Re: Predicting a champion in a sports league

On 25 Jan 2006 08:08:06 -0800, jdoiii@xxxxxxxxx wrote:

>I have various sports statistics (offensive and defensive performance
>measures) for multiple teams over multiple years in a single league.
>At the end of each year, a champion is determined by playoff among the
>top teams, thus the champion is not necessarily the team with the best
>winning percentage that year. I am trying to determine which measures
>are significantly associated with being the champion for any one year.
>teams, not sure if that makes a difference.)
(snipped)
>Would be glad to hear any recommendations/advice/references on how to
>best go about this.

A while back I had a useful model of the NFL.

The model assigned each team an offensive and defensive strength. It
assumed the scores were a function of these two strengths. The model
used a hill climbing algorithm to produce values of the offensive and
defensive strengths which best predicted the historical outcomes.
These results then gave a prediction for that week's games.

The problem with this model was determining how much history was
relevant to the current week. If a team looses a quarterback, its
offensive performance history becomes moot. Similarly for injuries.

In addition, the model assumed a linear (sort of) relationship between
score and strengths. In fact, teams will sometimes back off from
running up scores. Enough is enough may be the cause or rest the key
players for the next tougher game, etc.

Finally, individual statistics can be badly distorted--a loosing
quarterback may have is best game, measured in yardage thrown.

John Bailey
http://home.rochester.rr.com/jbxroads

.