<?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 Leonardo Acquaroli on Medium]]></title>
        <description><![CDATA[Stories by Leonardo Acquaroli on Medium]]></description>
        <link>https://medium.com/@leonardo.acquaroli00?source=rss-49e39e26175e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*_ww9rdYckSv-0V97iDTZ9Q.jpeg</url>
            <title>Stories by Leonardo Acquaroli on Medium</title>
            <link>https://medium.com/@leonardo.acquaroli00?source=rss-49e39e26175e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 02:31:21 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@leonardo.acquaroli00/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[Python and R libraries for soccer — 90° QUANTILE [8]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/python-and-r-libraries-for-soccer-90-quantile-8-3ff70fe76fed?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ff70fe76fed</guid>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Wed, 24 May 2023 15:06:39 GMT</pubDate>
            <atom:updated>2023-05-24T15:09:39.354Z</atom:updated>
            <content:encoded><![CDATA[<h3>Python and R libraries for soccer — 90° QUANTILE [8]</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*bYtP-Caj--uUO-oO.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme<br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>The data industry in football is growing and alongside follow technologies.<br>Today we are going to find out the<strong> most famous libraries </strong>used <strong>to analyse football data</strong> <strong>and create nice visualizations</strong> in order to better understand the <em>Beautiful Game</em>.</p><p>I will cover Python and R libraries and the reason is quite intuitive: R and Python, are, by far, the most widely used programming languages for any kind of data analysis and in the area of football they are no less.</p><p>So, let’s get the ball rolling!</p><p><strong>🐍 PYTHON:</strong></p><ol><li><a href="https://github.com/statsbomb/statsbombpy"><strong>StatsBombPy</strong></a>: We saw in a <a href="https://medium.com/@leonardo.acquaroli00/the-only-2-datasets-of-football-data-90-quantile-5-e403af1949ef">previous story</a> that there are only two open datasets with soccer events and one of these is made available by StatsBomb, probaby the most advanced sport-tech company in the world. This library can be used to parse those data from both the StatsBomb’s open database and the full complete database for customers through their API. And you can call it directly from your Jupyter notebook. The greatest obvious advantage it brings is that you don’t have to download any byte apart for the installation of the library itself. The second main point is that their data collections come from very recent seasons, for some leagues up to 20/21 and 21/22.</li><li><a href="https://mplsoccer.readthedocs.io/en/latest/index.html"><strong>mplsoccer</strong></a>: is the basic tool to start analysing events data beacause it is built on top of <em>matplotlib </em>(from which <em>mpl</em>soccer) and so it can create a lot of pitch representation and cool visualizations with just some known commands. It also offer a function to directly connect to and parse the StatsBomb datasets abovementioned.</li><li><a href="https://codeball.metrica-sports.com/"><strong>Codeball</strong></a>: is a library specifically designed to work with a software called Metrica Play and aims to create a centralized spot for both statistical and tactical video analysis. It is optimized to work with tracking data and can create very interactive videoclips with a lot of measurements, such as the one we watch on tv.</li></ol><p>Here it is an example of the usage of mplsoccer to load the StatsBomb open data set</p><pre>Input:<br>  from mplsoccer import Sbopen<br>  parser = Sbopen()<br>  competitions = parser.competition()<br>  competitions[competitions.season_name == &quot;2020/2021&quot;]<br><br>#Output:<br>#  FA Women&#39;s Super League<br>#  La Liga</pre><p><strong>🖥️ R:</strong></p><ol><li><a href="https://github.com/statsbomb/StatsBombR"><strong>StatsBombR</strong></a><strong>: </strong>As you can imagine this is the counterpart of the Python library. For the record, it has to be said that the R library was the original one and the StatsBombPy came later.</li><li><a href="https://www.rdocumentation.org/packages/soccermatics/versions/0.8.4"><strong>Soccermatics</strong></a><strong>: </strong>provides tools to visualise spatial information from soccer matches using x,y-coordinates of players. There are currently functions to visualise player trajectory paths, heatmaps of players’ position frequency, flow fields showing players movement over the pitch and much more. Its name is taken by a masterpiece of an authority in the field of football analytics, <a href="https://medium.com/u/be10d0434daf">David Sumpter</a>’s book “Soccermatics”.</li><li><a href="https://github.com/FCrSTATS/fc.rstats"><strong>FC_rStats</strong></a><strong>: </strong>a<strong> </strong>package with parsing, data preparation and plotting functions that help people working with OPTA and ChyronHego data.</li><li><a href="https://www.datofutbol.cl/blog/soccer-animate-r-package/"><strong>{socceranimate}</strong></a><strong>: </strong>to create amazingly smooth 2D soccer animations using tracking data. Unfortunately, as other packages we saw, it is highly binded to a specific data provider. Indeed, it works only with the data format of Metrica Sports, since they made public an example of tracking dataset.</li></ol><p>This article shows up some tools of the trade, each one with more or less slighlty different applications and so very useful to start playing with advanced soccer data.</p><p>Yet, what I would like to highlight is that still a lot of functionalities and data types (e.g. tracking data) are kept private for (legitimate) commercial interests.<br>This is perfectly fair to me but…it is a great limitation to the development of the soccer analysis industry, because it pushes talents away, towards more open sectors.</p><p>Again, my hope is that<strong> sport-tech companies and sport clubs will soon start collaborating with schools and universities</strong> because more awareness means more <strong>talents </strong>and also greater <strong>business opportunities </strong>in the future.</p><p>FINAL WHISTLE! 🕛</p><p><strong>References</strong>:</p><p><a href="https://www.linkedin.com/posts/mohamed-lemine-beydia_footballanalytics-python-codeball-activity-7059403690658398208-5mFk?utm_source=share&amp;utm_medium=member_desktop">Mohamed Lemine Beydia on LinkedIn: #footballanalytics #python #codeball #datascience | 104 comments</a></p><p><a href="https://www.linkedin.com/posts/mohamed-lemine-beydia_football-footballanalytics-rstats-activity-7057595418376822784-3NcF?utm_source=share&amp;utm_medium=member_desktop">https://www.linkedin.com/posts/mohamed-lemine-beydia_football-footballanalytics-rstats-activity-7057595418376822784-3NcF?utm_source=share&amp;utm_medium=member_desktop</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ff70fe76fed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Playerank, an algorithm to rate football players— 90° QUANTILE [7]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/playerank-an-algorithm-to-rate-football-players-90-quantile-7-87d8515a91eb?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/87d8515a91eb</guid>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Fri, 12 May 2023 09:05:15 GMT</pubDate>
            <atom:updated>2023-05-12T09:05:15.588Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*4H8WQl4j4mWzbzSM.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme<br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>Have you ever dreamt of having a way to rate football players like in Fifa? From 2019 you can (almost) do it by using <strong>Playerank</strong>.</p><p>This is an italian project developed mainly by two researchers of CNR (National Center for Research), <a href="https://medium.com/u/2e2cffbacaca">Luca Pappalardo</a><strong> </strong>and<strong> </strong><a href="https://medium.com/u/719a932755e0"><strong>Paolo Cintia</strong></a><strong>,</strong> with the help of some professionals of the sport-tech company Wyscout. It was published as an <em>outstanding article</em> on Transaction on Intelligent Systems and Technlogy (a very important data science journal).</p><p><strong>Playerank</strong> is an <strong>algorithm</strong> that takes in input event data (or event-logs) of one or more matches and <strong>computes the performance rating of each player in each match, eventually ranking them role by role</strong>.<br>It exploits two machine learning techniques and, in the end, returns a synthetic number, just like the overall score in Fifa.</p><p>Let’s dive into the details with the help of Paolo Cintia’s words collected in a brief interview (<em>quoted paragraphs</em>).</p><p><strong>THE ALGORITHM:</strong></p><p>To get to the Playerank ranking, four steps need to be fulfilled:</p><ol><li>Features weighting</li><li>Role detection</li><li>Performance rating extraction</li><li>Ranking role-by-role</li></ol><p><strong>(1) Features weighting</strong>: 🎯</p><blockquote>The problem of determining the contribution of a single component inside a team has always been a trend topic and it will continue to be.<br>In football (but not only) the result is achieved uniquely as a team but scouts, managers and fans argue on the<strong> “value” of each player</strong>.</blockquote><p>Starting from this observation, Playerank’s features need to be weighted based on their importance to win a game at a team level. But what is a feature in this case? Well, <strong>every type of action that players perform on the pitch is a feature to which a coefficient (weight) is associated </strong>according to how much that action increase/decrease the probability of winning.<br>Namely they used a <strong>LSVC (Linear Support Vector Classifier) </strong>to find the optimal weights in order to predict the outcome of a game (1 for a victory, 0 for a draw or a defeat).</p><p>💻 <em>Nerdy part (LSVC):<br>A </em><a href="https://scikit-learn.org/stable/modules/svm.html"><strong><em>LSVC </em></strong></a><em>is a Machine Learning technique that </em><strong><em>separates the observations </em></strong><em>in the dataset predicting whether they are a 1 (victory) or a 0 (non-victory) through an </em><strong><em>hyperplane</em></strong><em> in the d-dimensional space, (d = number of features). The equation of the separating hyperplane is given by a linear combination of optimal weights and features (w ⋅ x = k).</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rRzkClmYapQ7UC_Qtfcg_A.png" /><figcaption>Worst/top features and their weights.</figcaption></figure><p>It is important to notice that the authors left goals away because</p><blockquote>goal is the only event that matters in a match, yet it is the rarest of all.</blockquote><p>Anyway in the paper, linked at the bottom, they also provide an adjusted version that considers goals.</p><p>(2) <strong>Role detection: 🔍</strong></p><p>The role detection module of the algorithm was crafted on top of the assumption that a “role” is defined by the position a player keeps in a match or in a part of it and also implies that for different positions there are different tasks.<br>Hence the Role detector was developed with a <a href="https://en.wikipedia.org/wiki/K-means_clustering#:~:text=k%2Dmeans%20clustering%20is%20a,a%20prototype%20of%20the%20cluster."><strong>k-means</strong></a> clustering and the silhouette score <strong>over the average positions of each player</strong> to identify the optimal number of roles in the pitch (optimal for the learning pattern, without any tactical meaning).<br>Spoiler: the best number of clusters is 8.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/489/1*OWpmDjqty1CtobTlbcYiwg.png" /><figcaption>Map of the 8 roles identified by the 8-means</figcaption></figure><p>The detection strategy was also improved allowing for <strong>hybrid players</strong>, i.e. players who are associated to more than one cluster. This tuning was performed by fixing a <strong>threshold (δ) for the silhouette score </strong>and checking if the player fits also in other roles. This basically means comparing the fixed threshold with the <a href="https://towardsdatascience.com/silhouette-coefficient-validating-clustering-techniques-e976bb81d10c#:~:text=Silhouette%20Coefficient%20or%20silhouette%20score%20is%20a%20metric%20used%20to,each%20other%20and%20clearly%20distinguished.">silhouette score</a> of each cluster different from the one chosen by the 8-means and see if it is smaller then the user-designed δ.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/610/1*s4WXQmCZtk9-QEkfu4vjVQ.png" /><figcaption>Silhoutte score against the threshold (δ). <strong>d_k bar </strong>is the average difference between the player and all the data points of cluster k we are checking. <strong>d_i bar</strong> is the average difference between the player and all the data points of cluster i assigned by the 8-means algorithm.</figcaption></figure><p>To make an example imagine the 8-means algorithm classifying Joshua Kimmich in his first years as a <em>right-fielder </em>(C1 in the picture above) but then notice that if you choose δ high enough (expanding the grey area) you could probably detect that it was an hybrid <em>right-fielder</em>/<em>central-fielder</em>.</p><p><strong>(3) Performance rating extraction: ⚽</strong></p><p>Once the weights are calculated it is a piece of cake now to extract the player’s performance rating, that is the actual number that synthesizes the game of a single player.</p><p>It is sufficient to <strong>sum the products of the occurences of each feature by their weigth and divide all by a normalizing constant</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/1*Tsqo-D-PE2mePOZLCd6JDw.png" /><figcaption>Performance rating for player u in match m. R is the normalizing constant, w the weights, x the feature occurences.</figcaption></figure><p>For example if Messi in a match performs 2 simple pass assist, 1 key launch pass, 3 accurate shots and collects a yellow card the computation will be: <em>(1/R)*(w1*2 + w2*1 + w3*5 + w4*1) = r(Messi, m)</em></p><p><strong>(4) Ranking role-by role: 🏆</strong></p><p>Finally, since <strong>it only makes sense to compare players that have the same role</strong>, they are organized into 8 different rankings, one for each role, with the hybrids appearing in two or more.</p><p>This choice may appear as a limit but if you think at the mathematics behind the performance rating computation you will notice that, since assists and shots have the most positive weights and cards the most negatives, <strong>forwards and wingers will usually have higher performance </strong>ratings than mid-fielders and defenders so you need to compare those players role-by-role.</p><p>There is also a <strong>business motivation </strong>behind it: clubs don’t usually look for one champion independently of his role, they want a good player in a specific position. So having a role-wise ranking can shrink a lot the time spent for scouting.</p><p><strong>THE PROJECT: 📖</strong></p><p>Talking about business, the authors also built a <strong>web-platform </strong>to offer the possibility to try Playerank and found the homonymous <strong>startup </strong>to sell their services to clubs. Unfortunatly,</p><blockquote>Both the platform and the startup are undergoing a deep overhaul.<br>It will take some time to see them back again.</blockquote><p>Of course we are looking forward to find what amazing innovation they will bring into the football industry even if</p><blockquote><strong>the tangled mystery of how you win a game will remain unsolved forever.</strong></blockquote><p>FINAL WHISTLE! 🕛</p><p><strong>References</strong>:</p><p>Luca Pappalardo, Paolo Cintia, Paolo Ferragina, Emanuele Massucco, Dino Pedreschi, and Fosca Giannotti. 2019. PlayeRank: Data-driven Performance Evaluation and Player Ranking in Soccer via a Machine Learning Approach. ACM Trans. Intell. Syst. Technol. 10, 5, Article 59 (September 2019), 27 pages. <a href="https://doi.org/10.1145/3343172">https://doi.org/10.1145/3343172</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=87d8515a91eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data in football are, still, collected manually — 90° QUANTILE [6]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/data-in-football-are-still-collected-manually-90-quantile-6-21e5fdee7577?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/21e5fdee7577</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[soccer-analytics]]></category>
            <category><![CDATA[football-analytics]]></category>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Tue, 18 Apr 2023 09:47:58 GMT</pubDate>
            <atom:updated>2023-04-21T09:02:29.304Z</atom:updated>
            <content:encoded><![CDATA[<h3>Data in football are, still, collected manually — 90° QUANTILE [6]</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*A5ja9AI1J0_knBwL.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme<br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>💻 In very few cases thousands of data need to be collected by manually clicking on a button and the recording of football events is one of those.<br>In past articles we saw how to get from datasets to an xG model but how every shot (pass, tackle, etc.) is recorded?</p><p>⌨️ Even if a research branch is developing automatic tools, the standard procedure is still to click manually on the 2D pitch representation and insert event type (e.g. shot), team name, player name and other details. To complete this time-consuming task, the operators are assisted by a software like <a href="https://fcpythonvideocoder.netlify.app/">this</a> in which you can watch and annotate the game in a unique interface.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lJmyhFF1HnQP7ag_yXuh4g.png" /><figcaption>YouTubeCoder by FC Python</figcaption></figure><p>😍 This is YouTubeCoder by FC Python, an open website that offer a framework to annotate events by yourself, and, in my opinion, it is a great tool for begginers because it can allow you to do different stuffs like, for example, to analyse your local team matches or to start a small proprietary data collection of your favourite goals’ actions (mine is Kakà’s 1–2 against Manchester United in 2007 UCL).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/426/1*mUAfyPkr760gNxVvbqHvqw.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RjUJjYemCHm5K6CEfZ4YNg.png" /><figcaption>Kakà’s epic dribbling event recording</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jct5HS6zC04_53sgIHdQTg.png" /><figcaption>Kakà’s goal event recording</figcaption></figure><p>Of course, there are tools for professionals which are a bit more advanced but actually not so much (see pictures below).</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*nOj79BYLyvuTIF1MCPJ7Yw.png" /><figcaption>Screeshot from Pappalardo et al., (2019) A public data set of spatio-temporal match events in soccer competitions, Nature Scientific Data.</figcaption></figure><p>🤚With those advances framework the operator can add multiple details the the event like successful/not succesful tag for passes, goal/saved/out for shots or yellow/red card for a foul. But still manually!</p><p>Obviously this generates two possible issues:</p><ol><li>Errors in the data about (x,y) positions, players’ names, etc.</li><li>Disagreement between different annotators about the same action</li></ol><p>🔨 Yeah, I know what you are thinking, it’s 2023 and we are still collecting data like Charles Reep in 1957…<br>It takes hours of works to annotate an entire game but future is coming and there are studies in the field of Computer Vision (and also other approaches that use tracking data) that try to speed up this process by automatically detect events.</p><blockquote><em>“Ma questa, questa è un’altra storia…”</em></blockquote><blockquote><em>(“But this, this will be another story…”)</em></blockquote><blockquote><em>Cit. Federico Buffa</em></blockquote><p>FINAL WHISTLE! 🕛</p><p><em>The next story’s topic has still to be decided so it will be a surprise for me too!</em></p><p><strong>References</strong>:</p><p>Pappalardo et al., (2019) A public data set of spatio-temporal match events in soccer competitions, Nature Scientific Data 6:236, <a href="https://www.nature.com/articles/s41597-019-0247-7">https://www.nature.com/articles/s41597-019-0247-7</a></p><p><a href="https://fcpythonvideocoder.netlify.app/">https://fcpythonvideocoder.netlify.app/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21e5fdee7577" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The only 2 datasets of football data — 90° QUANTILE [5]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/the-only-2-datasets-of-football-data-90-quantile-5-e403af1949ef?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e403af1949ef</guid>
            <category><![CDATA[soccer-analytics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[football-analytics]]></category>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Mon, 03 Apr 2023 09:25:02 GMT</pubDate>
            <atom:updated>2023-05-14T10:41:32.446Z</atom:updated>
            <content:encoded><![CDATA[<h3>The only 2 datasets of football data — 90° QUANTILE [5]</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*Q_MLNkeyKfIx7Re2.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme<br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>In the <a href="https://medium.com/@leonardo.acquaroli00/a-mathematics-based-new-penalty-area-in-football-90-quantile-4-40ec01d9dee3">last story</a> I proposed a new penalty area with a curved shape that was calculated training a <strong>naive xG model </strong>with only <em>angle </em>and <em>distance </em>as variables. In this article we will see which data I used for the training.</p><p>🔒 Football analytics is still a subject for pioneers or, in other terms, is still a world open only to the football clubs and analytics companies who can afford to buy or collect the data.</p><p>🔬 Do you know <strong>how many open dataset </strong>are available as today? <strong>Only 2!<br></strong>The first one was released by StatsBomb and it is easily accessible by the <a href="https://mplsoccer.readthedocs.io/en/latest/gallery/statsbomb/plot_statsbomb_data.html#data">python library mplsoccer</a>.<br>The second dataset is the one I used. It was made available by the italian analytics company Wyscout and published with the help of two researchers of the University of Pisa, among the others, <a href="https://medium.com/u/2e2cffbacaca">Luca Pappalardo</a> and <a href="https://medium.com/u/719a932755e0">Paolo Cintia</a>, in the paper “<strong>A public data set of spatio-temporal match events in soccer competitions</strong>”.</p><p>📅 <strong>THE DATASET:</strong></p><p>The data are available in different json files containing info about:</p><ul><li><strong>Competitions</strong>: Serie A, Premier League, La Liga, Ligue 1, Bundesliga, European championship and the World cup;</li><li><strong>Teams</strong>: 142 teams, both clubs and national selections;</li><li><strong>Coaches;</strong></li><li><strong>Players;</strong></li><li><strong>Referees;</strong></li><li><strong>Matches</strong>: 1941 matches with all the macro-information like results, lineups, goal scorers, stadium, etc.;</li><li><strong>Events</strong>: these are the most useful data points since each (on-the-ball) event is the description of a single piece of a match, for example a shot, a tackle, or a pass. There occur about 1000 events per match.</li></ul><p><strong>COMPARING STATSBOMB AND WYSCOUT:<br></strong>Now, I want to compare those datasets highlighting their dis/advantages.</p><p>Statsbomb:</p><ul><li><strong>PROS: </strong>once installed the <em>mplsoccer</em> library, they are accessible without storing anything and so useful to create code that works in every machine; this dataset contains a long series of seasons and matches history, for some competitions even back to 1999.</li><li><strong>CONS:</strong> StatsBomb open dataset is usually slower to load data since it access them by api calls; their matches data contain less information than Wyscout’s ones.</li></ul><p>Wyscout:</p><ul><li><strong>PROS: </strong>once downloaded all the files locally (or in a GitHub repo), reading process is fast enough; every data point contains a lot of useful information; the x,y coordinates of each event in the pitch are easier to use in a customized field visualization since they are recorded as percentage over the lengths of a generic pitch.</li><li><strong>CONS:</strong> you need to store the data to access them; they refer only to one season; additional information are encoded as numerical tags that must be cross-referenced from a separate csv file.</li></ul><p>🔬 These datasets have a very important role in football analytics because they are the <strong>access-door for individual contributors</strong>. Indeed, I believe that, as in almost any other tech field, <strong>open innovation is the key</strong> to accelerate the development and, above all, the spread of this kind of data-knowledge.</p><p>🎮 This said, I invite you to play with those data and see how powerful they can be.</p><p>FINAL WHISTLE! 🕛</p><p><em>In the next story we will see how the event data in the datasets are recorded and how their collection can be speeded up with Artificial intelligence.</em></p><p><strong>References</strong>:</p><p>Pappalardo et al., (2019) A public data set of spatio-temporal match events in soccer competitions, Nature Scientific Data 6:236, <a href="https://www.nature.com/articles/s41597-019-0247-7">https://www.nature.com/articles/s41597-019-0247-7</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e403af1949ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A mathematics-based new penalty area in football — 90° QUANTILE [4]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/a-mathematics-based-new-penalty-area-in-football-90-quantile-4-40ec01d9dee3?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/40ec01d9dee3</guid>
            <category><![CDATA[football-analytics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[soccer-analytics]]></category>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Mon, 27 Mar 2023 16:29:41 GMT</pubDate>
            <atom:updated>2023-04-21T09:03:23.765Z</atom:updated>
            <content:encoded><![CDATA[<p>A mathematics-based new penalty area in football — 90° QUANTILE [4]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-bw-ABxexfuOUs1S_LOmA.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme<br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>😡 Have you ever got mad for a given penalty like this?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/426/1*PVv0715jNOt15AlBL4RMng.gif" /></figure><p>Yeah, I know Osimhen was just running towards the sideline with the back to the goal, in the most remote angle of the penalty area.<br>Was He likely to score or assist someone from that action? Well probably not.</p><p>👀 What if the penalty area was made like this?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/499/1*_anldrgOjnZBlSKI9JSJMQ.png" /><figcaption>Fig. 1</figcaption></figure><p>Clearly the referee would not have indicated the spot.</p><p>This week’s paper is called <strong>“A mathematics-based new penalty area in football” </strong>and propose to re-draw a penalty area based on the two most important features of the xG metric: <strong><em>distance </em>and <em>angle</em></strong><em> </em>to the goal<em>.</em></p><p>In particular the function used to plot the contour is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/349/1*KXtb6mdWV_lb9TDLdDX74g.png" /></figure><p>🧮 This is a function <strong>directly proportional to the angle seen by the shooter (<em>β</em>) </strong>and <strong>indirectly proportional to the distance from the mid-point of the goal (<em>d</em>), </strong>where <em>x </em>and <em>y </em>are the coordinates of the shot location in the pitch, <em>b=7.32 </em>is the goal-line length, <em>β</em> is calculated using the laws of trigonometry and <em>d </em>is the Euclidean distance.</p><p>As someone may have noticed this is a function in 2 variables — <strong>x</strong>, the position along the long side, <strong>y </strong>the position along the short side— thus to be represented in 2 dimensions it needs contour lines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/579/1*u9mls4FyoKaUumHD3i3YzQ.png" /><figcaption>Fig. 2</figcaption></figure><p>The author chose the curve (Fig. 1) which is about 22 m away from the goal, i.e. 3 times the length of the goal line (7.32m).</p><p>⚖️ You could also point out that the function is too simplistic because it gives the same importance to angle and distance. <strong>Two different weights or other functions may be used </strong>and so below I show my personal proposals.</p><p><strong>1 — Same function, different weights, different area:<br></strong>The first idea is to apply two different weights to <em>angle </em>and <em>distance</em>, but which to use? I took the coefficients of a logistic regression of <em>goal —</em> a dummy variable that is 1 for a winning shot and 0 otherwise — on <em>angle </em>and <em>distance⁴ </em>(to give more importance to ditance<em>).</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/988/1*RfSitYclb7w29F7sIgDSaw.png" /></figure><p>In particular <em>β = -0.037 </em>and <em>γ = -0.0000046 </em>which transform the original function into:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X7nyh2MU0tgwuoGdCEQjsw.png" /><figcaption>Function rescaled with logistic regression parameters</figcaption></figure><blockquote>This new rescaled function returns a curve area that, pivoted in the vertex of the bezel, measures 519.49 m², a number closer to the actual rules’ figure (544.5 m²) with respect to the area prompted by the researcher (473.83 m²).</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/391/1*Rq5HqJGh3h30UYjge2fpCg.png" /></figure><p><strong>2 — Different function, same area:<br></strong>The second approach I used is to directly substitute the researcher’s function with a <strong><em>sigmoid </em></strong>(i.e. a logistic function).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XXOj1SxHLarPFu0WrekpwA.png" /><figcaption>In this function the variables in <em>italic</em> seem not to be <em>x </em>and <em>y </em>but recall that <em>angle </em>and <em>distance </em>are both calculated from the (x,y) shots position.</figcaption></figure><blockquote>This time I <strong>fixed the measure of the New area to be the same as the Actual </strong>and the result is a more regular curve that is slightly closer to the goal than the original box. To obtain this equality I chose the level curve at <em>f(x,y) = 0.368339 </em>and, since the previous logistic regression can be intended as a very naive xG model, <strong>the line drawn can be interpreted as the bound of an area in which the probability to score is higher or equal to the value of the function (0.368339)</strong>.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/391/1*dlkL5ICSLNJwTxvMGyawfQ.png" /></figure><p>🥹 Yes, I know…a lot of math. Sorry for this but this time it was necessary.</p><p>If you want to replicate such kind of experiments you may want to know which data i used to train the logistic regression.<br>Well, stay tuned because next time we will see the largest open dataset on football. And guess what? It was made by italian researchers!</p><p>FINAL WHISTLE! 🕛</p><p><em>In the next story we will see the largest open dataset on football events. And guess what? It was made by italian researchers!</em></p><p><strong>References</strong>:</p><p>César A. Morales (2016): A mathematics-based new penalty area in football: tackling diving, Journal of Sports Sciences,<br>DOI: 10.1080/02640414.2016.1177657<br><a href="http://dx.doi.org/10.1080/02640414.2016.1177657">http://dx.doi.org/10.1080/02640414.2016.1177657</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40ec01d9dee3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The first xG model  — 90° QUANTILE [3]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/measuring-the-effectiveness-of-playing-strategies-at-soccer-90-quantile-3-3f1447708525?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f1447708525</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[soccer-analytics]]></category>
            <category><![CDATA[football-analytics]]></category>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Mon, 20 Mar 2023 17:32:19 GMT</pubDate>
            <atom:updated>2023-04-21T09:03:45.441Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The first xG model — 90° QUANTILE [3]</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-bw-ABxexfuOUs1S_LOmA.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme <br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>This week’s paper is called “<strong>Measuring the effectiveness of playing strategies at soccer</strong>” and it was written by Charles Reep (is that you, again?!) and Richard Pollard, the researcher who helped Reep in his late years developing his analysis with the use of the emerging technologies of ’90s.</p><p>I’d like to go straight to the core of this study: Reep and Pollard <strong>set the basis of Expected goals (xG) and Expected threats (xT) without naming them as such</strong>. These two metrics are among the most famous nowadays and several data science teams, both in football clubs and in data providers, developed proprietary models to estimate those values.</p><p>The aim of the paper was to create a measure of effectiveness of different action types and they called this measure the <strong><em>yield</em></strong>.</p><p>🧮MATH ALERT!<br>The elements (<em>i,j)</em> to calculate the <em>yield </em>are:</p><ul><li><em>j = 1 </em>(open play) or <em>j = 0 </em>(set play);</li><li>being the pitch divided into six zones, <em>i = 1, …, 6 </em>for each zone;</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/523/1*hVpN9s7SHEzXLVgqW4UeFw.png" /></figure><p>⬆️Now, described <em>pᵢⱼ </em>as the probability of scoring a goal for a possession of type <em>j </em>starting in zone <em>i, </em>the <em>yield </em>for a possession is defined as</p><blockquote><strong>YIELD: the estimated probability of scoring a goal minus the estimated probability of conceding a goal</strong>, based on the outcome of the possesion <em>(i.e. shot, set play gained, possession lost).</em></blockquote><p>The k-th <em>pᵢⱼₖ </em>is negative when the outcome of the possession is a lost ball, hence, for each possession of type j starting in zone i,<em> </em>the yield<em> </em>is the summation of all the positive scoring probabilities minus the negative ones, divided by the total number of occurences of that kind of possession.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/343/1*7Y7TdfvXCKJlPnF-UnbfgA.png" /><figcaption>Expected <em>yield for a given possession of type j starting in zone i</em></figcaption></figure><p>⏩Now, if you survived the maths, you can note that <strong>this approach is the very baseline reasoning behind xT</strong>, that is the increase (or decrease) in scoring probability generated by each single event (e.g. a defensive clearance or a smart pass).<br>Since xT was devised by Sarah Rudd in 2012, other advanced metrics have been developed on top of xT and the last one, for the time being, is the On Ball Value by StatsBomb who also cited Reep’s article <a href="https://statsbomb.com/articles/soccer/introducing-on-ball-value-obv/">in their announcement</a>.</p><p>⚽When the outcome is not a lost ball, we have to compute the probability of scoring associated to the shot or the set play gained.<br>In order to do this the authors developed <strong>the first xG model</strong>, without giving it such name:</p><blockquote><em>y = 1.245 — 0.219x¹ — 1.578x² + 0.947x³— 1.069x⁴</em></blockquote><p>To compute the probability they used a <strong>logistic regression </strong>with four variables:</p><ul><li>x¹ distance to goal,</li><li>x² angle of shot,</li><li>x³ a dummy equal to 1 if the player taking the shot was less than 1 yard from the nearest defender,</li><li>x⁴ a dummy equal to 1 for a set play.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/419/1*foynQkrwScsuWKCiLUs8tg.png" /><figcaption>Logistic probability distribution function</figcaption></figure><p>Obviously, this is a very basic statistical model but, as you can imagine, it shaped the football analytics world and it is still the baseline taught to newbies.</p><p>FINAL WHISTLE! 🕛</p><p><em>In the next story we will look at a curious but sharp proposal of a new redesigned penalty area based on the highest xG zone of the pitch.</em></p><p><strong>References</strong>:</p><p>Pollard, Richard, and Charles Reep. “Measuring the Effectiveness of Playing Strategies at Soccer.” <em>Journal of the Royal Statistical Society. Series D (The Statistician)</em> 46, no. 4 (1997): 541–50. <a href="http://www.jstor.org/stable/2988603.">http://www.jstor.org/stable/2988603.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3f1447708525" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Analysis of passing sequences, shots and goals in soccer — 90° QUANTILE [2]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/analysis-of-passing-sequences-shots-and-goals-in-soccer-90-quantile-2-d9198cdb0c66?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d9198cdb0c66</guid>
            <category><![CDATA[football-analytics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[soccer-analytics]]></category>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Sat, 11 Mar 2023 21:16:08 GMT</pubDate>
            <atom:updated>2023-04-21T09:05:31.164Z</atom:updated>
            <content:encoded><![CDATA[<h4>Analysis of passing sequences, shots and goals in soccer — 90° QUANTILE [2]</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-bw-ABxexfuOUs1S_LOmA.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme <br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>📊In this second story we will see the results of a paper which refines the intuitions of Reep and fixes the statistical error He made. <br>The title is: “<strong>Analysis of passing sequences, shots and goals in soccer</strong>”, by Hughes and Franks.</p><p>🏆The authors, experts in the field of performance analysis, replicated Reep’s study using data coming from 1990 and 1994 World cup.<br>They utilized the frequencies of <em>r</em>-pass chains but, unlike Reep, when it was about showing how many shots and goals came from those chains they normalized the data.</p><p>This means that:</p><ol><li>They counted how many 1-pass, 2- pass, 3-pass…, chains occured in all the 106 games recorded,</li><li>then took the number of shots and goals<strong> </strong>which came from those differently long passing sequences,</li><li><strong>divided the number of shots and goals by the number of occurencies of each <em>r</em>-pass sequence</strong>.</li></ol><p>⚠️In this way they obatained a shot-rate and a goal-rate, rather than just an absolute value, that can represent the degree of danger of each length of passing chains.<br>The results were different and more controversial than Reep’s idea of “long ball” but if you want the long story short:<br>defined <em>“direct play” </em>as a strategy focused on short passing sequences and <em>“possession play” </em>as a strategy focused on long passing sequences,</p><h3><strong><em>“direct play” </em>produces less goals than <em>“possession play” </em>(long passing sequences).</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hlx1_Fs7sk6Lq2GRBYk84Q.png" /><figcaption>The peak of goals per 1000 possession is for 6-pass chains.</figcaption></figure><p>Yet the results are more complex.<br>In fact, the claim above stems from two factors:</p><ul><li>⚽the <strong>number of shots per possession </strong>(for each of the differently long passing sequences),</li><li>🎯the <strong>shots-to-goals conversion rate</strong>, that is the<strong> </strong>ratio between shots and goals<strong> </strong>produced by each <em>r</em>-pass type of chain.</li></ul><p>⬆️⬇️The paper shows that the <strong>number of shots per possession </strong>is higher for long passing sequences but the <strong>shots-to-goals conversion rate is better (that means lower) </strong>for short ones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/534/1*Zly0z6ZBZAsaAwgjrTRTOQ.png" /><figcaption>Shots-to-goals ratios exploded. Alower value means a more effective type of action.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CDyRzIyAEGxBLWDZtd0I9w.png" /><figcaption>Longer passes chains produce more shots per 1000 possession</figcaption></figure><p>❓You could now wonder wheter <strong>is it better to choose <em>“possession play” </em>or <em>“direct play”</em></strong><em>. </em>In other words, which brings more goals between long and short passing sequences.<br>To answer this question Hughes and Franks presented a quick simulation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/539/1*r50_B7Wn5IHtdwMi4MuXAg.png" /><figcaption>Simulation for the number of goals from long and short possesions.</figcaption></figure><p>Don’t be scared if something is not clear from the picture above, I left for the conclusion the analysis about the differences between succesful (reached at least the semifinals) and unsuccesful teams (out at the group stage).<br>✅❎Indeed, they found that succesful teams are able to produce significantly more goals from long possessions than unsuccesful ones (8.5 vs 5.1) and, eventually, they used this figure to justify the final suggestion that “<strong><em>direct play” </em>could be good for low quality (unsuccesful) teams</strong> because they are less capable of playing and exploiting a long possession strategy, while <strong>for the high quality (succesful) teams “<em>possession play” </em>brings a lot of shots more and more goals too.</strong></p><p>FINAL WHISTLE! 🕛</p><p><em>In the next story we will look at another Reep’s work “Measuring the effectiveness of playing strategies at soccer” that gave birth to the concept of </em><strong><em>expected goal</em></strong><em>.</em></p><p>References:<br>Mike Hughes &amp; Ian Franks (2005): Analysis of passing sequences, shots and goals in soccer, Journal of Sports Sciences, 23:5, 509–514</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d9198cdb0c66" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Skill and chance in Association football — 90° QUANTILE [1]]]></title>
            <link>https://medium.com/@leonardo.acquaroli00/90-quantile-1-21de3416910e?source=rss-49e39e26175e------2</link>
            <guid isPermaLink="false">https://medium.com/p/21de3416910e</guid>
            <category><![CDATA[football-analytics]]></category>
            <category><![CDATA[soccer-analytics]]></category>
            <category><![CDATA[science]]></category>
            <dc:creator><![CDATA[Leonardo Acquaroli]]></dc:creator>
            <pubDate>Sun, 05 Mar 2023 12:53:31 GMT</pubDate>
            <atom:updated>2023-04-21T09:05:15.841Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-bw-ABxexfuOUs1S_LOmA.png" /></figure><p><em>Using a wordplay with the famous italian football tv-programme<br>“90° Minuto” these stories contains a series of informative analysis of scientific papers about football science.</em></p><p>👨🏻‍🏫 The first article tells about the paper that gave rise to Soccer analytics: <strong>‘Skills and chance in Association Football’</strong> by B. Benjamin and especially Charles Reep, the accountant considered to be the pioneer of sport analytics. Indeed, He hand-annotated the actions of more than 2000 English football matches in the 1950s-60s.</p><figure><img alt="Charles Reep — Former Royal Air Force accountant" src="https://cdn-images-1.medium.com/max/250/1*75Csxi29Zj2Kh9TCvx2Dzw.jpeg" /><figcaption>Charles Reep — Former Royal Air Force accountant</figcaption></figure><p>⛓️ In the short paper, Reep argues that as the length of a chain of passes increases, the probability of success of the next pass generally decreases. Thus, <strong>making an extra pass increases the probability of losing the ball on the next move</strong>.<br>This argument is brought forward by the fact that the distribution of <em>r-pass</em> chains in his data is well approximated by a negative binomial <br>NB(x=1,p=P(r)) and this is used to justify the claim above.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/393/1*am66xB1vdkz8bCWqCk9RgQ.png" /><figcaption>An example of Negative binomial distribution</figcaption></figure><p>⚠️ Below, the ex-accountant combines this information with an analysis of dangerous areas of the pitch identifying the <strong><em>shooting areas</em></strong>. These are the quarters of the green rectangle closest to the goals, <strong>where the probability of shooting and scoring increases significantly</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/881/1*aT7qxXNw2yRvrsT-DKym4g.png" /><figcaption>Shooting areas</figcaption></figure><p>⚽ The corollary of these observations is that <strong>it is better to throw the ball into the opponent’s <em>shooting area</em> with as few passes as possible</strong> in order to avoid losing it in one’s own danger zone and also to maximise the probability of shooting on goal.</p><p>💣Thus was born the <strong>long ball strategy</strong>, typical of English football and evangelised in those years by Charles Reep himself in several teams including Arsenal and Wolverhampton.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/748/1*7DlgyWjV76XymbM0G5YLhw.png" /><figcaption>An aerial duel — The long ball tactic implies a large number of those</figcaption></figure><p>The other side of the coin is that pressing into the opponent’s <em>shooting area</em>, with the aim of stealing the ball as close to the goal as possible, can increase exponentially the chance of scoring.<br>In particular the estimate based on Reep’s data highlights that more than 50% of goals come from regained possession into the <em>shooting area </em>and that the proportion of those advanced regained possessions to attacks starting behind is 1:3.<br><strong>Basically regaining possession into the opponent’s first quarter is three times more effective (in terms of goals) than starting an attack from behind</strong>.<br>Doesn’t it remind what we call now <em>gegenpressing</em>?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/181/1*yTp58iC96B2W1po3xzGVqA.jpeg" /><figcaption>Jürgen Klopp — The coach who most notoriously exploited <strong>gegenpressing</strong></figcaption></figure><p>In the following years (but also recently) many have criticised Reep and Benjamin’s statistical approach and numbers, which indeed show some flaws, but this paper remains of fundamental importance for the influence it had on the way football was played in the 1960s and especially on the discipline of Soccer analytics of which Charles Reep is still, figuratively, considered the father.</p><p>FINAL WHISTLE! 🕛</p><p><em>In the next story we will look at one of the papers criticising Charles Reep’s thesis “Analysis of passing sequences, shots and goals in soccer” by Hughes and Franks.</em></p><p>References:<br>Reep, C., and B. Benjamin. “Skill and Chance in Association Football.” <em>Journal of the Royal Statistical Society. Series A (General)</em>, vol. 131, no. 4, 1968, pp. 581–85. <em>JSTOR</em>, https://doi.org/10.2307/2343726. Accessed 9 Mar. 2023.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21de3416910e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>