Predicting National Team Football Match Outcomes

This repository contains the code for our Deep Learning: Architectures & Methods project (course by TU Darmstadt). This project aims to predict the outcomes of national team football matches using a novel deep learning approach. Our final evaluation is on the European Championship 2024 matches.

Motivation

• Existing work on football match prediction focuses on statistical models [1-5] and league games [1,2].
• Gap: Deep learning-based prediction for national team games
• Challenges:
  • Sparse data for national team matches
  • Players from different leagues with varying competitive levels
• Proposed solutions:
  1. Input: Per-player average match statistics from the past year in their national leagues
  2. League-dependent weighting of player statistics
  3. Two-step training process to leverage more data:
     • Step 1: Train on league data
     • Step 2: Fine-tune on national matches

Model Architecture

Our model consists of two main components:

1. Team Embedder: Processes average player performance to create team embeddings
2. Match Classifier: Takes two team embeddings as input and predicts the match result

Training Step 1: League Data

Training Step 2: National Team Data

Data

We retrieved all our data used for training and validation from the API-sports platform.
The relevant player statistics are:

• player_position
• minutes_played
• attempted_shots
• shots_on_goal
• goals
• assists
• total_passes
• key_passes
• pass_completion
• saves
• tackles
• blocks
• interceptions
• conceded_goals
• total_duels
• won_duels
• attempted_dribbles
• successful_dribbles
• cards

For Training Step 1, we used matches from the following leagues:

• Premier League
• La Liga
• Serie A
• Bundesliga
• Ligue 1

For Training Step 2, we used matches from the following national tournaments:

• UEFA Nations League
• World Cup 2018, 2022
• Euro 2020

Generate our learning data

To request the data from the API and preprocess it to work with our model, we used the python scripts in the 'data' folder REQUIRED: You need to have an API key from API-sports to use the scripts and use the API internal IDs for the leagues / tournaments

• get_training_data_with_avg.py creates a csv with the all matches from a given league (or tournament) in a given timeframe and the average player statistics of the last year for each team
• api_request.py creates a csv with the all matches from a given league (or tournament) in a given timeframe and the player statistics for each player (not average!)
• get_prediction_data.py creates folders for each participating team in a tournament filled with the average player statistics of the last year for each player

Example Generation for Training Step 1

Either use average data for training:

python get_training_data_with_avg.py <api_key> <league_id> <start_season> <end_season> <output_file>

Or (non-average) match data:

python api_request.py <api_key> <league_id> <start_season-end_season> <output_file>

Example Generation for Training Step 2

currently still needs the match data and the average player statistics of a given year separately, support for using data from get_training_data_with_avg.py tbd

First get all the matches of a tournament using api_request.py

python api_request.py <api_key> <tournament_id> <start_season-end_season> <output_file>

Then get the average player statistics for each team and each season in the tournament

python get_prediction_data.py <api_key> <tournament_id> <season> <output_folder>

Results

• Evaluated on Euro 2024 (51 matches)
• Achieved approximately 59% accuracy
• Slightly outperformed betting odds
• But: Needs more evaluation: Euro 2024 only has 51 games.

Authors

• Paul Andre Seitz
• Leon Arne Engländer
• Lars Damian Teubner
• Daniel Kirn

References

[1] Moustakidis et al. "Predicting Football Team Performance with Explainable AI: Leveraging SHAP to Identify Key Team-Level Performance Metrics". Future Internet, 2023.
[2] Choi et al. "Predicting Football Match Outcomes with Machine Learning Approaches". MENDEL, 2023.
[3] Pinasthika and Fudholi. "World Cup 2022 Knockout Stage Prediction Using Poisson Distribution Model". IJCCS, 2023.
[4] Schauberger and Groll. "Predicting matches in international football tournaments with random forests". Statistical Modelling, 2018.
[5] Groll et al. "Prediction of major international soccer tournaments based on team-specific regularized Poisson regression: An application to the FIFA World Cup 2014". Journal of Quantitative Analysis in Sports, 2015.