Selecting hidden Markov model state number with cross-validated likelihood (Q626234)

A hidden Markov chain model is considered in which the states chain is stationary and ergodic, and the observations are conditionally independent given the chain realization. The distributions of the observations are known up to unknown parameters which depend on the hidden states. The authors consider the computation of cross-validated likelihood criteria for selection of the number of the hidden chain states. Two algorithms are proposed. In the first one the training and the test subsamples are taken by considering, respectively, the odd and even indices of the original data set. In the second algorithm the elements of the test subsample are chosen at random and considered as missing data in the training subsample. A version of the EM algorithm is used for fitting the model by the training subsample. Numerical results are presented for simulated and biological real life data.