Back from the FA Cup weekend and I’ve been using the gap in the Premier League fixture list to further develop the EPL Model. Unfortunately, this time without much success.
An agent based model isn’t like the majority of statistical models, where if you make a change you can usually see fairly quickly whether or not your new model is better than the old one. When I change the EPL Model, I re-run a season (at least) of historical fixtures to see how the new iteration performs. That means simulating three hundred and eighty fixtures, two thousand times each. With around eight hundred ‘events’ per game, the computer is crunching through over six hundred million simulated passes, shots and tackles before I find out whether the idea was an improvement.
This weekend it wasn’t. Back to square one and the strikers in our simulations – which is what I was working on – aren’t any more accurately modelled than last week. It happens and to paraphrase Edison, I haven’t failed, I’ve just found another way that won’t work.
Last time out marked our first loss in a little while and overall the weekend can be chalked up as unlucky. The model called five results out of ten, which is normally not too bad, but two of them were Arsenal and Man City to win – both at home – and you don’t get many prizes for that. I say unlucky because we also had Southampton to win away at Sunderland and the Saints were 2-0 up and looking good before Sunderland pulled it back to a draw. Liverpool also very nearly pulled back to win after an uncharacteristic first half against Aston Villa.
That Liverpool result is also worth a quick note that the model isn’t good enough to understand when players aren’t in their best positions. Our modelled Steven Gerrard’s stats won’t change if you play him for the first half as a holding midfielder. In the simulations, he’ll always do what he usually does.
Here are the percentages for Tuesday and Wednesday evenings:
And the deeper passing and shooting stats from the model:
There’s usually an odd one in the list and this week it’s Liverpool; not the prediction that they’ll win, but that the model gives Everton so little chance. This is almost certainly unrealistic, but comes about as a result of the weight of shots that is expected to come from Liverpool’s midfield compared to Everton’s. Coutinho, Sterling, Gerrard and Henderson are expected to have a couple of shots each, in addition to a major threat from Suarez and Sturridge. Everton’s front line are potentially a match in terms of number of shots, but their midfield isn’t. This rounds out with the model not predicting a lot of goals for Liverpool, but a win in the majority of simulation runs.
And here’s who I’ll be backing. A few people have asked why I pick the bets that I do, given the model’s percentages, so I’ve written a post to explain here, to explain how it works.
Manchester United v Cardiff City – Home win
Norwich City v Newcastle United – very close. Away win.
Southampton v Arsenal – Away win
Swansea City v Fulham – Home win
Crystal Palace v Hull City – Draw
Liverpool v Everton – Home win
Aston Villa v West Bromwich Albion – Draw
Chelsea v West Ham United – Draw
Sunderland v Stoke City – Home win
Tottenham Hotspur v Manchester City – Away win
We’re backing a draw for West Ham against Chelsea, not because a draw is the most likely result, but because it’s probable enough to be worth backing. The most likely single result by some distance is a Chelsea win.
Stay tuned for a second post ahead of this weekend’s fixtures!