Background

Football is one of the most popular sports on the planet, it has been followed very closely by a large number of people. In recent years, emergence of new types of data collected from games such as play-by-play data including information on each shot or pass made in a match drove Data Science on the forefront of football.

In the race for footballing supremacy, teams are using an increasing amount of data and information to help give them an edge over the competition.

Data

Primary Data -

• The data was sourced from the pre-approved API https://www.football-data.org/documentation/quickstart

Secondary Data -

Additional relevant data relating to in-game statistics-

• Match data from European Leagues - http://football-data.co.uk .
• Team’s Performance - https://www.fifaratings.com/teams and https://www.fifaindex.com/teams/

Objective

Look at the play-by-play statistics involved in a game, which helped us evaluate a team’s performance, instead of just leveraging the actual number of “goals scored” metric from prior matches.

Combine key metrics of a team like their half time scores, team’s overall offensive and defensive ratings which are updated after each game to build model predicting the outcome of future matches.

Leverage in-game statistics to provide interesting and deeper perspective rather than the goals metrics, creating an opportunity to look beyond the match result itself.

Original Plan

Analytics Plan

• Leverage match data from the primary data source API https://www.football-data.org/documentation/quickstart
• Deriving insights from data and visualize using ggplot2().
• Use H2O to build the model from the primary data source.

Peer Comments Summary

Comments received from our peer teams suggested us to explore on many aspects of the game such as -

• Weather data - such data is not available from the API being used as well as weather data for each match is not readily available.
• Individual player performance/conditioning - player statistics for all the teams is not readily available also play-by-play statistics is the focus as we attempt to predict the outcome of the game for whole team.
• Team formation and statistics - We did incorporate formation of teams - - buying an additional subscription from the football API to get access to specific data - Indivdual team performance data from fifaratings.com
• Exploring secondary data - Explored and leverage from multiple other sources. These comments forced us to look more deeply into the kind of data that is accessible and available.

Ideas incorporated

• The match Odds for Win, Loss and Draw results for each team using an additional subscription to fetch data from the football API.
• Each teams in-game statistics from multiple external sources.
• Implemented for H2O’s AutoML - for automating the machine learning workflow.

Data Summary

Data from Primary source:

The primary source data from Competetion and Matches:

Data from secondary source:

Team performance data from secondary source: Team game Statistics

Dropping NA values from Team performance data to enhance our dataset.

Final dataset to be worked upon:

Exploratory Data Analysis

• Top 10 teams with most number HOME ground WINS

Top 10 teams with most number AWAY ground WINS

Most number of wins HOME/AWAY/DRAW

Top 3 winning team with HOME ground statistics

Machine Learning

Using H2O:

We have implemented four different models -

• Deep Learning Model 1
• Deep Learning Model 2 with tuning
• Deep Learning with Hyperparameter Grid search
• AutoML

Spliting for Training and Testing

Data wrangling for Machine Learning

Recipes and Baking

Deep Learning 1

Deep Learning 2 with tuning

HyperParameter Tuning for Deep Learning

AutoML

Summary for Model 1

Deep Learning M1

Summary for Model 2

Deep Learning M2

HyperParameter Tuning Results

Summary for AutoML

Summarizing the AutoML Results

Conclusion

• The deep learning modeling (model m2, with some tuning) gave good results in terms of a good RMSE score. Hyper-parameter tuning gave slightly better results as compared to m2 model.

• However, the results from the Auto ML were superior. The classifier is able to perfectly distinguish between the positive and negative class points perfectly.

• From the Auto ML confusion matrix, the total error across each class is 0.36. The RMSE is also very good. DeepLearning_grid_2_AutoML_4_20211208_140830_model_1 has the least error per class of 0.38.

• The predictors that made the impact were: hTAway, hTHome, Odds.homeWin, Odds.draw and Odds.awayWin.