>2. Can a person use your power rating to determine an expected point spread, like Sagarin's rating?
>3. Is your composite ranking an average of all six of your different rankings?
1. I'm not sure.
A. Disadvantages of models:
Model 1: sometimes high emphasis on few games because ignoring of blowouts
Model 2: 100% game margin model, underestimates teams that win by close margins (Ohio St 2002, Iowa 2004)
Model 3: counts last 5 games only, can overestimate a team based on a hot streak
Model 4: can overestimate latest games over previous games
Model 5: can be overdependent on schedule strength
Model 6: can greatly overestimate weak team in weak conferences (Pitt St Div II ranked as #20 in whole NCAA)
B. Advantages of models:
Model 1: ignores blowouts and has characteristics of Model 2
Model 2: best overall model at classifying weak teams as weak and strong teams as strong.
Model 3: ignores old games and focuses on recent games
Model 4: changes quickly to a teams sometimes recently found strength
Model 5: strong teams with strong opponents are identified well
Model 6: ranks teams with few losses higher than other models.
C. All models are unbiased an have no arbitrary elements. The RPI is the an example of a model with arbitrary elements.
http://kenpom.com/rpi_info.html
RPI = (1/4) X Winning Pct. + (1/2) X Average Opponent's Winning Pct. + (1/4) X Average Opponent's Opponent's Winning Pct.
The arbitrary elements are the coefficients for the variables 1. (1/4) 2. (1/2) 3. (1/4). Why not use (1/3), (1/3), (1/3)? or (1/2), (1/4), (1/4).
D. All models not to explicitly compute schedule strength first. All models calculate schedule strength after team strengths are discovered. Explicitly using schedule strength as an input to the discovery of team strength invites schedule strength bias and typically requires some sort of arbitrary weighting of records.
E. Models 4-6 do not have home-field advantage factoring. I would like to implement this in the future but have not found a way to do this with introduction an arbitrary weighting. The standard regression model (Model #2) does a beautiful job of calculating home-field advantage.
F. Game-margin models 1-3 are quite similar in structure but very independent of the W/L only models (Models 4-6). Additionally, models 4-6 are quite different from each other.
G. As mentioned is previous posts, a computer model can easily be made unbiased. However, that does not guarranty accuracy.
H. I think the key to success is to find as many different models are possible and average them. However, all sub-models must be unbiased and without arbitrary elements (points C and D). A average of unbiased models will produce more accuracy than any of the models will obtain independently.
2. Yes.
3. Composite is an average of the 6 subrankings with a compensation for the unequal ranges of the scores of the submodels.