Edd Webster CFG Junior Data Scientist Data Challenge Submission

This repository is the code and analysis submitted by Edd Webster for the CFG Junior Data Scientist Data Challenge submission, including a list of publicly available resources published by the football analytics community that are related to the project. For a summary pack in Google Slides, please see the following [link], or the PowerPoint version [link] or PDF version [link] saved locally in this repository.

👋 About this Repository and Author

For more information about this repository, I am available through all the following channels:

📋 Contents:

Notebooks
Data
- Data Sources
- Documentation
Resources and Further Reading
- Football Analytics:
- Data Science:

📔 Notebooks

For code used to produce the Chance Quality Models and engineered dataset of game two of the Metrica Sports sample data, see the notebooks subfolder, in which the workflow is divided into the following:

Metrica Sports - analysis of the major chances that each team created during game 2 with an exported dataset of shots for predictions to be made by the following Chance Quality Models developed from the shots data (static version).
Chance Quality Models from Shots Data to calculate the probability of a shot resulting in a goal. Two models have been created (so far), the first, a Logistic Regression model and the second using the Gradient Boosting algorithm XGBoost:
- Logistic Regression model - initial model created as part of the submitted task (static version); and
- Secondary model created using XGBoost - an improved model created after the initial submisssion on 21/03/2021, that explores trying to create a model with improved performance using Gradient Boosted algorithms (static version).

💾 Data

ℹ️ Data Sources

Due to the 100mb file size limitation in GitHub, all datasets used in this analysis, including the Tracking data, have been exported and made publicly available to view and download in Google Drive at the following [link]. However, all the important files are below 100mb, including the shots data provided, and the exported Metrica Sports shot data with calculated features including Interfering and Intervening players, both raw and engineered, are available in the data folder of this repository.

The provided sample of shots data; and
Metrica Sports Sample Tracking and corresponding Event data. For code to work with this data, see the LaurieOnTracking GitHub repo by Laurie Shaw and the corresponding Friends of Tracking tutorials;

📄 Documentation

All documentation saved locally in the documentation subfolder, including:

Provided notes and definitions for the a sample of shots [link]; and
Metrica Sports event definitions.

🏛️ Libaries

The Python libraries used in this notebook include:

📑 Resources and Further Reading

Please see my football_analytics repository for my attempt to create as concise a list of possible of publicly available resources published by the football analytics community.

The follow resources are those that were specifically used to inform and create my submission for the CFG Junior Data Scientist Data Challenge, specifically focusing on Expected Goals and Tracking data. I have also included links to other topics related to the role such as the application of Reinforcement Learning in football. Credits to all those cited below.

⚽ Football Analytics

👨‍🎓 Tutorials

Friends of Tracking YouTube channel [link] and Mathematical Modelling of Football course by Uppsala University [link]. The GitHub repo with all code featured can be found at the following [link]. Lectures of note include:
- Laurie Shaw's Metrica Sports Tracking data series for #FoT - Introduction, Measuring Physical Performance, Pitch Control modelling, and Valuing Actions. See the following for code [link];
- David Sumpter's Expected Goals webinars for #FoT - How to Build An Expected Goals Model 1: Data and Model, How to Build An Expected Goals Model 2: Statistical fitting, and The Ultimate Guide to Expected Goals. See the following for code 3xGModel, 4LinearRegression, 5xGModelFit.py, and 6MeasuresOfFit;
- Peter McKeever's 'Good practice in data visualisation' webinar for Friends of Tracking. See the following for code [link];
Soccer Analytics Handbook by Devin Pleuler. See tutorial notebooks (also available in Google Colab) that notably include: 3. Logistic Regression, and 7. Data Visualization:
FC Python tutorials [link];
DataViz, Python, and matplotlib tutorials by Peter McKeever [link] - I think his website is currently in redevelopment, with many of the old tutorials not currently available (28/02/2021). Check out his revamped How to Draw a Football Pitch tutorial;
McKay Johns YouTube channel;
Tech how-to: build your own Expected Goals model by Jan Van Haaren and SciSports.
Fitting your own football xG model by Dato Fútbol (Ismael Gómez Schmidt). See GitHub repo [link];
Python for Fantasy Football series by Fantasy Futopia (Thomas Whelan). See the following posts:
Building an Expected Goals Model in Python by Peter McKeever (using WayBackMachine);
An xG Model for Everyone in 20 minutes (ish) by Football Fact Man (Paul Riley).

🏛️ Libaries and GitHub Repos

Friends-of-Tracking-Data-FoTD;
SoccermaticsForPython - repo by David Sumpter dedicated for people getting started with Python using the concepts derived from the book Soccermatics;
LaurieOnTracking by Laurie Shaw - Python code for working with Metrica tracking data; and
Expected Goals Thesis by Andrew Rowlinson. See both his thesis [link] and the following notebooks:
expected_goals_deep_dive by Andrew Puopolo. See the following notebooks:
soccer_analytics by Kraus Clemens. See the following notebooks:
- Expected goal model with logistic regression
- Challenges using gradient boosters
xg-model] by Dato Fútbol (Ismael Gómez Schmidt)
soccer-xg by Jesse Davis and Pieter Robberechts - a Python package for training and analyzing expected goals (xG) models in soccer (not used this this assignment but referenced here); and
Google Research Football. See the Kaggle Competition alongside Manchester City [link] (ended October 2020).

✒️ Written Pieces

For a full list of Expected Goals literature, see the following [link].

Papers

The following Shiny App from Lars Maurath is a great tool for looking up publications [link].

Routine Inspection: A Playbook for Corner Kicks (2020) by Laurie Shaw and Sudarshan 'Suds' Gopaladesikan. Accompanying talk - 2020 Harvard Sports Analytics Lab];
Dynamic Analysis of Team Strategy in Professional Football (2019) by Laurie Shaw and Mark Glickman. Accompanying talks - NESSIS 2019, 2020 Google Sports Analytics Meetup;
Football Shot Quality: Visualising the Quality of Soccer/Football Shots by Andrew Rowlinson. See his GitHub repo for code [link]; and
Game Plan: What AI can do for Football, and What Football can do for AI (2020) by Karl Tuyls, Shayegan Omidshafiei, Paul Muller, Zhe Wang, Jerome Connor, Daniel Hennes, Ian Graham, Will Spearman, Tim Waskett, and Dafydd Steele, Pauline Luc, Adria Recasens, Alexandre Galashov, Gregory Thornton, Romuald Elie, Pablo Sprechmann, Pol Moreno, Kris Cao, Marta Garnelo, Praneet Dutta, Michal Valko, Nicolas Heess, Alex Bridgland, Julien Perolat, Bart De Vylder, Ali Eslami, Mark Rowland, Andrew Jaegle, Remi Munos, Trevor Back, Razia Ahamed, Simon Bouton, Nathalie Beauguerlange, Jackson Broshear, Thore Graepel, and Demis Hassabis;
Google Research Football: A Novel Reinforcement Learning Environment (2020) by Karol Kurach, Anton Raichuk, Piotr Stańczyk, Michał Zając, Olivier Bachem, Lasse Espeholt, Carlos Riquelme, Damien Vincent, Marcin Michalski, Olivier Bousquet, Sylvain Gelly. See the GitHub repo [link];
A Framework for the Fine-Grained Evaluation of the Instantaneous Expected Value of Soccer Possessions (2020) by Javier Fernández, Luke Bornn and Daniel Cervone;
Decomposing the Immeasurable Sport: A deep learning expected possession value framework for soccer (2019) by Javier Fernández, Bornn, and Dan Cervone. Accompanying talks - SSAC19, StatsBomb conference;
Ready Player Run: Off-ball run identification and classification (2020) by Sam Gregory;
Beyond Expected Goals (2018) by Will Spearman;
Wide Open Spaces: A statistical technique for measuring space creation in professional soccer (2018) by Javier Fernandez and Luke Bornn;
“The Leicester City Fairytale?”: Utilizing New Soccer Analytics Tools to Compare Performance in the 15/16 & 16/17 EPL Seasons (2017) by Hector Ruiz, Paul Power, Xinyu Wei, and Patrick Lucey;
Not all passes are created equal: objectively measuring the risk and reward of passes in soccer from tracking data (2017) by Paul Power, Hector Ruiz, Xinyu Wei, and Patrick Lucey. See Paul Power's talk [link] (downloadable MP4), and the webpage [link];
“Quality vs Quantity”: Improved Shot Prediction in Soccer using Strategic Features from Spatiotemporal Data (2015) by Patrick Lucey, Alina Bialkowski, Mathew Monfort, Peter Carr, and Iain Matthews; and
A Framework for Tactical Analysis and Individual Offensive Production Assessment in Soccer Using Markov Chains (2011) by Sarah Rudd. Accompanying NESSIS talk on Metacafe [link].

📼 Videos

For a YouTube playlist of videos collated around the topics of Expected Goals, see [link]. For a Tracking data in Football specific playlist, see [link].

Webinars and Lectures

Laurie Shaw's Metrica Sports Tracking data series for Friends of Tracking (see the following for code [link]):
- Introduction;
- Measuring Physical Performance;
- Pitch Control modelling; and
- Valuing Actions.
Demystifying Tracking data Sportlogiq webinar by Sam Gregory and Devin Pleuler;
Will Spearman's masterclass in Pitch Control for Friends of Tracking;
How Tracking Data is Used in Football and What are the Future Challenges with Javier Fernández, Sudarshan 'Suds' Gopaladesikan, Laurie Shaw, Will Spearman and David Sumpter for Friends of Tracking.
Introduction to tracking data in football by David Sumpter for Friends of Tracking;
Learning to Watch Football: Self-Supervised Representations for Tracking Data by Karun Singh. See accompanying blog post [link];
David Sumpter's Expected Goals webinars for Friends of Tracking (see the following for code 3xGModel, 4LinearRegression, 5xGModelFit.py, and 6MeasuresOfFit):
- How to Build An Expected Goals Model 1: Data and Model;
- How to Build An Expected Goals Model 2: Statistical fitting; and
- The Ultimate Guide to Expected Goals.
'Good practice in data visualisation' webinar by Peter McKeever for Friends Of Tracking. See the following for code [link];
The Ultimate Guide to Expected Goals David Sumpter for Friends of Tracking;
How to explain Expected Goals to a football player by David Sumpter;
What is xG? by Tifo Football;
Opta Expected Goals by The Analyst (formally Opta);
What are Expected Goals? by David Sumpter and Axel Pershagen;
Anatomy of a Goal by Numberphile Brady Haran);
Sam Green OptaPro Interview;
How Did These Goals Go In? - We Explain How Goal Probability Works by the Bundesliga;
Soccer Analytics: Expected Goals by Dan Altman; and
Anatomy of an Expected Goal by 11tegen (Sander IJtsma);
Anatomy of a Goal (with Sam Green) by Numberphile:
"Is Our Model Learning What We Think It Is?" Estimating the xG Impact of Actions in Football by Tom Decroos from the 2019 StatsBomb Innovation in Football Conference;
Statsbomb Data Launch - Beyond Naive xG by Ted Knutson;
Karol Kurach - Google Research Football;
Karol Kurach (Google Brain) "Google Research Football: Learning to Play Football with Deep RL;
Google Research Football by Piotr Stanczyk;
Google's AI Plays Football…For Science! by Two Minute Papers;
Changing the soccer transfer market with big data by Giels Brouwer;
Soccermatics: how maths explains football by David Sumpter;
Data Robot Opening Remarks & Keynote: Making Better Decisions, Faster with Brian Prestidge;

Miscellaneous

Jeff Stelling xG rant; and
Craig Burley xG rant.

YouTube Channels

Friends of Tracking with David Sumpter, Javier Fernández, Laurie Shaw, Sudarshan 'Suds' Gopaladesikan, Pascal Bauer, and Fran Peralta;
McKay Johns;
Mark Glickman – for NESSIS talks, uploaded to his personal channel. Old talks are available on his Metacafe channel. See the official website [link];
42 Analytics – for SSAC conferences;
StatsBomb;
Opta - including Opta Pro Forum talks; and
Tifo Football.

🔊 Podcasts

List of notable episodes:

🐦 Tweets

The benefits of including fake data in an Expected Goals model by David Sumpter [link].

🧪 Data Science

Mathematics

Classification Metrics

General

Overview

scikit-learn documentation classification metrics

ROC AUC

Log Loss

Log Loss
Accuracy, Recall, Precision, F-Score & Specificity, which to optimise on?
Understanding binary cross-entropy / log loss: a visual explanation
Intuition behind Log-loss score
Log Loss or Cross-Entropy Cost Function in Logistic Regression
[Lecture 6.4 — Logistic Regression | Cost Function]](https://www.youtube.com/watch?v=HIQlmHxI6-0) by Andrew Ng
Binary Cross Entropy/Log Loss for Binary Classification
A Gentle Introduction to Cross-Entropy for Machine Learning by Jason Brownlee
What is Log Loss? by Dan Becker

Modeling

Logistic Regression

scikit-learn Logistic Regression official docs
Logistic Regression wiki
Logistic Regression by StatQuest (Josh Starmer)
Logistic Regression for Machine Learning by Jason Brownlee
Logistic Regression Tutorial for Machine Learning by Jason Brownlee

XGBoost

Greedy Function Approximation: A Gradient Boosting Machine by Jerome H. Friedman
XGBoost: A Scalable Tree Boosting System by Tianqi Chen and Carlos Guestrin (the authors of XGBoost);
XGBoost GitHub repo;
Awesome XGBoost repo;
XGBoost official documentation. See the tutorials [link]:
- Parameter Tuning
- General Parameters
Gradient Boosting wiki;
A Gentle Introduction to XGBoost for Applied Machine Learning by Jason Brownlee
How to Develop Your First XGBoost Model in Python by Jason Brownlee
How to Configure XGBoost for Imbalanced Classification by Jason Brownlee
How to Visualize Gradient Boosting Decision Trees With XGBoost in Python by Jason Brownlee
Story and Lessons Behind the Evolution of XGBoost - brief history and backstory to the creation of XGBoost by Tianqi Chen;
XGBoost A Scalable Tree Boosting System - talk by Tianqi Chen at the LA Machine Learning Meetup Group on 02/06/2016;
Kaggle Winning Solution Xgboost Algorithm - talk by Tong He (author of the R XGBoost package) at the NYC Data Science Academy. Mo
Gradient Boosting Machine Learning - talk by Professor Trevor Hastie
XGBoost lessson, part of the Intermediate Machine Learning course by Kaggle
Hyperparameter Optimization for Xgboost by Krish Naik
XGBoost in Python from Start to Finish by StatQuest (Josh Starmer)
XGBoost + k-fold CV + Feature Importance by Prashant Banerjee
A Guide on XGBoost hyperparameters tuning by Prashant Banerjee
Hyperparameter Grid Search with XGBoost
Hyperopt the Xgboost model by Yassine Alouini
Using XGBoost in Python by Manish Pathak
Getting started with XGBoost by Kevin Lemagnen
A Beginner’s guide to XGBoost by George Seif
Boosting your Machine Learning Models Using XGBoost by Derrick Mwiti
XGBoost Algorithm: Long May She Reign! by Vishal Morde
Gradient Boosting and XGBoost by Gabriel Tseng
Gradient Boosting from scratch by Prince Grover
HyperParameter Tuning — Hyperopt Bayesian Optimization for (Xgboost and Neural network) by Tinu Rohith D
Complete Guide to Parameter Tuning in XGBoost with codes in Python by Aarshay Jain

Feature Interpretation

SHAP

Official GitHub repo
Official documentation
- Linear models
- Tree-based models
A Unified Approach to Interpreting Model Predictions by Scott M. Lundberg and Su-In Lee
True to the Model or True to the Data? by Hugh Chen, Joseph D. Janizek, Scott Lundberg, and Su-In Lee
Interpretable Machine Learning with XGBoost by Scott Lundberg (creator the of SHAP library)
Explain Your Model with the SHAP Values by Dr. Dataman
SHAP Values Explained Exactly How You Wished Someone Explained to You

About

⚽📊 All code and analysis by Edd Webster (@eddwebster) as part the submission for the City Football Group Junior Data Scientist role. GitHub repo also includes links to publicly available resources made available by the football analytics community that are relevant to this project.

Languages

Language:Jupyter Notebook 99.3%Language:Python 0.7%