A Journey Across Football Modelling with Application to Algorithmic Trading

  • Tarak Kharrat

Student thesis: Phd

Abstract

In this thesis we study the problem of forecasting the final score of a footballmatch before the game kicks off (pre-match) and show how the derived models canbe used to make profit in an algorithmic trading (betting) strategy.The thesis consists of two main parts. The first part discusses the database anda new class of counting processes. The second part describes the football forecastingmodels.The data part discusses the details of the design, specification and data collectionof a comprehensive database containing extensive information on match resultsand events, players' skills and attributes and betting market prices. The databasewas created using state of the art web-scraping, text-processing and data-mimingtechniques. At the time of writing, we have collected data on all games played inthe five major European leagues since the 2009-2010 season and on more than 7000players.The statistical modelling part discusses forecasting models based on a newgeneration of counting process with flexible inter-arrival time distributions. Severaldifferent methods for fast computation of the associated probabilities are derived andcompared. The proposed algorithms are implemented in a contributed R packageCountr available from the Comprehensive R Archive Network.One of these flexible count distributions, the Weibull count distribution, was usedto derive our first forecasting model. Its predictive ability is compared to the modelspreviously suggested in the literature and tested in an algorithmic trading (betting)strategy. The model developed has been shown to perform rather well compared toits competitors.Our second forecasting model uses the same statistical distribution but modelsthe attack and defence strengths of each team at the players level rather than ata team level, as is systematically done in the literature. For this model we makeheavy use of the data on the players' attributes discussed in the data part of thethesis. Not only does this model turn out to have a higher predictive power but italso allows us to answer important questions about the `nature of the game' such asthe contribution of the full-backs to the attacking efforts or where would a new teamfinish in the Premier League.
Date of Award1 Aug 2016
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorGeorgi Boshnakov (Supervisor) & Alexander Donev (Supervisor)

Keywords

  • Algorithmic Trading, Football models, Betting, Weibull, counting process
  • statistical arbitrage, Kelly betting,

Cite this

'