.webp)
Experience the power of Luzmo. Talk to our product experts for a guided demo or get your hands dirty with a free 10-day trial.
At Luzmo, we love regularly taking out our toolbox and hacking together data-driven apps. And rumor has it that the Luzmo team shares another passion besides data… We loooove football (or soccer - for the Americans). ⚽️
With the UEFA 2024 championship happening this year, of course we weren’t going to pass up the opportunity to build something fun on top of all that football data. What started as “a fun idea for a social media campaign” quickly escalated to our founders coding an app together.
The result is a Euro2024 AI pundit that predicts the final score for each upcoming game, winning odds for each team in every stage of the championship, based on historical data and performance. All of that rich data is visualized in interactive charts and commented upon by our witty resident octopus.
We didn’t expect the app to explode, with thousands of visitors in less than a week, and coverage on multiple press outlets. So today, we’re giving you a closer look behind the scenes, and show you how we built the predictive model and its resulting app.
For this app, we used 3 different data sources:
We added a simple calendar interface of all games, which lets you browse through a bunch of statistics related to the championship, updated in real time.
Let’s have a closer look at the insights you can find here.
In the “head to head”, you can compare how two competing teams stack up against each other. Compare their win probability, market value, player rankings and more to determine which team has the best scorecard to win the game.
You can also analyze more specific metrics regarding attack, defense and goalkeeping performance by browsing the different tabs. We used Luzmo to create the data visualizations and embed them in our Euro 2024 application.
To top it off, each head-to-head has a prediction for the final score, proclaimed by the all-seeing eye of our AI octopus, but more on that later!
While browsing the head-to-head comparison, you can further drill down into player stats. For example, let’s say you’re looking at goal stats for France. You can click on Mbappé (or any other player) to open up a separate dashboard with individual stats for that player.
In the above visualization, we presented the facts: historical data about player and team performance. But the beauty of a football tournament is in the speculation. Football is a sport with high variability, because goals are relatively rare (~2.8 goals per match during UEFA 2020). Even clearly advantaged teams regularly lose against a lesser side.
To predict who will win the championship, we used historical data in a predictive model that estimates the odds for each team to advance in each stage of the championship.
Besides data from the qualifiers, we’re updating the data and model every game day, which means the odds will shift as the tournament progresses. With a slider at the top, you can see how the odds for your favorite team are progressing during the tournament. For example:
With each game, we get more information, which will shift the odds for any team. Let’s have a closer look at the mechanics behind this predictive model.
Our predictive model relies on two main data points:
This means our model is heavily weighted towards recent performances, which causes a few surprises. For example, it initially hurt winning probabilities for fan favorites like England and Italy, as they didn’t perform great during the qualifiers.
For each team, we use the play-by-play data to calculate normalized defensive, midfield, offensive and keeper ratings. We then calculate an expected goals (xG) Poisson distribution of a team against every other team.
Each position rating affects every other position, but to varying degrees. For example:
We then calculate win/draw/loss and per-scoreline chances from the 2 Poisson distributions. Finally, we also apply a correction factor to align the overall average goals per game and draw percentage with historic data.
While we can use raw probability distributions for match probabilities, this is far less evident for tournament progression odds, because of the complex progression rules.
Instead, we use Monte Carlo simulation to play out the entire tournament 1 million times and tabulate how far each team progresses. For example, while France wins in 22 out of every 100 simulations, Georgia only ended up taking the win in 7 of the million simulations.
Monte Carlo methods are computationally less efficient, but lead to a richer 'data bycatch'. For example, we can also tabulate:
It's also easier to model effects like fatigue, and measure the sensitivity of the model to these. In our model, a team that has to play highly rated teams does build up some fatigue that lowers their ratings in later games. This effect is smaller if the team has a high-quality bench, as they can more easily cycle through players to rest them. The tournament is well-balanced though, as this only shifted chances by up to 2 percent.
We now have our probability calculations, but of course we wanted to present them in a fun and engaging way. That’s where large language models come in.
We used GPT-4-Turbo to interpret the position ratings and probability distribution. As we feed that data into the LLM, it understands the relative weaknesses and strengths of the various teams, and can identify surprising or outlier results. We chose to use GPT-4-Turbo because it proved to be superior at following instructions to GPT-4o.
We use the LLM to deliver a pundit’s commentary in a witty, humorous tone of voice. It’s a great example of how you can leverage generative AI to summarize insights from your data.
Our Euro2024 app is just an example of how you could use predictive modeling, embedded analytics and generative AI to build a next-gen data product in only a few days.
Besides placing some sports bets with friends, you likely won’t use this particular app for any life-changing decisions. However, any developer could take the same building blocks, and build AI-powered insights for their software users. To name a few examples:
In the context of your software application, these insights will empower your customers to make important business decisions faster. And more value for your customers means more loyal, happy users for your software.
Interested in adding AI-powered analytics to your software application? Look no further than Luzmo for stunning, interactive data visualizations. Hook up our platform to any AI model of your choice, and start building the analytics experience your customers desire.
Book a demo with our product experts today to learn more!
Experience the power of Luzmo. Talk to our product experts for a guided demo or get your hands dirty with a free 10-day trial.
