0
$\begingroup$

The question is about the equivalence between ARIMA models and hidden Markov models in the context of time series analysis/prediction. Specifically:

  1. Can any ARIMA(p,d,q) model bet represented by an equivalent HMM?
  2. Can any HMM be represented as ARIMA (or are HMMs a bigger class of models).

My impression is that the answers are 1. yes, and 2. no. However I am looking for a definitive answer. A reference to a reliable sources would be greatly appreciated as well.

$\endgroup$
2
$\begingroup$

In the standard HMMs, the state spaces of hidden variables are discrete. On the other hand ARIMA models can be represented as a Kalman Filter which is a continuous state space model (underlying hidden variables are continuous). This is the main difference. However there should be HMM extensions to manage continuous state spaces.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks! It is not in the original question, but I am also wondering if or to what extent ARIMA and HMMs are made obsolete by LSTM neural networks? $\endgroup$ – Vadim Aug 15 at 9:26
  • 1
    $\begingroup$ @Vadim, neural networks typically require longer series for training, while ARIMA can be fit quite effectively on just a handful of data points. This is not unique to time series; flexible machine learning methods in general have not made simple statistical models obsolete for the same reason. $\endgroup$ – Richard Hardy Aug 18 at 15:38
  • $\begingroup$ @RichardHardy this is a good point. But assuming that we have a sufficiently long sequence? I have encountered claims that there are situations where the models described by HMMs are poorly modeled by LSTMs. $\endgroup$ – Vadim Aug 19 at 12:33
  • $\begingroup$ @Vadim, I do not know LSTMs well enough, but assuming they are flexible enough to approximate any HMM to a high degree and given a sufficient large training set, I do not see why they would be inferior. But perhaps the assumption I just made does not hold? $\endgroup$ – Richard Hardy Aug 19 at 12:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.