Week 11: Spread Picks, Results and Analysis
I’m going to keep the early week posts short and sweet. Yes, unfortunately I’ve got a real job, and it’s only tangentially related to The Rotoquant. As a result, the early week posts will go as follows: one post to briefly analyze performance against the spread on Tuesday, one post to briefly analyze fantasy QB prediction performance on Wednesday, some mildly interesting and less brief musings and accompanying picks on Saturday – wash, rinse, repeat.
Let’s get into it.
Overall the picks for week 11 were a mixed bag. Outside of the three pushed games which I found annoying, there was the Patriots Panthers debacle on Monday night in which Cam Newton pulled a Tom Brady stealing a win away from the machine which had the Patriots picked plus 2.5 points.
As it pertains to this weeks picks the games served to highlight one key feature of an algorithmic approach over less mathematical approaches: rank ordering by a calculated confidence score. Each week when I break down what happened in the week prior, I’ll use the confidence scores as the barometer to measure performance. I expect that higher-confidence games are more likely to be picked correctly. When putting together a betting strategy, the confidence threshold might be viewed as one’s ‘appetite for betting risk’ – the lower the size of the SpreadCoverIndex (SCI), the more risky the bet and the smaller the presumed edge.
So, first let’s look at the picks as a whole if we ignore the index altogether and just pick every game:
Woa! 42% against the spread? Some edge huh?! What a chump I must feel like, having put these picks online so publicly with so much half-hearted braggadocio about being able to beat the odds makers… Or must I?
If we restrict our universe of games to those games for which the absolute value of the SCI is greater than or equal to 1.0, the week doesn’t look quite so bad:
Still, 56% against the spread, or 5 out of 9 non-pushed games is not miraculous. Better, but not great. If we consider the algorithm to be no better than flipping a coin at picking games, there is a 25% chance it would have flipped exactly 5 heads. So while 56% for 9 games is definitely better than 42% overall it’s not shockingly good. It isn’t such an improbable win percentage as to be considered outside the realm of random chance, particularly in a sample size this small. Sample size is a concept I’ll keep coming back to, both in this post and throughout the season, because it’s so easily and often overlooked, and so critically important when trying to tease out what is a truly predictive signal from what is merely statistical noise.
At any rate – it is a good sign that when the absolute value of the SCI increased, so too did the proportion of games picked correctly. This is the type of correlation we would expect, and even in small sample sizes it’s nice to have your expectations met.
Now let’s raise the bar once more and see how the algorithm performed on games at or above a 2.0 absolute SCI.
Interestingly, we see the win percentage increasing once again to 60% or 3 of 5. As before, this is definitely what we are hoping to see if we think the absolute value of the SCI has any predictive power for game outcomes. And to reinforce how important sample size is in understanding the likelihood of any given result, I’ll point out that while 60% certainly is a better win percentage than 56%, it’s also more likely to happen by chance in a 5 coin-flip sample than compared to 5 out of 9. The chances of exactly 5 out of 9 heads was 25%, whereas the chances of 3 of 5 heads is 31%. Sample size matters when interpreting results.
Lastly, we come to our most confident predictions, those with an SCI at 3.0 or above.
Wow! 67%, not too shabby! Most pros would probably be pretty excited about picking any number of games with a 67% win rate – and it once again provides evidence that an increase in SCI comes with an increase in the probability of picking the game correctly. So this is a very good result – especially when we consider that the TB and NYG games were not even close, and the one game that was missed in this group was the aforementioned rather incredible Panthers win. But, because I can’t reiterate the importance of sample size enough, I have to point out that the probability of flipping 2 out of three heads is 37.5% – making this smallest, most confident set of predictions also the set that has the highest likelihood of being produced by chance.
All in all I consider this to be a decent coming out party for the SCI algorithm, and it was ripe with opportunity to clearly demonstrate the meaning of the index and how we will interpret the results moving forward.
That’s all for now, be sure to check in tomorrow for the fantasy QB review.