Best of BP 2025: Adversarial Pitch Location

Pitch quality models frequently assess three components:

Stuff: The inherent nastiness of the pitch, considering features like velocity, movement, and spin

Pitch Context: The effectiveness of the pitch when used in combination with other pitches and game situations

Location: The pitch’s destination, because some locations are harder for batters to hit. 

A fourth aspect, the interplay between a pitcher’s collection of pitches, or arsenal, is the subject of ongoing analysis, but BP’s arsenal metrics should be available shortly. 

Thanks to gradient-boosting models, we’re measuring both the Stuff and the Pitch Context of pitches at a seemingly reasonable level. On our leaderboards, these calculations on a pitch-type basis are reflected in StuffPro and PitchPro, with the latter incorporating both context and general location. These and other metrics are a good start, but none of them consider an important reason specific locations are chosen: the batter’s individual weaknesses. Pitchers with effective command don’t just locate pitches in generally good spots: they locate pitches where they know a particular batter will struggle with them.

In other words, there are (at least) two layers of pitch location, and they should be quantified as such. The first is the general value, on average, of locating a pitch in a certain place. This is probably captured, more or less, by existing public approaches to measuring location quality: pitches down the middle are a bad idea, with pitches on the edges usually much better. But the second layer, addressed by this article, is the value, on average, of a particular location specific to each batter, as compared to some other location specific to that batter. Location analysis based solely on overall hitter tendencies ignores how teams choose pitch locations, and how great pitchers defeat hitters. (This could also explain why general location metrics currently struggle to replicate themselves from season to season). We call this second layer adversarial location, and we posit that it is a critical component of preeminent pitcher command.

To isolate adversarial location, we propose using a statistical method that is novel in the public baseball sphere. The method allows us to benefit from overall batter tendencies, while tracking batters who demonstrate unique strengths or weaknesses in different locations. And by incorporating prior information and spatial principles, we can better anticipate batter vulnerability even in locations with small or non-existent samples, using information we have about the batter in neighboring zones or, failing that, average performance from other players.

Along the way, we will confirm something long believed to be true, but to my knowledge never rigorously proven: that the success of “weak-contact” pitchers corresponds to their ability to locate pitches where individual batters least like to see them. In other words, your favorite command artist isn’t just working the edges in general, as reflected in our called strikes above average (CSAA) metric: these pitchers also target each batter’s specific areas of discomfort with ruthless efficiency, using adversarial location to neutralize the batter’s natural advantage in generating batted-ball outcomes. 

The Approach

On the surface, this task may sound easy: (a) divide the strike zone into defined locations, (b) track how each batter does in each location, (c) count how often a pitcher locates a pitch in those zones against each batter, and (d) keep track of how well the batter does overall. Some version of this sequence already drives the ubiquitous “pitch heat map.” Those heat maps, however, rely on raw results that are haunted by sample size and pitch variety. By sheer randomness, batters will never see the same number of pitches in each location, much less the same quality of pitches. These challenges make things unworkable in a hurry, particularly for players with small samples whose strengths and weaknesses we wish to forecast.

But, it can be done. Here’s how:

First, we divide the strike zone into a grid of acceptable granularity. A 5×6 (x, z) grid seems to be sufficient, giving you 30 sub-zones. You could choose a different grid, but the values don’t seem to improve (we’ll discuss why in a moment) and you are at risk of identifying fake trends (overfitting). Remember that the goal is not to model the pitch’s location (which we already know), but the role of the batter and pitcher in the location of that pitch. It is challenging for pitchers to repeat pitch locations within even a six-inch radius, and 12 inches is pushing it for many. And even if a pitcher has great control, the necessary precision is lessened by the batter’s limited ability to perceive the pitch’s location and execute on it. With only a few feet to cover in both horizontal and vertical directions, a 5×6 grid is plenty to get started. Our grid extends across both of what Baseball Savant calls the “heart” and “shadow” regions of the strike zone, while still distinguishing what is technically a strike from what is not. For the time being, it is not customized to individual hitters, but before long it will be.

Second, instead of the machine-learning methods that currently dominate public pitch modeling, we turn to an underappreciated area of statistical theory: spatial statistics. In particular, we take advantage of Tobler’s First Law of Geography: “everything is related to everything else, but near things are more related than distant things.” This matters greatly because it provides a prior distribution over the relationship of locations to one another, as diagnosed by the model. A batter who swings at high fastballs is likely also to swing at fastballs a bit higher than that, even if the tendency is lower. A batter who dislikes inside pitches probably dislikes even more a pitch thrown further inside, and so on. Whatever the result in a particular location, it presumptively gets less similar, in one direction or the other, as the location moves further away. This presumption allows us to predict batter results in locations we haven’t seen yet based on the locations we have seen, and makes it harder for counterintuitive results to stick. 

Third, we rely on what we already know about pitch quality from our PitchPro metric, which already accounts for the inherent pitch “stuff,” as well as the overall location of the pitch, the handedness of the batter, and the context (e.g., count) in which the pitch was thrown. This is our baseline, and only further improvements interest us. Controlling for these qualities helps ensure that we are looking at something new, not double-counting what we already see.

Fourth, we want to incorporate our favorite component of any rigorous model: skepticism. Modeled coefficients a/k/a random effects remain a go-to regularization method, and we use them here as well. If employed properly, batters who show unusual strength or weakness in a particular location will be credited or debited for it, but only if they do so consistently and against similar quality pitches in the same or similar locations. Otherwise, they are presumed to perform the way an average batter would perform from their side of the plate, which makes sense if the batter is an unknown quantity. (This is also why using a finer grid can make little difference: the random effects structure resists overreactions to noise, and converges toward the simplest, least-likely-to-be-wrong choice).

Fifth, we have to decide which pitch events merit this level of analysis. Our StuffPro and PitchPro metrics are built upon a familiar logical tree: is the pitch swung at? If there is no swing, is the result a called strike, a called ball, or a hit batter? If there is a swing, is the result a foul, whiff, or ball in play? And if there is a ball in play, what is the result? 

Although all these contingencies are interesting, balls in play dictate how virtually all runs are scored, and differences in runs scored determine which teams win games. Controlling the quality of balls in play is also how command artists survive, because many of them cannot limit the number of balls put into play: their stuff simply isn’t good enough to miss bats. So, for now we will score pitchers on how well their adaptation to individual batters improves likely results on balls put into play. Likely results are measured in expected run value as modeled by our PitchPro system for launch speed, launch angle, and spray angle.

This gives us the building blocks of our model: we regress (a) batted ball outcomes on (b) pitch identity, location, and general quality through PitchPro, and (c) a location grid, specific to the batter, that tracks the extent that individual batter deviates (plus or minus) from the general trend for their handedness in each location on the grid, mindful of the batter’s performance in nearby locations. For those who care, we use a Matérn covariance function over our location grid. 

To determine a batter’s deviation from the locational norm, we predict the expected run value of the ball in play at each grid location for the batter. We grade pitchers by how the pitch location affects the expected result of the BIP as compared to the batter’s average expected result for balls in play. 

The model runs quickly using glmmTMB, a speedy R package for multilevel modeling, which calculates the best rate of correlation decay across the entire strike zone grid for each season.

Illustrations

A few map comparisons will show why a hitter’s raw results are not only noisy, but sometimes misleading, at least when it comes to adversarial location.  Again, remember that we are not looking at the raw prediction of whether somebody is, in total, good or bad at hitting, although these tools allow us also to answer that question too.  What we want to know is where the hitter is vulnerable as compared to where a typical hitter from their side of the plate is vulnerable.  

A good illustration of the problem presented by raw data is Aaron Judge.  Here are his raw results in each part of the grid for this season, relative to league average:

The most obvious problem is the raw data’s suggestion that Judge is deadly in various areas of the zone except over the very heart of it, which is not very likely. There are multiple zones for which there is little to no data at all, because Judge tends not to put pitches in play from there. Although this provides a signal of a different sort, it doesn’t help us predict how that pitch would do if it did draw a swing. And all these locations have standard deviations so wide that they swallow the raw averages, a problem particularly acute with home run hitters, who present the widest range of potential run outcomes on any given pitch.

The adversarial location model, though, is not troubled by this: using its covariance matrix, and crediting the pitches for which it has the most information over the rest, it provides a more sensible read of Judge’s likely zone, and extends it across the entire region:

Interestingly, the model says that location (3,4) in fact is likely to be Judge’s strongest location, not a unique vulnerability.  It concludes that his strength continues toward the outer side of the strike zone but tapers off after that, particularly as the pitches move down or inside. Of course, Aaron Judge being Aaron Judge, there isn’t any area where the model finds him weak per se, but he is uniquely dangerous compared to other hitters in the yellow areas, even if his raw data suggests otherwise. Regardless, one thing the model is always able to do is to predict an outcome at any location, because even if it lacks data, it learns from how both the specific batter and batters in general tend to perform in nearby locations that are presumed to have relevant information.

Of course, it is also important to know when a hitter doesn’t have unique locational weaknesses. The raw data can be misleading here as well. Consider the raw run values of Alejandro Kirk’s balls in play by pitch location, as compared to the league average:

What a mess. By results alone, Kirk’s adversarial location is only good on pitches above the zone, and alternatingly good or bad every few inches you move.  This doesn’t make sense, unless his hitting zone is largely uniform and the deviations we see are largely deviations from a similar baseline.  Let’s see what the model predicts:

Indeed. Considering all the information, and the overall trends of nearby locations, we have a hitter who gets somewhat better further inside, and somewhat weaker as you trend outside, but overall is pretty consistent.  

Let’s do one more, Stephen Kwan. Kwan’s raw run values on balls in play, as compared to average, provide another patchwork quilt that hints at some curious trends, but not in a way that provides confidence:

Kwan’s “strengths” are a bit all over the place by the raw data.  Pitches trending lower and outside seem to be favorites, while pitches up in the zone are not.  Yet, in location (1,2) there is a possible anomaly that, in addition to liking pitches lower and more outside, he thrives on the pitch up-and-in. You don’t see that every day. Small sample wonder in that location, or does he have a few very-different areas of strength? Our model helps answer that question:

Interestingly, even applying our model’s skepticism, there appears to be something to it.  Low pitches in general are an area of strength, particularly when they are more outside. The map grows cold as you move up in the zone but tilts back toward productivity as you get up and in.  

I like this example because the model is not simply shrinking everybody, everywhere: if the hitter does something unusual, even in an area with less general plausibility, the model can still pick it up, thereby telling you that pitches on this map can benefit from being high, just not too high.

The Results

Having discussed the batter profiles, let’s talk about the results for pitchers.

First, let’s make sure we are answering the right question. There are (at least) two ways to summarize a pitcher’s performance from a model like this. Method 1 is a standard with or without you (WOWY) / marginalization analysis, and summarizes the pitcher’s expected BIP results relative to the league-average expected result against those same batters across those 30 locations. Method 2 compares a pitcher’s results in each location only to their opposing batter’s average performance across their individual predicted grids. The choice makes a big difference:

Table 1: Reliability of Two Summary Methods, 2023-2024 MLB pitchers(Spearman Correlation, weighted by pitch volume over the entire grid)

Comparison Method
Spearman Correlation

Relative to League Average Location for Batter
-0.10

Relative to Individual Batter’s Average Across Locations
+0.50

Correlations close to zero are worthless. Correlations in the .5 range are moderately strong, and for a skill as complicated as pitching, they are a powerful signal that we are onto something. Table 1 suggests that pitchers do not show unique skill at simply targeting areas where batters are generally above or below average (this may be a denominator issue, defined more by the overall quality of the hitter than the location per se).  But they show a definite difference in how they target locations where batters are individually uncomfortable.

This finalizes our definition of adversarial location: the pitcher’s ability to locate pitches where a particular batter gets below average batted-ball outcomes relative to other locations for that batter, not simply locations where a batter gets below average results.

Do pitchers typically thought of as command artists tend to excel at adversarial location, as we suspected? Indeed they do. Consider this summary of 2024 adversarial command results, providing the weighted percentile for the rate at which pitcher minimizes the risk of runs on balls in play by finding the locations where individual batters least like to see a pitch, best and worst:

Table 2: Pitcher Adversarial Location, 2024 MLB Season

(minimum 150 pitches, higher percentile is better)

150 pitches seems to be a good general-purpose threshold: the results tend to make consistent sense, plus we clear out the position players and small-sample wonders. As for the names, having Jose Quintana and Kyle Hendricks at the top is good to see. Likewise, having José Alvarado in the second percentile for purposeful location should not surprise anybody. It also illustrates the impossibility of Paul Skenes.

Here is how the leaderboard has shaped up so far in 2025:

Table 3: Pitcher Adversarial Location, 2025 MLB Season

(minimum 150 pitches, higher percentile is better)

Jose Quintana grabs the crown again, although Merrill Kelly and Zack Wheeler make sense as well. Emerson Hancock has in fact had something go right this year. On the opposite end, we have two Rockies (perhaps a Coors issue?), top Rule 5 pick Shane Smith, and Kumar Rocker bringing up the rear.

How many runs are saved (or lost) by a pitcher’s skill in adversarial location? For 2024, here is a conservative estimate:

Table 4: BIP Runs Saved by Adversarial Location, 2024 MLB Season

(BIP pitches only)

The results are conservative because they include only pitches actually put into play, not all pitches that were or might have been swung at. Expanding into counterfactuals should increase the affected run volume—perhaps to a significant degree—recognizing that while batters see many pitches, they still conclude every plate appearance with a single event. These estimates also don’t reflect the variance that will cause pitchers on both ends of the spectrum to have more extreme results. 

The value of adversarial location should already be reflected in a pitcher’s overall expected runs allowed. Nonetheless, adversarial location is critical to understanding one way that pitchers can and do succeed. Pitchers that have poor adversarial location, either due to lack of overall command or team directive—“just throw it down the middle!”—could be costing their team a win or two over the course of a season, to the extent the pitcher can do better. (And irritating the advanced scouting/R&D team, which might as well not bother printing out those fancy reports). By contrast, pitchers with excellent adversarial location can, and do, succeed with lower pitch quality. And pitchers who can combine great stuff with generally good location and outstanding adversarial location seem to offer the greatest potential for long-term success. 

We have calculated adversarial location values back to 2017, so it’s worth looking at who has had the most success—and failure—over that time period. I suspect the names will not surprise you:

Table 5: Career Runs Saved by Adversarial Location, 2017–2025 MLB Seasons

(BIP pitches only)

The top of this leaderboard is a “who’s who” of the best command artists from the past decade, many of whom featured a terrific change-up (in Kyle Hendricks’ case, he featured two). At the other end is a much different list, although some of it may be by choice, with pitchers who either can or must rely instead on raw stuff or some other aspect of good pitching to get where they want to go. (To stick around long enough to rack up that many lost runs, the pitcher obviously has to be good at something, or just pitch for the Rockies).

Conclusion

Baseball is a zero-sum game. That means every action that undergoes analysis requires the context of the actions of the opponent. We analysts have been lucky, compared to the poor souls in other sports, because more of the variables we’ve historically studied have worked as independent variables. But no variable is truly independent: The path of the ball that leaves the hand and the path of the ball that meets the bat are the same ball. As we continue to dig deeper in our research, and grow more precise with our tools, we’re able to treat these variables as dependent, and thus model them more realistically. Adversarial location is just one example of this next step, but the early returns demonstrate the potential significance of this vein.

We will continue to refine the concept of adversarial location, but welcome any thoughts you have on how it might be better presented or discussed. 

So far, further topics of interest include: 

Subdivision by pitch type, 
Subdivision specific locations on the grid (there are 30 of them, after all), 
Investigating adversarial angles, not just locations,
Publishing our own modeled batter heat maps, and
Grading batter swing decisions based on batter-specific zone locations, rather than simply being in or out of the strike zone.

For the time being, the ratings of all pitchers from 2017 through the present can be found in this Google Sheet.