The next 10 match days are played in the German Bundesliga and so it is time for the next summary of the current betting history of my Poisson prediction model. During my summary about the first 10 match days I faced a really big loss. This trend continued also for the next match days, but some interesting observations can be made.
Before I start, I have to admit, that there was an error during my last analysis of match days 1 to 10. I used the wrong filter for my simulations of the past matches. During the optimisation of the predictive model, I defined to select a bet as an underestimated bet, if it offers a value >= 0,2. But for the first summary I used a filter for value > 0,2. This was a small error, with a big difference. Without the error match days 1 to 10 provided a loss of -201,1 units instead of -256,1 units.
As already mention the next 10 match days in summary also resulted in a loss. We can see a minus of -182,6 units, which leads to an overall loss of -383,7 units. In comparison to the last 4 years, this looks worse and worse. Every past year provided a positive profit at this stage of the season. The best season (2014/15) provided a profit of 439,1 units.
But this time I don’t want to analyse, which bets caused such a big loose. I also already mentioned the weak points of my Poisson model. Now I want to draw your attention to an interesting point of the current betting history. Until match day 12 there is a steady downward trend, which results in a minus of -328,9 units. After match day 12 the continues trend stops. Between match day 13 and 20 there is still a minus of -54,8 units, but there are more ups and downs. This is more evident, if you just take a look at these specific match days. The profit-loss graph jumps between the positive and negative area. This indicates, that since match day 13 the Poisson prediction model is now more succesful in recognising underestimated bets.
There are 2 possible explanations for such a behaviour. At first, the attack & defence strength should now be more adapted to the team performance of the current season. As more games are played in the current season, the average is more influenced by the performance of the current season than past seasons. Again, using a weight factor for the predictive variables should help to faster adapt to the current performance. The second explanation could be the something, which is called “regression to the mean”. For the first 10 match days I recognized, that Cologne caused a huge amount of the whole loss, because until then they have played well below their possibilities. So they hugely underperformed. Taking a look at the past bets, Cologne caused a profit. Since the change of the manager Cologne’s performance has increased. So they converge back to their mean performance. That is called “regression to the mean”. If a team hugely under- or overperformes, you can expect an opposed phase, as a team performance fluctuates around the average team performance.
Of course there is also a simpler explanation for such a behaviour: luck and bad luck. This is something, which I still was not able to clarify, whether my predictive model in the past just performed because of luck, or whether I really can expect a steady profit.
The increasing performance of my predictive model can also been seen at my current Kicktipp group and the published bets at Pyckio.
As every year we started a Kicktipp group at my company. To see, whether my predictive model is able to compete with the human estimations, I take part with two accounts. For both accounts I use one of the simulated Kicktipp strategies. My main account “Andre” is currently at the first place. This account uses an adapted 2:1 model. But more interesting is the BeatTheBookie account, which uses the Expected Goals model. A long time the model was not able to correctly estimate the results of the single matches. It was often placed at the last or second last place for a single match day. But since match day 10 you can see an increasing performance. It gained a better and better ranking. The last 5 match days it was able to gain the first rank 2 times and the second rank 3 times.
A similar increasing performance can be seen at my Pyckio betting history. There was a dramatic loss until matchday 10 (Nov 18). After that, you can see the similar behaviour as my model simulation. The continues loss ended and some up and down started. In contrast to my simulation, the Pyckio betting history offers a profit of about 20 units the last 10 match days.
The difference to my simulation is just simple: Pyckio offers Pinacle odds, whereas I use Bet365 odds for the simulation. The second reason is, that Pyckio does not offer all markets, I use for the simulation, so that I am not always able to place every suggested bet.
As the persistent loss has now ended, I am excited, how my model will behave the next part of the season. But I am still not really optimistic, that the season will end with a profit.
If you have further questions, feel free to leave a comment or contact me @Mo_Nbg.