Too Many Draws

A prediction model needs Draw refinement

My database of football matches across Europe currently totals around 8,000 matches (and it is still growing as I add more leagues and seasons). One thing that has recurred with my prediction model is the problem of draws.

Anybody familiar with football modelling literature will know that representing the draw in the model output is one of the trickiest parts of the whole prediction modelling process. This is because, typically, a football model is ultimately trying to identify the best of two teams. It’s A vs B. The models don’t often account for something in the middle (in other words, a drawn match) because we’re too busy siding with one team or the other.

When poring over the outputs of my predictive model recently, it was clear that reasonably large number of predicted results were falling into the Draw bucket but without necessarily hugely over-achieving on the total Draw percentage for the entire match database. It felt like an optimisation opportunity.

Some Draws are better than others

A dig around in the Draw predictions revealed some interesting information. While a match might end up looking like a potential Draw, because the ratings for each team cancel each other out, it became clear there was a subset that is a more hardcore likely Draw and a subset that probably should belong to the Home win bucket.

• 30.8% of all original match predictions were Draw predictions
• Of this total, 28.2% of the predictions were accurate
• The overall Draw percentage for the entire match database is 25.1%, so 28.2% correct predictions is a small edge (equal to 12.4% uplift versus the population norm)

Upon closer examination of the model mechanics, it transpires that matches where Home attack is cancelled out by away defence, there is an increased likelihood of a draw. Interestingly, I don’t find the same necessarily with a home defence cancelling out an away attack.

Within the prediction engine I have built attack ratings and defence ratings which contribute to the overall final team rating. By comparing home attack to away defence, I identified an entire subset of matches that were even more likely to be a draw: namely, where a home attack rating is more or less cancelled out by a similar away defence rating.

Conversely, where home attack appeared primed to overwhelm away defence, despite being in the Draw bucket, these team were instead moved into the Home win bucket.

The results of this shift are useful indeed:

• Fewer teams in the Draw bucket increases Draw prediction accuracy to 30.6% (from 28.2%)
• This is an uplift of 21.9% from the population norm,. This is a meaningful change

Another noteworthy aspect of this model change is the improvement to profitability. While this prediction model is not necessarily being created for betting purposes, it’s reasonable to say that bookmaker odds are a good yardstick for accuracy because they are a representation of market sentiment. If every man and his dog thinks Team A should be a hot favourite, for example, the wisdom of the crowd is usually correct.

So, using that logic and referring to bookmaker odds as a confidence signal, we can assume the changes to the Draw bucket are positive. The model tweaks make the entire model around 33% more profitable, working on a level stakes basis.

What did I learn?

The fact I was able to find optimisations within the Draw bucket is good. The teams that moved from the Draw bucket to the Home win bucket materially improved the profitability of the entire model. It also suggests that I probably didn’t build enough home advantage into the original model in the first place, so this remains an area of research for me to work with.

The model is now achieving a ‘Correct’ prediction strike rate of 49.4% using data from across five different countries. Fingers crossed that, with a bit more digging, I can get over that magic 50% threshold.