Introduction

This post is to summarize the 2016 results to date (round 05) and compare this to the bookmakers predictions. Considering that round 6 is now going to be my first betting round (in lui of round 5 - purely due to the minimal time I have had to work on this), I thought it important to track how we are going to date. I am also going to discuss the results of a model re-run since the results of round 4 and 5 were less than favorable.

Summary of results to date (r05)

We observe from the summary table/s and figures below that:

• overall model accuracy/precision to date is 56%

• overall bookmaker accuracy/precision to date is 66%

• the model had higher accuracy than the bookmakers in rounds 2 and 3

• the models round 4 and 5 accuracy was below 50%

• Bookmaker precision was very poor (<30%) in round 3

• In round 5, there were 6 away team wins vs 2 home team wins

• In round 4 there were and even number of home and away team wins (4 home and 4 away)

• In all other rounds the number of home team wins outweighed the number of away team wins

• poor accuracy in round 4 is largely attributed to predictions being based on round 2 results (without knowing round 3 results) and high away team wins (4)

• poor accuracy in round’s 5 is largely attributed to high number of away team wins (6)

• round 3 contained a drawn match which is historically rare (there have only been 8 drawn games since the inception of the golden point rule representing 0.5% of all games to date)

## [1] 0.564

## [1] 0.657

Discussion of results

We can see from the results that the model performed particularly poorly through rounds 4 and 5 as compared to the bookmaker. This dramatically affected overall precision which was 56% as compared to the bookmakers 66%. After reviewing the model, I determined that the main reason for poor performance through rounds 4 and 5 is due to the high weight of the importance of feature ‘home_away’ in the model. Because the model (in most circumstances) preferentially picks home teams to win, the model suffered from a dramatic increase in away teams winning through rounds 4-5.

We also observe poor accuracy from the bookmaker in round 3 where there were a number of very close games, and a draw. As the model was likely ignoring additional feature the bookmaker was relying on in round 3 (and using home/away as a highly weighted feature) the model performed well by comparison.

Generally, I would have expected model accuracy to be increasing as rounds progress (based on historic analysis), however the importance of home/away in the model was such that it overwhelmed the importance of other features such as summary of results to date, previous match ups etc. Re-running the model with less weight (less feature importance) on home/away confirms that accuracy would be higher by boosting the importance of other features. We have seen that in 2015, the importance of home away teams was much lower than historic (2009-2014) and 2016 may turn out to be similar in this regard.

Re-modelling

Since I didn’t get an opportunity to use round 3 results for round 4 predictions I firstly wanted to re-model this to determine if round 4 accuracy would have been improved. Secondly, because I am generally not satisfied with the average precision to date, and the degradation in precision (we should actually be improving in precision) I wanted to re-model all rounds with an improved predicative model. Specifically, I am concerned that the original model/s place too much significance on the home/away feature and that (like in 2015) we may see reduced significance of this feature in 2016. Therefore I chose to ‘de-weight’ this feature so that it was not an overwhelming predictive feature in the model/s. As you will see, the new model improved average precision to date, generally improved precision across all rounds as compared to the current model, and had slightly better overall precision than the bookmaker

Financial modelling

Finally, I wanted to determine how we would have went using various staking strategies to bet on predictions to date. As stated previously one of the main reasons for choosing to wait until round 5 or 6 to place bets was due to historically poor (50-60%) model performance in rounds 1-5. So it is worth finding out if we would have lost money betting in these rounds this year. I was also interested to know if my preferred staking strategy would have mitigated losses (avoided losing bets). i am going to keep it relatively simple and just compare a fixed staking strategy to my preferred staking strategy.

Conclusion

I have shown that my strategy to wait until round 6 to make a bet was a good choice, as the original model suffered from low overall accuracy (56% to date). Despite this I have also shown that my preferred staking strategy had high accuracy (71%) and was able to mitigate most losses, however would have only broke even in terms of profit. I have also shown that my latest model would have produced much better results through rounds 1-5 (67.5% accuracy) and we would have made a profit no matter which staking strategy used. I am going to utilize the latest model for round 6 predictions and at this stage, I will utilize the risk adverse preferred staking strategy.

Appendix - Summary of all predictions to date