Thursday, 1 June 2017

Lawro: Master Bayesian

With another season over I wanted to look again at the accuracy of Mark Lawrenson's predictions in the Premier League.  I've been doing this for a number of years now, starting with an analysis of the 2012/13 season and I'm always amazed at how well he does.

At face value, Lawro's table based on his weekly predictions for the BBC look a bit ridiculous, not least his prediction that Liverpool would go through the season unbeaten:
League Table of Lawro's predictions (via BBC Sport)
If you'd used his selections and placed £10 on his predicted result (Home/Draw/Away), you'd have made £501.50 profit from 368 matches (I've excluded the 12 rearranged games) and actually made a small £7.45 profit on his Liverpool predictions.

Lawro's 92 points prediction for Liverpool may seem a bit extreme but you have to consider that this is built up from a series of individual predictions rather than a single prediction of their points total.  In any single game it may make sense to not predict Liverpool will lose (even if you put aside the fact he's an ex Liverpool player).

I previous years most of Lawro's profit has been derived from sitting on the fence more than the average punter and taking advantage of the fact that draws are often overpriced as people in general want to back a winner rather than a stalemate.

This year however, his profit (and I'm sure what has been a bit of a hit on the Bookies in general) comes from the fact that the bigger teams have got their act together and this season has very much been a two-tier league (or 3 if you consider Everton in no-man's land on their own).

If I had to summarise what I thought made Lawro successful all these years, I think it's the fact that he isn't too swayed by short term noise and to some extent you could largely predict his predictions from early on in the season (not least the scorelines used)

For years now, Lawro's kept a pretty simple system and almost always predicts a draw as being 1-1 rather than 0-0 or 2-2, as mentioned in previous years this makes sense as it makes you more likely to be closer to the true result as although betting wise predicting 1-1 when a match finishes 2-1 is no closer to being right than predicting 0-0 it 'feels' closer
The difficultly of any prediction system or model is how much to weight the recent past vs. more historical data. e.g., Is form over the last 1/3/12 games more predictive than what happened last season or even further back.

A great example of this would be Leicester, although a serious title defence was always going to be incredibly unlikely, their results under Ranieri could also be considered to be somewhat surprising.

Obviously clubs don't work in a vacuum and changing managers impacts things (both positively and negatively I can tell you as a Swans fan), but trying to judge whether something is an outlier or the new normal is hard.

People often talk about things 'regressing to the mean' but the whole issue with that is it assumes you've predicted the right mean to start with but as Lawro's predictions have proved over the years, having an opinion and sticking with it isn't the worst strategy in the world.

I'd like to thank MyFootballFacts for collating the Lawro results each week and for collating results and betting odds, both are great resources.

It's still a work in progress but I've used the info from the Football-Data site to create an interactive dashboard of shots taken/faced by each team:

Twitter: @we_r_pl