The study of time series prediction is pervasive in various fields.We propose a cluster-based hidden Markov model to approach the multi-step prediction problem in time series.As multi-step time series prediction problem is not fully addressed from a system angle

we utilize the hidden state of hidden Markov model to represent the inner state of a time series production system.We also promote a cluster algorithm combining the temporal and similarity criteria to address the distance calculating issue in time series clustering.This non-trivial criterion proves effective in multi-step time series prediction.Through a non-parameter approximate method we estimate the inner hidden state distributes from every single state.And we also prove the correctness of an iteratively refinement of the cluster-based hidden Markov model(HMM).Experimental results on authentic data indicate the effectiveness and accuracy of this approach.