I've long been fascinated by football results forecasting and have spent untold hours experimenting with all kinds of strategies and ratings. The Holy Grail for me however remains elusive - but I do so enjoy the journey.
Recent back testing various incarnations of ELO ratings and similar approaches, it's apparent that the degree of influence that the most recent result is allowed to contribute, whilst critical to future accuracy is also just about impossible (for me) to pin down with any degree of certainty.
This raises a conundrum I've struggled with for some time regarding the construction of traditional rating methods, ELO included, which is as follows;
If, say, our ratings indicate a home win is highly likely. And the game finishes as a home win, that's great isn't it - just what we want.
However, almost inevitable with such calculations, the home sides rating will increase and the away team's rating will decrease.
But why? The method indicated a home win, it WAS a home win, the ratings were good - so why change 'em? The indications were that the home team were superior in this contest, the result bears this out. In such a case doesn't increasing the rating almost suggest that the home team were MORE superior than was indicated?
Which brings me around to my latest tinkering, over the past 12 months or so. Rather than putting too much emphasis on the most recent game, I'd examine all recent games as a whole to get a handle on what each team may be capable of next time out. I cobbled together my own generic algorithm then processed a group of recent games representing each team's most recent ten matches, 'evolving' ratings for each team which best fitted this group of past games. Once completed, apply these genetically evolved team ratings to the next round.
Various combinations of parameters were tested ...
Most recent 6 games, 8, 10, 12, 14 ... were all tried, the best combination of which was ALL games. Using the full set of season's results had the best outcome.
Single rating per team, or different home & away ratings? Different home and away ratings proved the better predictor.
Modelling goal differences or straight WDL. The WDL approach appears more effective here.
Also found to be productive was a slight bias toward more recent results, inasmuch being correct with a result from 1 week ago was ranked higher than one from 1 month ago. A graduated scale ensured more recent results would always be more significant.
Using data from football-data co uk I'm able to process whole seasons on a day-by-day basis (starting from when each team has played a minimum of 10 games) using all previous results from the season to evolve best-fit ratings, which are then used to forecast the result of the following day. Also tracking success/failure and a profit/loss account using both the average and best odds in the Joseph Buchdahl results files.
Results from 2016/17 season, all 5 English leagues;
number of predictions = 1988
43.0% correct predictions
+52.81 return at best odds
-50.05 return at average odds
15/16 season, 40.9% correct, +30.70 best odds (-80.95 average odds)
14/15 season, 42.8%, +66.86 best, -52.62 average
... but can't win 'em all
13/14 season, 40.9%, -81.31 best odds, -184.69 average odds.
Trials of the same process with Euro & overseas leagues have proved unprofitable.
I'm not really sure where to go from here, or whether this is worth pursuing further at all. Although the >40% strike rate and best odds profit does favourably compare with other WDL predictions I've looked at. Checking the figures I can't detect any worthwhile correlation between higher or lower HR-AR differences.
I can share the output files, but not sure what admin/mods will allow/approve of. My postings are still being individual moderated having already been banned on here for allegedly spamming with my first (innocent) post. I can even share the .exe file which will allow others to process league seasons, but sharing .exe files from unknown new bods like me is an obvious minefield!
Recent back testing various incarnations of ELO ratings and similar approaches, it's apparent that the degree of influence that the most recent result is allowed to contribute, whilst critical to future accuracy is also just about impossible (for me) to pin down with any degree of certainty.
This raises a conundrum I've struggled with for some time regarding the construction of traditional rating methods, ELO included, which is as follows;
If, say, our ratings indicate a home win is highly likely. And the game finishes as a home win, that's great isn't it - just what we want.
However, almost inevitable with such calculations, the home sides rating will increase and the away team's rating will decrease.
But why? The method indicated a home win, it WAS a home win, the ratings were good - so why change 'em? The indications were that the home team were superior in this contest, the result bears this out. In such a case doesn't increasing the rating almost suggest that the home team were MORE superior than was indicated?
Which brings me around to my latest tinkering, over the past 12 months or so. Rather than putting too much emphasis on the most recent game, I'd examine all recent games as a whole to get a handle on what each team may be capable of next time out. I cobbled together my own generic algorithm then processed a group of recent games representing each team's most recent ten matches, 'evolving' ratings for each team which best fitted this group of past games. Once completed, apply these genetically evolved team ratings to the next round.
Various combinations of parameters were tested ...
Most recent 6 games, 8, 10, 12, 14 ... were all tried, the best combination of which was ALL games. Using the full set of season's results had the best outcome.
Single rating per team, or different home & away ratings? Different home and away ratings proved the better predictor.
Modelling goal differences or straight WDL. The WDL approach appears more effective here.
Also found to be productive was a slight bias toward more recent results, inasmuch being correct with a result from 1 week ago was ranked higher than one from 1 month ago. A graduated scale ensured more recent results would always be more significant.
Using data from football-data co uk I'm able to process whole seasons on a day-by-day basis (starting from when each team has played a minimum of 10 games) using all previous results from the season to evolve best-fit ratings, which are then used to forecast the result of the following day. Also tracking success/failure and a profit/loss account using both the average and best odds in the Joseph Buchdahl results files.
Results from 2016/17 season, all 5 English leagues;
number of predictions = 1988
43.0% correct predictions
+52.81 return at best odds
-50.05 return at average odds
15/16 season, 40.9% correct, +30.70 best odds (-80.95 average odds)
14/15 season, 42.8%, +66.86 best, -52.62 average
... but can't win 'em all
13/14 season, 40.9%, -81.31 best odds, -184.69 average odds.
Trials of the same process with Euro & overseas leagues have proved unprofitable.
I'm not really sure where to go from here, or whether this is worth pursuing further at all. Although the >40% strike rate and best odds profit does favourably compare with other WDL predictions I've looked at. Checking the figures I can't detect any worthwhile correlation between higher or lower HR-AR differences.
I can share the output files, but not sure what admin/mods will allow/approve of. My postings are still being individual moderated having already been banned on here for allegedly spamming with my first (innocent) post. I can even share the .exe file which will allow others to process league seasons, but sharing .exe files from unknown new bods like me is an obvious minefield!