In this article I want to demonstrate why I believe using the 4 football models will generate a profit in the long run. After one month of betting some of you might be wondering why I am still trusting value suggestions from a model like ECPO for instance which has been consistently losing money in its first 33 bets… Hopefully this should complement this earlier post and help clarify my thought process.
The evaluation exercise I will undertake here for each model is the same as in last week’s article, where we take 10 seasons of past results (2005-2015) and repeatedly go through the following steps:
- Set season i aside for evaluation (with i varying from 1 to 10)
- Build a model using the other 9 seasons
- Use the model to simulate betting odds for the entire season i
- Compare modeled odds to bookies’ historical odds and bet on every game where the model sees value with a stake following the Kelly criteria.
This process allows assessment of the model behavior with data it has not seen during its building phase and by having it run through all the games in a given season we get a feel for potential losing and winning streaks we might have to face.
Let’s start with the model currently performing the worst in this first month of activity then:
As explained previously each line represents the state of our bank throughout an entire season, assuming we start with 500$. First we can see that for 6 of the 10 seasons the model would have allowed us to finish with a profit (final bank over 500$). It is also clear that huge variations occur during the season and important losing phases of ~50 bets are very likely to occur at some point during any season. Fortunately however for most seasons these are balanced by equally long if not longer winning phases.
One key point behind my idea to use multiple models covering different leagues and different markets is to try and level up these losing phases with (mostly) uncorrelated models. We will now look at our second worst performing model so far: EPLO.
Once again we can see that more often than not the season long balance would be in profit (7 out of 10 seasons) and although the model seems to endure less peaks and troughs some long losing phases are still to be expected.
Next one up is EPLH which also operates in the premier league but in the home win market instead of the over / under like all the other 3.
In this case it is important to note that odds can be much larger (eg. up to 10) than in the over/under market where they rarely occur outside the 1.5 – 3.0 odd range. This explains the steeper ups and downs that can be observed above. Unlike for the first two models we here see the first significant season loss (2012-2013) where we would have almost gone bankrupt; we also have one huge profitable season (2005-2006) where the bank would have more than doubled. Overall 9 of the 10 season would have registered a profit.
Finally the performance of our current leader SP1O is shown below:
Here again the majority of seasons (8 out of 10) would have been profitable but one season (2005-2006) would have actually thrown us into bankrupt.
Looking at these it seems pretty clear to me that no model should be judged on its short term performance. Massive ups and downs are to be expected throughout a season and a lot of patience is needed. As mentioned above this variability is one of the key reasons I chose to operate with 4 models covering different leagues and markets. Here is a summary of how the end of season bank would have looked if using the 4 models together:
As this orange/brown curve shows every single season would have generated a profit if all 4 models had been used together, and for the seasons where SP1O or EPLH had registered important losses we could have counted on the other models to balance the sheet.