0

I've going on a little competition with friends: We're writing models for the upcoming worldcup to see whoms model gets the most points out of a tip game.

So my approach would be to write a neural network and train it with previous worldcup results regarding to the anticipated wining rates (back then), to maximize the score of the tip game (e.g. 6 points for the exact score, 4 correct goal difference, 3 correct winner).

Scrap rates from various sites (bwin etc.) and let the network tip for me.

I'm familiar with linear algebra, probability calculus etc., but have never programmed a neural network yet.

Since I have not much time left, could someone help me by picking the best approach (like which concept/algorithm I should use) or link me a tutorial to a similar problem or the approach that would fit?

Best, Hannes

Gerard Júlio
  • 55
  • 7
  • http://cs229.stanford.edu/proj2014/Ben%20Ulmer,%20Matt%20Fernandez,%20Predicting%20Soccer%20Results%20in%20the%20English%20Premier%20League.pdf : search for "machine learning model to predict soccer" - it's too broad for SO – Mitch Wheat May 17 '18 at 03:03
  • 2
    Welcome. Please see the [How do I Ask a Good Question?](https://stackoverflow.com/help/how-to-ask) page to see how you can improve your chances of getting meaningful help here. Good luck :) Voted to close as too broad and/or requesting off-site resources. – code_dredd May 17 '18 at 03:06
  • I am totaly aware that the question is very broad, I just want some ideas/input of other users to my question. Since I don't know exactly where to start, everyones opinion could help and searching won't help much. I just need a direction. – Gerard Júlio May 17 '18 at 03:43
  • @GerardJúlio I totally understand and can relate to it, but questions that are overly broad are too difficult and impractical for anyone to even attempt to answer in a meaningful way. Consider [this](http://meta.stackoverflow.com/q/284236) meta post. – code_dredd May 18 '18 at 16:44
  • @ray but I already have a very interesting answer and the thread was closed shortly afterwards because of YOU. That makes me think that you are/were wrong – Gerard Júlio May 18 '18 at 22:29
  • @GerardJúlio Hi. I don't have time for your accusatory attitude. I tried to help you by pointing you in the _right_ direction, but I see you're not here to listen yet. So, I'll just reiterate that you should take the time to _read and understand_ the linked post on how to ask _good_ questions in SO. The fact that an answer was posted does not make this a good question. Lastly, this was not closed because of me; it was closed because a large enough number of higher-rep users also _agreed_ with my assessment. Since SO is a Q&A site, not a discussion forum, you need to be _specific_. Good luck. – code_dredd May 18 '18 at 22:53
  • @ray thanks for your advise, reporting people on subjective opinions won't help anybody though. Sadly you did not help me at all, I still have the opposite feeling – Gerard Júlio May 18 '18 at 23:22
  • @GerardJúlio As you yourself recognize, your feelings are just as subjective as mine. In my admittedly subjective opinion, this simply needed to be narrowed down. If votes from other users showed anything, it's that while my opinion is clearly subjective, it appears to be _less subjective_, and thus _closer_ to objective reality, than yours is. You can look at [this question](https://stackoverflow.com/q/605845/4594973) to get a better idea of what a "good" question looks like, etc. – code_dredd May 19 '18 at 02:23
  • @ray I guess the fact that the first meaningfull answer showed up after approx. 5min, shows that your subjective opinion was not really close to the reality.. To close the post you just need one vote and an admin that has the same opinion as you do. If you really wanted to help, you could just have posted your opnion without voting for anything... – Gerard Júlio May 19 '18 at 04:38
  • @GerardJúlio Your comment shows you don't understand the facts of how the site works; your attitude won't get you very far either. Go look at other questions that have been closed down and see how many actually even bother to explain why the question was closed or stick around trying to help people improve their future chances like I tried to do here with you... I'm done with this conversation; trying to speak some sense to you is a complete waste of my time. This is not the youtube comment section and I don't have time for it. Best regards. – code_dredd May 19 '18 at 04:52

1 Answers1

1

I did this for the 2017 UK Premier league season. However I accumulated the first 19 (out of 38 Games), in order to try and help with my predictions.

I would attack this in the following manner

  • Get Data on the Teams (What you consider data I will leave up to you)
  • Get the History of previous matches (Personally I think this is not going to help, as teams change so much)
  • Python
    • Pandas
      • Create New Features
    • Keras
    • Model away

When I was playing with the UK Premier league, got a prediction accuracy of aprox 62%. So how to make it better ?

  • Distance travelled
    • Improved it by 1.2%. Some teams do not like traveling it seems.
  • Weather
    • I got the weather forecast per ground @ Kick-Off-Time (What a pain)
    • Accuracy improved by 0.5%

What I did not do - was get a list of Players who were playing per game, per day. And attribute things like distance run, passing percentage, goals, fouls, yellow cards etc

Tim Seed
  • 5,119
  • 2
  • 30
  • 26
  • Sounds cool! Thx for the hint with Pandas and Keras (prob also going to use Tensorflow)! So I would just take the rates from betting sites, try to convert them into scores (e.g. 1:2) and try to maximize the score of the game (e.g. for the correct result you get 6 points, correct tendency 3 points, etc.) – Gerard Júlio May 17 '18 at 03:30
  • Interesting! If you accumulated the first 19 games and then applied your ML on the data, did you consider using something like Bayesian? Although the dataset may be too small for that. – Haris Nadeem May 17 '18 at 03:43
  • I did not consider using Baysian, mainly as I was trying to use Keras/ML and I thought this would have been an interesting project. I fear the world cup dataset will be too small, and that any ML will be subject to massive over fitting. – Tim Seed May 17 '18 at 08:01