In the past I already posted some summaries in my pick history for different models. So everybody could get an impression, how a real life betting using my models could look like and to test, whether the profit, indicated by the backtesting, can also be reached in the future. With this post I want to start such a series again for my ML Poisson model and additionally compare it to the performance of the other models. So let’s start…
As already mentioned in my last post, I will use the fair Bet365 odds for the profit calculation, as the bookie margin has a big influence on you value and it’s basically impossible to beat a margin of 5% or more in the football markets.
Since the start of the season the ML Poisson model shows a bad performance, when simulating a 1 unit flat stack blind betting strategy. Ignoring Draw-bets and the Seria A, as these are 2 weak spots of my model, we got an average profit of -5%. The Vanilla xG Poisson model looks similar. Both models are based on xG data. Suprisingly the old ZIP Poisson model shows the best season performance, although the model is based on goal data. There we got a current profit of crazy high 13,2%.
But goal data contains way more noise. So over the long term, I expect less profit between 3%-4%. For both xG models the backtesting indicated a higher profit. But that’s how predictive analytics works. You are just training or creating a model from past data. Nobody is able to know, what will happen in the future and whether a model will keep the performance from the past.
My real life betting experience using the ML Poisson model looks really similar to the simulation. Using a 2€ flat stack (I am betting just for fun đŸ˜‰ ) provides a yield of -4,15% over a course of 221 bets. All the bets were settled using the best odds at a broker. My loss is not as high as in theory, as I of course do no blind betting. Some bets I skipped, if the value is just too small as well as I used Asian Handycap markets to increase the hit rate of my bets. But this did not change the trend, which was already indicated by the proft simulation. Let’s hope for the best and a change of this trend. Currently the sample size is relatively small.
If you have further questions, feel free to leave a comment or contact me @Mo_Nbg.
Beat the Bookie 
Did you try creating an ensemble model by running XG Boost on all the separate models? I am sure all the models contain both signal and noise to varying degree. The goal model will capture something of the team’s ability to finish off high leverage situations. A shots model captures some of the work rate and the XG model the overall quality of build up play. Apologies if you already tried this in a previous post.
LikeLike
No, I did not yet tested an ensemble model. To use such different features as xG, Shots and more other stats, I created the ML Poisson model, which is based on a Random Forest. So I am able to use multiple features in comparison to the classic Poisson implementation, which it limited to either usage of goal or xG data.
LikeLike