Subvertadown models are built from a special version of multivariate statistical regression, with cross-validation.

That means we don’t choose our favorite defensive stats from our own idea of what should be useful. Instead, we let the computer test what data would have been most helpful for predicting past games. By asking "What data (data which existed before each game was played) could have improved our ranking of D/ST picks?", in a very precise and scientific way (to be sure we don't fool ourselves) the models extract the best predicting drivers.

Hundreds of variables are tested, dozens of variables end up used in some way. A more comprehensive list of tested parameters is listed in the ”What’s in the Model?” article.

Although this process could be used to create a single model in one go, today’s final Subvertadown D/ST model is actually based on a number of sub-models. For example, some factors are more or less predictive in the early season; therefore there are different models tuned to the time of season. Some factors are only relevant for a high range of scores— which helps us to better sort among the top teams each week; therefore there are different models for low scores. And my studies show that it additional helps to include separately-tailored models that predict the D/ST score components, such as sacks and interceptions. Each of these sub-models typically contain a dozen critical variables, and all the sub-models finally combine into the final model.

D/ST fantasy models typically include points-allowed, yards allowed or rushing yards allowed, pace of game, interception rates, betting lines, and different weather factors; as well as opponent factors such as sacks allowed by the opposing QB/OL and factors of the opposing QB skill. 

One key additional detail is the processing used on all the above data, to account for historical strength of schedule (the influence of prior opponents on the data that the D/ST faced), and to identify predictive trends.

Finally, there is a procedure for adjustments to account for roster changes, on a weekly basis. That means, the output of above model typically gets modified to account for injuries, acquisitions, etc.