This is the first of the topics that I’ll be covering looking at some of the questions subscribers have asked me over this Summer. There are probably going to be half a dozen posts or so over the next 10 days covering the outstanding topics. Not all of the topics will be looked at in great depth but I’ll hopefully be able to give a view on everything.
The first topic is a question Les posed to me earlier this Summer. I know he was thinking solely about the fact that last season, my ratings got off to a dreadful start in the first month of the season. It actually goes down as the worst month the ratings have ever had in 6 seasons in the sense the losses weren’t due to the fact the selections were unlucky or they were drawing. No, the losses were due to the fact the selections were getting beat and in most cases, they were getting hammered!
The question Les asked me was ‘after what happened last season, do you think it would be better to wait until October to follow the bets this season?”.
Being honest, the topic of seasonality has been touched upon by me before but usually, it is discussing the fact that historically, the first half of the season has been better than the second half of the season. The other topic that sometimes gets discussed is whether we should follow the ratings blindly at the end of the season as there tends to be a lot of meaningless games taking place and clearly, my ratings can’t take this into account. Well, the point that Les has made is a new one and therefore, is there anything in the fact that the ratings started last season so badly?
I think to answer this, it’s probably worth touching on all of the systems. I know Les was only thinking about the established systems and in particular, the combined systems 6-21 to 8-22 but I think it’s worth looking at the results of all the systems to see if we can see any sort of pattern emerging to do with time of the year.
What I’ve done is grouped the results for each set of system sby calendar month and then grouped them by first half/second half of the season. Now, the first half/second half isn’t an equal split in all leagues and as you’ll see when we look at Europe, it’s definitely not an equal split but being honest, the number of games in each split isn’t important. The fact that I am drawing a line at the end of December and starting the second half of the season from Jan 1st is the important bit. This gives us a consistent approach to looking at seasonality and this in itself provides a big enough sample in each subset to make any conclusions we can draw meaningful.
So, to get us started, here are the results for the Established Systems:
As always, there is a lot of data to look at and compute but I’ll leave those of you reading to go away and look at the results in depth. I’ll pick out some highlights.
Without a doubt, we can see a bias towards the returns from the first half of the season. The first half of the season is about twice as profitable as the second half of the season. This isn’t a shock as I’ve touched upon it before and people should be aware of this for these systems.
Interestingly, if you work along the table, you can see the discrepancy drops as you start to cover more and more of the draw. Hence, when you get to AH0.5, there isn’t that big a difference in returns between each part of the season. This would imply that we see more draws in the second half of the season for the selections. I witnessed this first hand in the first season of live bets when backing outright made a loss of 21pts in the second half of the season but backing AH0.5 made a profit of 154pts!
Of course, those following last season will be aware that covering the draw didn’t work anywhere near as well and in the first half of last season, AH0.5 lost 2pts whereas backing outright made a profit of 146pts. Almost a complete reversal of what we saw the previous second half of the season!
In terms of the profits by month, February and April tend to be tougher months to make big profits but they remain profitable. November and December tend to be when the big profits are made and this has definitely been the case in the two live seasons to date.
August only has a handful of bets, so we can’t draw too many conclusions from this month but September looks to be an OK month. When you take into account it made a substantial loss last season (first ever time it had been a losing month), then these results are still very acceptable. October looks a trickier month than September and therefore, I would personally find it hard to say to anyone that they should skip September and start in October. Nothing in the data backs this up as a good thing to do.
Likewise, there is always a lot of talk about May and whether we should have bets in May. Well, it was profitable last season (just!), it was profitable in the first live season and looking at these results, although May will never have a lot of bets, it is the 3rd most profitable month of the season historically. Hence, I think there tends to be an overreaction to the impact that the closing games have on the systems. Yes, it’s never ideal that you are backing teams that have nothing to play for but then again, if it was simply a case of backing teams to win who needed to win, why bother playing the game and just pay us our winnings now Mr Bookie.
Overall then, to directly answer the question from Les, I can’t say that starting following the systems in October is a better idea than starting to follow in September. I’d probably go as far to say that based on the historical data, it would cost you money in the long-run.
What about the other systems in the UK? Do we see the same trends?
Here’s the results for the New Systems:
Well, the same story is true here about the fact the first half of the season is best. February and April again stand out as being difficult months and February in particular looks a very tough month for these systems. 2,274 bets is a big enough sample and it’s scary to think that has only generated 1pt profit in the month! When you think of the overall returns on the system, it can’t all be down to variance and I would suggest there may be something about that month. Covering the draw doesn’t help at all in this month too.
Interestingly, May and August stand out as the best months (August barely has any bets though) and again, it backs up my point that we shouldn’t be running scared of the end of season games. November is a stunning month again and clearly, this month is a critical month when it comes to the performance of the UK systems. Both sets of systems so far rely on November being a great month to boost profits.
Here’s the last of the UK systems. This is the Misc systems results:
A very similar trend at a high level although interestingly, if you look at last season, these systems lost in the first half of last season. I know it was down to TOX, STOY and STOZ as they started the season terribly last season. Saying that, you can’t base a lot on one season and you would still come to the conclusion the first of the season is probably a safer bet than the second half of the season.
February and April again standout as being difficult months to profit from and October looks a tricky month too. November again jumps off the page as being a good month and I’m starting to see that November is going to be an important month next season!
Overall then, that’s 4 individual rating algorithms, 3 Similar Game models and everything points to some very strong conclusions. Aside from variance playing a part, in the long-run, we would expect the first half of the season to be more profitable than the second half of the season for the UK systems. We would expect September and October to be OK months, November to be a great month, December and January to be OK, February to be a tough month, March to be an OK month, April to be a tough month and for the season to finish well in May. This is based on 6 past seasons experience, covering over 45,000 games in total. Interestingly, the live results to date have mirrored the experience to a great degree, with the first half of the season being most profitable, November being the best month and February and April being the toughest months to profit from.
Before we draw this to a close, a quick look at the European systems:
You can see clearly that 70% of the games for these systems take place after the New Year period. This will be because the leagues start later than the UK leagues in general and I think they may also have a little winter break over the Festive period in some of the leagues.
Based on the experience to date, there doesn’t appear to be any bias towards either half of the season. I wouldn’t bet it will stay like this when the systems go live as you can never be sure but at the same moment, there appears to be no seasonal bias.
I laughed to myself when I saw the results for September and June on the Euro systems. September has somehow managed to have 47 wins in 56 bets! I’d be willing to bet September loses this season when I’m following the bets. Likewise, June also has an ROI above 100% and although based on small samples, all we can say is that there doesn’t appear to be any great concern with the ratings at the start of the season or the end of the season.
I find it very interesting that February and April appear to be the two strongest months of the season! Considering everything I’ve said to date about the UK systems, it would be heaven if we found out that the Euro systems did best live when the UK systems struggled. April has a scary number of bets compared to some other months but then again, it is the most profitable month that has a decent sample of bets.
Overall, nothing stands out too much if I’m honest apart from the fact the best months for the Euro systems appear to be when the UK systems have their worst months!
Before I go, I should probably touch upon the Under/Over systems. I honestly believe that any variation we see in performance here is likely to be due to variance as there aren’t a great deal of bets most months and therefore, anything can happen I think!
Here’s the results of these systems:
Nothing stands out if I’m honest and even last season when the systems were live, the first half of the season was much like the return in the second half of the season. November appears to be the best month which gives me a little bit of comfort and if my ratings are working best in this month (as they appear to do above), then it’s nice to see it reflected in these results too. Hence, not only can the ratings determine the best value bets, they also know if the games are going to have many goals in them too!
Hopefully this post answers any concerns that people had about the impact that various times of the season may have on the results of the systems. No one knows what will happen next season but at this game, you have to make rational decisions based on the evidence in front of you. If you do that, you can have no regrets about how things turn out.
This problem can be answered by performing a Chi Square Test for independence.
ReplyDeleteFor the established systems we have the following distribution:
Observed
W L Total
1H 4162 4472 8634
2H 3679 4698 8377
7841 9170 17011
Expected
W L Total
1H 3980 4654 8634
2H 3861 4516 8377
7841 9170 17011
ChiSqr: 2.04979E-08
Under the hypothesis of independence the number of observed and expected winners
shall be approximately the same.
Here we have observed 4162 winners which is far more than the 3980 we expected.
The statistical test has a p-value of 0.00000002.
This is very, very significant.
Odds are 1 : 200 millions that this result has occurred by chance!
SecondWalz
Hi Secondwaltz.
ReplyDeleteAlways nice to see your mathematical take on things. ;) Quite like the odds you quote there and hopefully the systems continue to live up to their past performance next season!
It would be quite interesting to get your mathematical view of whether you think February’s performance for the New systems is simply variance or something more underlying based on the tests you do. I personally don’t believe it can be anything other than variance since the month can’t really affect the system performance at all but the results are quite damning compared to other months.
Cheers,
Graeme
I had a look at this so may as well add my bit.
ReplyDeleteI did a test for difference between 2 binomial proportions and found got a test statistic of -5.6. We use the normal dist to get a p-value of 1.02489E-08, so in line with SecondWalz findings.
One thing I wanted to note is that, by my calcs, we get a slightly better than average price in the second half of season (2.612 vs 2.609), and these tests completely ignore the price taken.
-Martin
Hi Martin.
ReplyDeleteThanks for the comment.
I think what you guys are doing is interesting (I know SW is a mathematician) but considering I’m not from a mathematical/statistical background myself, don’t expect a detailed reply back! ;)
Graeme