Best of BP 2025: Adversarial Pitch Location

Pitch high quality fashions continuously assess three parts:

Stuff: The inherent nastiness of the pitch, contemplating options like velocity, motion, and spin

Pitch Context: The effectiveness of the pitch when utilized in mixture with different pitches and recreation conditions

Location: The pitch’s vacation spot, as a result of some places are tougher for batters to hit. 

A fourth side, the interaction between a pitcher’s assortment of pitches, or arsenal, is the topic of ongoing evaluation, however BP’s arsenal metrics ought to be accessible shortly. 

Because of gradient-boosting fashions, we’re measuring each the Stuff and the Pitch Context of pitches at a seemingly cheap degree. On our leaderboards, these calculations on a pitch-type foundation are mirrored in StuffPro and PitchPro, with the latter incorporating each context and basic location. These and different metrics are an excellent begin, however none of them contemplate an essential motive particular places are chosen: the batter’s particular person weaknesses. Pitchers with efficient command don’t simply find pitches in typically good spots: they find pitches the place they know a specific batter will wrestle with them.

In different phrases, there are (not less than) two layers of pitch location, and they need to be quantified as such. The primary is the basic worth, on common, of finding a pitch in a sure place. That is in all probability captured, roughly, by present public approaches to measuring location high quality: pitches down the center are a nasty thought, with pitches on the perimeters normally a lot better. However the second layer, addressed by this text, is the worth, on common, of a specific location particular to every batter, as in comparison with another location particular to that batter. Location evaluation based mostly solely on general hitter tendencies ignores how groups select pitch places, and the way nice pitchers defeat hitters. (This might additionally clarify why basic location metrics at the moment wrestle to duplicate themselves from season to season). We name this second layer adversarial location, and we posit that it’s a crucial element of preeminent pitcher command.

To isolate adversarial location, we suggest utilizing a statistical methodology that’s novel within the public baseball sphere. The tactic permits us to profit from general batter tendencies, whereas monitoring batters who display distinctive strengths or weaknesses in numerous places. And by incorporating prior data and spatial ideas, we are able to higher anticipate batter vulnerability even in places with small or non-existent samples, utilizing data now we have in regards to the batter in neighboring zones or, failing that, common efficiency from different gamers.

Alongside the way in which, we’ll verify one thing lengthy believed to be true, however to my information by no means rigorously confirmed: that the success of “weak-contact” pitchers corresponds to their skill to find pitches the place particular person batters least wish to see them. In different phrases, your favourite command artist isn’t simply working the perimeters typically, as mirrored in our known as strikes above common (CSAA) metric: these pitchers additionally goal every batter’s particular areas of discomfort with ruthless effectivity, utilizing adversarial location to neutralize the batter’s pure benefit in producing batted-ball outcomes. 

The Strategy

On the floor, this process could sound simple: (a) divide the strike zone into outlined places, (b) monitor how every batter does in every location, (c) rely how typically a pitcher locates a pitch in these zones towards every batter, and (d) hold monitor of how nicely the batter does general. Some model of this sequence already drives the ever-present “pitch warmth map.” These warmth maps, nonetheless, depend on uncooked outcomes which are haunted by pattern dimension and pitch selection. By sheer randomness, batters won’t ever see the identical variety of pitches in every location, a lot much less the identical high quality of pitches. These challenges make issues unworkable in a rush, notably for gamers with small samples whose strengths and weaknesses we want to forecast.

However, it may be completed. Right here’s how:

First, we divide the strike zone right into a grid of acceptable granularity. A 5×6 (x, z) grid appears to be enough, supplying you with 30 sub-zones. You could possibly select a unique grid, however the values don’t appear to enhance (we’ll focus on why in a second) and you might be liable to figuring out faux tendencies (overfitting). Keep in mind that the aim is to not mannequin the pitch’s location (which we already know), however the position of the batter and pitcher within the location of that pitch. It’s difficult for pitchers to repeat pitch places inside even a six-inch radius, and 12 inches is pushing it for a lot of. And even when a pitcher has nice management, the required precision is lessened by the batter’s restricted skill to understand the pitch’s location and execute on it. With just a few ft to cowl in each horizontal and vertical instructions, a 5×6 grid is lots to get began. Our grid extends throughout each of what Baseball Savant calls the “coronary heart” and “shadow” areas of the strike zone, whereas nonetheless distinguishing what’s technically a strike from what just isn’t. In the intervening time, it isn’t custom-made to particular person hitters, however earlier than lengthy it will likely be.

Second, as an alternative of the machine-learning strategies that at the moment dominate public pitch modeling, we flip to an underappreciated space of statistical idea: spatial statistics. Specifically, we reap the benefits of Tobler’s First Legislation of Geography: “every part is said to every part else, however close to issues are extra associated than distant issues.” This issues significantly as a result of it offers a previous distribution over the connection of places to 1 one other, as identified by the mannequin. A batter who swings at excessive fastballs is probably going additionally to swing at fastballs a bit larger than that, even when the tendency is decrease. A batter who dislikes inside pitches in all probability dislikes much more a pitch thrown additional inside, and so forth. Regardless of the end in a specific location, it presumptively will get much less related, in a single course or the opposite, as the situation strikes additional away. This presumption permits us to foretell batter ends in places we haven’t seen but based mostly on the places we have seen, and makes it tougher for counterintuitive outcomes to stay. 

Third, we depend on what we already learn about pitch high quality from our PitchPro metric, which already accounts for the inherent pitch “stuff,” in addition to the general location of the pitch, the handedness of the batter, and the context (e.g., rely) by which the pitch was thrown. That is our baseline, and solely additional enhancements curiosity us. Controlling for these qualities helps be certain that we’re one thing new, not double-counting what we already see.

Fourth, we need to incorporate our favourite element of any rigorous mannequin: skepticism. Modeled coefficients a/ok/a random results stay a go-to regularization methodology, and we use them right here as nicely. If employed correctly, batters who present uncommon power or weak spot in a specific location can be credited or debited for it, however provided that they achieve this persistently and towards related high quality pitches in the identical or related places. In any other case, they’re presumed to carry out the way in which a mean batter would carry out from their facet of the plate, which is sensible if the batter is an unknown amount. (That is additionally why utilizing a finer grid could make little distinction: the random results construction resists overreactions to noise, and converges towards the only, least-likely-to-be-wrong selection).

Fifth, now we have to resolve which pitch occasions advantage this degree of study. Our StuffPro and PitchPro metrics are constructed upon a well-recognized logical tree: is the pitch swung at? If there isn’t any swing, is the end result a known as strike, a known as ball, or successful batter? If there’s a swing, is the end result a foul, whiff, or ball in play? And if there’s a ball in play, what’s the end result? 

Though all these contingencies are attention-grabbing, balls in play dictate how nearly all runs are scored, and variations in runs scored decide which groups win video games. Controlling the standard of balls in play can also be how command artists survive, as a result of a lot of them can not restrict the variety of balls put into play: their stuff merely isn’t ok to overlook bats. So, for now we’ll rating pitchers on how nicely their adaptation to particular person batters improves doubtless outcomes on balls put into play. Possible outcomes are measured in anticipated run worth as modeled by our PitchPro system for launch velocity, launch angle, and spray angle.

This offers us the constructing blocks of our mannequin: we regress (a) batted ball outcomes on (b) pitch id, location, and basic high quality by means of PitchPro, and (c) a location grid, particular to the batter, that tracks the extent that particular person batter deviates (plus or minus) from the overall development for his or her handedness in every location on the grid, aware of the batter’s efficiency in close by places. For individuals who care, we use a Matérn covariance operate over our location grid. 

To find out a batter’s deviation from the locational norm, we predict the anticipated run worth of the ball in play at every grid location for the batter. We grade pitchers by how the pitch location impacts the anticipated results of the BIP as in comparison with the batter’s common anticipated end result for balls in play. 

The mannequin runs shortly utilizing glmmTMB, a speedy R bundle for multilevel modeling, which calculates the most effective fee of correlation decay throughout the whole strike zone grid for every season.

Illustrations

A number of map comparisons will present why a hitter’s uncooked outcomes usually are not solely noisy, however typically deceptive, not less than on the subject of adversarial location.  Once more, keep in mind that we aren’t trying on the uncooked prediction of whether or not anyone is, in complete, good or dangerous at hitting, though these instruments enable us additionally to reply that query too.  What we need to know is the place the hitter is susceptible as in comparison with the place a typical hitter from their facet of the plate is susceptible.  

An excellent illustration of the issue offered by uncooked knowledge is Aaron Decide.  Listed here are his uncooked ends in every a part of the grid for this season, relative to league common:

The obvious drawback is the uncooked knowledge’s suggestion that Decide is lethal in varied areas of the zone besides over the very coronary heart of it, which isn’t very doubtless. There are a number of zones for which there’s little to no knowledge in any respect, as a result of Decide tends to not put pitches in play from there. Though this offers a sign of a unique type, it doesn’t assist us predict how that pitch would do if it did draw a swing. And all these places have customary deviations so extensive that they swallow the uncooked averages, an issue notably acute with dwelling run hitters, who current the widest vary of potential run outcomes on any given pitch.

The adversarial location mannequin, although, just isn’t troubled by this: utilizing its covariance matrix, and crediting the pitches for which it has essentially the most data over the remaining, it offers a extra wise learn of Decide’s doubtless zone, and extends it throughout the whole area:

Apparently, the mannequin says that location (3,4) the truth is is more likely to be Decide’s strongest location, not a singular vulnerability.  It concludes that his power continues towards the outer facet of the strike zone however tapers off after that, notably because the pitches transfer down or inside. After all, Aaron Decide being Aaron Decide, there isn’t any space the place the mannequin finds him weak per se, however he’s uniquely harmful in comparison with different hitters within the yellow areas, even when his uncooked knowledge suggests in any other case. Regardless, one factor the mannequin is at all times capable of do is to foretell an consequence at any location, as a result of even when it lacks knowledge, it learns from how each the particular batter and batters typically are likely to carry out in close by places which are presumed to have related data.

After all, it’s also essential to know when a hitter doesn’t have distinctive locational weaknesses. The uncooked knowledge might be deceptive right here as nicely. Take into account the uncooked run values of Alejandro Kirk’s balls in play by pitch location, as in comparison with the league common:

What a large number. By outcomes alone, Kirk’s adversarial location is barely good on pitches above the zone, and alternatingly good or dangerous each few inches you progress.  This doesn’t make sense, except his hitting zone is essentially uniform and the deviations we see are largely deviations from an identical baseline.  Let’s see what the mannequin predicts:

Certainly. Contemplating all the knowledge, and the general tendencies of close by places, now we have a hitter who will get considerably higher additional inside, and considerably weaker as you development outdoors, however general is fairly constant.  

Let’s do yet another, Stephen Kwan. Kwan’s uncooked run values on balls in play, as in comparison with common, present one other patchwork quilt that hints at some curious tendencies, however not in a means that gives confidence:

Kwan’s “strengths” are a bit everywhere by the uncooked knowledge.  Pitches trending decrease and out of doors appear to be favorites, whereas pitches up within the zone usually are not.  But, in location (1,2) there’s a potential anomaly that, along with liking pitches decrease and extra outdoors, he thrives on the pitch up-and-in. You don’t see that day-after-day. Small pattern surprise in that location, or does he have just a few very-different areas of power? Our mannequin helps reply that query:

Apparently, even making use of our mannequin’s skepticism, there seems to be one thing to it.  Low pitches typically are an space of power, notably when they’re extra outdoors. The map grows chilly as you progress up within the zone however tilts again towards productiveness as you stand up and in.  

I like this instance as a result of the mannequin just isn’t merely shrinking all people, in every single place: if the hitter does one thing uncommon, even in an space with much less basic plausibility, the mannequin can nonetheless decide it up, thereby telling you that pitches on this map can profit from being excessive, simply not too excessive.

The Outcomes

Having mentioned the batter profiles, let’s speak in regards to the outcomes for pitchers.

First, let’s make sure that we’re answering the appropriate query. There are (not less than) two methods to summarize a pitcher’s efficiency from a mannequin like this. Technique 1 is a normal with or with out you (WOWY) / marginalization evaluation, and summarizes the pitcher’s anticipated BIP outcomes relative to the league-average anticipated end result towards those self same batters throughout these 30 places. Technique 2 compares a pitcher’s ends in every location solely to their opposing batter’s common efficiency throughout their particular person predicted grids. The selection makes a giant distinction:

Desk 1: Reliability of Two Abstract Strategies, 2023-2024 MLB pitchers(Spearman Correlation, weighted by pitch quantity over the whole grid)

Comparability Technique
Spearman Correlation

Relative to League Common Location for Batter
-0.10

Relative to Particular person Batter’s Common Throughout Places
+0.50

Correlations near zero are nugatory. Correlations within the .5 vary are reasonably robust, and for a ability as sophisticated as pitching, they’re a strong sign that we’re onto one thing. Desk 1 means that pitchers don’t present distinctive ability at merely concentrating on areas the place batters are typically above or beneath common (this can be a denominator difficulty, outlined extra by the general high quality of the hitter than the situation per se).  However they present a particular distinction in how they aim places the place batters are individually uncomfortable.

This finalizes our definition of adversarial location: the pitcher’s skill to find pitches the place a specific batter will get beneath common batted-ball outcomes relative to different places for that batter, not merely places the place a batter will get beneath common outcomes.

Do pitchers usually considered command artists are likely to excel at adversarial location, as we suspected? Certainly they do. Take into account this abstract of 2024 adversarial command outcomes, offering the weighted percentile for the speed at which pitcher minimizes the chance of runs on balls in play by discovering the places the place particular person batters least wish to see a pitch, finest and worst:

Desk 2: Pitcher Adversarial Location, 2024 MLB Season

(minimal 150 pitches, larger percentile is best)

150 pitches appears to be an excellent general-purpose threshold: the outcomes are likely to make constant sense, plus we filter the place gamers and small-sample wonders. As for the names, having Jose Quintana and Kyle Hendricks on the prime is sweet to see. Likewise, having José Alvarado within the second percentile for purposeful location shouldn’t shock anyone. It additionally illustrates the impossibility of Paul Skenes.

Right here is how the leaderboard has formed up thus far in 2025:

Desk 3: Pitcher Adversarial Location, 2025 MLB Season

(minimal 150 pitches, larger percentile is best)

Jose Quintana grabs the crown once more, though Merrill Kelly and Zack Wheeler make sense as nicely. Emerson Hancock has the truth is had one thing go proper this 12 months. On the alternative finish, now we have two Rockies (maybe a Coors difficulty?), prime Rule 5 decide Shane Smith, and Kumar Rocker mentioning the rear.

What number of runs are saved (or misplaced) by a pitcher’s ability in adversarial location? For 2024, here’s a conservative estimate:

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

(BIP pitches solely)

The outcomes are conservative as a result of they embrace solely pitches truly put into play, not all pitches that had been or might need been swung at. Increasing into counterfactuals ought to enhance the affected run quantity—maybe to a big diploma—recognizing that whereas batters see many pitches, they nonetheless conclude each plate look with a single occasion. These estimates additionally don’t replicate the variance that may trigger pitchers on each ends of the spectrum to have extra excessive outcomes. 

The worth of adversarial location ought to already be mirrored in a pitcher’s general anticipated runs allowed. Nonetheless, adversarial location is crucial to understanding a technique that pitchers can and do succeed. Pitchers which have poor adversarial location, both because of lack of general command or crew directive—“simply throw it down the center!”—may very well be costing their crew a win or two over the course of a season, to the extent the pitcher can do higher. (And worsening the superior scouting/R&D crew, which could as nicely not hassle printing out these fancy reviews). In contrast, pitchers with glorious adversarial location can, and do, succeed with decrease pitch high quality. And pitchers who can mix nice stuff with typically good location and excellent adversarial location appear to supply the best potential for long-term success. 

Now we have calculated adversarial location values again to 2017, so it’s price who has had essentially the most success—and failure—over that point interval. I think the names won’t shock you:

Desk 5: Profession Runs Saved by Adversarial Location, 2017–2025 MLB Seasons

(BIP pitches solely)

The highest of this leaderboard is a “who’s who” of the most effective command artists from the previous decade, a lot of whom featured a terrific change-up (in Kyle Hendricks’ case, he featured two). On the different finish is a a lot completely different listing, though a few of it could be by selection, with pitchers who both can or should rely as an alternative on uncooked stuff or another side of fine pitching to get the place they need to go. (To stay round lengthy sufficient to rack up that many misplaced runs, the pitcher clearly must be good at one thing, or simply pitch for the Rockies).

Conclusion

Baseball is a zero-sum recreation. Which means each motion that undergoes evaluation requires the context of the actions of the opponent. We analysts have been fortunate, in comparison with the poor souls in different sports activities, as a result of extra of the variables we’ve traditionally studied have labored as impartial variables. However no variable is really impartial: The trail of the ball that leaves the hand and the trail of the ball that meets the bat are the identical ball. As we proceed to dig deeper in our analysis, and develop extra exact with our instruments, we’re capable of deal with these variables as dependent, and thus mannequin them extra realistically. Adversarial location is only one instance of this subsequent step, however the early returns display the potential significance of this vein.

We’ll proceed to refine the idea of adversarial location, however welcome any ideas you’ve got on the way it is perhaps higher offered or mentioned. 

To date, additional matters of curiosity embrace: 

Subdivision by pitch kind, 
Subdivision particular places on the grid (there are 30 of them, in any case), 
Investigating adversarial angles, not simply places,
Publishing our personal modeled batter warmth maps, and
Grading batter swing choices based mostly on batter-specific zone places, moderately than merely being in or out of the strike zone.

In the intervening time, the scores of all pitchers from 2017 by means of the current might be discovered on this Google Sheet.