There are many factors other than just luck that can influence the final result of a game making it just too complex to number-crunch. This makes football match prediction an ideal test-bed for researchers trying to come up with computer decision-makers that can deal with real-life uncertainties.
The data that is freely available for football match analysis is very limited and typically restricted to facts such as the number of goals scored, shots, red/yellow cards, corners, and so on. While some of these stats are very useful, the most important factors that can determine the outcome of the match are missing from such data. Some of these factors are:
- Team spirit/psychology
- Team form
- Team fatigue
- New manager
- Key player missing/returning back to action
More importantly, some of these missing factors cannot be represented by hard evidence. Instead of ignoring these factors on the basis that no data are available, we incorporate them into our models in terms of expert knowledge. Bayesian networks are particularly suitable for this kind of uncertainty analysis.
How good are the models?
The first thing to realise is that for a football forecast model to make a profit, the model must be able to assess probabilities that are noticeably more accurate than bookmakers' odds. This is not as easy as many football fans think. This is because the bookmakers publish odds which incorporate an in-built profit margin (like roulette), so that a fixed percentage of profits is guaranteed in the long-run over all the bets received.
However, our recent research was based on the idea that sports gambling markets, and particularly football, publish odds that are biased towards giving the bookmakers the most profits. This means the publicised odds suffer from a certain amount of built-in inaccuracy. A gambling market is inefficient if there is a way to consistently generate profit against published market odds. Based on the above hypothesis, the bookmakers' intended inefficiency can potentially be exploited for profit with a sufficiently accurate model.
We used our Bayesian network models to make predictions, giving probabilities for wins, losses and draws about premiership matches. To keep them honest, all of our predictions were published online at www.pi-football.com before the kick-offs. Our results demonstrated profits against the published market odds of around 10%, and over the period of two consecutive football seasons. This was the first academic study to demonstrate profitability that is consistent against football market odds.
For more info, visit www.pi-football.com