Reconciling Pitcher (x)BABIP and Hard Contact Allowed

This is a long one. I appreciate your patience in advance.

Mike Podhorzer, I and sporadic others have — but primarily Mike has — carried the torch on developing ‘expected’ metrics, such as xBABIP (expected batting average on balls in play), xHR/FB (expected home run-to-fly ball ratio) and xK% (expected strikeout rate), all and the rest of which can be found here. For the uninitiated, these xMetrics help describe how a hitter or pitcher should have performed based on various measurements of the events that unfolded and typically are more predictive of future performance than the original metric. They’re not perfect, but, like other advanced metrics, they give us a better understanding of player performance and ability.

Each metric — xHR/FB, xK%, etc. — has formulas for both hitters and pitchers, with the hitter metrics typically having stronger correlations than those for pitchers. Unfortunately, pitcher xBABIP has always eluded us. It’s inappropriate to repurpose hitter xBABIP for pitchers, but it’s because the model coefficients (weights) would be different, not because the theory underpinning the model is flawed.

That’s the problem, though: hard hits, line drives, infield fly balls — these all should affect a pitcher’s BABIP allowed. Our intuition begs it to be true. Yet there’s a resounding lack of evidence that suggest otherwise. The correlation between BABIP and hard-hit rate (Hard%), line drive rate (LD%) and infield fly ball rate (IFFB%), among others, borders on nonexistent:

• LD%: Line drive rate correlates very weakly with BABIP, producing a 0.12 R². It’s better than nothing, but, unfortunately, LD% has virtually no correlation from year to year, something that rplunkett97 documented here. In other words, line drives do affect pitcher BABIP, but they’re worthless to use in a predictive manner, which is typically the reason why we use these metrics.
• IFFB%: Infield fly ball rate does not correlate with BABIP, producing a 0.05 R². Using IFFB% in a vacuum ignores how frequently a pitcher allows fly balls in the first place; converting it to pop-up rate (PU%, hypothetically) by multiplying IFFB% by overall fly ball rate (FB%) generates a metric that correlates very weakly with BABIP, producing a 0.10 R². It should have a much stronger correlation, given pop-ups are basically free outs.
• Hard%: Hard-hit rate does not correlate with BABIP, producing a 0.03 R². This will be the impetus for the rest of this post.

All together, these metrics do very little to tell us anything meaningful about pitcher BABIP allowed. Line drives result in hits almost 70% of the time and infield flies, for all intents and purposes, are automatic outs. Yet we struggle to demonstrate any kind of meaningful relationship between them and pitcher BABIP. Likewise, you’d think a pitcher allowing frequent hard contact should struggle more than one who doesn’t, yet no dice.

Enter Statcast. It has revolutionized, or at least begun the new revolution about, how we consume player baseball data. (You may or may not be aware of the alleged Launch Angle Revolution in progress.) There are so many tools at our disposal — exit velocity, launch angle, barrels, and so on and so on — that we can basically predict a batted ball’s hit probability based on the results of previous batted balls with similar exit velocities and launch angles. (Statcast’s expected wOBA, or xwOBA, can tell us the probability of a hit but also the probable value of that hit as well.)

The granularity of this data has seemingly demystified many lingering sabermetric arguments. Yet… yet. I’m still struggling to grapple with the enigma that is Robbie Ray. (Yeah, it’s [yet] another post about Robbie Ray and BABIP.) Somehow, it always comes back to Ray.

Ray got BABIP’d to death last year, allowing a historically bad .352 mark in a full season’s work. Now, his BABIP is 75 points lower than last season (50 points lower than his career mark) — and Statcast validates it, suggesting his .206 expected batting average (xBA), per Statcast, stays true to his actual .203 batting average allowed despite allowing hard contact at a rate that almost paces the league (as sorted by 95 MPH+ percentage). Statcast, with its robust and granular data, accounting for exit velocities and launch angles and all that, defends Ray’s performance.

I want to defend it, too, but I’m still not totally sold. It’s really hard for me to swallow the fact that it’s somehow OK that he allows as much hard contact as he does. Statcast’s leaderboard, as previously linked, helpfully provides many of its cornerstone metrics in shorthand as percentages or averages. I’ll use a snippet from Tony Blengino’s contact survivors piece a month ago to further my cause:

Well, first off, let’s not get overly excited. Ray has been extremely lucky across all BIP types, with his Unadjusted Contact Score well below his adjusted mark on flies (92 vs. 119), liners (62 vs. 110), grounders (104 vs. 132) and overall (81 vs. 115). Hitters are batting only .500 AVG-.690 on liners, compared to the MLB average of .650 AVG-.869 SLG. Not a lot of skill in that; Ray’s average liner authority of 96.3 mph is second highest of the above hurlers.

I wanted to strip away all of the extra context and focus exclusively on exit velocity. Such focus assumes that pitchers don’t necessarily have control over where hitters put the ball in play. I think this is more truthful than not. Data-driven pitchers probably understand how their offerings fare by hitter handedness, pitch location in or out of the zone, etc., but they can’t reasonably coax a particular batted ball type, let alone a particular batted ball type in a particular direction, at any given moment. Statcast’s xwOBA and xBA credits the pitcher for these phenomena — credit I don’t think pitchers necessarily deserve in full.

Statcast’s leaderboard simplifies its robust database, but its shorthand representations made it easy for me to run a quick regression of BABIP to the percentage of batted ball events (BBE) of 95+ MPH (among pitchers in 2017 who have allowed at least 190 BBE). Nothing revolutionary — just putting what I perceive to be the collective intuition into code. The correlation was weak bordering on moderate, registering a 0.16 R²; incorporating barrel percentage (barrels per BBE, or Brls/BBE) improves the R² to 0.19, which is encroaching upon legitimate “moderate correlation” territory. (There’s a moderate-to-strong correlation between 95+ MPH% and Brls/BBE.) It’s not a lot, but it’s the best I’ve seen. The correlations are weaker for 2016, which suggests to me MLBAM’s tools and technologies have improved (unsurprisingly).

Accordingly, I calculated new expected BABIPs (hey, xBABIP!) for each pitcher based on both specifications. The former (95+ MPH% only) for Ray: .325 xBABIP. The latter (95+ MPH% and Brls/BBE): .332 xBABIP. These are deliberately simplified models, but they fundamentally betray what Statcast is telling me, telling us. Ray has the 7th-largest discrepancy between his BABIP and xBABIP; the other six have BABIPs lower than .230, so it’s understandable why Statcast doubts them (looking at you, Ervin Santana).

I fully support having more data than less — goes without saying. But I think we lost the forest for the trees. This isn’t to say these results override Statcast ‘expected’ metrics. It’s just that these results are tangible and support what I think many of us intuit about pitchers who allow hard contact: that if we see it on one side of the coin (for hitters), we should somehow see it manifest on its other side as well. It can be theoretically divisive, though, if you think, like I do, that pitchers have less control over where batted balls go upon contact than we may think, or if you disagree. The easy but probably most honest answer is the truth is likely somewhere in the middle, blurred by variance.

So, that’s it. It’s not a big revelation; it just makes me feel a little better about believing that contact quality allowed actually means something. Although I had defended Ray and his inflated BABIP previously, I have had a hard time defending his low BABIP (and remarkably high strand rate) now.

All that said, I can’t leave you hanging without a table. Here’s your table of BABIPs versus xBABIPs.