Optimal Prediction with Conditionally Heteroskedastic Factor Analysed Hidden Markov Models