DomModel

Our AI evaluates past teams against a metric known as SRS, which can loosely be thought of as a net points statistic adjusted for strength of schedule. The machine creates a multi-dimensional, nonlinear “mesh” on which teams are placed throughout the season. Many seasons ago this mesh had been more sophisticated (and the resulting output more noisy) but the current incarnation settles on the use of DVOA as a summary “quantitative” statistic and our in-house power rankings as a summary “qualitative” statistic.

SRS has the convenient property of being subtractable in order to derive the difference, in points, between two NFL teams. But it is imperfect in that it 1) weighs much more heavily towards favorites compared to real world NFL spreads, and 2) is site-neutral, which the NFL (barely) is not. Accordingly, the AI makes its own judgements about how much bias to give to home field and how much bias to give to underdogs, but we do have control over the data we feed it to reach that conclusion. For example, there are obvious reasons not to feed the AI data from the 2020-2021 season in its computation of home field advantage.

The wager selection threshold is now driven quantitatively. In an ideal world, we would take any wagers above the mean wager strength, but the mean wager strength is itself also subject to variance. We thus only take wagers above a certain threshold confidence interval.