Our colleague, Bálint Kulcsár, drew inspiration from Dr. Hannah Fry's book "The Indisputable Existence of Santa Claus" to create BALLER, arming it with extensive knowledge to predict the outcomes of the UEFA EURO2024 matches. BALLER's impressive preparation led it to become one of the top performers in our company’s prediction contest. If you're curious to learn how this remarkable result was achieved, check out the details!
---
As I sit here, I’m looking at one of my favorite books on my ever-growing bookshelf: Dr. Hannah Fry’s The Indisputable Existence of Santa Claus. Dr. Fry, a UCL maths professor, combines her brilliant mind with a touch of British wit, making her books a joy to read.
One chapter, in particular, had me laughing out loud. Dr. Fry used statistics to develop a strategy to "crush her relatives" in Monopoly at Christmas. She meticulously details which properties to prioritize, dismissing train stations as "worthless" and recommending hotels on the red fields. Her point? Stats and models work, even if they need a little twist.
I love when humor and maths blend, and I like to think that my AI project does that too, albeit to a lesser degree. Enter BALLER (Brilliant AI Logic for League and European Results). While I don’t claim to fully understand ChatGPT, I’ve been experimenting with it for a few months and have some intriguing insights to share.
Every four years, my company hosts a Euro outcome prediction competition. This year, with BALLER’s help, I finished 3rd out of 77 participants—a huge leap from my 40th place out of 72 in 2020. The twist? I know nothing about football and don’t care much for it. All my predictions came from BALLER.
Here’s how I turned an AI model into a football prediction champ:
1. Training the Base Model:
BALLER’s foundation was a base AI model (GPT-4.0o) trained - out of the box - on extensive football data, providing essential knowledge for predictions. But I needed more.
2. Data Scraping with RPA:
I used Robotic Process Automation (RPA) to scrape up-to-date football data from various websites—player stats, team performance, historical outcomes. Feeding this data into BALLER made its predictions more relevant.
3. Incorporating Real-Time News:
I added real-time news updates, like Mbappe’s broken nose, to BALLER’s data. I even integrated free football sports news RSS feeds to keep BALLER updated.
4. The Outcome:
The result? BALLER placed 3rd. It wasn’t perfect, but impressively accurate. In the final game between Spain and England, a 1-1 draw—right up until the 82nd minute—could have secured BALLER the top spot.
I was curious how well the base model held up for other games: predictions for international and US games seemed better, but it struggled with niche markets like Japan, Kazakhstan, and even Australia. When I found myself scraping football data in Japanese in the middle of the night, I knew I had to stop this fun experiment.
This project taught me the flexibility and potential of AI. Generative AI like ChatGPT, though not designed for football predictions, can perform exceptionally well with the right tweaks and data. So the point isn’t that Gen AI was designed for this—it wasn’t. It’s that with creativity, it can be repurposed effectively.
Much like Dr. Fry’s statistical strategies for Monopoly, my journey with BALLER highlighted the power of adaptability and creativity in tech. Models and algorithms, when twisted just right, can achieve amazing results beyond their original purpose. It makes me wonder what else AI will be capable of with the right ideas, creativity, data, and people.