The other day, I came across a blog from about a decade ago of some VIP who was a GE Six Sigma Black Belt who was doing hardcore sabremetrics for hockey. http://hockeyanalytics.com/ Very interesting stuff. Much of what he did could be converted to NLL. One thing he mentioned that I had forgotten: Using the Poisson Distribution for making predictions about goal occurrences and probability of victory. I never liked the Normal Distribution because it theoretically could account for "negative goals". Again, I had forgotten Poisson which is better suited to our analysis tasks. I found this website too: https://hockeyanalysis.com/

The problem I have with Poisson Distributions is if the variance is less than the expectation or it is greater than the expectation aka over/under dispersion. I kind of lean towards Negative Binomial distributions for this reason - after few beers possibly even Poisson-inverse Gaussian models.

Can you elaborate? Even that is currently beyond me. That said, I don't think a Poisson holds either but it holds better than a normal.

Vin I do agree that a Poisson can hold better than normal in some circumstances. But Binomials take into account past success and probability of future success - in sports think of it a team (guy) going on a run. Can someone else here chime in on this? No way you get this kind of analysis on a Facebag group page.

i kind of feel that binomials are a bit simplistic for betting/predicting sports because they inherently assume that the odds of the two possible results are constant for each trial in a set. I do not see that as being the case if you make a complicated switching function for 'p' for each game based on home vs away, result of last game, first or second game in a weekend, key players all healthy vs not all healthy, better goals for or not, better goals against or not, better winning percentage adjusted to strength of schedule or not. then i guess you have to numerically integrate that over the remainder of the season to predict wins and losses. WFinMA modelled the whole season once, i think with binomial, but, with the p updated based on every week's result. let's revive that thread.

Sorry, found that 2009 season simulation thread. it looks like he used lognormal distribution to predict the number of goals a team would score in each of their games, and, used that to simulate 25,000 seasons for each team to estimate probability of them finishing in each position, in the playoffs, etc. Was apparently based initially on 2007-2008 scoring data. Then, as 2009 season goes on, it replaces that data with current-year data, weighted to most recent weeks. Included a correction factor for scoring in home games and a special Rochester 0.5 goal per game correction factor for John-Grant's-knee-is-NFG-this-season

cant spend all my time surfing porn, looking at pictures of cats, and whatever else it is that people do on the Interwebs

I deleted Twitter so I could spend less time looking at pictures of cats, and more time looking at porn.