@Sam: When I say "flat", I mean just unroll your data for each example into a simple vector, and use a model that doesn't do anything special with time series. XGBoost could be a good baseline here, you can just throw some data in and get a quick-and-dirty predictive model, without worrying too much about details. The point would not necessarily be to use that model, but to get a sense of what a "good" model has to beat in order to of been worthwhile.
