If you follow us religiously (and damn it, you should) you’ll have noticed our strong preference towards underdogs the past few weeks.
Theoretically, the balancing pass we undertook in mid-October should have eliminated any inherent bias the model holds towards favorites or dogs. The model actually evolves, mildly, each week of the season, taking into account week to week changes in its predictions compared to where spreads ultimately end up. This weekly “education session” has resulted in a model that, contrary to my intentions, is tilting itself towards dogs.
It’s not a bug. It’s a feature.
The weekly adaption has served us well. Underdogs have dominated this season, posting an incredible 133-107 record (55.4%) against the spread. This tendency has not gone unnoticed in media coverage either, with thoughtful posts in the sportsbetting community contributing to the subject.
This deviation is statistically significant. If these wagers were true coin flips, the resulting standard error of n = 240 such flips would come to about 3.2%. We can say with ~90% certainty that this outperformance from underdogs is not just due to mere chance.
Sportsbooks underestimated dogs this year, and we stood to benefit.
Whether the mispricing is due to the strangeness of the 2020 season or due to deeper inefficiencies in the sportsbetting markets we cannot say. It is easy to dismiss this as we have everything else in 2020, “because of covid.” But this default looks increasingly suspect when we discover that 2019 saw underdogs overcome favorites 131-116-9.
The model will continue to update throughout the playoffs, and I expect we see a continued bias towards underdogs. The smaller slate of games and more efficient lines found in the postseason NFL will likely result in only a handful of additional wagers.
For 2021, I expect we maintain what has been a highly successful and consistent process. We will need to re-evaluate our expectations for home field advantage, which as of this writing has fallen beneath 1.5 points. Considerations around site design are also on our docket. I expect offseason tweaks to the model will largely focus on back-end tasks, like maintaining an efficient codebase and increasing automation.