You would have lost your shirt on Leo last month. If I'm using
a big Alpha, I'm pretty close to the long term average, anyway.
But, with many thousands of games in the database, it's possible
to choose an Alpha which results in least variance when testing
a new model against history.
I've also experimented with trend (which can also be smoothed)
to see if I can squeeze out something more, e.g., a rookie's improvement
over the first few years, but if there's anything there, it's
overwhelmed by recent performance.
