[DC 86]

[quote author=CrazedZooChimp link=topic=15920.msg176253#msg176253 date=1476726398]

I'm actually surprised the "recent" model has us ranked so high, seeing as we're 1-2 in our past three games with an ugly score differential.

Indeed, sometimes the ratings Sagarin posts are bizarre enough that I suspect actual calculating errors.
Of course since he doesn't publish exactly what he does, no one can check.
[/quote]

Sagarin's highly flawed models have been reverse engineered by various groups over the years. The "Predictor" model basically consists of weighted linear regression using almost-uncapped score margins. Sagarin's season ending ratings have retrodictive value, but any objective analysis (measuring deviation from the final score) concludes that the consensus opening and closing betting lines are far stronger predictors. The problem with Sagarin's (and others') predictive models is primarily that they are purely statistical with no adjustments for factors that may have distorted previous results (e.g. lucky plays, officiating errors, personnel changes) or will impact future results (e.g. injuries, weather conditions).

[DC 86]

[quote author=CrazedZooChimp link=topic=15920.msg176253#msg176253 date=1476726398]

I'm actually surprised the "recent" model has us ranked so high, seeing as we're 1-2 in our past three games with an ugly score differential.

Indeed, sometimes the ratings Sagarin posts are bizarre enough that I suspect actual calculating errors.
Of course since he doesn't publish exactly what he does, no one can check.
[/quote]

Sagarin's highly flawed models have been reverse engineered by various groups over the years. The "Predictor" model basically consists of weighted linear regression using almost-uncapped score margins. Sagarin's season ending ratings have retrodictive value, but any objective analysis (measuring deviation from the final score) concludes that the consensus opening and closing betting lines are far stronger predictors. The problem with Sagarin's (and others') predictive models is primarily that they are purely statistical with no adjustments for factors that may have distorted previous results (e.g. lucky plays, officiating errors, personnel changes) or will impact future results (e.g. injuries, weather conditions).