This. If the 2009 model worked well, did you try letting your algorithm to 'forget' the 2010 data and see if the model worked better?