Win a copy of OCP Java SE 8 Programmer II Exam Study Guide this week in the OCP forum!
  • Post Reply Bookmark Topic Watch Topic
  • New Topic

PSG vs Real Madrid (football)  RSS feed

 
Ranch Hand
Posts: 449
7
Android Open BSD Slackware
  • Mark post as helpful
  • send pies
  • Quote
  • Report post to moderator
While I was seeing the football this evening, I had a huge doubt, and started to panicking. So far in the ANN(artificial neural networks), Classification and regressions  I met as newbe consider some dependent variables, but not the time when they appear.

Let's say I want to predict next week who between Real Madrid and PSG will win the football match with an ANN, so that I have a dataset with all the matches won by the two teams. But since the physical shape of the two teams is really important, the last matches in time order are more representative than the one played long time ago( although I am going to use them)

I know that possibly this is a really complex model , but which K words i can find on the internet to understand which studies should I follow to reply to the next 2 questions?



1) How can I ponder the model so that more I go far in the time and less determinant is the result. I looked in the internet and cannot find nothing, There is a thing called Poisson distribution, but does not seems what I want to do, I was even thinking at a Markov Chain model



2)More important, how can I on the same time ponder/weight all the games the two teams play, basing my predictions with another weight, a coefficient of difficulty of every single match, on the base of the weekly ranking of the teams. Example I will consider more easy for the PSG  the game played with the last team in the ranking that week, than  the match they played against the first one. I could even make things worst, adding to my coefficient of difficulty also the number of goals they scored as a positive thing, and make this coefficient diminish if the number of goals conceded is low.
 
Ranch Hand
Posts: 966
9
IBM DB2 Java Netbeans IDE Spring
  • Mark post as helpful
  • send pies
  • Quote
  • Report post to moderator
Interesting question. My humble opinion is that no ANN or other deep learning algorithms are really needed to try to predict the outcome of a football match. What do you need, generally speaking, is a way to measure how much a football team is strong in a given moment of the championship, and, because I don't think that the ability to play football can be measured in absolute terms, you need a way to measure a team strength with respect to competitors. Somehow similar to ELO points used in chess. ELO points are based upon the concept that the more the difference of ELO points between two players is,  the more likely is that the player with higher ELO score will win the match. ELO score is adjusted after each match: you get an increment or a decrement of your score proportionally to the difference of your ELO and your opponent's, so that you won't get many points if you are strong and defeat a weak opponent, while you will loose more points if you are defeated by a weaker opponent.
Building an ELO-like score may be enough, and you may try to create such rating by a) assigning  to each team an initial score b) update for each team its score using the recent historical series of match outcomes.
 
Giovanni Montano
Ranch Hand
Posts: 449
7
Android Open BSD Slackware
  • Mark post as helpful
  • send pies
  • Quote
  • Report post to moderator

Claude Moore wrote:Interesting question. My humble opinion is that no ANN or other deep learning algorithms are really needed to try to predict the outcome of a football match. What do you need, generally speaking, is a way to measure how much a football team is strong in a given moment of the championship, and, because I don't think that the ability to play football can be measured in absolute terms, you need a way to measure a team strength with respect to competitors. Somehow similar to ELO points used in chess. ELO points are based upon the concept that the more the difference of ELO points between two players is,  the more likely is that the player with higher ELO score will win the match. ELO score is adjusted after each match: you get an increment or a decrement of your score proportionally to the difference of your ELO and your opponent's, so that you won't get many points if you are strong and defeat a weak opponent, while you will loose more points if you are defeated by a weaker opponent.
Building an ELO-like score may be enough, and you may try to create such rating by a) assigning  to each team an initial score b) update for each team its score using the recent historical series of match outcomes.

`
I resolved the first one aspect using an exponential move average, the second point make sense what you say about ELO, an individual score, but also I can assign some weights depending  by the position, or even better bayes theorem to enquire about the state t+1 represented by a new match
 
Don't get me started about those stupid light bulbs.
  • Post Reply Bookmark Topic Watch Topic
  • New Topic
Boost this thread!