I am having a hard time trying to make/pick a prediction model in R.
The data: I have information on 40 different players, with all their recorded performances(training loads) over the last season. These instances are not evenly spaced over time, and each instance has 24 variables:
- 1 column with player-id's
- 1 column with the instance date
- 2 columns with binary variables: injured/not injured, and gets injured/not (the instance before the player is injured is the 'gets injured' instance).
- 4 measured numerical independent variables (including sprint and total distance), and 16 numerical variables that are made using the first 4.
The goal: give a risk percentage that a player injured (possibly if the player exceeds distance x) in the next /match/event (i.e. a future, not yet recorded event).
- My suspicion is that the cumulative load on a player causes him to have an injury, so the model should take previous instances (load) of thát player into account, but not of other players. The resulting model should however be generalizeable and able to give a risk for a previously unsean player (if there is enough load information available on that player)
- Getting injured is a rare event.
I was thinking towards a series of time series, but I do not know how to split/combine the different players with records on different dates in order to make the generalized model. I also read about sliding window, could that be a good solution? I have also tried the zelig package, but this does not result in any actionable outcomes (probabilities 0 or 1 occurred).
Question: which predictive modelling technique takes this interrelateability into account yet allows for generalizeability in predicting rare events?