Well that was a lot of fun. Starting out with a subscription to EPL Index statistics, I decided in January this year to have a go at building a predictive model for Premier League football games. Wallpapering Fog's been a record of the development and predictions (all posts here) and has done a great job of keeping me motivated. It would be easy to skip a week's predictions, when you've got a hangover on Saturday morning and real-life stuff to do, but publicly committing to posting has worked. It's also turned up a fascinating community of analysts, who I'd probably never have read otherwise. Isn't Twitter brilliant?
From a personal point of view, it's been a massively satisfying project. If only because I've wanted to build a proper agent-based model for ages, but couldn't figure out something really interesting - with good data available - to throw the technique at.
The good news is, this all means the model will definitely be back next season!
I didn't originally build the model to bet with and had only ever placed a couple of fun bets in my life until this year (you know, Exeter City to score against Man U, that sort of thing. Geddon Ciddy.) But then I lined up some simulations of past games from the model against the bookies, as a test to see how it compared. It won. So naturally, I started betting...
The first post with actual results picks in it, as opposed to just percentage chances, was the 2nd March and then we had every game until the end of the season. Richard Whittall has written an interesting post today about the psychology of forcing yourself to bet and that's exactly what I've been doing. No attempt to identify systematic errors in the bookmakers odds, or to find the best prices in the market, just pick a winner in every game and bet on it.
Here's what happened, with a £10 bet on each game*
*Technically, I wish it had. I actually ran for a few weeks at a quid a game, because I'm a coward.
Overall the model called 55% of results correctly, which is a fraction higher than back-testing suggested and would probably settle to something closer to 52% over more games.
That number at the end's nice though, right? With a £100 bank to start and betting £10 a game, three months later you've still got your original £100, plus £166.50. And plus £100 bonus for opening a new account too, if you've got any sense.
Loads of development to come over the summer. But for now, the drinks are on Wallpapering Fog.
From a personal point of view, it's been a massively satisfying project. If only because I've wanted to build a proper agent-based model for ages, but couldn't figure out something really interesting - with good data available - to throw the technique at.
The good news is, this all means the model will definitely be back next season!
I didn't originally build the model to bet with and had only ever placed a couple of fun bets in my life until this year (you know, Exeter City to score against Man U, that sort of thing. Geddon Ciddy.) But then I lined up some simulations of past games from the model against the bookies, as a test to see how it compared. It won. So naturally, I started betting...
The first post with actual results picks in it, as opposed to just percentage chances, was the 2nd March and then we had every game until the end of the season. Richard Whittall has written an interesting post today about the psychology of forcing yourself to bet and that's exactly what I've been doing. No attempt to identify systematic errors in the bookmakers odds, or to find the best prices in the market, just pick a winner in every game and bet on it.
Here's what happened, with a £10 bet on each game*
*Technically, I wish it had. I actually ran for a few weeks at a quid a game, because I'm a coward.
Overall the model called 55% of results correctly, which is a fraction higher than back-testing suggested and would probably settle to something closer to 52% over more games.
That number at the end's nice though, right? With a £100 bank to start and betting £10 a game, three months later you've still got your original £100, plus £166.50. And plus £100 bonus for opening a new account too, if you've got any sense.
Loads of development to come over the summer. But for now, the drinks are on Wallpapering Fog.
7 comments:
Are the bet365 odds the odds of the correct result or the prediction?
What would the balance be showing if you had only bet on games where the model gave a result with longer odds of say 2 to 1 or longer? Would that have done better overall than having bet on every game?
Odds shown are on the actual result, not my prediction.
Not sure on the longer odds thing - I need to test a load of different strategies now the season's finished.
Have you heard of the Kelly Criterion?
http://en.wikipedia.org/wiki/Kelly_criterion
What would have happened if you were betting with the amounts suggested by it from the beginning until now?
I've only heard of it fairly recently (keep saying I'm not a gambler!)
Really interesting stuff and I'm going to run that analysis at some point this summer to work out what would have happened.
Gut feel is that I can't increase the stakes very much across a 2-3 month period without risking losing my bank.
Yeah, it would only decrease your stakes... Betting 10% of your bankroll (like you did in the start) is huge.
I'm not a gambler as well but a mathematician...
Can you please upload an Excel file with your odds, Bet365 odds, and the output?
Thanks,
Amir
Yep, if I was doing it properly, there's no way I'd stake 10% at the start like that. I had a matched deposit from the bookies, so 5% really, and it was an experiment more than a full-on strategy!
Not yet on the Excel book, sorry. There's quite a bit more work I'd like to do before making detailed data public (in an easy to work with way at least).
All the game results and past odds are here though www.football-data.co.uk/englandm.php
Interesting £105 of profit comes from draws, very rare that the bookies would ever predict a draw based on odds alone.
Post a Comment