Using binary time series regression to predict horse racing outcomes

Abstract

Time series regression analysis is a statistical technique that arises when the dependent variable and its predictors are regarded as time series variables. Therefore, the assumptions of classical linear regression are violated since the observations in the study are no longer independent. In this dissertation, we examine how binary horse racing outcomes are a ected by a number of time series predictors. When the dependent variable has a binary form, this has a conditional Bernoulli distribution. We therefore study the properties of generalized linear models in the context of a time series, particularly for the binary response context which use the logit, probit, log-log and complementary loglog link functions. The generalized linear models are tted on datasets of two drivers, whose response variable could be used on betting markets which include the winning position, or the top three positions. This will determine the important predictors for these outcomes, while betting strategies are applied for improving the proportion of correctly guessed bets.