This paper proposes a novel classification approach to carrying out sequential data classification. In this approach, each sequence in a data stream is approximated and represented by one state space model – liquid state machine. Each sequence is mapped into the state space of the approximating model. Instead of carrying out classification on the sequences directly, we discuss measuring the dissimilarity between models under different hypotheses. The classification experiment on binary synthetic data demonstrates robustness using appropriate measurement. The classifications on benchmark univariate and multivariate data confirm the advantages of the proposed approach compared with several common algorithms.