The Future

Problems and things to correct

  1. I think the most glaring issue is that the most likely result calculators logic is not wholly sound, as I think it has some dependent variables being treated as independent and thus the probabilities are not quite correct. For example: goal difference and probability a team scores and concede the corresponding values - if GD is 1 then probability of Team A scoring 2 and conceding 0 is clearly 0 which is clearly not the same as P(A scores 2)*P(A concedes 0). The same goes as for one team to score and other team concedes, they are clearly related. 
  2. Result and score are very different, it’s much easier to say a team will win a game than to say by how much, and the current system doesn’t really reflect this at all. It gives a predicted GD but if say team A is 0.5 xG better than team B does this mean to expect team A to win or draw. It may be better to predict the result first and followed by a score based on predicted goals. 
  3. There is a pretty big assumption that the relationship between attributes and results is linear; while this may be the case there is no proof to back up this claim, or at least it hasn’t been tested. 
  4. Certain attributes still have fixed assumed relationships, like all the different type of forms - where it is assumed that all games in the past have equal weight on form. Also especially with form it may be better to look at the change in form rather than just how many points have been gained in the last 5 games as the current system ranks WWLLL and LLLWW as the same level of form.
  5. I’m not certain that the current way of testing success is the most optimal. It’s certainly the most natural and easiest to interpret, because at the end of the day you want it to get results right. However, each model will have its own cap that it can’t perform above and if it’s performing below, means the model doesn’t reflect reality very well and there’s a problem with the model. For example, in the easiest model which gives all 3 results equal weight you’d expect to get ⅓ results correct if it was a good model. If the predictor is performing significantly above or below this level then the predictor’s model doesn’t reflect how the data actually works and a new model should be considered. 
  6. Many more subtle issues which I’ll come to as I fix them. 


Solutions

  1. Tidy up the maths - which will involve basically rewriting the whole system
  2. Create Version 3 using a Bayes classifier for result and then xG scores and conceded by each team
  3. Try using a non-linear regression algorithm, to see if that improves the system 
  4. Change the attributes; try and input data which is as simple as possible into the prediction so it can assume relationships.
  5. Create a grouping of games with similar likelihoods of HW, D and HL and see if the percentages of actual results from these games matches the predicted. However, this system still has its flaws and is really testing something else, whether your model reflects the real data and the max/expected performance of the model. So while this can be used to test the performance and suitability of a model, it will only tell us how well it reflects the data, not how accurate or good the model is - how much it’s percentages match real life not how confident it is.

Comments

Post a Comment

Popular posts from this blog

The problem with xG totals

There has been several reports in the media suggesting that despite Town’s impressive first half performance against West Ham they were in fact lucky to get a point in the end due to a number of high quality chances given away. Now when you look at the xG totals from the game it would seem to back this theory up with the xG score being: 1.19-2.04. The only problem here with these totals is that they take all context out of the individual chances when summing them up, it assumes all chances are independent for the relation to hold. Take Hernandez’s double chance just after half time where he takes a shot from close range which is saved before putting a header over the bar. Now he generates an xG of 0.49 and 0.57 for these two chances respectively, which would suggest that he has a 22% chance of not scoring, 50% chance of scoring once and a 28% chance of scoring twice. Although this can’t be true because how could he have scored two goals in that time there was only 1 second between ...
Read more

Premier League Predictions 11

Image
Each week I am going to try and remember to post my ML premier league predictor's predictions for the round of fixtures and keep track of how well it is doing in terms of correct results and scores. I will also be including a track of money won and lost if each predicted result was bet on. This will be done using my v2 for now but will transition to v3 when it is complete. I am starting 10 weeks in to the season as most teams have now played 5 home and 5 away games and thus the attributes fed to the machine learning algorithm should be fully populated and give more sensible predictions. Due to the way it is set up in very early weeks attributes like form can behave weirdly if a team has 0 points and another max even if it has only been 1 game. For comparison, so far the predictor has 51 results correct of which 7 of these have been correct scores. Betting £1 on each game's predicted result would have yielded a profit of £5.01 so far not including stake. Week 11 Prediction...
Read more

Does time of shot affect affect xG?

A major decision for managers of smaller clubs, who can’t just blow opposition teams away, is whether to go out quick at the start and try and get an early lead, then defend it or to keep it tight, before throwing everything at the wall for the last period of the match to try and snatch it at the death. Both methods have their benefits; the early attack gives more time to get the goal, but there’s no definitive end point as to when to start shutting up shop if the goal can’t be found as it’s unlikely a team would be able to keep up the intensity for a whole match. The late attack gives you more defensive security, as you prioritise not conceding in the early stages but can backfire if your opposition is fitter and have more in the tank at the end. What if it was possible to tell when is best to attack based purely on xG, if it could be shown that the time in the match affected how likely a shot was to be scored? It could help to decide when to attack to get the most out of it, or wh...
Read more