Monday 1 July 2013

Introduction to the TFA Value Ratings

As many of you reading this will be aware, one of my big tasks this Summer was to incorporate the TFA value ratings into my results sheets and consequently, to start releasing the TFA value ratings with my bets each week.  Before I start discussing the results of this project, I think I should make it clear how much work, time and effort has gone into this. To say it hasn’t been easy is a small understatement!

Basically, my value ratings are only one of 16 ratings I generate for each game. Ultimately, in terms of importance, the value ratings would be the 3rd or 4th most important rating factor I generate as I have better indicators within my ratings which aren’t actually linked to the price of the selection.  The value rating is generated as a combination of some of the other rating factors I’m using but on its own, I’ve never really thought the value rating was too powerful.

My issue with doing this project has been a simple one. My base ratings sheets I use each week to generate the ratings before transferring the data to the next set of sheets to generate the systems contain the value ratings. However, when I generate the system bets, the value ratings (and all the other ratings) just drop out of the process and the only thing that remains are the odds as I obviously need the odds to work out whether or not the bet appears on a particular system as many of the systems have odds restrictions.

Going back and trying to tie up my system results with the value ratings doesn’t sound like it should be a hard task but unfortunately, given the fact that teams name change between datasets (due to human input in places), dates of games change sometimes and the fact that lookups in Excel have plenty of limitations when it comes to big datasets, it’s not been easy to get the data to tie up. Stupid stuff like me using the first 4 characters of a team’s name in a lot of my data but when it comes to attaching value ratings, I was finding lots of crazy numbers thrown up and it was linked to the fact that when Dundee played Dundee United last season for example, my lookup was recognising Utd as the value bet and given I was backing them at 7/1 and they were odds on, it was saying they were massive value! Obviously, Dundee were the bet.

I’ve tried my best to cleanse the data the best I can but I’ll be honest and say I wouldn’t be surprised if there were a few gremlins within the data. When you are dealing with 7 seasons worth of data and human input into all of the live results, it’s easy to see how things can go awry at times!

In terms of data, I have decided to only use the last 3 years worth of data.  I could have went back further but this would be using 100% backtested/backfitted data and therefore, I was determined to try to stick to live results where possible. Obviously, I have ratings with different periods of live results but anyone following the service will be aware of how long each system has been alive.

In terms of systems, the value ratings are available for systems 6,7,8,21,22,31,32,33,41 and 42. Obviously, 6,7,8 have the same value ratings, 21 and 22 have the same value ratings, 31,32 and 33 have the same value ratings and 41 and 42 have the same value ratings. 4 sets of value ratings, one set for each algorithm.

As soon as the service closes this Summer (will be 8th July), I will release this additional dataset to members so they can do their own analysis. I have also spent a lot of time over the last 2 weeks building the new TFA bet sheet for the upcoming season and incorporating in the value ratings and the changes I promised last season, so hopefully subscribers will be impressed with the new sheet and the increased user friendly nature of it. I will also send this out as soon as the service closes.

Rather than analyse each algorithm independently on the blog, I’m just going to look at the overall results of all 4 algorithms combined. Obviously, people tend to follow different systems from each algorithm and therefore, they will need to analyse the results by algorithm if they wish to do so. 

One limitation with this analysis is that there is no right or wrong way to look at the value ratings. There are unique ratings for every game here and therefore, I need to group them in some way to get some meaningful analysis. I have chosen to look at the ratings in 7 groups.

The groups are <5% value, 5%-10%, 10%-15%, 15%-20%, 20%-25%, 25%-30% and 30%+.  Hence, I’ve used 7 distinct groups. There is no right answer here and some of the conclusions may change slightly if you look at the ratings in different grouped sets but I would think the overall conclusions would be similar no matter how you look at the ratings.

Anyway, enough of the caveats…..

Here’s the overall results for all the games over the 3 seasons:


Some people will find this amazing reading this comment but this is the first time I’ve ever looked at the performance of my value ratings like this. I’ve never really been interested in the value ratings until now as I didn’t have the data in this format to allow me to do any analysis but I find these results fascinating.

I really like the fact that as the value increases, the ROI increases. Nothing too surprising as if you look at the strike rates, it tends to be the case that the strike rate reduces as the value increases and therefore, the ROI increases likewise. This is basically saying that my ratings find the most value on bigger priced selections and anyone who understands my results over the last few seasons will be aware that this is the case.

I think the interesting aspect to all of this work I’ve been doing is the interaction between the systems and the value ratings. As most of you will be aware, as the TFA systems filter the bets as you move up systems, the average odds reduce. Therefore, seeing the strike rate reduce and the average odds increase as the TFA value % increases is an interesting phenomenon I suspect for many.

One point I’ve always made clear from the start of this footie project is that the systems don’t use the value ratings when filtering the bets. Therefore, it’s not totally surprising this occurs but I suspect the interesting thing we probably need to look at is what happens if we include the TFA value % into the filtering of the systems.  For example, if we were to remove all value bets less than 10% from systems 7,8 and 22, what does this do to the performance of systems 7-21 thru to 8-22? 

Anyway, park that thought for the moment and we can revisit it later this Summer. It’s the sort of thinking that I want all of you reading this to be doing now going forward. I’m giving us another tool to use and it’s up to us how we use it.

Getting back to the table above then, I think there is maybe an argument for missing out all system bets where the value is 5% of less. However, a profit of 3.7% from 2,280 bets is still a decent profit and at this game, I’m not one to be cutting out bets and reducing turnover if there appears to be an edge.  I would need to look at the bets in a bit more detail before making this sort of decision myself.

Here’s the performance by season:


 I think this table highlights the issue with just deciding to cut out chunks of bets based on small samples of data. After the first season, you would have probably dropped all bets below 20% value and think you’ve got this game cracked. 

In the second season though, the same bets were nowhere near as good and the smaller value bets performed much better. Last season, the lowest value bets did poorly and created a hefty loss over the season but the 5%-15% bets had a great season. I know myself that there is a thin line between a bet that may be 4.99% value and a bet at 5.01% value and therefore, it’s just variance that these results have been thrown up I suspect.

Last season, the very best bets did really well and interestingly, this wasn’t the case with the TFA systems where the very best bets struggled last season! Again, quite an interesting phenomenon and gives myself (and probably everyone reading this) some food for thought with regards defining the best bets that my ratings throw up. Until now, it would be the bets on the higher systems but knowing these aren’t necessarily the highest value bets means it’s probably not clear cut.

Here’s the split by Home and Away:


Now, I do find this quite interesting and this backs up my own thoughts about my ratings. The low value home bets appear to be unprofitable.  I’m not surprised by this at all as I think it’s harder to find an edge with Home bets and I think my ratings are really good at pinpointing some strong Home bets but when it comes to the weaker bets, I do think the home bets struggle a little. In addition, finding value with odds on home bets is much more difficult and I’m not surprised to see the average odds being lower and the fact a loss is made.

I already try my best to not play too many low odds bets but at least we have a quick filter we can apply if we wish to weed out these unprofitable bets.

Here’s the same results split by Season:


I think this points out what we already know in the sense that homes struggled last season. However, by looking at this, you can see it was actually only the low value homes that struggled. When my ratings were finding a fair bit of value, the home bets were doing OK.

The aways were hard to make out last season. The big value aways made massive profits but we know the aways were easy to make money last season as there was a freakish number of away winners at good odds, so my ratings caught this benefit last season. The aways with less than 5% value made a loss but again, the 5%-10% had an amazing season!

Hard to draw any conclusions over a season in these groups as there isn’t enough data and last season is probably not going to indicative going forward as it was a strange season with a low number of Home wins and a high number of Away wins.

Here’s the results by League:


I’ll just pick out some interesting observations here which will give us some ideas for going forward.

  • Bsq Prem bets where there is less than 25% of value are only break-even over the last 3 seasons.
  • Bsq Prem bets with 25%+ value have an ROI of 55%+!
  • League One bets with an ROI of 5% or less are break-even.
  • League Two’s highest value bets are heavily loss making. This makes perfect sense as the systems have always struggled in League Two and we now know why! 
  • The lower value bets (<15%) in the Premiership are loss making.
  • The lowest value bets in the SPL are better than the highest value bets.

I think having this new data available adds another dimension to the TFA service and as I said before, I think this is maybe the missing piece of the jigsaw for us.  I think it’s important that we learn more about these ratings before we all dive in and start adjusting our staking plans or systems to follow

Here’s the results split by League and Home and Away:


Again, there are lots of interesting little snippets of info here but in isolation, with the very small data samples, not sure we can pay too much attention to these results.

OK, I think that’s enough of an introduction to the TFA value ratings for the time being. There is a lot of new information available in this post and I suggest those following next season take some time to read this and try to get the juices flowing around what this means about your portfolio of TFA bets going forward.

The service is now open for new subscriptions although I will close the service next Monday (July 8th). Looking this morning, there are only a handful of places still available although to give as many people as fair a chance as I can, I will try to keep the service open until next Monday. I would rather have a few extra members who are keen to follow the service and have shown their commitment by joining in a week window as even after last season, I lost more members than I thought which is amazing considering the season I had! 

I’m having a little break away from things this week (after today) to try to recharge my batteries after the last few weeks and the intense workload this value ratings project has brought.  The next step is to finish off building the new European ratings for next season. Going to try to finish off these today as I’ve been working on it when I got bored doing data analysis from the value ratings project.  Not got too much more to do although I’m still not sure how I plan on sharing the results on the blog as I don’t want people following the Euro systems too closely and showing amazing backtested results I don’t really believe in is maybe not a great idea! 

2 comments:

  1. Great work Graeme. Not much experience with football betting but plenty with value filtering with horse racing and if it were me I would ONLY be betting those at >30% for more selectivity and to maximise ROI. Would have liked to see breakdown of Value %s by each system both overall and over each season to see if any specific system(s) were responsible for the higher ROI to give even more selectivity. Also A/E and Chisq%'s for each system would be most revealing and may even convince a football betting 'virgin' like me to sign up :-)

    ReplyDelete
  2. Hi Dave.

    Thanks for the comment. The data you need to do these tests is freely available in Excel downloads on the blog. You can get the data here. http://the-football-analyst.blogspot.co.uk/2013/06/season-review-sheets-to-download.html

    Looking to maximise ROI is maybe the biggest mistake that most football punters make I think and anyone who’s been following the service over the last few seasons won’t be agreeing with your comment above mate! I definitely don't and I think it's a typical trap newbie football bettors would fall into.

    What you really want to do is maximise your ROC over the season and ultimately, as I’ve shown a lot on the blog over the past 18 months, this won’t be achieved by aiming for a high ROI. The higher the ROI, the lower the ROC generally as to achieve a high ROI, you need to filter the bets accordingly and by doing this, you reduce the target ROC. Not only that, by filtering the bets too much, you fall into the trap of taking something that is statistically significant and then running the risk of making it statistically insignificant or backfitted which can lead to even more problems.

    Turnover and maximising ROC is the key to successful football betting and therefore, ROI is irrelevant in many ways. I’ve shown this before on the blog through numerous examples. ROI for show, ROC for dough was my term (I should have trademarked it!) that has now been used by nearly everyone in the industry.

    I think anyone joining TFA now (bearing in mind they are 3 seasons behind most others) will learn a lot and it will hopefully open up their minds to a different way of betting.

    Being honest, I’m not sure TFA is for football virgins unfortunately Dave. 3 years ago it was but I’ve taken a lot of people on a journey with me and not sure I have the appetite to continually start people off again at a beginner level with TFA. Happy for people to join and do their own reading but I suspect TFA isn’t for anyone who doesn’t want to really invest the time and effort into understanding how the systems work.

    Cheers for the comment.

    Graeme

    ReplyDelete