allen29 Posted February 13, 2015 Share Posted February 13, 2015 Hello, Few months ago I started a thread on this forum about weather someone mined ClubElo probabilities to see if they can provide value bets. For the past 2 months i think, I have mined 610 games from Premier league, Championship, France Ligue 1, France Ligue2, Italy Seria A, Italy Serie B, Bundesliga, German League 2, Spanish Primera, Spanish Segunda, Belgium League, Portuguese League, Belgium League, Eredivisie, Israel League, Scottish league, Greek league. The odds I gathered from bookmakers were the maximum odd for 1X2 market from Pinacle, Marathon and Matchbook, those being my 3 bookmakers that I have an account with. The probabilities are gathered from ClubElo, but not on a continuum basis, because I have missed 1 round or 2 of games. I will try to describe my methodology briefly: 1.Determining the accuracy of the probabilities provided by Clubelo and the bookmaker (adjusted to match 100%) within my sample . Mainly saying, out of all the instances when either the bookmaker or Clubelo said there was a x% chance of that event (1,X,2) to happen, on how many occasions did that actually happen? These are the plots and average errors for the 2 competing predictions in my sample. I used 20 intervals, ranging from 1% to 99%, with the length of an interval being 5%. If I had a decent number of observation for each probability point I would have not used intervals. This is what I got: The RMSE and MAE errors were computed from the deviations from each interval mean, where there are at least 5 realizations. As you can see, the bookmaker did have a slight advantage over ClubELO, but even without applying a statistical test, you can say that it could be due to randomness or small sample, not particularly accuracy. In general, the predictions look similar when plotting them against actual realizations, so my conclusion and what I am trying to test is: When you have prediction models with similar accuracy, large deviations of predicted outcomes from each other should provide profitable betting opportunities (i.e.- value). 2. Then, I compute the % which the odds differ from each other (one could say that I should have used the difference between the probabilities, not the odds and can have a strong argument about this, as this issue confused me to). For example, if ClubElo published a HW probability of 65% - odds of 1.538 and the bookmaker odds are 2.1 (48%), the bookmaker forecast would be (2.1/1.538)-1 = 37.15% higher than ClubElo's. The next part was determining what the minimum % of difference in odds should be. I experimented with this and will post results based on different parameters. This are the results, divided for each bet type, with different parameters. Flat stakes of 1 point. One of the my main surprise came after asking the question: if there is a more than x% difference between the 2 predictions, in each way (both negative and positive deviations), does that makes the draw more likely? Well, from my sample, it seems like yes...there were 69 duplicates because if a team A classifies as overvalued(difference between odds is at least 20%), then there is a chance that team B will classify as undervalued (difference between odds is If you had backed all the draws, the result would have been: http://tinypic.com/?t=postupload So, the % of draws in my sample of 212 is 31.13% compared to 27.54% (bet on all draws). But of course, you must always look with a pinch of salt at numbers like that and leave room for the possibility that you may have been lucky. Of course, a much larger sample is needed and time to pass for this to be more than just a lucky guess. I will continue to gather ClubElo probabilities and test them and see if this pattern still continues. harraroni 1 Quote Link to comment Share on other sites More sharing options...
methuselah Posted February 13, 2015 Share Posted February 13, 2015 Re: Can ClubElo predicted probabilities outline value bets? Very interesting. Look forward to further posts. Quote Link to comment Share on other sites More sharing options...
harraroni Posted October 31, 2016 Share Posted October 31, 2016 Hi, just read your post and was very interested, did you continue mining the data? Do you have the results? Also would you be able to give me the spreadsheet you used so i can do a test of my own. The results are very interesting to me. Quote Link to comment Share on other sites More sharing options...
allen29 Posted November 3, 2016 Author Share Posted November 3, 2016 (edited) Hi @harraroni. I didn't test that further, I'm using something else now. But I still have clubelo scores from clubelo. Here is an export from my database if you want to play with. It contains elo scores from some european leagues, as well as odds from oddsportal. It's not all the games played in a season, just those where I had elo scores in the database. https://www.mediafire.com/?1591l99du0iczz0 Edited November 3, 2016 by allen29 Quote Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.