WIP: Rating systems

algorithm

Rn - The new rating of the player (After the match) Ro - The old rating of the player (Before the match) S - The actual score/outcome of the match E - The expected score/outcome of the match K - The K-factor

Rn = Ro + K * (S - E)

The K-factor essentially represents the “Maximum” number of points that a player can gain during a match (If the players are of equal skill level, they will gain k/2 ELO points).

The n score estimation parameter represents the ELO point differential for a player that is 10x better than another player (meaning that a player with an ELO of 1400 is 10x better than a 1000 ELO player).

This means that, if a player always plays against players of equal ELO level, they would have to win n/(k/2)

K=40 for new players until they play 30 games K=20 for players with > 30 games and never had an ELO > 2400 K=10 for players with > 30 games and have had an ELO > 2400

If few matches will be played (such as American Football, with ~16 games), then the rankings should be more volatile. Higher K factors [relative to n] would be better, so a K factor from K = (2*n)/20 to K = (2*n)/25 is more appropriate. If many matches will be played (such as Baseball, with ~162 games), then the rankings should be less volatile. Lower K factors [relative to n] would be better, so a K factor from K = (2*n)/30 to K = (2*n)/45 is more appropriate.

-> we can split football into seasons and only ever need to look at the current batch - all older games can be dismissed -> can have a json file for each month and will never need a db bc it never grows too much!

resources

https://blog.mackie.io/the-elo-algorithm http://pena.lt/y/2013/01/31/understanding-elo-ratings/ http://pena.lt/y/2013/02/07/applying-elo-ratings-to-football/ http://www.glicko.net/research/acjpaper.pdf

yolo

https://www.csie.ntu.edu.tw/~cjlin/papers/online_ranking/online_journal.pdf https://www.csie.ntu.edu.tw/~cjlin/papers/online_ranking/