In a message dated 11/5/2004 1:33:56 AM Central Standard Time, "Paul
Kislanko" _kislanko@..._ (mailto:kislanko@...) writes:

I'd agree that when a game is played is always useless to derive a
metric that orders all times in the field, but it makes sense to
include it as one factor if the purpose of the system is to preedict
future results (which as was pointed out is not the purpose of the BCS).

You'd need to be pretty sophisticated to use it (normalize
quality-of-opponents and game location out of it come to mind) but for
sure if you're trying to predict Purdue's next game you'd want some
kind of trending built into the formula.


Taking when a game happened into account is actually a very difficult thing
to do. The systems that do that best are the sequential systems that use
margin of victory, much like what I suspect Vegas uses, and although these
systems
tend to predict pointspreads very well if they use MOV in the first place,
they tend to not predict straight-up winners well and they're horrible at
retrodiction -- explaining what has already happened. These systems take every
game into account based only on what they know when the game happened, and then
they move on. The only BCS system that works this way now is Billingsley,
and without MOV (and even with some of the kinds of things that Billingsley's
MOV system had), the system loses whatever advantages it might have in terms
of accuracy.

It's easy to mess a system up by using trending. Dunkel's system, which was
a component of the BCS for a long time, tends to overcompensate for what
happens in very recent games, and you will get plenty of results that look
ridiculous on their face -- such as the time that an NFL playoff team that had
a
three game losing streak at the end of the year was in the bottom five in the
league. It is also possible to lose context by overrelying on recent games --
Dunkel's basketball system is completely unable to determine that the NBA's
Western Conference has been head and shoulders above the East for several
years, and the New York Times, which uses a momentum factor, tends to overrate
teams from weak conferences late in the year after a long conference season.

For the most part, systems that do well in accuracy tests do not care when a
game was played. Most do care where a game is played, but realistically
that's not a huge factor. If you know how to do the calculations, it doesn't
particularly matter whether or not you're using MOV when you're determining a
home field advantage.


[Non-text portions of this message have been removed]