An investigation of the state formation and transition limitations for prediction problems in recurrent neural networks