From betting tipsters to financial advisors, past data is widely used to predict future events. In this article, Dominic Cortis discusses the pitfalls and advantages of this method and explains why prediction models failed to predict Leicester’s chances of winning the league.
Prediction is not an easy business to be in as from my short experience it can sometimes lead to two extreme impressions. When everyting goes as planned, they treat you as if you have prophetic abilities. When extreme events you considered aren’t even close to reality – then expect a charlatan status.
Independent of what is being predicted, there is one great likelihood – the exact prediction made will not occur,especially if there are a multitude of possible outcomes.
Using the past to predict the future
A useful method is to anchor to the past or general information. For example, we know that home teams score more than away teams. So when faced by two teams of equal strength, the best guess would be that the home team will win. Which half will have most goals in a soccer match? Well the second. Will England win on penalties? No.
What could give us more insight is the measure of dispersion, such as the standard deviation, as this shows the amount of possible discrepancy from best estimate.
If we had to anchor however, we could never predict Leicester FC’s ascent (yes, all my articles will have to mention the city) at the start of the season, or would we? Taking the last few games of 2014/15, we could have interpreted Leicester as having a shot at doing well. Although, I must admit, it seemed fanciful to the extreme to even consider top 6.
The limitations of modelling
Modelling should never be done in a silo and must consider the intricacies of each case. I simply adore this analogy of how to use prior probability. The article on Hans Solo and Bayesian Priors explains the Star Warsscenario of the Millenium Falcon flying through an asteriod field where C3P0 tells Han Solo that he has a 1 in 3720 chance of successfully navigating through the field. Yet C3P0 didn’t consider that the normal stats do not apply to Han, and should be updated accordingly for his scenario.
This is what has been happening in Leicester’s case. Firstly some models don’t really apply here. For example a typical model is to predict the number of goals to be scored as a Poisson process.
Yet, this would place a higher chance on a team with a higher scoring rate to win. But Leicester is 1-0 winning type of team, although this would be counteracted by their defense strenght in the model.
Secondly, even if proper stats were being used – the market seemed illogical. Everyone has been expecting Arsenal to put on the pressure and Leicester to deflate. While that was a plausible scenario, the most likely one was that they both stay on the same course.
The middle way
In conclusion, anchor but don’t forget that the past performance is not an indicator of future results. Hence you need to be subjective, creative, judgemental and consider the case at hand, however don’t be overly whimsical. No, it isn’t easy.