Martin Ingram, University of Melbourne, Australia

Applied statistics workshop 
Martin Ingram, University of Melbourne, Australia
"Gaussian Processes for Paired Comparison Modelling"

Paired comparison models such as Elo and Glicko are popular for predicting the outcome of sporting events. Their closed-form updates are fast to compute, but it is not always clear how they can be extended or modified. I show that the Glicko model is a special case of a Gaussian Process model with a random walk kernel and a paired comparison likelihood. Framing the model as a Gaussian Process allows modification of model assumptions, such as replacing the assumption of a random walk over time with a smoother function. I fit a Gaussian Process model to several years of ATP tennis data and evaluate its performance against Elo and Glicko to investigate whether different modelling assumptions translate to improvements in predictive performance.