How to use a set attributes of an entity at different time snaps to make predictive analysis? The problem is to come up with a classifier for any task based on a set of attributes of an entity having different values at different times. For instance think about football players and their match stats varying from match to match (accurate shoots, passes, possessions etc.)
Then you can train a model for any task using these stat values. However there are different perspectives to use such a data. One is to use each match values of a player as a different instance, the second is to taking the mean of all stats or the last is to using the mean of latest match stats. 
I am threading this question, maybe there is something different or more suitable for such time varying data problems and I want to hear the ideas from you guys? 
For being particular, again you are given a set of players and their match stats on hundreds of matches. Some of thesee players are also ranked with their abilities on Attacking, Defending with points changing from 0 to 10. Your task is to have a model that predicts a novel player's ability points based on his match stats. How would you evaluate such a problem from your vantage point ?
 A: Dynamic models that update the abilities at each time step are generally better, but hard to implement. If you want to use lots of things like number of passes, then you will generally want to be using some sort of regression somewhere in your model.
http://onlinelibrary.wiley.com/doi/10.1111/rssa.12015/abstract is a paper that predicts a player's goal scoring ability with a mixed effects model.
State-space models give a way of estimating the abilities of football teams, given the number of goals scored in a match, over time. An example is this paper http://onlinelibrary.wiley.com/doi/10.1111/rssa.12042/abstract.
There are a number of ordinal regression models. A dynamic one, which aims to rate teams, is http://onlinelibrary.wiley.com/doi/10.1111/1467-9884.00236/abstract.
There is the also pi-football model, which uses a detailed Bayesian network that uses lots of information: http://www.sciencedirect.com/science/article/pii/S0950705112001967.
Hopefully these papers will give you some good ideas on how to tackle the problem.
