If you can generalize on similar problems and be correct more often than random chance, then you have a useful model regardless of the data backing that model.