hey dudes,

How do you guys go about projecting pass attempts for a team/QB in a given matchup? Factoring in the expected score and the opponent of course. I feel like ive been burned with some good QBs against bad pass defenses who end up throwing the ball like half of their average. Any thoughts?

depend your project

Game flow dictates the amount of passes thrown. Projections aren't helpful. Teams can get up early and run it to kill the clock and for ball security.

Also, randomness takes over in the redzone. Does the WR get tackled at the one? Does the QB then have to throw it in? Or do they run it in?

Seems like you've fallen victim to variance.

so does anyone have any helpful advice or? Id be willing to pay/trade for something useful.

How much?

Dunno, see what the answer was when he asked the same question in sports betting

its hard to say because every 'answer' ive gotten so far hasnt even been remotely close to what Im looking for. I think the people that know have very little/no incentive to help me, and those that dont, dont even really know what im asking. So it would be hard to commit to a figure before seeing at least some of your work.

Im more interested in trade though. I could show someone how to beat a ton of other stuff, esp in nba and nhl. I also have access to a few juicy books and could possibly give someone action on some nhl correlated parlays (maybe we agree on a set number of bets for a set amount beforehand.)

Ive been betting on sports very seriously for a long time now and have access to all kinds of ways to make money at it. Some stuff you wouldnt even really think of. So if can answer this, it cant hurt to send me a PM and see if we can work something out that works for both of us. I dont expect a free handout.

If you want an idea of how much this costs, I charge $150/hour for statistical consulting. I'm not sure exactly how long it would take to complete but I can't imagine it coming in at fewer than 20 hours. If you're just looking for a report on what the correct answer is, then the hours would be on the lower side. If you also want a proprietary code library (an R package) that scrapes and formats data, runs the models, and produces the output, the hours would be on the higher side.

I have no idea how much noise exists in pass attempts but you should strongly consider the possibility that getting burned by low attempts was just noise and not really explainable by any model.

This season Brady has gone a lowest of 26, high of 45
Brees, 22-49
Mariota, 13-43
Carr, 21-58
Mahomes, 24-53
goff, 24-53 also.

Qb's achieving half their average is hard to do. Pretty unlucky to see that in a small sample of bets.

Over the last 5 years Brady has averaged 37 attempts per game. Roughly 74% of those games had between 30 and 44 attempts. The data over 5 years, consists of 77 games, it looks quite normally distributed. (not sure if it should be as making down allows you more opportunities and more downs, I guess if you arent making downs, you'll tend to throw rather than run too)

Using a simple normal distribution for Brady it looks like he will;
attempt 36 passes or more 46% of the time
37 or more 51%
38 or more 56%
39 or more 60.5%
40 or more 65%.

These %'s would change depending on your opposition, the most important thing is how often they occur relative to each other.

If your model says 37 is most likely and the bookie has o35.5, unless its a low Vig book, you will still lose.

75% of the time Brady gets from 30-44 attempts per game, how does that vary compared to game totals? Also you really want your model attempts to exceed the bookies line by 2.5+ attempts, unless you are getting a good price...... could be hard finding books that far off

As to building a model that works, I have no idea........

If it was a no vig book he would have lost over this sample. Not sure if that's the point you were making or not.



I'm pretty skeptical about using a normal distribution. Even if it's actually normal I'd at least run it out as binomial since it's discrete. In reality, the mean > median means it's slightly right-skewed, and I'd propose that's probably a real thing and not just noise for various football reasons. If I'm correct about skewedness and mistakenly model it as Gaussian, I'd lose solely due to that as illustrated above. The best stock fit is probably negative binomial from which you could run out a NB regression with the relevant cofactors.

And then if you're doing it for Godgers you can chuck everything out of the window given the upgrade from Fat Mike's playcalling to literally anything else

The point I was making about vig is that your edge needs to exceed vig.


I dont think the distribution shape really matters. We are never going too far from the mean/median anyway. You are never going to be able to take say Brady over 55 attempts.

How much more normal does your histogram get if you are using bin sizes of 2 or 3?

Well no kidding but the example you picked is strange because 35.5 was the actual median and he would have lost at any vig over that sample. But you're wrong about if the real number is 37 and the book says 35.5, because

the distribution matters a lot. If I run it out as a binomial with median 37, the o35.5 comes in 63.52% of the time. You would absolutely hammer that.

If I run it as a negative binomial with median 37, the o35.5 comes in 56.18% of the time. Not something I'd bet a house on with high vig.

The binomial model is only supposed to be used until you expect at least 10 successes and 10 failures. Then, it follows the rules of the Normal Model.

Plus, I think what was trying to be said is that the shape of the distribution of all QB Pass attempts (or even just by Brady) doesn't matter. If we take an infinite amount of 16 game samples, the averages (of pass attempts) will be normally distributed, centered around the mean number of pass attempts of all (or even just all of Brady's) QBs.

The distribution of the sample means will always be normally distributed around the population mean.

That is not quite right. Pass attempts are a finite and discrete quantity so using a discrete distribution will always be the most accurate answer. Any normal approximation to the binomial for finite n will be biased even if only slightly so. Since betting edges are small, we should use the most accurate methods in place of ones that are merely convenient.

We aren't trying to find the mean of the sampling distribution though. We want the median and we can't invoke CLT for that. It can get very ugly depending on the underlying distribution and/or how noisy the data is (I did my master's thesis on this). Quickest approach is usually to use simulation if we just want a confidence interval for the median, but that isn't actually what we want here. We're trying to build a prediction model for the median with covariates which is quite another thing.

True and another argument in favor of just doing straight-up simulation.