Hidden Markov Chain Random Probabilities

Hi friends, I am working on a fx forecasting algorithm, which uses discrete hidden markov chain to predict next day's return (rise fall zero).

What ı understood from hmm is that it should converge to the best sequence. I assign random prorabilities on state, observation, distribution matrixes, and end up with different sequences. this is also the case when ı use ready matlab functions.. can u please clarify what ı miss here?