> > Yes, if seen stuff on exponentially weighted moving average, and it looks > > like a reasonable way to go. > > From what I've seen though, they all deal with time decay. > > > > What I want is time decay AND games decay to be taken into account. Any > > ideas how to combine the two decay variables in the single average? > > I think Steve Simon answered that very explicitly. What is it about his > solution that does not fit?
His suggestion was fine, but rereading my last post I see I didn't make myself clear. I wondered if there were any more sophisticated ways available to combine the two variables. The two scales time and games are (potentially) very different, as Steve noted, and just adding will probably decay the bias very strongly towards one or other - days will be a greater variable than games. I know I can transform - by say taking the root of the days - but I just wondered if a more sophisticated method might yield a better result. For example, producing an estimate for each separately and combining the result in some way - a naive average, or perhaps some other way? Again perhaps it's useful to note that the two definitions of time are of course correlated. However they are not correlated across teams eg 2 games ago for one team may be 10 days, for another it may be 100. It might also be worth saying that for many cases there may only be a few prior performance figures to work on (or none at all !) eg a new team. The whole point of the exercise is to make an estimate of the *real* current performance rating of a team or individual from its/his previous performance figures which can only be estimates (in my case anyway) with the ambition of ranking teams. Again it is a given that the more *recent* a figure, the more weight that should be given to it - for example running a 4 minute mile 10 years ago should have relatively little impact on whether the athlete is capable of doing so now. I think this is a reasonable aim. Cheers. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
