The NFL season is quickly approaching and many of you have begun asking what my plans are when it comes to sharing my NFL model plays. For those of you that followed along last season, we enjoyed a 75-47-6 record (61% hit rate) for +43.1 units (18% ROI). To keep things short and sweet, I will be charging for the NFL model plays this season. However, I will still be doing a free weekly writeup during the regular season that will cover one model play in-depth. Those write-ups can be found on this site every Thursday afternoon.
Model Play Package Details
Full refund issued if the model plays do not generate a 5 unit profit ($100 bettor breaks even, or is refunded)
Includes model plays for regular season and playoffs
Includes model-recommended preseason and in-season futures
I made a change in how the model derives larger spreads. Last year the model struggled with spreads of +7 or higher last season (11-11-2 for -2.72 units) and plays on such spreads were discontinued Week 10 onwards. I implemented a change and backtested it, and the record on those plays were now 7-5-0 for +1.73u. I was more pleased with the increased selectivity (12 plays vs. 24) as opposed to the better performance, as the sample is obviously too small to make anything of the increased performance.
Last year each point of disagreement equaled one unit of play which made it so that the lowest possible play was two units, given that the model required two points of disagreement for a play to qualify. This year each point of difference will equal a half unit of play, so that the base play is one unit. I’ve went ahead and made this change across the site and in last year’s model tracker, as the previous net was +86.2 units but now stands at +43.1 units. The scale does not change the ROI for last year, which still stands at 18%.
I didn’t follow along last year, how does it work?
There are so many words that I could use to describe the entire experience I had with my NCAAM model. It was humbling to have nearly 1,000 people donate to Doctors Without Borders, raising $25,378 in the process. It was fulfilling to have the model finish 77-49-2 (60.94%) for +34.43 units (13.76% ROI) and provide an incredible return to those who contributed their money to a good cause. I decided to take a look at how donators fared given the day they contributed, here is a look at those splits:
If you donated on 2/12: 56-37-2 (60.00%) for +21.34u (11.52% ROI)
If you donated on 2/13: 55-36-2 (60.22%) for +22.53u (12.40% ROI)
If you donated on 2/27: 28-16-0 (63.64%) for +16.27u (19.74% ROI)
If you donated on 2/28: 25-14-0 (64.10%) for +15.29u (21.67% ROI)
Looking at the above splits from the perspective of a $100 unit bettor, donators had a profit yield that was 75 to 110 times larger than the $20 minimum donation. In other words, you would have to be a 25 cents per unit bettor to not have made enough profit to cover the minimum donation. You get the idea, so I’ll move on some of the model’s key performance splits.
Home/road and favorite/underdog splits
If you recall, the NFL model from last season performed best with underdogs and with road teams. Interestingly enough, the same was largely true with the NCAAM model. Here’s a full look at those splits:
Like with the NFL (and any sport), I attribute this to the psychological tendency for bettors (especially “square” bettors) to favor favorites and home teams. Because of this, oddsmakers typically shade their lines in order to capture more value on that action. The effect is very small on a game-to-game basis, but provides opportunity and value over the course of a season. Next lets take a look at when we combine home/road and favorite/underdog splits:
Obviously combining the larger edges on road and underdog teams into a singular road underdogs split very clearly demonstrates where the model found the most opportunity and success. Road underdogs accounted for 60% of all model plays and those plays hit at a 64.47% rate, generating a 25.55 unit profit (17.38% ROI). The other splits contain far too small of sample sizes to make any definitive statements, but are there for you to interpret any particular way you’d like.
Unit size splits
If there’s one split of the model’s performance that was most concerning, it would certainly be how the model performed as the disagreement level grew. One sign of a good model is that the model generates profit at a higher rate as the size of disagreement increases. Given that the model had a 13.76% ROI on all model plays but a 9.08% ROI on model plays that were two or more units is admittedly a bit concerning. The only counter I can offer is that flipping two or three of the results in that split from losses to wins does make that split more profitable than the overall ROI. Either way, it will obviously be something I keep an eye on for future seasons.
Frequently backed/opposed teams
There were certainly teams that the model deemed undervalued and overvalued and took repeated action on or against. Of the teams that the model backs three or more times (Binghamton, Columbia, New Hampshire, Notre Dame, Portland, San Diego State, and William & Mary), the model went 12-9 for +5.82 units. Of the teams that the model played against three or more times (Brown, Drake, Florida Atlantic, Fresno State, Louisville, Murray State, Saint Mary’s, Stony Brook, Utah State, VCU, Virginia, and Virginia Tech), the model went 29-21-1 for +9.80 units. I thought the positive results for both were a good sign for the model as it signals that the model is proficient on some level of identifying which teams are overvalued and undervalued in the market.
It goes without saying that these splits have way too small of sample sizes to make any definitive conclusions, but I included it just for fun and it was a split I got asked to include on several occasions. The Ivy League obviously stands out as the model’s best conference with a 6-1 record for +6.96 units. Outside of that, I thought the model’s performance with the Power 5 conferences would be interesting to look at given that those games get the most attention in terms of handle. The Big 12, SEC, and Pac 12 combined for a 5-0 record for +8.75 units whereas the Big 10 and ACC went a combined 9-9 for -1.99 units.
Takeaways and closing thoughts
First and foremost, I think it’s safe to say that I should have put more time and attention into the NCAAM model. For the entire first month of sharing plays, I only checked lines once in the morning and that was it. It comes as no surprise that once I started checking lines more frequently, including looking at overnight lines, the model found more opportunity and more success.
In the past, I’ve typically been very hesitant to take any action any earlier than February. I don’t think I’d go as far to say that I would use the model for the non-conference play that starts the season, but I definitely think there is an opportunity to use the model in December and January.
All in all, I’m incredibly satisfied with what the NCAAM model accomplished. The profit it was able to generate was of course a great feat, but nothing I’ve done in my life comes close to the collective contribution we were able to generate for Doctors Without Borders. I don’t think there has been a day since that I haven’t thought about those two donation periods and the rush of happiness I experienced as donation after donation came in. Thank you all again for making that happen.