<?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 Ryan Miele on Medium]]></title>
        <description><![CDATA[Stories by Ryan Miele on Medium]]></description>
        <link>https://medium.com/@ryancmiele?source=rss-0815fc24bf98------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*DWR-dvC0nyIR14qqIpI7qQ.jpeg</url>
            <title>Stories by Ryan Miele on Medium</title>
            <link>https://medium.com/@ryancmiele?source=rss-0815fc24bf98------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 13:34:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ryancmiele/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[What Never Happened: Measuring Defensive Value Through Spatial Suppression]]></title>
            <link>https://medium.com/@ryancmiele/what-never-happened-measuring-defensive-value-through-spatial-suppression-b7f1953c3323?source=rss-0815fc24bf98------2</link>
            <guid isPermaLink="false">https://medium.com/p/b7f1953c3323</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[sports-analytics]]></category>
            <category><![CDATA[hackathons]]></category>
            <category><![CDATA[hockey]]></category>
            <dc:creator><![CDATA[Ryan Miele]]></dc:creator>
            <pubDate>Thu, 02 Apr 2026 19:18:09 GMT</pubDate>
            <atom:updated>2026-04-02T19:18:09.343Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pYF7ppKukJrRFxxuFtMh4A.png" /></figure><p>HALO Hackathon 2026</p><h3><strong>Hockey Context</strong></h3><p>Modern hockey analysis has made tremendous strides in the past decade in terms of quantifying shot quality through the use of expected goals models. These advances have fundamentally changed the way players are evaluated. However, a key limitation in using these frameworks is that hockey is a dynamic sport where multiple moving parts interact simultaneously which makes it hard to distinguish how much credit each player deserves in a sequence of plays. Specifically in evaluating defense, public models assess the shot quality that is given up while the player is on the ice, but far less attention is paid to the defensive processes that suppress shots before they can occur in the first place.</p><p>The way defenders take away space relative to passing and shooting lanes directly impacts which offensive options are actually available. This impact is difficult to capture using strictly event data. By looking at the spatial relationships between the puck carrier, potential pass options, and nearby defenders, we can start to evaluate defense not just by what happens, but by what <em>doesn’t</em>.</p><p>This project aims to introduce a spatial framework that uses player tracking data to measure defensive positioning and show how effectively defenders limit expected goals by denying passing and shooting lanes in real time.</p><h3><strong>Process</strong></h3><p>To attack this problem, I utilized nearly 500 games of 2023–24 AHL player tracking and events data provided by Sportlogiq. The data was filtered to only include events where the puck carrier was in their offensive zone and removed all events that occurred less than two seconds after an offensive zone faceoff. The logic behind this choice is that the time immediately following a faceoff is usually chaotic, with players worried about recovering the puck rather than their structured defensive responsibilities like blocking passing or shooting lanes.</p><p>The next step was to create a snapshot in time at each event. For every event, all players on the ice were taken into account and separated by team. Potential passing lanes were derived by calculating the distance and angle between the puck carrier and each of their teammates. A passing lane would be considered to be blocked if a defender was positioned within 3 feet on either side of the direct line from the puck carrier to the potential pass receiver. The same logic was incorporated into finding the potential shooting lane, which also would be considered to be blocked if a defender was located within the 6 foot width of the direct line to the center of the net.</p><p>To evaluate the value of each potential pass, a simple expected goals model was created which assigned a predicted value based on the location of each potential pass target. For sake of simplicity, this model assumes that the potential pass target would shoot immediately after receiving the pass. This is a lofty assumption but predicting where the player would skate after receiving the pass while also accounting for how the defense would react would open up a new can of worms that could be addressed in a future part two of this project. To factor in the possibility of a bad pass turnover, the expected goal value of shots from potential pass targets with a clear passing lane was reduced based on the length of the pass.</p><p>Let’s look at the following visual to get a better understanding of this process. The blue triangle depicts the offensive puck carrier, and the blue circles represent his offensive teammates. The blue dotted lines represent the potential passing lanes between the puck carrier and his teammates, while the orange dotted line represents the potential shooting lane for the puck carrier. The brown circles represent the defensive players. The model will flag if any of the defenders are blocking either the passing or shooting lanes. The defender at the top of the screen is blocking the short pass towards the wall.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/956/1*7wh1liHEqvEeRJ0NdmutXw.png" /></figure><h3><strong>Results</strong></h3><p>Using the framework described above, I calculated the total expected goals that were individually suppressed as a result of the defender’s positioning. To account for the differences in defensive zone usage across players, I created two different rate statistics: <strong>Expected Goals per Blocked Lane </strong>and <strong>Expected Goals Prevented per Event.</strong></p><p>As I stated previously, each event in the offensive zone is captured as a snapshot in time The model is binary, classifying a defender as either occupying a passing or shooting lane or not. <strong>Goals per Blocked Lane</strong> is calculated by dividing Total Expected Goals Prevented by the number of times the defender is blocking a passing or shooting lane. This essentially shows the quality of the lanes the defender disrupts.</p><p><strong>Expected Goals Prevented per Event</strong> is calculated by dividing Total Expected Goals Prevented by the total events the defender participates in. This statistic provides a better overall measure of a defender’s positioning, since it includes the instances where the defender is out of position and not blocking a lane. Here are the leading Defensemen in Expected Goals Prevented per Event:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*0w9zblyV9IhMhlltfxv5WA.png" /></figure><h3><strong>Team Building Impact</strong></h3><p>This model exposes a structural blind spot in the current defensemen evaluation market. Public metrics and traditional counting stats undervalue defenders whose primary contribution is eliminating high value options before they become events. By quantifying this space taken away from the offense, front offices gain a tool to find undervalued players whose impact is invisible to box scores, but critical to the outcome of the game.</p><p>In a league constrained by a hard salary cap, the ability to acquire defenders who are consistently in the optimal position to suppress scoring chances at below market value represents a direct competitive advantage. This framework reinvents defensive evaluation from surface level outcomes to a process related approach which enables organizations to exploit inefficiencies in the market from a cost value standpoint. Ultimately, it will lead to a deeper understanding of how successful teams prevent offense before it materializes.</p><h3><strong>Limitations and Next Steps</strong></h3><p>While this model is valuable, there are many limiting factors based on the data and approach that was used. As it’s stated in the competition instructions, location data for every player on the ice wasn’t always available for each event. Oftentimes players couldn’t be identified and were listed as an “NA” value or simply weren’t included at all. To provide the most accurate model possible, the precise location of every player on the ice is crucial. Another limitation of the data is since it’s event based, random gaps of time occur between events which causes there to be no evidence of the skating route that players took to get from point A to point B. It would paint a complete picture of each player’s positioning and decision making if there were consistent data points every second or so.</p><p>There could be inherent biases for this model based on the coaching staff’s defensive scheme. This could benefit a player with a great coaching staff who is instructed to be in the right place at the right time, and detrimental to a player who has a role which routinely pulls them away from key passing or shooting lanes. It’s important to distinguish these two cases apart from each other.</p><p>As I touched on before, there are many avenues to build upon and improve this model. Finding out the exact probabilities of a pass being completed based on the defenders in the area is a logical next step. This could lead to predicting the movements after the player has received the pass based on their spatial relation to nearby defenders which would enhance the predictive power of the expected goals model. Big picture team analysis can be conducted to figure out the optimal positioning of each defender for different clusters of alignments and styles that the offense could potentially throw at you.</p><p>Although this project focuses on the defensive side of the puck, the view of this project can easily be flipped to the offensive perspective. With the current framework that’s in place, the puck carrier’s decision making can be evaluated based on whether they choose the pass or shot attempt to maximize the potential expected goals of the possession. This offensive approach would display if the player is making fundamentally sound decisions on a consistent basis.</p><h3><strong>Appendix</strong></h3><p>Code: <a href="https://github.com/RyanMiele/Colorado-Hackathon/">https://github.com/RyanMiele/Colorado-Hackathon/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b7f1953c3323" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cutting Edge: Calculating Change of Direction Skating Metrics]]></title>
            <link>https://medium.com/@ryancmiele/cutting-edge-quantifying-change-of-direction-skating-metrics-2c0716bed623?source=rss-0815fc24bf98------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c0716bed623</guid>
            <dc:creator><![CDATA[Ryan Miele]]></dc:creator>
            <pubDate>Fri, 25 Apr 2025 02:17:17 GMT</pubDate>
            <atom:updated>2025-04-25T02:21:18.424Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tYV9-fmbIAN7t-rxNhxMyA.png" /></figure><h3><strong>Hockey Context</strong></h3><p>Hockey’s playing surface makes it one of the most unique sports in the world. Skating requires a one-of-a-kind skillset that takes years to develop. While analytics have started to creep into most decisions made in modern day hockey, one area of the scouting process that remains old-school (at least in the public sphere) is evaluating a player’s skating. Most skating evaluations rely on the eye test to judge whether a player is fast and smooth. With the availability of tracking data, it’s possible to dig deeper than just watching film.</p><p>The NHL has recently released tracking data on NHL EDGE regarding the speed of players. Although this is valuable, it’s only scratching the surface of the capabilities that skating metrics can provide in player acquisition and allocation.</p><p>This project aimed to generate new change of direction skating metrics to gain a better understanding of a player’s strength and weaknesses. Additionally, I explored the overall impact that these measurements had on overall team performance.</p><h3><strong>Process</strong></h3><p>To create these new statistics, I used the data from the tracking, events, and shifts datasets provided by the Big Data Cup. The data was from three separate games where the names of the teams and players were generalized (For example Team A vs Team B).</p><p>Here are the five key metrics I produced to evaluate skating.</p><h4><strong>Agility Score</strong></h4><p>- What it measures: The average speed that a player changes direction.</p><p>- Calculation: Mean product of instantaneous velocity and angle changes.</p><p>- Significance: Higher scores indicate the ability to maintain speed while changing direction.</p><h4><strong>Angular Fatigue</strong></h4><p>- What it measures: How much the ability to change direction decreases over time.</p><p>- Calculation: slope(mean angle change vs. time intervals).</p><p>- Significance: Quantifies how sustained effort affects directional sharpness. A steeper negative slope indicates quicker fatigue during play.</p><h4><strong>Small Direction Change Recovery</strong></h4><p>- What it measures: Time to regain speed and balance after slight direction changes.</p><p>- Calculation: Time required to return to 90% of pre-direction change speed after turning at an angle greater than 45°, but less than 90°.</p><p>- Significance: Measures balance and recovery capabilities following gradual turns.</p><h4><strong>Large Direction Change Recovery</strong></h4><p>- What it measures: Time to regain speed and balance after sharp direction changes.</p><p>- Calculation: Time required to return to 75% of pre-direction change speed after turning at an angle greater than 90°.</p><p>- Significance: Measures balance and recovery capabilities following sharp turns.</p><h4><strong>Multi-Directional Explosiveness</strong></h4><p>- What it measures: A combined measure of how quickly a player changes direction, their speed, and smoothness of movement.</p><p>- Calculation: mean(turn acceleration during significant direction changes &gt; 45°)×peak jerk.</p><p>- Significance: Evaluates how explosively a player can execute high-impact directional changes.</p><p>I believe that these metrics cover a broad spectrum and can help uncover strengths and weaknesses that we are not able to uncover with public data.</p><h3><strong>Results</strong></h3><p>Using these statistics, we can compare the skating ability of different players in the dataset. Here is one of the best skaters: Player 90 on C.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/778/1*QSY9muCiB5PxK1hd3Jl8vQ.png" /></figure><p>Player 90 has elite agility and Multi-Directional Explosiveness. This shows that he has elite burst and control during rapid transitions which is essential in evading opponents in tight spaces.</p><p>On the other side of the spectrum, Player 56 on Team G is one of the worst skaters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/778/1*nDzzzW5OLg32jqYzoxCQ6w.png" /></figure><p>He possesses poor metrics across the board which is a huge red flag in the evaluation process.</p><h3><strong>Team Impact</strong></h3><p>Having a few new skating metrics is fun, but when push comes to shove, can we use these to win? If so, which are the most important in winning games? To research this, I created a simple expected goals model to assess the quality of chances that were created. This was only a three-game sample, so the sample size is limited. After building the model, I calculated the expected goals percentage of all players while they were on the ice, factoring in strength state, and found their correlations to each of the skating metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*DIN_ZC8WNQE9ZRoxXIkR0Q.png" /></figure><p>As you can see, the Multi-Directional Explosiveness is the most correlated and the Large Direction Change Recovery is the least correlated by a wide margin. This shows that we should take more stock in the top four metrics. An additional way to show how powerful Multi-Directional Explosiveness could be in player evaluation, 73.3% of the players in the top half of xG% possessed a Multi-Directional Explosiveness above the 50th percentile.</p><h3><strong>Conclusion</strong></h3><p>To wrap things up, this project shines a light on the untapped potential of skating analytics in hockey. While traditional scouting has long relied on subjective assessments of speed and smoothness, tracking data opens a whole new world of insights into a player’s skating ability. With metrics like Agility Score, Angular Fatigue, Small and Large Direction Change Recovery, and Multi-Directional Explosiveness, we can dive deeper into a player’s movement and skill, beyond just how fast they skate.</p><p>In the future, I hope more tracking data becomes available to the public so that change of direction skating statistics can be more widespread. With a larger sample size, I would love to explore specific scenarios such as how players with higher change of direction scores force turnovers on the forecheck or how defenders suppress the opponent’s chances in tight spaces behind the net.</p><p>By incorporating these metrics into the evaluation process, teams can gain a more well-rounded, data-driven understanding of players. It allows for identifying hidden strengths or weaknesses that might otherwise slip through the cracks. This approach has the potential to reshape how we assess skating in hockey and give teams a leg up when it comes to player development and finding under-valued players.</p><p>Sources:</p><p>Matthew Barlowe 2017 — “NHL Expected Goals Model”</p><p><a href="https://rstudiopubsstatic.s3.amazonaws.com/311470_f6e88d4842da46e9941cc6547405a051.html">https://rstudiopubsstatic.s3.amazonaws.com/311470_f6e88d4842da46e9941cc6547405a051.html</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c0716bed623" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Optimizing Defensive Matchups Using Machine Learning]]></title>
            <link>https://medium.com/@ryancmiele/optimizing-defensive-matchups-using-machine-learning-02e23447a8dc?source=rss-0815fc24bf98------2</link>
            <guid isPermaLink="false">https://medium.com/p/02e23447a8dc</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[nba]]></category>
            <category><![CDATA[basketball]]></category>
            <category><![CDATA[predictive-modeling]]></category>
            <category><![CDATA[r]]></category>
            <dc:creator><![CDATA[Ryan Miele]]></dc:creator>
            <pubDate>Wed, 15 Jan 2025 17:31:06 GMT</pubDate>
            <atom:updated>2025-01-15T17:31:06.893Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Pen6Qxujz7ppZwxseXTllw.png" /></figure><h3>Addressing a Problem</h3><p>Basketball strategy has evolved over the past decade to enhance player performance, with a strong focus on improving shot selection. While analytics have crept into almost every decision made in the modern era of basketball, one area that remains medieval is choosing defensive matchups. Since the inception of man-to-man defense, the common philosophy has been to match players with similar heights to create the most balanced defensive lineup. While this is the standard approach, there are certainly slight tweaks made, such as adjusting matchups based on defensive ability and assigning the best defender to the opposing team’s top scoring threat.</p><p>I’m not naïve to the fact that numbers play a role in determining matchups especially at the NBA level, but it’s not to the degree that it should be. Some offensive players struggle to score against certain physical measurements that are different from the typical body type that they’re accustomed to competing against, such as a short and quick defender or one with long wingspan.</p><p>My goal for this project was to develop a model that identifies the optimal physical attributes needed to minimize offensive production against each offensive player in the NBA. Then, I identified the current player whose attributes most closely align with these optimal dimensions, making them the best candidate to effectively defend the opposing player.</p><h3>Matchup Impact Model</h3><p>My first step was attempting to answer the question of how well a defender guards an offensive player strictly through numerical data. There is a very valuable function in the hoopR package that allows me to load detailed data for every individual matchup. This dataset included the offensive and defensive players involved, the amount of time they matched up, and key box score stats such as shot attempts, points scored, fouls, blocks, turnovers, and more.</p><p>I loaded every offensive/defensive matchup over the past 5 seasons with a minimum 25 possessions for that season. I chose this threshold because I wanted to exclude small sample sizes, but still leave a low enough bar to include non-typical matchups such as a smaller defender guarding a larger offensive player and vice versa. I separated the matchups by year and made each of the box score stats on a per possession basis.</p><p>To quantify the effectiveness of the defender, I built a multiple linear regression model using <strong>points per possession</strong> as the target variable. After testing for multicollinearity and removing a few box score stats, the final model included the normalized values of <strong>field goal percentage</strong>, <strong>shooting fouls per possession</strong>, <strong>blocks per possession</strong>, and <strong>turnovers per possession.</strong></p><p>After running the model, I used the coefficients as weights to create a new metric called <strong>Matchup Impact Score</strong>. This statistic consolidates all aspects of the matchup into one number based on the importance of each input. I believe this is a better way to evaluate a matchup compared to stand-alone metrics such as points or FG% allowed since it gives more context.</p><p>For perspective on the scale, Matchup Impact Scores span from -5 to 12 (A lower score means better job defensively) with a median of -0.05.</p><p>Here are the best individual offensive performances against a specfic defender over the last 5 seasons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*46I2XU3l_py1muEkbfoojQ.png" /></figure><p>Here are the worst individual offensive performances against a specific defender.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zPMIHgPA9LZr6l4DxO0vSA.png" /></figure><p>All of the members of the bottom 10 scored 0 points against their defender, so the main separator was how many times turnovers they committed.</p><h3>Data Preparation</h3><p>I wanted to add more context to the offensive and defensive players in the matchups data frame. Therefore, I loaded all NBA Combine physical measurements from 2000–2024 to extract each player’s height without shoes and wingspan. I noticed that there were two main issues with this data: some players, such as Victor Wembanyama, did not participate in the Combine, and players may have experienced slight growth spurts in the years after they were measured during their rookie season. To address this, I downloaded the heights and wingspans for all players for the current season from<a href="https://craftednba.com/"> CraftedNBA</a>.com. The matchup data I was using in this project was for the last five seasons, so between the data from CraftedNBA and the Combine, I was able to gather information for about 99% of the players.</p><p>Next, I wanted to describe the shot profile and quality of each offensive player. I extracted shot locations for each player and categorized each shot as either a Rim, Short Mid-Range, Long Mid-Range, or Three-Point attempt. Then, I calculated the frequency and efficiency with which they shot in each zone for that particular season.</p><p>I then joined this data to the matchups data frame, adding the height and wingspan characteristics for both offensive and defensive players, as well as the offensive statistics for the offensive players only.</p><h3>Predicting Matchup Impact Score</h3><p>After wrangling the data into the proper format, I created an xgboost model to predict Matchup Impact Score based on the shot profile and physical measurements of the matchup. Here is how important each of the inputs are in determining Matchup Impact Score.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vFg5GtLJxHkzfYbJ2ls8gg.png" /></figure><p>The defender’s height is the most important, followed by wingspan and length. The most important aspects of the shot profile are the Rim FG% and how frequently they shoot from short Mid-Range which is interesting. The least important factors are Rim frequency and the offensive player’s height. Rim frequency being the lowest is surprising to me since I assumed that players who got to the rim more would have a higher overall efficiency.</p><p>These predictions were added to the larger data frame so we now have the expected Matchup Impact Score as well as the actual.</p><h3>Optimizing Matchups</h3><p>I’m pretty familiar with creating predictive models at this point, but this next section was uncharted territory (shoutout to Nick Kalinowski for introducing me). Using <strong>Bayesian Optimization</strong>, I sought to find the optimal height and wingspan of the defender to minimize the expected Matchup Impact Score. Essentially, I created a function where you input an offensive player, and it finds the physical dimensions that would defend the player most effectively. Running this machine learning function is a relatively lengthy process, as it takes roughly 5 minutes to calculate for each individual player.</p><p>After acquiring the defender’s optimal height and wingspan, I wanted to see which current NBA player mirrors these measurements most closely. To accomplish this, I assembled another function. <strong>This function does not account for the quality of the defender; it simply compares the physical characteristics.</strong> I chose not to include defender quality because the model would overly favor better defenders when I wanted the focus to remain on the opponent’s size.</p><p>Below is an example output of the optimization function for Trae Young in 2025.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PrY7_tareGGUMFsR" /></figure><p>The model makes it clear that the best way to guard Trae is with a smaller defender who has tremendous length. Again, when it states that Terry Rozier is the optimal defender, it is simply considering the physical measurements and not the quality of the defender.</p><p>For another example, here is the output for Kevin Durant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3Cilf6p6pgI9btHhj3VX-w.png" /></figure><p>It’s best to guard Durant with a defender a few inches shorter than him who possesses a long wingspan to contest his pull-up jumpers.</p><p>Another way we can visualize how well defenders match up is with an aggregated hex chart. Remember, lower values indicate better defensive performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EvJHuE6RH3dUhJf1Mg16Zg.png" /></figure><p>As we can see in this chart, the defenders who have had the most success are around 6&#39;0&quot; tall with a wingspan between 6&#39;7&quot; and 6&#39;8&quot;. Trae Young has excelled the most against defenders around 6&#39;4&quot; tall with a 6&#39;4&quot; wingspan.</p><p>In addition to player deployment strategies, these models and visualizations can also be used for identifying players to acquire. For example, let’s say a team wants to beat the Thunder but doesn’t have a great on-ball guard defender to guard Shai Gilgeous-Alexander. They can use this model to get creative and acquire someone via trade or free agency to attempt to stifle the MVP candidate. You can also view this more generally and pursue free agents or draft prospects who fit the mold of an archetype that players tend to struggle against.</p><h3>Next Steps and Ways to Improve</h3><p>Moving forward, the logical next phase would be to build an application (or simply a list) of each offensive player’s optimal defensive physical measurements and the corresponding player who aligns most closely. There could also be a team filter to help solve the problem of “How should the Brooklyn Nets guard Jayson Tatum?” — where you choose a team, and it will match Jayson Tatum’s optimal defender measurements with someone on their roster instead of the entire league. The major roadblock to building this application is how long the model takes to run, so I may need to come up with a faster way to make it work.</p><p>There’s definitely room for improvement in building the models. For the Matchup Impact Score metric, the weight in the formula for turnovers was much lower than I anticipated, since forcing a turnover should be worth more than allowing a shot. Ideally, I would determine how much value forcing a turnover in a specific matchup has on winning the game (please let me know if you have any ideas). This would make the Matchup Impact Score a much better representation of defensive value.</p><p>Here is a link to all of my code for this project: <a href="https://github.com/RyanMiele/NBA/blob/main/Optimizing%20Defender%20Matchups">https://github.com/RyanMiele/NBA/blob/main/Optimizing%20Defender%20Matchups</a></p><p>Thank you so much for reading. Like always, any constructive criticism or comments are encouraged and genuinely appreciated. Quick shoutouts to @kalidrafts and @Tejfbanalytics for help with the models.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=02e23447a8dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modeling the Goal Threat of Point Defensemen on the Power Play]]></title>
            <link>https://medium.com/@ryancmiele/modeling-the-goal-threat-of-point-defensemen-on-the-power-play-51da94769c64?source=rss-0815fc24bf98------2</link>
            <guid isPermaLink="false">https://medium.com/p/51da94769c64</guid>
            <dc:creator><![CDATA[Ryan Miele]]></dc:creator>
            <pubDate>Mon, 26 Aug 2024 14:49:59 GMT</pubDate>
            <atom:updated>2024-08-26T14:49:59.730Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e6xof0h1qjkHMgzXHQkdzQ.png" /></figure><h3><strong>Hockey Context</strong></h3><p>Power plays are one of the most substantial organized advantages in all professional sports. Teams averaged 3.02 power plays per game in the 2023–24 regular season, which means that roughly 1/10th of the game was spent with an extra man. It is critical to capitalize on these opportunities by gaining any competitive edge possible. Schematically, most NHL teams apply the same structure with slight variations to their 5v4 power play. They utilize a 1–3–1 shell where a mobile, puck moving, defenseman controls the point, 2 elite scoring forwards who present a one-timer threat position themselves on the wings, an opportunistic goal scorer who can provide support for board battles sets up in the high slot between the faceoff circles called the bumper, and a big, gritty player positions in front of the net looking to clean up the scraps and block the goalie’s vision.</p><p>Based on the eye test, it seemed like the best scoring chances on the powerplay are created by the shots of the 4 forwards. Yes, the point defenseman is key to making quick, smart passes, but, to me at least, it seemed like they didn’t have much of an impact in terms of shooting. Shots from the point appear to have a low likelihood of reaching the back of the net, which gives the impression that teams are better off cycling down low to the forwards.</p><p>This project aimed to assess the effectiveness of point defensemen on the power play by evaluating how well their shots contribute to goals. Additionally, I explored the overall impact of having a higher point goal threat on power play performance.</p><h3><strong>Process</strong></h3><p>To investigate this premise, I used 5v4 NHL power play data from the last 5 regular seasons. I strictly used 5v4 power plays since they’re the most common, and I didn’t want players who had more opportunities at 5v3 to skew the results. I used Dan Morse’s hockeyR package to extract play by play data, and moneypuck.com for individual stats. These are both free data sources that are open to the public. The primary focus of the project is on shooting since each shot attempt has an (x,y) location in the play by play data. The only passing information that can be taken would be assists which would only paint a partial picture. By only including assists, it would ignore players who made excellent passes into high danger areas that were missed by the recipient.</p><p>To quantify the point defenseman’s ability to create goals from their shot, I knew that simply counting goals would be too straightforward, and shooting percentage involves a ton of context and possibly luck. Therefore, I knew a new metric would be necessary. There are 3 main ways that a player’s shot can result in a goal: a normal goal from the shooting player, a shot that is deflected in by a teammate, and a shot that leads to a rebound goal. To eliminate some of the randomness with rebound goals, I used expected goals from expected rebounds. This takes the characteristics of the shot to determine the odds that it would create a rebound. Again, this combats the considerable amount of luck involved in generating rebounds, and thus does a better job at predicting goals in the future. Here is the formula:</p><p><strong>Point Goal Threat</strong> = (Indiv Goals Scored+ Primary Assts on Deflections + xGoals on xRebs) / Time on Ice</p><p>The final number can be interpreted as goals that the player has generated from his shot per 60 minutes. Here are the best and worst defensemen in this metric last season.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/818/1*C7nqYJPuimO-TKhTfXHhJA.png" /></figure><p>It’s refreshing to see that we have some of the usual suspects near the top, such as Bouchard, Makar, Josi, and Quinn Hughes. I knew Noah Hanifin was a great player, but I didn’t realize just how dominant he was on the power play.</p><h3><strong>Impact on Other Positions</strong></h3><p>After creating this new metric, I wanted to explore how it affects other players on the ice. Using play by play data from the hockeyR package, I was able to categorize power play positions based on shot locations (There’s more detail on how I grouped these locations in my code). Before testing, I predicted that a greater threat from the point defenseman would create more opportunities for others, leading to a higher shooting percentage. Let’s see if this holds true. Here’s a look at shots from the wing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/846/1*dl0YHrEHxCQm8hyks-vLRA.png" /></figure><p>The left column groups all defensemen the last 5 seasons who played at least 50 minutes on the power play based on their Point Goal Threat percentile. Power plays who had a point defenseman in the 75th-100th percentile shot 1.5% higher from the wing position than the 0–25th percentile.</p><p>Now, let’s look at the bumper position.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/870/1*a0JjJfSZcfeDbPJAjA8-Rw.png" /></figure><p>Power plays who had a point defenseman in the 75th-100th percentile shot 2% higher from the bumper position than the 0–25th percentile.</p><h3><strong>Overall On Ice Impact</strong></h3><p>It’s fun to create a new stat, but when push comes to shove, is it useful? Can we use it to anticipate future power play results? The Point Goal Threat of the defenseman on the power play has a 0.44 correlation to goals scored per 60. This is a decently strong correlation for a specific stat and shows how important the point defenseman truly is. I also created a model to predict power play goals per 60. Here are the model inputs and their feature importance. As you can see, Point Goal Threat has the largest impact by a wide margin.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1008/1*k2wXYU3Cz9uu-cNm4o8a0A.png" /></figure><p>Next, I looked to see how well it can predict next year’s results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*xswwq4gmO5Sfwrfv4246LA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/628/1*BLvqfWkVuRxqkh3aACB4QQ.png" /></figure><p>The model is better at predicting next year’s power play goals than actual goals scored. This shows how powerful Point Goal Threat is and how valuable it can be.</p><h3><strong>Conclusion</strong></h3><p>Point Goal Threat is an interesting new tool that can be utilized by organizations to improve their power play. It’s a more accurate way to measure goal creation than the standard stats that are typically used such as goal totals and shooting percentages. Again, it should not be used as the “end all be all” of who is better, but when used correctly, it can be incredibly valuable for player evaluation.</p><p>This analysis also shed some light on how important it is to have a defenseman on the power play who presents a serious threat to the defense. Better defensemen improve shot quality for their teammates which leads to more goals scored.</p><p>Thank you so much for reading and I hope you were able to take something from my research! As always, please reach out if you have any questions. Here is my <a href="https://github.com/RyanMiele/NHL/blob/main/NHL%20Utah">github</a> if you want to take a look at the code.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=51da94769c64" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Modern Big Man: How Stretch Centers Influence Shot Selection and Outcomes]]></title>
            <link>https://medium.com/@ryancmiele/the-modern-big-man-how-stretch-centers-influence-shot-selection-and-outcomes-c09b202a0360?source=rss-0815fc24bf98------2</link>
            <guid isPermaLink="false">https://medium.com/p/c09b202a0360</guid>
            <dc:creator><![CDATA[Ryan Miele]]></dc:creator>
            <pubDate>Mon, 05 Aug 2024 14:50:30 GMT</pubDate>
            <atom:updated>2024-08-05T14:50:30.557Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-GuOpehd_sKTbZ-CbvoTcQ.jpeg" /></figure><h3>Basketball Context</h3><p>It’s no secret that having a big man who can shoot 3-pointers is a huge advantage. Although it’s becoming more prevalent, having a 7-footer who can stretch the floor is rare. This transfers stress to the defense by opening up space for other offensive players to operate and allowing the opportunity to add new wrinkles to the playbook. I wanted to explore how the presence of a stretch 5 influences the defense and if it enhances the team’s shot selection.</p><p>This idea came to mind while watching the Dallas Mavericks’ defense during the playoffs. In the Western Conference Finals, the Timberwolves starting lineup featured Rudy Gobert at center who is not a threat to shoot the 3 ball. This allowed the Dallas centers to sag off and help with protecting the rim.</p><p>This is displayed in the following clip:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*KJNBy8HZegFPwMCSmslIWQ.gif" /></figure><p>Since Gobert isn’t a threat to shoot from deep, Lively can take his eyes off of him early and roam to the rim to execute a weak-side block. This ability to clog the paint is a huge advantage for defenses and is always in the back of the mind of offensive players while driving.</p><p>Let’s compare this to when Dallas played against the Celtics in the Finals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*Vch33SZk4CEc0jZgObns9A.gif" /></figure><p>Here, Jaylen Brown drives to the paint and draws help from the Maverick’s rim protector, Gafford. This leaves Boston’s center, Horford, wide open in the corner and he sinks the 3.</p><p>Let’s look at a similar scenario in another possession.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*MuJRODASyVa_R8yM92v92A.gif" /></figure><p>Lively can see Jrue Holiday blow by his defender and he has to make the calculated decision to stick on Porzingis to negate the 3 attempt instead of helping on the drive. (This is also genius by Mazzulla to put a guard in the dunker spot to force a smaller defender to protect the rim, but that’s a story for a different article) I know these are only a few possessions, but I think they do a good job of highlighting the conflict stretch 5s can give opposing centers.</p><h3>Why Does This Matter?</h3><p>I think the general notion of the value of stretch 5s stems from being a kick-out 3 option when the defense collapses on drives. However, the more a center excels from behind the arc, his defender is further inclined to stick on him. <strong>My theory is that having a center who is a consistent threat from 3 removes the defense’s rim protector from the paint which would result in more frequent, and more efficient, rim attempts. </strong>I know this isn’t a mind-blowing hot take of a theory, but I wanted to investigate whether there was evidence of this occurring.</p><p>Going back to the examples above, Minnesota had a rim attempt rate against Dallas that was 3% more than Boston’s rim attempt rate against Dallas. Although the team with the stretch 5s did not get to the rim more in these series, Boston shot a whopping 6.1% higher FG% at the rim (stats via <a href="https://www.pbpstats.com/">pbpstats.com</a>). This supports the idea that the lack of shot-blocking presence in the paint makes the life of the driver easier.</p><p>These two series are a small sample size, so let’s explore whether this trend exists in the entire NBA.</p><h3>Process</h3><p>To dive into my hypothesis, I took every 5 man lineup from the last 5 NBA regular seasons that played at least 50 minutes together. Next, I defined a stretch center to be any center that had at least 100 3-point attempts in that season. For example, Alperen Sengun was not considered a stretch center for the 2022–23 season since he did not meet this threshold, but he was in 2023–24. I considered adding a 3P% minimum as well, but the results weren’t affected so I kept it simple.</p><p>A tricky problem to navigate was lineups that didn’t include a center. Some guys who frequently played center in small-ball lineups, like Chet Holmgren and Kevin Love, are listed as PFs by NBA.com. I had to do some data wrangling, but I was able to combine these stretch PFs with the stretch centers under the term “stretch big”.</p><h3>Results: Rim Shot Location</h3><p>Why don’t we look at some results? In the last 5 regular seasons, here’s the difference between lineups that included a stretch big vs lineups that did not.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CRyCJScNfpNlmMvMI1WI_A.png" /></figure><p>We can see from the table that <strong>lineups that included stretch bigs had a higher rim attempt frequency, rim shooting percentage, and overall shot quality while having a lower percentage of their shots at the rim blocked.</strong></p><p>Let’s take a look at if this holds for only last season.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h3qG3fI874Z7hO6_lnXV2Q.png" /></figure><p>During the 2023–24 season, we saw an increase in the difference between rim attempt frequency, rim shooting percentage, and blocked rim attempts.</p><p>Does this trend hold true in the playoffs?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RHXMN-Q0cjQauXweq_nFBg.png" /></figure><p>I found these results to be very interesting. In the last 2 playoffs there was a microscopic difference between the two types of lineups in rim attempt frequency, blocked rim attempts, and shot quality. We do still see the gap which favors lineups with stretch 5s in rim shooting percentage.</p><p>Another interesting takeaway is the number of occurrences in each type of lineup. <strong>There was a huge disparity during the regular season, but the split between lineups that featured a stretch big or not dwindled to almost 50/50 for the playoffs.</strong> Can we infer that the teams who make the playoffs are more likely to deploy this big-man archetype?</p><h3>Results: Overall Effectiveness</h3><p>Okay so lineups with a stretch 5 have an easier time getting to the rim, so what? Are they superior in all facets of the game, or only in that one specific area?</p><p>I’m glad you asked!</p><p>Here are the last 5 regular seasons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NQE3ZwDwMKGEZZbvlsg-8g.png" /></figure><p>Lineups with a stretch 5 have an Offensive Rating that is 0.7 points higher while having identical Defensive Ratings. There’s a slight edge here, but nothing too extreme. Both Net Ratings are positive which shows there’s some selection bias since more successful lineups are more likely to hit the 50-minute threshold together.</p><p>Things get more interesting when looking at the playoffs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1018/1*IQyAxv9ixjFxoi6yhdok-A.png" /></figure><p><strong>In the playoffs of the last 5 seasons, lineups with stretch 5s were significantly better. </strong>They had an Offensive Rating that was 2.5 points higher as well as a lower Defensive Rating.</p><p>Things get even more drastic when you look at only the last 2 playoffs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1014/1*8NeD1qyFIW5rEvQDT_X70A.png" /></figure><p><strong>Lineups that included a stretch 5 had a 5.9-point higher Net Rating in the playoffs of the last two seasons.</strong> That’s a HUGE gap in performance and something that should stick out to every basketball fan.</p><h3>Application to Team Building</h3><p>Although there was only a slight edge in overall results during the last 5 regular seasons, the playoff numbers show how impactful a stretch big can be when it matters the most. I believe that a stretch big is a necessity if you’re trying to win an NBA championship. <strong>My philosophy in centers is that unless they are an elite defender/rim protector (top 3–5ish in the league), then they need to at least present the threat of scoring from 3.</strong> This is one of the main reasons I was so high on Kel’el Ware in the pre-draft process. Although he only attempted 1.3 3s per game last season, he knocked down an impressive 42.5% of them. A 7-footer averaging 1.9 blocks per game, with the ability to spread the floor is exactly what you want in today’s game.</p><p>As I discussed earlier, one of the main advantages that the stretch 5 creates is that it pulls the main rim protector out of the paint. To combat this, <strong>I predict there’s going to be a defensive schematic trend for the defensive rim protector to use matchup zone principles to not get pulled out to the 3-point line.</strong> It would take a lot of switching and communication, but I think it could help to battle the increased rim FG% that we saw above.</p><p>If you made it this far, thank you so much for reading! Please let me know if you have any questions/comments/constructive criticism. Special shout-out to Sauurabh Rane for introducing me to pbpstats.com (follow him on twitter @SaurabhOnTap)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c09b202a0360" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Weak Link Theory: Does the Worst Defender Have a Greater Impact than the Best Defender?]]></title>
            <link>https://medium.com/@ryancmiele/the-weak-link-theory-does-the-worst-defender-have-a-greater-impact-than-the-best-defender-9869bc0b0634?source=rss-0815fc24bf98------2</link>
            <guid isPermaLink="false">https://medium.com/p/9869bc0b0634</guid>
            <dc:creator><![CDATA[Ryan Miele]]></dc:creator>
            <pubDate>Mon, 22 Jul 2024 14:56:54 GMT</pubDate>
            <atom:updated>2024-07-22T14:56:54.219Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AU5PmDo-cSlPyegBDoaKxg.jpeg" /></figure><h3>Basketball Context</h3><p>Automatic switching has become one of the most common defensive principles in the NBA over the last 10 years in an attempt to neutralize the 3-point attempt. While switching does an effective job at this, it leaves the defense vulnerable to mismatches.</p><p>An offensive strategy to counter the automatic switch is “hunting” for advantages. Using a screen, a great scorer will force a defensive switch to be matched up with the opposition’s worst defender to create an isolation opportunity. Offenses forcing switches to create these favorable matchups have become so prevalent that it seems as if the best defenders aren’t guarding the best offensive players very often.</p><p>Due to the frequency of being “hunted”, it made me wonder <strong>if the worst defender on the floor had a larger overall defensive impact compared to the best defender.</strong></p><h3>Introducing the Weak Link Theory</h3><p>I first read about the Weak Link Theory in a Football article written by Eric Eager (read it <a href="https://sumersports.com/the-zone/the-weak-link-nature-of-football/">here</a>) where he discusses the idea that in pass coverage, the worst pass defenders have greater ramifications than the top pass defenders. Similar to hunting a matchup in basketball, Quarterbacks can identify the inexperienced or inadequate defender and exploit that mismatch while avoiding the threat of the best pass defender. This suggests that the defensive unit is only as good as its weakest link.</p><p>The 2024 NBA Finals was the perfect illustration of this concept. On one side you had the Mavericks who, although having a very solid defense in the second half of the season, had a clear weak link in Luka. The Celtics routinely attacked Luka who was battling injuries and was clearly a step or two behind. This strategy ended up paying dividends as it got them many easy buckets.</p><p>On the flip side, the Celtics’ defense was incredibly balanced, showing no signs of vulnerability. Even if Luka or Kyrie wanted to hunt for a mismatch, there wasn’t an option. Holiday and White are two of the best defensive guards in the world. Tatum and Brown are long, athletic, and bring intensity on the defensive end. Even Porzingis and Horford are mobile for their size and contest shots well. There was no clear Achilles heel to exploit. I believe this was one of the main reasons the Celtics were so dominant all season.</p><h3>Process</h3><p>To investigate this theory, I used the NBAr package to extract every 5 man lineup from the last 5 regular seasons that played at least 40 minutes together. I know that 40 minutes is a very arbitrary minutes filter, but I thought around 50 defensive possessions was a solid balance of eliminating the high variability of small sample sizes while not excluding too many unique lineups. To quantify each player’s defensive ability, I used Basketball Index’s DLEBRON metric. This is an all-in-one defensive metric that estimates a player’s defensive impact per 100 possessions (read more about it <a href="https://www.bball-index.com/lebron-introduction/">here</a>). I joined this to the lineup data frame and calculated the best and worst defender in the lineup based on their DLEBRON. From there, I explored their relationship to the lineup’s overall defensive rating.</p><h3>Exploration</h3><p>Over the last 5 years, 1828 different lineups played over 40 minutes together in a season. Here is the breakdown of the best defenders in those lineups.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/824/1*Oy3e1xuX0HyYkKB8yK9tng.png" /></figure><p>Centers are most frequently the best defender on the floor at 50.8%. Although guards are only the best defenders 11.6% of the time, they have the best defensive rating for the overall lineup. Having an elite guard defender is very beneficial for the overall team’s defense.</p><p>Now, let’s look at the worst defenders in the lineup.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/848/1*iJyspd-axtnyV70GQ5bzew.png" /></figure><p>It’s clear that a guard is the worst defender on the floor an overwhelming majority of the time. Centers are practically never the liability on defense which is very interesting. The sample size for the center is too small to draw a conclusion for the defensive rating.</p><h3>Results</h3><p>Now let’s get to the real interesting part: the correlations. <strong>Is the quality of the best defender more correlated to the lineup’s total defensive rating than the quality of the worst defender?</strong></p><p>Drum roll, please….</p><p><strong>Over the last 5 NBA regular seasons, the best defender had more of an impact on the lineup’s defensive rating.</strong> The DLEBRON of the best defender has a correlation of <strong>-.209</strong> and the DLEBRON of the worst defender has a correlation of <strong>-.156</strong>. I’m slightly let down by these results since they disprove my original hypothesis.</p><p>The next question would be how does the quality of the 2–4 best defenders on the floor affect the lineup’s defensive rating? Is there a stronger correlation between them compared to the best/worst defender? Let’s take a look.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4o3pD1ilvySHip27xFSQNQ.png" /></figure><p>As you can see, the 2nd and 3rd best defenders are more important than the best defenders in preventing points. Obviously, the best defender is still important since to have a highly rated 2nd best defender you need to have a highly rated best defender. I was very surprised to see that the worst defender has the least impact compared to the other 4 defenders.</p><h3>Change Over Time</h3><p>I was curious about whether this relationship is shifting over time. Let’s take a look.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tXppZt860DM5UgYETyq67g.png" /></figure><p>Although the best defender has a stronger correlation on average throughout the last 5 seasons, the worst defender has been on the rise recently. <strong>The worst defender has had a greater influence in each of the last 2 seasons.</strong> In the last 2 NBA regular seasons, the DLEBRON of the worst defender has a correlation of <strong>-.231</strong>, and the DLEBRON of the best defender has a correlation of <strong>-.198</strong>.</p><h3>Application to Team Building</h3><p>I believe that the NBA is trending towards a time where the 5th defender is crucial to success. I think we’ll see teams attempting to replicate the Boston model of not having a weak link. Boston’s “worst” defender in their starting lineup according to DLEBRON was Jaylen Brown who was in the 49th percentile across the league. So their worst defender was right around league average. They even had a deep defensive bench that featured Kornet (88th percentile), Horford (86th percentile), and Hauser (54th percentile).</p><p>Sam Presti and the Thunder have already hopped on the trend this offseason by acquiring Hartenstein (98th percentile) and Caruso (97th percentile). Similar to Boston, OKC had a very strong 5th defender in Josh Giddey who was in the 47th percentile. I’m sure it was just a coincidence that both number 1 seeds prioritized a defense without a weak link. Presti saw an opportunity to bolster a point of strength and now he can flaunt a starting unit whose worst defender by DLEBRON is Jalen Williams who was in the 62nd percentile. Pretty scary!</p><p>I understand that building a team this way is easier said than done. It’s difficult finding 5 guys who are above-average defenders while also being capable on the offensive end. Therefore, something we can take away from my research is that at minimum <strong>teams should attempt to have 2 very good defenders in their starting lineup</strong>. As discussed previously, the DLEBRON of the 2nd best defender in the lineup has the strongest correlation to the lineup’s defensive rating. For that reason, I think GMs should make this a major point of emphasis when constructing their roster.</p><h3>Future Steps</h3><p>My research in this article looked solely at the defensive end of the court. I realize that having offensive shot creators is extremely important, and a lot of teams will live with a player who may be a liability on defense due to their offensive talents. I’d say Trae Young is a perfect example of this. He gets routinely attacked on the defensive end, but his elite offensive playmaking allows him to overcome this.</p><p>In the future, I want to see if I can find the threshold of how good a player has to be offensively to overcome being a liability on defense for the lineup. Similarly, I want to see how good the other teammates have to be defensively to shelter a defensive liability.</p><p>Thank you for reading, and I truly hope that this research added some value to you. This is my first time writing an article like this, so all feedback and constructive criticism is encouraged.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9869bc0b0634" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>