search

larsvancutsem / piratings

mirror for the R package "piratings" https://cran.r-project.org/web/packages/piratings/index.html, calculate dynamic performance ratings for association football teams based on the paper of Constantinou and Fenton which can be found [here] (http://www.constantinou.info/downloads/papers/pi-ratings.pdf)
13 stars 2 forks source link

Learning rate for lambda and gamma #3

Closed NazimR closed 5 years ago

NazimR commented 5 years ago

Hi,

Firstly, I would like to thank you for this awesome package and I really appreciate it :). Im try to recreate the pi-rating from work of Constantinou and Fenton (2012). They use 20 EPL season (1992/1993 to 2011/2012). The first 5 EPL season (1992/1993 to 1996/1997) are for generating the initial ratings for the competing teams, 10 season (1997/1998 to 2006/2007) to learn the parameters lambda and gamma and the last 5 EPL season (2007/2008 to 2011/2012) for prediction.

The question is when I run the optimization of lambda and gamma, why the result that I obtained is not exactly the same as the value of minimum squared error of expected goal difference observed in Constantinou and Fenton (2012) using the 10 EPL season (1997/1998 to 2006/2007) which I extracted from same source (Football-Data.co.uk)?.

Constantinou and Fenton (2012) result: lambda = 0.035. gamma = 0.7. error squared = 2.6247.

My result: lambda = 0.045. gamma = 0.45. error squared = 2.555431.

I really appreciate your help and thanks in advance, Regards. NazimR

larsvancutsem commented 5 years ago

Hi NazimR,

First of all, thank you for liking the package and, thanks for bringing this discrepancy to my attention. Although I haven't been able to validate the code extensively on the data used by Constantinou, the code should accurately reflect the approach taken in the paper. A comparison of squared goal differences might yield a clue to where the discrepancy might be. I did however test predictive power of the pi-ratings created using this package & they were statistically significant (p<0.0001) and yielded better results than elo-ratings in various cases.

However, I will test it using Constantinou's data and, update the package as soon as I find a discrepancy between any of the plots.

Kind regards, Lars

NazimR commented 5 years ago

Hi Lars,

Thank you for your response. Here, I attach the data that I extracted for your reference and thank you again for this awesome package :). RAW - EDITED.zip

Kind regards, NazimR