Five Things I Learned at Saberseminar 2026

Once a year, the baseball statistics world converges on Chicago for a weekend of presentations, discussions, networking, and a frankly shocking number of job interviews. I’m talking about Saberseminar, if you’re not familiar. The conference has long been a source of baseball learning, where a mixture of leading analysts and up-and-coming students present their latest research. Every time I go, I leave with my head overflowing with new avenues of study, new things I’m curious about, and a bunch of cool knowledge that I won’t expand on but am nevertheless happy to have learned. I went two weeks ago, and I have a notebook full of ideas to investigate now. I can’t tell you everything that I learned – if you want that, you’ll just have to go to Saberseminar yourself. But there were five talks in particular that I loved. They weren’t the only good presentations by any means. But I have a column for highlighting five things, and it felt only appropriate to highlight five of my favorites from the weekend.

1. Injury Research by Scott Powers et al
The first presentation of the weekend is usually a good one, and this year didn’t disappoint. Scott Powers is a professor at Rice University who has a long history of good presentations at analytical conferences. His past research runs the gamut from batted-ball modeling to volleyball serve optimization. This year, his presentation focused on injury prevention. Powers, Rose Graves, Marina Vannucci, James Buffi, Daniel Barrueco, John D’Angelo, and Amanda K. Glazer set out to come up with some realistic injury guidance based on pitch characteristics and biomechanical data from the major leagues. As the name of the talk implies, they took a central idea into their work: Don’t pitch an injury prevention plan that boils down to “stop throwing so dang hard.”

The first part of their work focused on what they could glean from Statcast data. They identified a number of characteristics that are correlated with increased likelihood of elbow injury. Most of those won’t surprise you: fastball velocity, fastball break, and slider usage were the top culprits. Interestingly, they also included some measures of effort – the ratio between average and 90th-percentile fastball velo, for example – and found no linkage there. That’s a great proof of concept for the idea, but unfortunately it doesn’t really present a workable plan for pitching coaches. “Hey, just, uh, make your fastball slower and give it less movement.” As the title of the pitch implies, that’s no good, so they went deeper by using biomechanical data to look for more granular data.

I won’t pretend to understand how to turn three-dimensional motion capture data into workable data sets, but hey, that’s what a team of PhDs is for. One trait jumped out as a marker of reduced injury risk but not reduced velocity: off-arm momentum generated. In other words, after holding a number of things constant (hey, they’re the research geniuses, don’t ask me exactly how that part works), pitchers who generated more momentum with their glove arm saw reduced incidence of injury without declining velocity. That’s a notable and actionable finding, because most of what we have on arm injuries is “no one knows why this happens and that’s bad.”

Is that the answer to our prayers? Surely not, because if a single change could reduce injuries, some team would have figured it out by now, and the knowledge wouldn’t stay private for long. Going from “swinging your glove arm more is good” to turning that into actual pitching technique is a big leap. But research like this can only help advance the state of pitching development, and anything that can help slow the epidemic of elbow injuries is a good idea in my book. I also appreciated that the group published some null results, most notably: spin rate, slider break, and their various measures of effort. It’s good to know what we don’t know – and in this case, we don’t know whether spinnier pitches carry a higher likelihood of injury, despite what you might hear from an armchair analyst.

2. Changeup Biomechanics by Ben Lerch, Harry Pavlidis, Stephen Sutton-Brown, and Gretchen Oliver
There was a lot of biomechanical talk at this year’s Saberseminar, which mirrors the way that team-side (and lab-side) research is heading. Candidly, I’m not a biomechanist, and so I spent a lot of these sessions scribbling down cryptic notes and moving my arm around experimentally to try to understand the motions that people were discussing. But this presentation, “Biomechanical Predictors of Fastball-Changeup Velocity Delta in Division 1 Collegiate Pitchers,” spoke to me. It asked a question that I assume we’ve all wondered: How do pitchers produce such varied speeds on changeups?

Obviously, plenty of changeup speed variation comes down to grip. But that’s not the only determinant, and pitchers with the same grip can produce meaningfully different speed differentials. Lerch, a PhD student at Auburn, gathered a ton of the team’s own markerless biomechanics data, as well as pitch-level data on the same pitches. He and the rest of the researchers used that data to hunt for mechanical markers that were associated with either less or more speed differential between a pitcher’s primary fastball and changeup. Grip isn’t measured in these, and so it’s out of the experimental treatment by default; instead, this was about the other things that pitchers do to manipulate their changeups.

The primary suspect, both listed by the researchers and what I thought of in my seat, was torso velocity. Arm speed is part of the deception inherent in a changeup, but I thought that using a normal arm swing but with less forward transfer of momentum, measured by torso velocity, would be a good way to make a changeup look fast and yet go slow. But there was essentially no correlation to be found. In fact, most of the factors Lerch and company considered had no measurable effect.

That said, one marker worked. The more a pitcher allowed their lead knee to collapse during their delivery, the more velocity differential they tended to create. This is a different channel than the one I expected, but it works fairly similarly. When you keep your lead knee stiff, force gets more efficiently transferred from ground strike to your throwing motion. The opposite is true when you keep that knee poised – the force transfers more readily. By allowing more of a collapse, the same grip and arm action produce less velocity.

Not everyone wants a larger gap between their fastball and changeup velocity. Some pitchers would be well served by a pitch that’s hardly slower than their fastball; Félix Hernández’s elite changeup is my favorite example of this. Some pitchers succeed thanks to huge gaps. There’s no one right answer. But as we learn more about the physical movements that create separation between pitches, the toolkit for creating the perfect changeup for each individual pitcher is getting filled out. I don’t know how easily this will be incorporated into pitch design, but I’m confident that independent pitching labs and teams will both be trying it.

3. Modeling Uncertainty by Aidan Resnick
This one was perfectly calibrated to get my attention. Resnick works in financial markets, and he took some of the concepts of his job – namely option pricing and theory – and applied them to pitch modeling. As a former market maker (among other hats), I already look at innocuous data and see connections to the way securities and derivatives are priced. Someone else doing the same thing was music to my ears.

The basic concept here relates to what are called “Greeks,” so called because they’re all represented by Greek letters. They denote derivatives of various types. In finance, they’re things like change in value per change in price (delta), change in delta per change in price (gamma), change in value per change in interest rate or time, etc. In Resnick’s version, he set velocity as the “underlying” and made derivatives off of that. Change in expected pitch result per change in velocity is delta. Gamma is change in delta per change in velocity, or in other words, whether the amount your fastball improves by with an additional mile an hour changes based on how fast you were already throwing. There are other Greek characters too, like change in value relative to sweep (vega-x) and (ride) vega-y. The point is that the central idea of measuring change in expected run value for change in some underlyings is a new and interesting project.

I’ll get to the big takeaway from this project in a moment, but there are some fun little takeaways too. Not every pitcher gets the same benefit from a change in fastball velocity. A tick on Mason Miller’s fastball is worth four times as much as a tick on Kyle Hendricks’ heater. That shape is generally true across the league – the harder you throw, the more gains you make from adding a marginal mile per hour. That interplay between speed and the effectiveness of adding speed is intuitive, and it also explains why these concepts are interesting. Universal takeaways like that are useful even outside the confines of a specific model, and even if you don’t buy the pitches-as-options state of the world Resnick is pitching.

A corollary of all of this is what I’d consider the bigger takeaway, and it’s driving some behind-the-scenes research at FanGraphs as I write this. To make this kind of differentiable pitch model, you have to break from the style used by PitchingBot or Stuff+. I don’t claim to know the inner workings of every public pitch model, but they mostly work using advanced statistical techniques, particularly gradient boosting. That’s a method of applying flexible rules that is useful in hunting for interactions between pitch characteristics; maybe this much vertical break matters more from one arm slot than from another, or something along those lines.

That flexible handling of interactions makes these models hard to interpret, though. You can’t “bump” a stuff model by a mile an hour and get a coherent answer about how good the pitch would be. It just doesn’t work that way. There’s no “velocity slope” where every mile an hour improves the model’s estimation of a pitch; every pitch falls into the equivalent of a very complex Plinko board with many different rules for how the different characteristics will interact. That means that the concept of Greeks can’t exist in such a model, because you can’t really calculate a derivative.

Resnick created a three-factor model for his presentation, but you could create a different version yourself. The interesting part for me is that the style of pitching model we’ve most frequently seen isn’t the only way to do it. There are tradeoffs associated with complexity, and not all of them are immediately evident. I also really enjoy the idea of measuring rate of change, rather than just a steady state, and I imagine teams do as well. Knowing how good a pitcher is? That’s table stakes these days. Knowing how good they can become? More fruitful, and modeling some version of their pitch Greeks feels like a good way to make progress on that front.

4. Historical Reframing by Jun Hee Kim, Adrian Burgos Jr., Shen Yan, Ryan To, and Daniel J. Eck
At a past Saberseminar, Eck presented something called Full House Modeling, a method of adjusting statistical records based on the cohort of players currently in the major leagues. To simplify, the method relies on producing a hypothetical talent pool for each year and adjusting performances based on that talent distribution. In other words, players who succeeded against a smaller subset of competition – less global population and less non-American talent characterized the early years of baseball – receive proportionally less credit than players who produced similar results against wider competition.

That formulation of this model has been much discussed, and you can see some formal explanations here. This presentation, spearheaded by Kim with the assistance of some of the professors who wrote the initial full house modeling paper, focused on modifying and extending the method to put the statistical accomplishments of Negro League players in context.

To make this model work, the researchers added some demographic data to their population-level splits. If you’re trying to figure out how competitive the Negro Leagues were in a given year, or how competitive the AL and NL were for that matter, knowing the proportion of players of each race in the league, as well as their representation in the U.S. population, is necessary. With that data, Kim was able to take existing statistics and create adjusted WAR totals that put everyone on a level playing field.

That was only half of the presentation. The other half covered testing whether this method passes the smell test. Luckily, there’s a great sample of players to use here: players who first played in the Negro Leagues and later transitioned to AL or NL clubs. The model makes a true-talent prediction for every player in every year, with aging thrown in. By comparing the talent predictions and actual results of players who switched between leagues, they got good confirmation that the rough contours of the method worked; predictions for league-switching players did a good job of matching their actual results.

I don’t think this research is going to revolutionize my understanding of baseball. But I appreciated Eck’s first pass at the subject – I’ve long considered Barry Bonds to be the greatest player of all time relative to his era, and he tops Eck’s leaderboard comfortably – and I always have my eye out for ways of putting past eras into broader context. I also think that this style of modeling has applications for league translations, though I’ll leave the implementation to someone who’s a little handier with math. I thought this research was fascinating and something of a throwback compared to a lot of the nitty-gritty pitching mechanics presentations.

5. Pitcher/Batter Game Theory by Owen St. Onge
St. Onge, a master’s student at Syracuse, came up with an idea that made me jealous. I’ve looked into swing timing a bit recently, and St. Onge took the same data and asked a novel question: Can we measure how pitchers respond to late and early swings? He took Statcast bat tracking data and pitch data, turned the bat tracking data into timing information, then correlated it with future pitcher behavior.

The question at hand: Do pitchers and catchers respond to notably late or early swings? To define notably early or late swings, St. Onge took a data set of barreled contact and then predicted intercept points for optimal contact after controlling for all the usual suspects like pitch type, location, game state, handedness, and player identity. From that, he then tagged swings as early, on time, or late.

Strong relationships jumped out immediately. When hitters were early on fastballs, pitchers were far more likely to throw them something slower on the next pitch. When they were late on a fastball, the pitcher was more likely to double up. Likewise, when hitters were early on slower pitches, the pitcher was more likely to follow up with another slow pitch – but when they were late on slow pitches, pitchers were more likely to follow up with a fastball.

That alone would be a solid finding, but there was more. There was a notable difference between hitter behavior depending on whether they were very late or very early on the previous swing. Being early on one pitch is associated with being early on the next, and the same is true for being late. But the association on the late side is meaningfully less strong; it almost looks like it has a floor, in fact. One way of thinking about that is that hitters can catch up to anything, but they have a harder time with the inverse.

This project gave me a ton of ideas for extensions. Could we measure catcher game calling by looking for their ability to wrong-foot hitters? Could we produce individualized hitter reports that highlight specific talent or lack thereof when it comes to adjusting swing timing? Beyond timing, what about testing for poor swings, by one definition or another, based on pitch pairings? There’s a big world to investigate here.

That’s the best part about Saberseminar. It’s only two days long, and the presentations are packed in so tightly that there’s simply no way you’ll remember them all perfectly, even if you’re sitting there jotting down notes all day. But almost every presentation has depth to it. There are always avenues for more research. If you leave Saberseminar without wondering about some fundamental truth of baseball, you’re not doing it right. I got more research ideas in two days in that auditorium than in the previous month of pondering baseball data on my own. This list of five only scratches the surface. I can’t recommend Saberseminar enough – and I hope to see you there when I’m greedily taking notes on another batch of presentations next year.