Meet the winners of the Laduma Analytics Football League Winners Prediction Challenge
Meet the winners · 1 Dec 2022, 09:37 · 6 mins read ·
9

Meet the Winners of the Laduma Analytics Football League Winners Prediction Challenge, a collaboration between Laduma Analytics and Zindi, where participants were to predict the outcome of a football match, based on historical match and player data.

The challenge attracted 699 participants and over 4 000 submissions from 75 countries, all vying for a \$2 000 prize pool. Winners João Luiz Bunoro, Ernest Paris and Team Dummies (Yohannes Melese and Stephen Kolesh) recount their experiences and share what they learned.

### 1st place: João Luiz Bunoro, Brazil

My name is João Luiz Bunoro (Bunoro), I live in São Paulo, Brazil. I have a BSc and MSc degree in Physics from the University of São Paulo. After university, I started working in the banking industry with Credit Risk, and I am still working in this area as a risk manager. Analysing data and extracting useful information from it is part of my job, especially when I need to conduct some studies or solve a business problem. However, today Data Science is much more of a hobby than part of my job and I like to study AI, ML and programming in my spare time.

Tell us a bit about your solution, and the approach you took.

In this challenge we had access to two types of data: i) the results of each game and ii) the list and description of the events of each game (e.g. shots, passes, tackles). Especially, the second dataset included the position (X and Y coordinates) where the events occurred and the time when each event started and ended. The solution was based on a simple approach built after an exploratory analysis of the train dataset.

Instead of using the data to predict the probability of possible outcomes ("home/away win" and "draw"), I used the data to find (and count) the number of goal-related events and the teams that scored them. After this step and based on a simple count of the goals events, it was possible to predict how each match ended.

The combination of i) exploratory analysis and ii) focus on finding goal events first rather than directly predicting the outcome of matches helped me to find the final solution. In this sense, I explored, for example, several heat maps showing the position of passes and goals and statistics of the variables, but the most important information was the duration time of the events (that is, the difference between the "end_minutes" and "start_minutes "). I found that the events related to goals lasted longer than the other events. After that, I took some examples and I found that the pattern of events related to "Goals" and "Own Goal" were different, which allowed me to create a simple rule to identify goals and the team that scored.

An assumption used was that the X position of the team who scored the goal was greater than 52.5 (i.e. after the middle of the field).

What set your winning solution apart from others? Share some tricks and tips here:

In addition to exploring the data, I followed the discussions on the contest page and watched the webinar,which helped me a lot.

Words of encouragement for others, or advice that has helped you?

My main motivation for participating in ML competitions is to improve my data science skills. More than being in the first position, the whole process of understanding the problem, interacting with other participants, and looking for new methods or tips is a good way of learning and eventually appearing in the top positions of a competition.

### 2nd place: Ernest Paris, Spain

My name is Ernest Paris (Ernest-P) I’m based in Europe and I lead AI projects for a pharmaceutical company.

From time to time I love to get into an AI competition. At Zindi we have some people who get into all or almost all competitions, but that's not me. So far I'm taking one competition at a time.

Tell us a bit about your solution and the approach you took.

I have written my solution on the Zindi discussion boards, please access my solution explanation, questions and answers here:

What set your winning solution apart from others?

My solution is based on identifying attacking players (strikers and other players helping strikers to score) versus goalkeepers.

In the train set, I could identify some attacking players and some goalkeepers, so on the test set, I could start iterating to identify more players (e.g. known strikers from train scores so the player conceding the goals which were previously unclassified becomes identified as a goalkeeper).

How do you prepare for a challenge?

For me, the key issue is to keep expectations under control, try to learn something new, and have fun!

Words of encouragement for others, or advice that has helped you?

Learn AI through a course or in theory, and join a competition to test what you just learnt.

What do you like about Zindi?

The community is nice, there's always one (or more) competitors that catch my attention. Some for technical reasons, others because of their helping nature, and others because of the domain.

I also appreciate some "extras" about Zindi that if they weren't there it could ruin the whole experience - like platform availability and reliability. Zindi has an easy and intuitive interface, I can just focus on the AI and not on how to use the platform.

### 3rd place: Team Dummies, Ethiopia & Kenya

We are team Dummies, a duo of Yohannes Melese (Jonny) from Ethiopia studying a BSc in Electronics and Communication Engineering at Adama Science and Technology University (ASTU) and Stephen Kolesh (Koleshjr) from Kenya a fourth-year Computer Science Student at Multimedia University of Kenya (MMU).

Tell us a bit about your solution, and the approach you took.

First, we built an action prediction model, in which we added seven features related to goals scored and goals conceded to the original features we were given. These features were the key features of our model.

Second, we used the modified data from the first approach and the main game statistics files to generate more features like pass accuracy, total shots, team attacking and defence strength from the previous seasons of goals scored and goals conceded. We then trained catboost and lightgbm models on 10 StratifiedKFold.

What set your winning solution apart from others?

Having a separate model from the main one which was predicting the action type of the game statistics was the key. It gave us confidence for our result prediction model.

How do you prepare for a challenge?

We were participating in other competitions on Zindi, and also we did research on papers that are written on this area.

Words of encouragement for others, or advice that has helped you?

Keep trying, it's all about trying new things even if it won't give an improvement, you will learn something from it.