Ask HN: Football statistical analysis | Hacker News Reader

Ask HN: Football statistical analysis

7 pointssocrates19986 comments

Hi HN, I have asked around my local network, but no one seems to be able to point me in the right direction.

I was a High School football coach for about 10 years, but now I am self employed and am doing some analytical work for some local football teams.

Anyway, in football when breaking down an opposing teams offense, you have a bunch of data. What I think is happening is that the opposing coach is calling plays to the right and left in almost 50/50 split, you can see it in the data.

I want to be able to compare the other teams decision with a truly random data set.

So, if the coach calls three Right plays in a row, how does this compare to his decisions in the past AND how likely is he to call either right or left on the next play.

I was told it had to do with some Bayesian statistics, but I am not sure.

Here is some example data of a recent breakdown I did last season:

https://www.dropbox.com/s/3m9mr24rahwwzoz/St%20Thomas%20Aq%20Break%20down%20data%202016%20copy.xlsx?dl=0

Looks to me like you're trying to figure out the conditional probability of L/R given a bunch of other factors, and you're looking to tune the model using historical data.

Your factors look to be both categorical and numeric. E.g. off-play and yard line. This given, I'd start by training a conditional inference tree with 60% of the data. Amongst other things it'll give you a good idea of how your factors are related to each other and which matter (ctree is not difficult to interpret). You can then use the other 40% to test how well your tree predicts. Then, you might try other in a more informed manner.

If the above paragraph made no sense it may be advisable to out-source the task. There's quite a lot to the modelling effort.

Thanks for the feedback. Ideally, I would like to do it the way I think you are describing it. Get a bunch of factors and see how they all interact with each other. And try to get the probabilities of all the different scenarios.

For example, if the coach runs the ball to the right in the red zone, what is the probability that he runs it again to the right on the next play.

The problem with trying to input a bunch of factors and trying to get the conditional probabilities of all those situations is that it can get exceedingly complex and I don't think it will be particularly useful given my small data sets.

I normally only have about 100 offensive plays to analyze and use data from.

My theory is that if I compared the R/L direction of the plays that the opposing coach has historically done in the last two or three games with a truly randomization of R/L data, that the coach will severely regress to the mean because he doesn't like the idea of going 5 times to the right in a row.

However, I need some type of percentage comparison or data analysis to show this instead of a subjective picture.

Have a look at http://rattle.togaware.com/ .

It's a UI you can use with R and it's fully capable of doing what I described. You'll need to massage the format of your files a bit. Once you've got some outputs going, post back and I'll try and help you avoid the analysis pitfalls.

What usgroup said is correct. To give a little more background, this is a classification problem (predicting L or R) and it is a supervised learning problem (since you are training it on historical data).

There are many supervised learning models that can be used for classification. The simplest ones that first come to mind are logistic regression and decision trees. If you want to get into more complex models, look into boosting models and random forest.

I'm not sure what your programming background is, but creating classification models is very easy in Python using the scikitlearn library. Creating the features that are used to train your model would be more difficult, but you can always start with something simple and iterate.

It's also important to know how to measure the performance of your model. In this situation, plain accuracy will probably be OK assuming predictions are roughly 50/50 chance. But AUC is typically the standard used to measure model performance.

While I'm a novice in this area, you may want to look into Markov Chains. It only looks at the last state before making a decision but you can determine how big your state is. You could have 3 plays be a single state and go after which three plays come after, for example.

Great, thanks for the idea. I will look into Markov Chains.

Reply on news.ycombinator.com