Project information

  • Problem: Soccer Match event forecasting
  • ML Areas: Multi-Variate Time Series Forecasting, MLOps
  • Learning technique: Supervised Learning
  • Tools: Python, Pytorch, Git, Pandas, Flask
  • Project date: September 2020 - February 2021
  • Project URL: Soccer Match Prediction project

Description

Soccer Match Prediction

Licence

Libraries:

Pytorch Flask Tkinter Pandas

Dependences:

Python

Contents

Description

The idea is to create a custom deep learning algorithm for soccer match events, in particular the double chances events, that are 1X and X2. This events represent respectively the cases of home team winning/drawing and away team winning/drawing.

Goal: Training a neural network able to predict with high precision the double chances in soccer matches and building up a MLOps architecture, based on API, able to manage the training/inference of model in a automatic way.

Dataset

It was extracted from football-data uk, which permits to access to many national data tournaments, fully updated after every match. This website provides free data and available to everyone.

The given data contains many details for each match results:

  • match data
  • home team and away team
  • final score
  • half time score
  • betting odds
  • some match statistics (such as shots, corners, fouls, yellow/red card, etc...)

All details of the given dataset are available here

I use my model to train over the following national tournaments:

  • Italy (Serie A)
  • England (Premier League, Championship)
  • Germany (Bunsdesliga, Bundesliga 2)
  • Nederland (Eredivisie)
  • Spain (Liga, Liga 2)
  • France (Ligue 1, Ligue 2)
  • Belgium (Jupilier)

Task

Multi-variate Time-Series Classification

Solution

The provided solution is divided in two part: the first one regards the architecture of the project and the second one is a technical part, related to the deep learning model for the event prediction

Architecture

It is structred with Flask API on top that manage the possibilties of training and infering the DL model. Basing on MLOps theory, the training is triggered through some API, it chooses the best hyperparameters for that model and evaluate it. Finally set the model up to be processed. The prediction is also triggered through API.

Model

I create a neural network model that consists on merging two neural network branches: the first one processes data optimizing the favourite chance for home team (1X) and the second one takes care of optimizig the favourite chance for away team (X2)

Two networks were tested for solving this problem:

  • Each neural network branch is composed by LSTM network Fully Convolutional
  • Each neural network branch is composed by LSTM network Fully Connected

The two branches are then concateneted.

Author

Daniele Moltisanti

Linkedin

Portfolio