I have constructed a simple HMM (Hidden Markov Model) with 2 states on the Vol (stdev) of a time series of currency returns.

The state vector I produce looks reasonable, in the sense that it appears to have identified periods of high or low vol.

However, I appear to be able to generate a very similar state vector just by applying a simple rule like:

if (Vol) > x then 1 else 0.

Is there any advantage/difference to using a HMM?

  • $\begingroup$ Maybe that in the HMM case you let the data decide the best set of regimes for you? $\endgroup$
    – Lisa Ann
    Aug 12, 2013 at 13:55
  • 1
    $\begingroup$ But maybe HMM just splits states in middle of range? if vol goes from 5% to 15% perhaps it just says if Vol > 10% then 1 else 0. If this is case why bother why more complicated HMM? $\endgroup$ Aug 12, 2013 at 14:19
  • 1
    $\begingroup$ According to my experience with RHmm package, it usually does only when the two states have got the same st. dev. As instance, try that with the VIX index: you will get the "panic" state quite above 24 but with a huge variance (which explains the over 40 spikes). If you splitted in the middle range, you'd get different values. $\endgroup$
    – Lisa Ann
    Aug 12, 2013 at 14:24
  • $\begingroup$ OK I see. So, the informational value of this simple HMM is not very different to my simple rule. BUT the fact I have not picked(optimised it) is useful right? $\endgroup$ Aug 12, 2013 at 14:42
  • 1
    $\begingroup$ Cross-posted to NP. $\endgroup$ Aug 12, 2013 at 17:57

2 Answers 2


Have a look at the following paper:
Regime Shifts: Implications for Dynamic Strategies by Kritzman, Page and Turkington

From the paper (p. 25):

But why go through all the trouble? When dealing with regime shifts, we expect Markov-switching models to perform better than simple data partitions based on thresholds. For example, in Figure 1, if we had simply classified the observations that were in the highest quartile as being associated with Regime 2 (the high-mean regime), we would have misidentified the actual regime 40 times out of 200 observations. In contrast, a well calibrated Markov-switching model would have misidentified the actual regime only three times. Arbitrary thresholds give false signals because they fail to capture the persistence in regimes as well as changing volatilities

There doesn't seem to be a free version available any more. If you find one, please let me know: I will then update the link.

  • $\begingroup$ @user2675052: Thank you, I would appreciate if you could upvote and accept my answer then :-) $\endgroup$
    – vonjd
    Aug 13, 2013 at 16:29
  • $\begingroup$ @vonjd Link appears to be broken. $\endgroup$
    – htd
    Oct 6, 2015 at 10:29

HMM allows to get transition matrix that provides additional information itself about probabilities of switching.

As HMM looks on complete state path it allows to identify, for example, short periods of low volatility in high volatility regime that were not a result of regime switching. If we apply some simple rule we have a larger number of switches and thus have biased transition matrix. If the results of this analysis are used in some forecasting or simulations, error in transition matrix would be crucial.


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.