<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Matt O&#39;Conor on Medium]]></title>
        <description><![CDATA[Stories by Matt O&#39;Conor on Medium]]></description>
        <link>https://medium.com/@Matt_OConor?source=rss-b676e04dcfdb------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*P1eNE_YFH4BHrZiO44X0yg.png</url>
            <title>Stories by Matt O&amp;#39;Conor on Medium</title>
            <link>https://medium.com/@Matt_OConor?source=rss-b676e04dcfdb------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:10:40 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Matt_OConor/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[ABS Strategy Development]]></title>
            <link>https://medium.com/iowabaseballmanagers/abs-strategy-development-87eb6c96538c?source=rss-b676e04dcfdb------2</link>
            <guid isPermaLink="false">https://medium.com/p/87eb6c96538c</guid>
            <dc:creator><![CDATA[Matt O'Conor]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 16:01:59 GMT</pubDate>
            <atom:updated>2026-11-20T16:01:59.892Z</atom:updated>
            <content:encoded><![CDATA[<h3>Intro</h3><p>The ABS system is potentially coming to college baseball in 2027, which means it’s crucial for us at Iowa Baseball to have the best strategy we possibly can. What makes a strategy good is sound logic, reasonable assumptions, and easy-to-understand takeaways. Our goal with this project was to come up with ways to extract the most value with our ABS challenges throughout the season to give us another competitive advantage.</p><p>Now I’ll explain the blueprint for our 2027 ABS strategy. The final equation that we use to deem whether it’s a good or bad decision to challenge will be derived from the expected value formula:</p><p><em>EV = (Probability of Success * Potential Gain) — (Probability of Fail * Potential Loss)</em></p><p>From here, you can calculate the minimum rate of success you need to at least break even on your decisions:</p><p><em>Probability of Success = Potential Loss / (Potential Loss + Potential Gain)</em></p><h3><strong>Strategy Development</strong></h3><p>This will be the formula we will reference throughout the rest of the blog, as P = L / (L + G). When applying this formula to baseball, and more specifically the use of ABS challenges, the variables will be: G = change in win probability after a successful challenge, L = the impact of losing a challenge on win-probability.</p><p>Note: Calculations were run on unrounded values. Intermediate figures shown in this article are rounded for readability, so recomputing by hand may differ slightly in the last digit.</p><p>Calculating G is the easy part, as we already have an in-house win-probability model that uses the score, inning, number of outs, baserunners, and count. For this, we are not taking team strength into account. Using an app, you can switch around all of these factors and get the value of G instantly from any game situation you want, which is the increase in win-probability if a challenge is successful.</p><p>Now we have the tricky part to solve for, which is L (cost of challenge lost). For pricing the utility of a challenge, you would definitely say that the current inning, score, and how many challenges you have left are things you should consider, but how do you whittle all of this down into a singular number? Since there isn’t prior data of ABS from college baseball, we cannot see its effects on win-probability. So, how can you compute its value? First, we need to find the number of missed calls per D1 game. Using the college ABS rules from the National Strike Zone Initiative, which defines the strike zone as 19 inches wide (1 extra inch on each side of the plate), with the top of the zone being 56% of the batters height and the bottom of the zone being 24% of the batter’s height.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ns_jhVNbKon-Mh2V4yKIkw.png" /></figure><p>Given these strike zone parameters, we found that in 2026, 55.8% of pitches were called strikes or balls, and out of those, 11.5% were called incorrectly based on the new strike zone. And in 2026, there were 50,616 innings, and 35.7 pitches per inning, giving an average number of pitches in a nine inning game of 321. Then you take (321 * 0.558 * 0.115) / 2 to get 10.3 expected wrong calls per game for one team.</p><p>But you cannot assume that a team will challenge every single wrong call throughout a game, and that they will be 100% accurate with their challenges. This means we need to make an adjustment. Using 2026 MLB data, teams successfully overturned 1.17 calls per game each. Also, we need to adjust for the fact that MLB umpires are much more accurate than college umpires, so we take our 1.17 and multiply it by 10.3 (average number of D1 umpire miscalls) / 4.97 (average number of MLB umpire miscalls). With that adjustment, we estimate 2.43 successful challenges per game for a D1 team. This serves as our baseline estimate for the number of successful challenge opportunities a team can expect over the course of a game. Then to make this dynamic, you need to make an adjustment to the number based on how deep you are into the game. You do this by taking 2.43 * (outs left / 54). This way, at the start of the game, you’re expected to have 2.43 correct challenges left in the game, whereas in the top of the ninth, you are expected to have 0.27 remaining.</p><p>Because college ABS has not yet been implemented, this number should be viewed as an estimate rather than a ground truth. We used MLB challenge outcomes as a starting point and scale them based on the observed difference in umpire accuracy between MLB and Division 1. The resulting 2.43 gives us a reasonable baseline for strategy development, but we expect to update this value as college ABS data becomes available.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*84F8psZ5M8-PpUCty3kT0g.png" /></figure><p>Now that we have an estimated quantity of successful challenge opportunities remaining (Q), we need to estimate how much those future call flips are worth.</p><p>We will do this by finding the average gap in win probability between a ball and a strike on called pitches from the current point of the game through the end of the game (denoted as A). The score is held where it is, as it would be hard to estimate the timing and amount of scoring throughout the game. Tied at the first pitch of the game, A = 2.45, tied at the first pitch of the bottom of the ninth, A = 4.74. A is essentially the average impact future challenges will have.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9RGf1EaGdc6sDOZ7s_6hKg.png" /></figure><p>Next we have to account for how many challenges we have left. Correct challenges are kept, so one challenge can take every future success as long as you still hold it. With one left, losing it gives up the whole remaining stream: L = A * Q. Tied at the first pitch, that is 2.45 × 2.43 = 5.94.</p><p>With more than one left, the one you would lose is a spare. You still hold the challenge that protects the rest of the game. Q is only an average, so let R be the number of expected successful challenges still coming, a Poisson count with mean Q. A spare is worth one flip, and only if at least one success is still coming. With two left, L = A times the chance that R is at least 1. Tied at the first pitch, that chance is 0.91, so L = 2.45 × 0.91 = 2.23. Tied with two left in the top of the fifth, Q is 1.34 and A is 3.21, so that chance is 0.74 and L = 3.21 × 0.74 = 2.38. This methodology accounts for the differing value of losing your spare challenges versus losing your last challenge.</p><p>Going back to the original break-even formula, since we now know how each component is calculated you can see why it works. P = L / (L + G). One more adjustment made to the final number was accounting for who is challenging. Since our pitchers will not be challenging, we are looking at this from a perspective of hitter vs. catcher. In MLB during 2026, catchers challenged successfully 59% of the time, whereas hitters were successful only 49% of the time. That success rate is an imperfect measure of challenge skill, but we still believe there is an inherent bias that should be accounted for in our strategy. Anyone around the game will tell you the catcher knows the zone better than the hitter: the catcher sees far more pitches, and has the better vantage point. So to account for this, we adjusted the probability of success needed by around +5% for hitters and −5% for catchers. After that slight tweak, that’s everything that goes into the calculus behind a college baseball team’s ABS strategy. Again, this formula gives us the minimum level of confidence a player should have when challenging a pitch, under any game situation. Playing around with the inputs on the interactive app, you can see the effect each game state has on whether the player should challenge or not.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PBxNO8sfja39hzYdrUhbpQ.png" /><figcaption>First pitch of the game, requires a confidence level of at least 64.4% that the umpire’s call was wrong.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UBihPpUmvM6WQQ4TA9fX5A.png" /><figcaption>High leverage situation, two challenges remaining, requires a low confidence level to challenge.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iEQya1XF8fdKkMNuyLmHvA.png" /><figcaption>Low leverage situation, one challenge remains, requiring a high confidence level to challenge.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LoY_dHin2u8F9pFdE4vTjA.png" /><figcaption>Very high leverage situation, one challenge remains, 0.0% says you should challenge the pitch even if it was down the middle.</figcaption></figure><p>This is great and all, but players and coaches are not going to be calculating these probabilities from the dugout or batters’ box, which is why we need an implementation plan. Right now, with rules being unknown (everything here is subject to change), we’ve set these buckets for players to challenge: Green light is 0–25%, yellow light 25–75%, red light 75–100%. These guides will need to be relayed before the pitch, which is fine because we can calculate the value before every pitch very easily. The Brewers experimented with something like this in Spring Training, where they put colored cards on the dugout railing notifying players when an opportune time to challenge is. This was immediately banned however so depending on the rules the Big Ten lays out, this could be axed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eSrLMD6uiLSPrEtyguPRoA.jpeg" /><figcaption>via MLB.com</figcaption></figure><h3>Key Takeaways</h3><p>My first takeaway is the observed tradeoff between the value of flipping a call and the number of expected successful challenges. As you can see, they move in the inverse direction. As the game goes on, you’re expecting less successful challenges, but those challenges become more meaningful. The green line shows the cost of losing your last challenge in win-probability, as you can see it becomes less-costly to lose your last challenge as the game goes on, which makes sense.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f0SkX8rxFeOyZ16imYBjxQ.png" /></figure><p>My second takeaway is that the cost of losing a challenge is determined mostly by how many you have remaining. You’ll notice that losing your first challenge in the early game doesn’t have a huge impact on win-probability, but losing the last becomes very costly. This gap shrinks as the game goes on as you’re likely not going to have many opportunities to overturn a call.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j8Zo_Ut6kFsFGbD8UX1gJw.png" /></figure><p>My final and the most interesting takeaway is the difference in strike zones. The solid blue line is the observed strike zone in Division 1 from 2026. The blue line and the dashed black line are where the middle of the ball must be for it to be a strike. The corners and the bottom of the strike zone will increase substantially, while the outer part of the zone will shrink. I think this will be very interesting to see if pitching strategies change by throwing more pitches at the bottom of the zone and throwing outside less.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mdMEN-y9IeYiIPv4Up2Hcw.png" /></figure><h3>Conclusion</h3><p>In summary, using a probability break-even analysis with the impact of a challenge won versus the impact of a challenge loss will give you the minimum level of confidence you need to challenge. This gives us a framework for maximizing win probability from our ABS challenges, with the goal of turning a bunch of small decisions into wins over the course of a season. It’s also cool to see how projects like this one will start out with a stated goal that is accomplished, but open up other doors for new projects like a new pitching plan to take advantage of the new strike zone.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=87eb6c96538c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/iowabaseballmanagers/abs-strategy-development-87eb6c96538c">ABS Strategy Development</a> was originally published in <a href="https://medium.com/iowabaseballmanagers">Iowa Baseball Managers</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Closing the Gap: Using MLB Statcast Data to Model Collegiate Arm Angles]]></title>
            <link>https://medium.com/iowabaseballmanagers/closing-the-gap-using-mlb-statcast-data-to-model-collegiate-arm-angles-dec11f6f7904?source=rss-b676e04dcfdb------2</link>
            <guid isPermaLink="false">https://medium.com/p/dec11f6f7904</guid>
            <dc:creator><![CDATA[Matt O'Conor]]></dc:creator>
            <pubDate>Fri, 13 Feb 2026 15:55:36 GMT</pubDate>
            <atom:updated>2026-02-13T15:55:36.770Z</atom:updated>
            <content:encoded><![CDATA[<p>Arm angle has become a vital piece for biomechanical research and development across baseball. MLB recently made arm angle data public for all MLB games dating back to 2020, leading to great public research. However, this data currently exists only at the MLB level. At the college level, there is no direct arm angle measurement available, creating a meaningful gap in player development, scouting, and research applications. Without a reliable estimate, college programs are limited in their ability to analyze mechanical consistency, pitch design relative to slot, or arm slot variability across pitch types. This gap made it a priority to derive a reliable arm angle metric by leveraging historical data and building models that can be applied to college baseball data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bCDV-aJlE6xQH8Y_tdhE7Q.png" /></figure><h3><strong>Model Design</strong></h3><p>The goal of the project was simple: create an accurate arm angle statistic. First, we gathered MLB Statcast data going back to 2020, since that is the earliest point where arm angle data is available. I used a Generalized Additive Model (GAM) to predict arm angle, allowing for non-linear relationships between arm angle and key release variables. Feature selection was guided by both domain knowledge and causal direction. Release height, release side, and extension were included and smoothed within the GAM, as these variables directly influence arm angle. Variables such as pitch movement were intentionally excluded, as they are outcomes of arm angle rather than inputs, and including them would confound the modeling objective. Separate models were built for each pitcher handedness and each pitch type to account for structural differences across groups. Model performance was evaluated using validation error on the held-out dataset, with model complexity controlled to ensure the predictions generalized beyond the training data.</p><h3><strong>Model Results</strong></h3><p>The statistical performance was impressive given the fact that we deliberately used a small number of variables to avoid overfitting. Across all of the models the RMSE = 5.28º and R² = 0.80. The RMSE of 5.28° represents the typical scale of prediction error, with larger errors contributing more heavily to the metric. Rather than an average miss, it reflects how tightly predictions cluster around true arm angles. An R² of 0.80 shows that the model explains roughly 80% of the variation in arm angle. We prioritized lowering RMSE rather than maximizing R², as minimizing absolute prediction error was more important than explaining additional variance, since even small increases in error could meaningfully impact how a pitcher’s arm slot is interpreted or classified.</p><p>Another check we made was validating the structure of the college data with the MLB data. The average range of a pitcher’s arm angle across pitch types is 5.93º in MLB and 7.67º in college. The difference is large, but it makes sense because MLB mechanics are much more refined than the typical college pitcher. Also, the mean difference is 1.4º, but the medians are only off by 0.34º, indicating the model’s core accuracy is sound. The larger mean is driven by a few outliers in the college data rather than systematic bias.</p><p>Another statistic I looked at is the standard deviation of a pitcher’s arm angle on different pitch types. In the MLB subset, the standard deviation was 2.64º, while the college subset was 4.55º. However, the college subset has many more entries, so we can use a weighted standard deviation for a more accurate representation, which is 3.27º. The differences are noticeable, but again MLB pitchers’ mechanics are much better and consistent between pitch types, so the differences make sense. These descriptive statistics from the two data subsets reflect real-world mechanical variability rather than model bias, so they do not indicate a structural problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZLsiCtMARE7TEyv386vutA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uquzKJrUmxADCaPC9L9NDQ.png" /></figure><p>Looking at these visualizations, the prediction error distributions are normally distributed and centered near zero across all pitch types. The model performance chart demonstrates strong predictive accuracy across all pitch types, with predicted values clustered tightly along the actual values. The consistent performance across both left-handed and right-handed pitchers, as well as across all pitch classifications, validates that the models are reliable and unbiased for practical application.</p><h3><strong>Model in Use</strong></h3><p>To further validate my findings, I used video analysis to compare Iowa pitchers with MLB pitchers. Here are a few examples:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*TvHG5Z14jXyebfTpak9tdA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*pj_javz5acrAK1rrlXaf0A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-fQQUzPvjDF_4obzawVORQ.png" /></figure><p>I then compared recent college pitchers who were drafted and debuted in MLB to compare their collegiate arm angle to their MLB arm angle. Of course, this isn’t a 1:1 comparison, since all of these players have likely modified their mechanics since being in college. A few of these stand out to me. Grant Taylor’s difference could be explained by the amount of time from his final season in college (2022) and his debut season (2025). Skenes’ difference stands out as well, but there is a reason for the large discrepancy. He’s noticeably dropped his arm slot throughout his career as a pitcher dating back to his time at Air Force. This breakdown from MLB Network shows it: <a href="https://youtu.be/6ZlGNZTYguY?si=zI02iKhZhN56AabB">Paul Skenes Mechanical Evolution</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WbWHfzg25MVXOLhETvg5Eg.png" /></figure><h3><strong>Next Steps</strong></h3><p>Now that a reliable method for estimating a pitcher’s arm angle exists at the college level, this metric can be directly applied to the types of analysis discussed throughout this project. This work enables projects such as building a college-level version of a dynamic dead zone, incorporating arm angle into Stuff+ models, optimizing pitch movement relative to a pitcher’s arm slot, and monitoring mechanical consistency across pitch types and over the course of a season. Because the metric is derived from inputs already available in college data, it can be scaled across multiple seasons at the pitch level. With all the advancement that could come out of this, it’s also important to acknowledge the limitations. This metric is not a direct measurement of arm angle, and therefore it is not perfectly accurate on an individual pitch or pitcher. Instead, it represents the closest possible estimate given the information available. With some of the current data constraints, there is an upper bound on achievable accuracy, and this model is designed to approach that ceiling rather than eliminate all error. Ultimately, this project transforms arm angle from an MLB-only metric into a scalable tool for college and other lower-level player development/evaluation. More broadly, it establishes a framework for creating reliable proxies when key measurements are unavailable at lower levels of play.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dec11f6f7904" width="1" height="1" alt=""><hr><p><a href="https://medium.com/iowabaseballmanagers/closing-the-gap-using-mlb-statcast-data-to-model-collegiate-arm-angles-dec11f6f7904">Closing the Gap: Using MLB Statcast Data to Model Collegiate Arm Angles</a> was originally published in <a href="https://medium.com/iowabaseballmanagers">Iowa Baseball Managers</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>