ABSTRACT
What makes this paper so interesting is that it presents a very thorough analysis of the complexities involved in making predictions sufficiently accurate to permit profitable betting in the AFL prediction market. Discussed are the various explanatory variables and optimization of their measurement. For example, the decomposition of home-ground advantage into home-player familiarity with the ground, visiting team fatigue in traveling interstate and other factors evidently adds to profits. There is also testing of the optimal way to predict explanatory variables such as past performance: does one use moving averages of exponential smoothing or some other technique? The problem from an academic viewpoint is that the paper suppresses actual coefficient values and other details for commercial reasons. The subsequent literature is far more specialist in nature, and a very brief description of three papers will suffice.1 Grant and Johnstone (2010) predict game outcomes and simulate betting by pooling forecasts of winning probabilities derived from a web-based football “tipping” competition, which has been conducted by the computer science faculty at Monash University in Melbourne since 1995. They present exhaustive tests of different pooling and betting methods and show that statistically significant, although not large, profits may occasionally be made using this approach, although in the long-term average losses prevail. Ryall and Bedford (2010), on the other hand, claim that long-run profits are available in this market if a ratings-based forecasting model is adopted. The model used is that of Elo (1978), originally designed for ranking chess players. Over the 2001-2008 AFL seasons they generate a return of investment of 8.8 percent, betting a constant amount on each game and 10 percent using a Kelly system. These returns are greater than those presented here, but the method adopted is highly computer intensive and may be impractical if rankings are to be updated after each round.2 If this model is indeed successful, it would presumably yield even better results if rankings were regularly updated. Sargent and Bedford (2010) show how exponentially smoothed, one-step forecasts of AFL player performance data are improved by first applying a nonlinear smoother to the raw data. In this respect, their paper builds upon Bailey and Clarke (2004) in its analysis of exponential smoothing as yielding improved forecasts over simple and moving averages. Player performance is defined as an index based upon several player-level statistics of the kind used in the paper (kicks, handballs, etc.), but no use of the predictions in simulated (or real) betting in the AFL prediction market is presented. The central feature of the analysis presented here is its attempt at simplicity, if not naivety. Thus, the regressions run are of the simplest kind and the variables
used are extremely basic: no attempt is made to index player performance, the emphasis being on the raw data. Further, home-ground advantage is represented by a dummy variable, thus precluding any degrees of advantage. Finally, in order to predict player performance, reliance is upon simple means alone. The reason for this approach is two-fold. First, it is interesting to ask whether profits are obtainable, however modest they may be, without resorting to complications – and the answer turns out to be positive. Second, testing the methodological hypothesis that the better the regression, the more profitable will be the predictions it yields, requires that as many confounding factors as possible be removed from the analysis.
