Can hidden Markov models be used to model any time series data?

Question

Can HMMs be used to model any time series data? Or does the data have to be that of a Markov process?

In HTK documentation, I see that the first few lines state that it can model any time series

HTK is a toolkit for building Hidden Markov Models (HMMs). HMMs can be used to model any time series and the core of HTK is similarly general-purpose.

Answer 1

Yes, you can fit any time series (with or without external variables) using HMM, but there are some constraints:

1. It should follow the Markov property.
2. There is some variance that other models are not able to capture (in other words, the system is partially observable).

Adding to point 1, for HMM, it should hold true, but the way Baum Welch algorithm works, indirectly it considers the values of more than the previous state for HMM (order-1). The state $t-1$ depends on $t-2$, which in turn depends on $t-3$. The calculation of parameters (transition, emission, starting probabilities) happens over multiple iterations and it finds parameters in such a way that holds Markov property true.

I think that, when they say 'any', they mean even when you don't have all variables needed to forecast future values.

Answer 2

HMMs can be used to model sequential data, which is composed of discrete tokens, and it should generally follow the Markov property, which is the assumption that the probability of a class/label given observation depends only on the preceding class/label (rather than on some longer sequence).