<?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 Eugene Tulyagijja on Medium]]></title>
        <description><![CDATA[Stories by Eugene Tulyagijja on Medium]]></description>
        <link>https://medium.com/@etulyagijja?source=rss-f6e1efbbe3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*BhCi0DlO9NfalR_RtoMVRg.jpeg</url>
            <title>Stories by Eugene Tulyagijja on Medium</title>
            <link>https://medium.com/@etulyagijja?source=rss-f6e1efbbe3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 14:15:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@etulyagijja/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[Building The “Perfect”​ March Madness Bracket Using Artificial Intelligence]]></title>
            <link>https://medium.com/@etulyagijja/building-the-perfect-march-madness-bracket-using-artificial-intelligence-683bcc679230?source=rss-f6e1efbbe3------2</link>
            <guid isPermaLink="false">https://medium.com/p/683bcc679230</guid>
            <dc:creator><![CDATA[Eugene Tulyagijja]]></dc:creator>
            <pubDate>Fri, 05 May 2023 22:23:29 GMT</pubDate>
            <atom:updated>2023-05-05T22:23:29.288Z</atom:updated>
            <content:encoded><![CDATA[<h3>Intro</h3><p>According to the NCAA, there’s a 1 in 9.2 quintillion chance of making a perfect bracket yet every March, millions of people fill out March Madness brackets to compete against their friends, family, and a chance to create that “Perfect” Bracket. Everyone has different ways of picking their bracket; some people pick higher sides, upsets, alma mater, and experts who rely on AP rankings, “eye tests”, and head-to-head records.</p><p>As a Sport Analytics major, I felt like it was an obligation to implement my methodology to attempt to make a perfect bracket, and what other method to implement right now than Artificial Intellgience, with all the hype from ChatGPT, I thought would be a cool idea to see if I can quantify different categories to create a perfect bracket.</p><h3>Understanding How to Model with Artificial Intelligence</h3><p>Artificial Intelligence (AI) — The easiest way to explain AI is that it is developing a computer to understand, comprehend and perform tasks that usually require human intelligence.</p><p>Implementing an AI model in this context is perfect because this model can make a decision based on the dataset we feed it without human interference/ bias being involved.</p><p>My model will utilize a set of NCAA data that will enable it to recognize certain patterns, like upsets, seed factors, etc. This will allow it to reach a conclusion or make a decision on each game and the whole result of the tournament based on historical data. This requires millions and millions of data sets from the past. Hence, AI models are particularly suitable for solving complex problems while providing higher efficiency/cost savings and accuracy compared to simple methods. (<a href="https://viso.ai/deep-learning/ml-ai-models/">Viso.ai</a>)</p><h3>Methodology — Getting the Data/ Data Cleaning</h3><p>A huge advantage of any modeling is depending on the quality of historical data used to train the AI. <a href="https://www.sports-reference.com/">Sports Reference</a> had data going back to 1960 that included key stats and team averages for the year leading up to each tournament as well as matchups and outcomes. We also incorporated “CBBpy” and scraping data from <a href="https://s3.amazonaws.com/static.akk.io/Akkio-NCAAHistoricalData.csv">NCAA data</a></p><p>The data we collected include field goal percentage, 3-pointers, free throws, blocks, rebounds on both sides, etc. The data also includes the match-up results for each team in each round of the tournament, the score, and who won the game. Finally, the data includes each team’s tournament seeding.</p><h3>Methodology — Model Building</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tzSq3MagR47f_-w4" /><figcaption>Example of an Deep Neural Network — Picture Source: <a href="https://www.bmc.com/blogs/deep-neural-network/">https://www.bmc.com/blogs/deep-neural-network/</a></figcaption></figure><p>I began the model by implementing a Deep Neural Network technique to identify and find patterns in input and output layers of each individual stats that will factor into the outcome of each game. This includes running a simulation of any “upsets” matchup a 1,000 and letting the AI predict the odds of the “seeding” upset happening again in the future.</p><p>All the steps so far have led to this where we simulate the outcome of each game and matchup in the 2022–23 tournament a thousand times each with the training data set that the AI created from the historical dataset we collected. Result 1 Result:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TE5wU2OdiNNgsbX0" /><figcaption>Result 1 (Before Adjustment)</figcaption></figure><h3>Back Testing</h3><p>To determine how good any model is or get a sense of where the model is at, I used the same model to predict the 2022 NCAA Tournament:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-hGUiiS2G94OLW0B" /><figcaption>Back Testing Results #1 (Elite 8)</figcaption></figure><p>I correctly predicted Kansas to win, and the AI predicted that North Carolina(Correct), Duke(Correct) and Houston would round out the final four. With limited testing and computing power, I’d say the model was pretty successful.</p><p>I wanted to get all four of the final four results right so I went back and re-loop the Deep Neural Network to find other combinations of parameters that can be contributed to a more “accurate” model. I also developed an analysis function that automatically evaluates the backtest to determine if it’s accurate to the actual results. After doing that he’s the updated results of the Elite 8 matchup:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NUfcqZN3p3QWRt-x" /><figcaption>Back Testing Results #2 (Elite 8)</figcaption></figure><p>This time we still correctly predicted Kansas to win, as well as all the final four teams (Kansas, Villanova, Duke, and North Carolina) I’m really happy with the results of the updated model, so now it’s time to implement this model to make a bracket for this year(2023.)</p><h3>2023 Final Bracket Predictions</h3><p>Here’s what the AI Predicts,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jQ6r4QyDFPGMNtp3" /><figcaption>Final Prediction</figcaption></figure><p>My model has Texas beating Alabama in the finals to become National Champions, even though it is a little boring that the one and two seeds are playing each other for the national championship. The road to it is filled with some upsets, like 10 seed USC making a run to the sweet sixteen, as well as number 11 seed Pitt, knocking out number 3 seed Xavier. We will see in the coming weeks how my model stands.</p><h3>Conclusion</h3><p>After a year’s worth of work experimenting with different techniques this one is the one I ended up with, excited to see how the model holds up in the coming weeks as we will be able to determine how accurate this model is.</p><h3>Future Improvements</h3><ul><li>Testing out the model with other years to create more back testing opportunities.</li><li>Run the simulation more than 1,000 times to get more accurate results. (limited with computer capabilities)</li><li>Combine the multiple “Weak” back testing data and models to produce better data for the actual model to train off</li></ul><h3>Source</h3><p>Model Inspiration — <a href="https://www.akkio.com/post/crushing-your-march-madness-bracket-with-ai">https://www.akkio.com/post/crushing-your-march-madness-bracket-with-ai</a></p><p>Artificial Intelligence Information — <a href="https://viso.ai/deep-learning/ml-ai-models/">https://viso.ai/deep-learning/ml-ai-models/</a></p><p>Improvements — <a href="https://www.analyticsvidhya.com/blog/2015/12/improve-machine-learning-results/">https://www.analyticsvidhya.com/blog/2015/12/improve-machine-learning-results/</a></p><p>Dataset — <a href="https://www.sports-reference.com/">Sports Reference</a> , <a href="https://s3.amazonaws.com/static.akk.io/Akkio-NCAAHistoricalData.csv">NCAA data</a>, <a href="https://github.com/topics/ncaa-basketball">https://github.com/topics/ncaa-basketball</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=683bcc679230" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>