Modelling football match results and the efficiency of fixed-odds betting John Goddard* University of Wales Swansea Ioannis Asimakopoulos University of Wales Bangor * Corresponding author Address: Department of Economics University of Wales Swansea Singleton Park Swansea SA2 8PP UK Tel: +44 (0)1792 295168 Fax: +44 (0)1792 295872 E-mail: [email protected] Key words: Sports betting, efficiency, ordered probit Abstract An ordered probit regression model estimated using 15 years’ data is used to model English league football match results. As well as past match results data, the significance of the match for end-of-season league outcomes; the involvement of the teams in cup competition; the geographical distance between the two teams’ home towns; and the average attendances of the two teams all contribute to the model’s performance. The model is used to test the weak-form efficiency of prices in the fixed-odds betting market, and betting strategies with a positive expected return are identified.1Modelling football match results and the efficiency of fixed-odds betting 1. Introduction The predictability of match results is the main concern of research on the efficiency of sports betting markets. The recent applied statistics literature has focussed primarily on modelling goal scoring (Dixon and Coles, 1997; Rue and Salvesen, 2000; Crowder et al., 2002). Recently some econometricians have suggested modelling match results directly (rather than indirectly through scores) using discrete choice regression models (Forrest and Simmons, 2000a,b; Koning, 2000; Kuypers, 2000; Dobson and Goddard, 2001). A focus on match results rather than scores can be justified partly on grounds of simplicity: fewer parameters are required; estimation procedures are simpler; and the resulting models lend themselves to the inclusion of a variety explanatory variables. This paper presents a regression-based model to explain and predict match results that is more extensive and comprehensive than any previously developed, as regards both the range of explanatory variables that are incorporated, and the extent of the data set used to estimate the model. Research into the efficiency of prices set by bookmakers in betting markets has provided a small but increasing contribution to the financial economics literature on market efficiency. Much of this literature focuses on racetrack betting, but betting on team sports match results has also attracted attention. Early researchers sought evidence of inefficiencies in the form of systematic biases in bookmakers’ prices, such as home-away team or favourite-longshot biases. More recently forecasting models have been used to establish whether historical information available in previous match results can be extrapolated to formulate profitable betting strategies. This paper uses the model outlined above to test the weak-form efficiency of the prices quoted by high street bookmakers, and to identify potentially profitable betting strategies.2The paper is structured as follows. Section 2 reviews the literature on modelling and forecasting football match results. Section 3 describes the specification and estimation of an ordered probit regression model to explain and predict match results. Section 4 reviews the literature on betting market efficiency. Section 5 investigates the efficiency of the prices quoted by high street bookmakers over four English football seasons from 1998-9 to 2001-2, using both regression-based tests and direct economic tests of the profitability of betting strategies based on the model’s evaluation of expected returns. Section 6 concludes. 2. Modelling football match results A limited but increasing number of academic researchers have attempted to model match results data for football. Early contributions by Moroney (1956) and Reep et al. (1971) use the poisson and negative binomial distributions to model at an aggregate level the distributions of the numbers of goals scored per game. This aggregate approach precludes the generation of specific forecasts for individual matches based on information about the respective strengths of the two teams. By comparing final league placings with experts’ pre-season forecasts, however, Hill (1974) demonstrates that individual match results do have a predictable element, and are not determined solely by chance. Maher (1982) develops a model in which the home and away team scores follow independent poisson distributions, with means reflecting the attacking and defensive capabilities of the two teams. A full set of attacking parameters and a set of defensive parameters for each team are estimated ex post, but the model does not predict scores or results ex ante. A tendency to underestimate the proportion of draws is attributed to interdependence between the home and away scores, and corrected using the bivariate poisson distribution to model scores. Dixon and Coles (1997) develop a forecasting model capable of generating ex ante probabilities for scores and match outcomes. The home and away team scores follow independent poisson3distributions, but for low-scoring matches an ad hoc adjustment allows for interdependence. Using a similar framework, Rue and Salvesen (2000) assume that the time-varying attacking and defensive parameters of all teams vary randomly over time. The prior estimates of these parameters are updated as new match results information is obtained. Markov chain Monte Carlo iterative simulation techniques are used for inference. Crowder et al. (2002) develop a procedure for updating the team strength parameters that is computationally less demanding. Researchers who have examined the impact of specific factors on match results include Barnett and Hilditch (1993), who investigate whether artificial playing surfaces, used by several clubs during the 1980s and early 1990s, conferred an additional home-team advantage. Ridder et al. (1994) show that player dismissals have a negative effect on the match result for the teams concerned. Clarke and Norman (1995) use a range of non-parametric techniques to identify the effect of home advantage on match results. Dixon and Robinson (1998) investigate variations in the scoring rates of the home and away teams during the course of a match. The scoring rates at any time depend partly upon the number of minutes elapsed, but also upon which (if either) team is leading at the time. Recently,