Persistence and Recurrence in Atmospheric Circulation

We report preliminary results of an eeort to use variants of the Hidden Markov Models developed by speech researchers to characterize persistence and recurrence of atmospheric circulation patterns in a 36 year record of Northern Hemisphere 700-mb geopotential heights. Using a cross validation scheme, we t autoregressive hidden Markov models (ARHMMs) with a range of complexities , varying the autoregressive order and the number of hidden states. As the autoregressive order increases from zero to four, the optimal number of hidden states decreases from more than eight to one, i.e., simple autoregressive models are better than multi-state models when the autoregressive order is at least four. We believe that the failure to nd regime-like behavior is an artifact of a temporal low-pass lter used in preprocessing the data and that more careful work is indicated.