I don't use exponential smoothing in estimating player/team skill
level (i.e., point-win proportions in my system), prefering the
long-term estimate. Tiger and I have discussed this in the past.
Suffice to say I agree philosophically with J.A. Tech in the sense
that it well may be self-defeating. The money boys at the simulcasts
and in the fronton know who is hot; better to try to find someone
who is "cold" and could be undervalued given their "true"
skill level.
I' would like to hear more about your empirical analysis that
resulted in the choice of a .10 alpha, El Tigre. While I am a
fan of empiricism, I was trained to put theory first. And if I
am simulating a Yanks/Sox game circa 1948, Ted Williams is going
into the mix as a .340 hitter whether he was 0 for 5 or 4 for
5 the day before.
I do have an elaborate way of estimating team point-win percentages.
It involves coming up with an estimate of the theoretical skill
level of the team using the skill estimates of the individual
players via a regression formula I developed, and combining that
with the actual performance data of the team. Bascially, as more
observations come in, the theoretical estimate is weighed less
and the actual performance data more. It is especially useful
when a new player comes on board, like Lopez in November.
For fun over the holidays, I worked on some time series models
of the a priori simulation probabilites of the winning exacta
combination and got some interesting (alas, not from a finaincial
standpoint!) results. Can't go into them at the moment -- perhaps
at the next Geek Forum.
NF
