Estimation and forecasting team strength dynamics in football : investigation into structural breaks

Sellitti Rangel Junior, J 2019, Estimation and forecasting team strength dynamics in football : investigation into structural breaks , PhD thesis, University of Salford.

PDF - Accepted Version
Download (16MB) | Preview


This PhD thesis studies the dynamics of team strengths in football. It investigates the presence of structural breaks, which occur when there is a change in parameters that govern dynamics in a time series. In football, such structural breaks occur because of events such as squad changes during transfer markets as well as managerial or ownership changes.

Team strengths are estimated across seven seasons of the Premiership and Championship football leagues and then analysed through a time series perspective, based on the double Poisson model with an added dependence parameter for lower scores and an exponential decay factor that adds more weight to more recent matches. This weighting scheme means that a pseudo-likelihood is used to estimate strength parameters. A rolling window approach is used to obtain a time series for the attack and defence strengths of teams in order to investigate the presence of structural breaks. We show that structural breaks are present in the majority of the time series. These present a challenge for the prediction of match outcomes. By not taking parameter discontinuity into account, one is in essence forecasting team strengths for the next match using incorrect parameter values.

We then carry out a forecasting exercise. This involves comparing the mean square error of the one-step ahead forecast of team strengths for all teams, using the two most recent seasons as the out-of-sample forecasting period. We find that different models have a smaller mean square error for different teams, but in particular two models stand out as the best ones: a simple random walk and forecasts made by model averaging. Even though the time-varying parameter model performs quite poorly according to the mean square error, it provides the best match predictions for one of our sub-samples. We conclude that different forecasting models that account for structural breaks can certainly improve forecast accuracy, although our findings are consistent with the econometrics literature that no one model forecasts best all the time. Given the prevalence of structural breaks in determining the dynamics of team strengths, this research has important implications for bookmakers and punters in the betting industry to take these matters into consideration when modelling football match outcomes.

Item Type: Thesis (PhD)
Contributors: Scarf, PA (Supervisor)
Schools: Schools > Salford Business School
Depositing User: J Sellitti Rangel Junior
Date Deposited: 05 Apr 2019 10:29
Last Modified: 27 Aug 2021 21:20

Actions (login required)

Edit record (repository staff only) Edit record (repository staff only)