Brady and I have been betting the NFL for years now, and since starting First and Thirty we’ve posted countless articles discussing how our process works, why we make each wager, and the relevant game information bettors need to be thinking about. We pride ourselves on total and complete transparency.
Unlike most handicappers, who have either entirely subjective recommendations or who show only the results of black-box models, we make clear exactly how our process works and encourage you to bet with us because that process is consistent.
And you, my dear punter, want to know that process has positive expectation. I am here to tell you that it does.
The results
Since putting into place our current model two seasons ago, our record is a phenomenal 82-60-3, a 57.7% winrate, grossing +15.2% ROI per dollar wagered.
But nothing in life is free, and that includes your bookmaker. Most bookmakers profit by charging vigorish, or “vig,” in the form of a markup on your losses. Estimating our profitability requires us to make an assumption about the vig, which will vary depending on where you get your lines. Below, I assume the industry standard vigorish of -110.
After accounting for the vig, our edge still looks absolutely exceptional at +10.0%. Based on the 5,000 simulations bootstrapped below, we are over 90% confident in our ability to beat the vigorish.
I’ve always been a more visual kind of guy, so I hope the plot below helps illustrate how meaningful this is. Each point below is a two-season simulation, and the simulation is run 5,000 times. The dense block of points is concentrated above 0, which is extremely compelling given that these simulations account for the vig.
To close out my thoughts on our results, I want to point out that our record for 2020 was consistent with our record for 2021. We recommended 86 wagers in 2020 with an edge of 10% and 59 wagers in 2021 with an edge of 10%.
Some reasons for caution
These results are extremely positive, and we are happy about them. But, there are some reasons why we want to treat our results with caution.
First, some of the inputs to the model may be non-stationary, meaning that they change over time. In particular, the model self-adjusts for home field advantage, which has tended to decline in recent years, and for the generally strong performance of underdogs, which persisted throughout 2020 and 2021. Like all sports, the prevailing meta in the NFL is likely to change over time, and changes in these trends could impact the success of the model. For example, the introduction of an expansion team in London could lead to higher home field advantage as a result of increased travel times and jet lag.
Second, while our performance has been very strong, we are still only working with two seasons worth of go-forward data. The interquartile range of our bootstrap is the same size as our mean. While I am cautiously optimistic about our results, we are still losing bettors in about 10% of our simulations, and it will take a continued level of exceptional performance over an increasingly large sample to reach a 95% confidence interval.
Third, I decline to analyze our unit-based ROI here, both because it is highly leptokurtic and because we have been inconsistent in its analysis and application. Still, we might at times recommend larger than normal wagers, and I encourage bettors to be sensible in applying their bankroll appropriately. A separate analysis on wager "strength" may be needed in the future, but the rarity of such bets presents sample size problems.
Finally, in relation to the point above, we made substantially more wagers in 2020 than we did in 2021. Towards the end of 2021 we applied a statistical approach to determining our wager strength cutoff, which will be updated and applied throughout the upcoming 2022 season. While we believe we are tuning this parameter appropriately, there may be a tradeoff between taking stronger wagers and reaching the “long run” more quickly.
About our friend Kelly
If you aren’t familiar with the Kelly Criterion, I encourage you to get acquainted. In the interest of brevity I will conclude by pointing out that the Kelly-optimal wager size will be about 10% of your bankroll, and a more conservative half-kelly sizing will be about 5%. So, for every $1,000 in your bankroll, a $50 per-game wager is more or less appropriate.
Estimation results of bootstrap
I'm mostly including this for the sake of transparency.
Wessa P., (2019), Bootstrap Plot for Central Tendency (v1.0.16) in Free Statistics Software (v1.2.1), Office for Research Development and Education, URL https://www.wessa.net/rwasp_bootstrapplot1.wasp/