How often do you update the rankings?
Once a week, generally on Wednesday.
Do you use the kind of subjective inputs that many systems use, especially at the beginning of the season?
Beginning with the 2012 season, we are using “priors” in our rankings, but they are completely quantitative and based on the same methodology we use during the season. To arrive at our priors we regress the rankings from the end of the previous season to the mean (0). We do this separately for each of the 8 statistics that comprise our rankings, and the amount of the regression varies. How much it varies depends on what history tells us – we rely on the priors as much as we can to predict this year’s games, but no more.
These priors matter less each week of the season, as new current-season data come in. How quickly they taper off again varies by statistic and is determined by what is optimal for prediction. Priors for some stats are gone by mid-year while others are reliable enough to improve our predictions deep into the season.
Do your rankings assume that a team has a single, fixed ability/quality level throughout the season?
They do not. In fact, we estimate the degree to which team ability changes over time and incorporate those estimates into our ranking. This means performance in the recent past weighs more heavily than performance in the distant past. In technical terms, our rankings allow for “non-stationarity”.