You are viewing a single comment's thread from:
RE: How I created a WINNING fantasy football prediction algorithm using Bayesian Modeling - The first stop on my journey [~7 min. read, original content and analysis]
It looks interesting but I don't fully get what you are trying to predict. Could you explain me a little bit better what you want to predict ? I might be able to help you with some future ideas.
I want to use the set of prior information on fantasy performances to build a model to predict future performances. I can use those predictions to influence my daily fantasy lineup I enter on a daily fantasy website, hopefully outperforming the statistically challenged.
This approach I outline is a good start, but does have its limitations that you can improve upon when considering other factors. I have found, it is pretty darn good at forecasting the consistency of a player.
So you are trying to model the following, right ?
P(A|B) = P(Fantasy points prediction next week | Fantasy points this week)
Correct. Modeling the training set with the gamma distribution and selecting a likely to happen quantile value as their "B" value. Generating "A" values with many other high leverage football factors (red zone %, snaps, etc.).
Ok , cool. The gamma distribution looks quite weird though. It isn't really fitting the density in the region with highest probabilities. So the region that occurs most frequently is completely underestimated by the gamme distribution. Any idea why this is ? Is this an artefact of having chosen a small sampling size ?
Tried to reply to the last comment of yours below but couldn't. Yes, this is completely due to sample size. This was trained from 103 performances, which is a lot, but not as robust as, say, all performances since 2014 for all players (the lead picture from the story).
Ok. Things are getting clearer to me :-).
You know what might improve the model ? At this moment you are modeling the following:
P(Performance this week | Performance last week).
Did you already try to model the following ?
P(Performance this week| Performance last week, Performance 2 weeks ago, Performance 3 weeks ago, ...)
You could take more past performances into account. This is what control engineers do when they model a system. The more you go backwards, the higher the order of the system. In order to determine amount of weeks you have to go back, you need gradually increase the order of your model and test its performance on a test set that you didn't use for training. Normally, in the beginning increasing the order could improve your model. But at some point increasing the order, will make your model worse. Than is when you know you don't have to increase the order anymore.
What do you think about this idea ?