International Journal of Forecasting 26 (2010) 445–447 www.elsevier.com/locate/ijforecast
Editorial
Sports forecasting There are fifteen papers in this special issue. They are a testament to both the universality and the diversity of sport. The 39 authors come from three continents (and many more countries), and their papers cover five sports and the Olympics, as well as a survey which covers an additional two sports. In some of these studies the authors have sought to determine whether it was possible to forecast either the outcomes of sporting events or the relationships that are involved in the management of sports, while other papers are related to the betting market for their particular sport. This latter group of papers can be further categorized in two groups. One set of papers questioned whether the forecasts were sufficiently accurate to be profitable if bets were placed on the basis of the forecasts, while the other group viewed the betting market as providing an input into the forecasting process. We highlight some of the interesting features of these papers, including the sports that were investigated, the questions that were examined, the methods that were used, the new methodologies, and the profitability of betting strategies based on the forecasts. 1. Topics associated with each sport Four authors investigated various questions associated with soccer. Franck et al. compared the forecasting accuracies of two soccer betting markets: bookmakers and a betting exchange. Hvattum and Arntzen examined the value of ratings based on past performances for predicting the outcomes of matches. Leitner et al. also used ratings in predicting the outcomes of matches, but their research was extended to forecasting the results of the EURO 2008 tournament.
Strumbelj and Sikonja use bookmaker odds as forecasts for analyzing the matches of six major European soccer leagues. The topic of one of the three papers investigating Australian football differs considerably from the other two. Sargent and Bedford are concerned with forecasting Australian Football League player performances. On the other hand, Grant and Johnstone combine the probability forecasts of individuals to calculate the overall probability of a given outcome of a match. Similarly, Rydall and Bedford use an optimized ratings model to predict the outcome of a game. There are also three papers dealing with various aspects of horse racing. Lessmann et al. demonstrate that a state of the art machine learning model, a ‘random forest classifier’, can be used to make substantial profits from horserace betting markets, and also show that its predictions outperform those from traditional statistical techniques. Schnytzer et al. use a new variable based on estimates of insider trading to forecast the outcomes of horse races and show that this new variable has value. Smith and Vaughan Williams confirm the existence of a favorite-longshot bias, and find that the degree of bias declined significantly over the period studied, with important implications for the forecasting of horse race outcomes. Tennis drew the attention of del Corral and PrietoRodriquez, and Easton and Uylangco. The former paper tested whether the differences in rankings between individual players was a good predictor for Grand Slam tennis outcomes, while the latter used a model that predicted the probability of a player winning a tennis match, with the prediction updated on a point-bypoint basis. These forecasts were compared with the implied probabilities embedded in the betting odds.
c 2010 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved. 0169-2070/$ - see front matter doi:10.1016/j.ijforecast.2009.12.005
446
Editorial / International Journal of Forecasting 26 (2010) 445–447
Other sports attracted a smaller number of papers. Forrest et al. predicted the number of Olympic medals that would be won by different countries and Boulier et al. investigated managements’ abilities to forecast the future performances of athletes in the American National Football League. Finally, Stekler’s survey paper also included results from baseball and basketball. 2. Forecasting methodologies In many cases, the methodologies employed in these papers are conventional and have been used previously. These obviously include regressions, probits and logits,2 and will not be discussed further. There were a number of analyses that generated predictions using ratings and rankings, namely Boulier et al.; del Corral and Prieto-Rodriquez; Hvattum and Arntzen; Leitner et al.; and Rydell and Bedford. Some of these rankings were derived from the Elo ratings system, which was originally developed to measure performance in chess matches, but has been applied to team performance in the papers in this issue. Since the quality of a chess player cannot be measured precisely, it must be derived from data based on a player’s wins, losses and draws.3 A similar approach is used when measuring team quality. In any match, the answer as to whether a player (team) is expected to win is obtained by comparing the ratings of the two competing players (teams). If that player (team) exceeds these expectations, points are added to the rating, while the opposite occurs when a player (team) fails to meet these expectations.4 A number of methods were utilized for measuring player performance. Sargent and Bedford use non-linear Tukey smoothing to evaluate a player’s performance over the course of a season. Boulier et al. studied whether or not football executives were successful in evaluating the future performances of players who were entering the draft conducted by the
National Football League. However, the careers of some of the individuals had not been completed, and thus their entire performance was not known and could not be evaluated. In order to correct for this situation, the authors used censored regressions and life tables. Lessmann et al. introduce a new methodology for effectively predicting the outcomes of competitive events. This is achieved by integrating the random forest classifier with a conventional statistical technique, a conditional logit model. The effectiveness of this approach is demonstrated by the substantial profits which are achieved using an investment strategy based on the predictions. Hopefully, in the future, some of these newer techniques can be applied to sports analyses other than those that appear in this issue. 3. Are forecasting methodologies profitable? Although the focus of these fifteen papers was on predicting either the outcome of a sporting event or the probability that a specific event will occur, a number of these studies also sought to determine whether betting strategies based on these forecasts could be profitable. Using a Kelly method of betting, an ex post application of various strategies revealed that they could have been profitable, see Grant and Johnstone; Lessmann et al; and Rydall and Bedford. Other authors reached the opposite conclusion, specifically Hvattum and Arntzen; and Schnytzer. Finally, Easton and Uylangco compared strategies based on their model with the betting market, and concluded that the market is highly efficient. From our perspective, it does not matter whether the betting strategies were profitable or not. If individuals develop new techniques to try to “beat the market” that is all to the good. Our ability to forecast will thus continue to improve. Acknowledgements
2 When there are more than two possible outcomes, as in soccer matches, ranked probability scores are used rather than Brier Scores (see Strumbelj). 3 This measure is similar to the power scores that have been used to evaluate football teams. 4 Rydall and Bedford use an optimization procedure on the Elo ratings system.
We would like to thank everyone who contributed their time and effort to making this issue successful. The authors and referees all cooperated and met the tight deadlines that we imposed. In particular we would like to thank the many referees, whose
Editorial / International Journal of Forecasting 26 (2010) 445–447
Herman O. Stekler ∗ George Washington University, United States E-mail address:
[email protected].
comments definitely improved the quality of these papers.
Leighton Vaughan Williams 1 Nottingham Business School, Nottingham Trent University, United Kingdom E-mail address:
[email protected].
447
∗ 1
Corresponding editor. Tel.: +1 202 994 1261. Tel.: +44 0 115 848 6150.