Prediction Model: Wk 14 Forecasts | Arsenal, Liverpool, Man Utd to win

Prediction Model: Wk 14 Forecasts | Arsenal, Liverpool, Man Utd to win

Let’s kick off with a quick round-up of Saturday and Sunday before we get onto the midweek game predictions…

The past weekend went fairly well, with a small 5% profit if you were betting on the EPL Model’s calls. We got six out of ten results correct, with the games at Chelsea, Man City, Cardiff, Everton, Norwich and Newcastle all running as forecast.

Premier League Predictor Results

Fulham was the big miss this week, with a predicted draw coming out as a comfortable West Ham win. The model’s been struggling with Fulham this season because some of their play has been so poor, it actually lies outside the norms that it uses. This excellent piece from Colin Trainor illustrates some of the issues the model’s been having. It was written only six games into the season but clearly shows that even back then, Fulham had established as a side that concedes many shots from high risk locations, but then shoots from dreadful areas itself.

There’s no reason the model couldn’t handle that, if we had data available every week on the locations shots were taken from, but we don’t. I’m forced to use more aggregate goal conversion stats – because that’s what there is unless you’re lucky enough to work with the fully detailed Opta data feed – and so the model adjusts more slowly than I’d like to poor shot selection and conceding shots from bad areas. When goal scoring performance looks really poor, but with low numbers of shots to use for the assessment, the stats are moderated back to a more ‘normal’ (but well below average) figure. This is fine and it works well, until you’re dealing with a Fulham, who really have had appalling stats and it’s not just noisy low frequency data that needs to be tuned out!

These sorts of problems will stabilise as the season progresses and we get more data but it wouldn’t be any fun if I only did these posts from the beginning of February.

Here are the mid-week predictions:

3-12

Liverpool to bounce back and hammer Norwich, even without Sturridge? We’ll have to wait and see. Being agent-based, the model does know that he’s not playing…

Here are the possession and shooting stats for each game:

3-12 detailed

I’ve had a some questions these past few weeks about why I call bets the way that I do, so a quick word on the method before we get to the list. Firstly, these are definitely not based on my judgement, they’re 100% taken from the model using a simple rule of backing the most likely result, unless the draw percentage is over 25%, in which case back that. As a strategy this was profitable last season and modelling historically, should have been profitable the season before that, too.

Draws are almost never the most likely single outcome of a game (Pulis’s Stoke managed it occasionally), but you can get a decent number of correct predictions on them, so it’s profitable to back a draw when the model indicates that it’s a more likely result than usual.

A few people have pointed out that this is quite an unsophisticated betting strategy and I make no claims otherwise. It is! It works though. It’s partly come about as a result of this model not having been built primarily for betting, but with the aim of getting as many EPL game results correct as possible and as a tool that could predict if one starting line-up was better than another. It’s not designed to do things like go looking for good value long-odds bets.

You might think a better strategy would still be to look for ‘value’ in the bookies odds – games where the model disagrees strongly with the bookies – but this will not work consistently. You have been warned. The model does have an edge, but chasing differences to the bookies very often leads to betting on the occasions where the bookies know better than my forecasts. It’s quite a stable model, but all models have errors and this strategy unerringly finds them.

Finally, a word on correct scores, which I’ve mentioned before. It’s not terrible at these, but it’s also not good enough to profitably use to bet. The model doesn’t know about ‘score effects’ – factors like teams taking their foot off the gas at 3-0 up, or committing players forward to chase the game at 1-0 down – so the “most likely score” column is for interest more than anything. Also, even my standard two thousand simulation runs per game isn’t always enough to stabilise the scoreline in high scoring games. All in all, betting on correct scores seems like a mugs game to me – they’re far too random… Could you use that prediction to bet on both teams to score? Maybe, but I haven’t checked yet.

This quick mid-week update has turned into a bit of a ramble! Here’s what I’ll be betting on Tuesday and Wednesday. See you on Friday for this weekend’s post.

Crystal Palace v West Ham United – Away win
Arsenal v Hull City – Home win
Liverpool v Norwich City – Home win
Manchester United v Everton – Home win
Southampton v Aston Villa – Close one this. Home win.
Stoke City v Cardiff City – Draw
Sunderland v Chelsea – Away win
Swansea City v Newcastle United – Away win
Fulham v Tottenham Hotspur – Away win
West Bromwich Albion v Manchester City – Away win