproven above.

The caveat: most of the fingerprint applies to visitors too — guest and resident effects correlate .49–.74 across parks. Residents show a real additional signature on top (the venue-by-residency interaction is statistically unambiguous — conservatively p ≤ 7×10⁻³ in our cross-reviewed record, with the final model-based tests far stronger), but the shared component means we cannot fully separate the park changing behavior from the park’s tracking system measuring swings slightly differently. Nobody has publicly audited per-stadium bat-tracking calibration; until someone does, the airtight version of this article’s claim is “swings measured in this park bend this way, and the hitters who live there bend furthest.” The resident excess is the component a simple venue-wide calibration offset would struggle hardest to explain — which is different from proof against every measurement story, and we’re saying so.

One clear takeaway for tonight: when the broadcast shows a Rockie slapping a flat line drive into the gap, you may well be watching a measurable, repeating, building-shaped bend in the bat path — one the data can now see coming a season in advance. When it shows a Fenway righty lacing one off the Monster, remember the strange part: his swing, measured, is aimed the other way. And when someone tells you the new acquisition has already changed his swing for the park — that story is still inside the wedge.