this k-factor simply rescales the top end of the ELO spectrum. It simply reduces the apparent contribution of a vet (ie, someone would end up with an ELO of 2800 instead of 3000 due to this k-factor). It imposes some sort of a "soft limit" on ELO, ie while ELO can always increase, it increases less and less. This is good because I dont think that anyone would be worth that high of an ELO (a 3500-pointer is not worth a 2000 and a 1500, but he might be worth 2 1500s, and so reducing his ELO to 3000 could make sense). This also prevents people from ever increasingly raising their ELO up to a point where it doesnt actually reflect their skill, but rather the fact that they just played 2000 1-point games.
As long as the k-factor works in both directions (ELO increasing and decreasing), then it is mathematically sound.
In other words, this k-factor is mathematically sound, and seems to make sense with regards to correctly defining someone's rank. It plays no factor on convergence, but might help in the accuracy (by broadening the spectrum, which I suggested in another post...even though it only broadens the higher-part of the spectrum, it still works soundly).
ELO Reset
doesnt actually help convergence.
For example, consider a player (with high ELO; lets say k-factor = 0.25, ie he gains 1/4 of regular ELO) who plays many games during a week.
without k-factor, he would have gained 400 points.
with k-factor, he would have gained 100 points.
If this converged, he would have gained 0 points. While it appears the the k-factor helped in the convergence (100 is closer to 0 than 400), this is merely an illusion. Taking this to the extreme, putting a k-factor of 0 would be perfect, as it would bring immediate convergence. However, we both know that this makes no sense.
edit: well, in a sense, it will indirectly help convergence. By increasing the accuracy on his rank, the ELO prediction will be better and make him converge faster to 0. But this is an indirect result.
For example, consider a player (with high ELO; lets say k-factor = 0.25, ie he gains 1/4 of regular ELO) who plays many games during a week.
without k-factor, he would have gained 400 points.
with k-factor, he would have gained 100 points.
If this converged, he would have gained 0 points. While it appears the the k-factor helped in the convergence (100 is closer to 0 than 400), this is merely an illusion. Taking this to the extreme, putting a k-factor of 0 would be perfect, as it would bring immediate convergence. However, we both know that this makes no sense.
edit: well, in a sense, it will indirectly help convergence. By increasing the accuracy on his rank, the ELO prediction will be better and make him converge faster to 0. But this is an indirect result.
Last edited by tmc on Sat Sep 02, 2006 8:07 pm, edited 1 time in total.