I have the following problem I consider to model with HMM. However, it seems to me like a non-standard application or it possibly makes no sense employ HMM. I already have some practical experience with HMM used on time series and quite good theoretical knowledge.

Problem: I am observing a continuous, time series-like process, minute data $\{y_t\}$, with bunch of regressors. I want to predict the mean for every hour given the past hour, more accurately, I can use circa past 45 minutes data.

So I was thinking of defining the states as whether the hourly mean is greater then zero and otherwise. I would have then a binary state variable. But like this, I can't see how to define a state space model

$y_t|x_t \sim p(y_t|x_t)$

$x_t|x_{t-1} \sim p(x_t|x_{t-1}),$

while in my setting, e.g., $y_1, \ldots, y_{60}$ are generated by one state, so the state should last 60 minutes and the transition shall occur at the end of every hour and one state generates 60-observations sequence with 0 probability of transition. The closest problem I can think of is for instance regime detection in financial data, bull and bear market.

Thus I don't see some straightforward application of HMM, if at all it is possible, but I would like to use it due its probabilistic interpretation. I wonder, if it is better to define the states like above and transform the problem to classification.. I can also provide more informations if this grabs your attention. Thanks in advance.

  • $\begingroup$ Tell me if Im thinking about this correctly: your dependent variable is, say, the temperature, and we have a bunch of other independent variables like cloud coverage and wind speed? $\endgroup$ – redress May 24 '17 at 20:47
  • $\begingroup$ Yes, let's say it is like that. $\endgroup$ – reicja May 24 '17 at 20:50
  • 3
    $\begingroup$ Sounds like a classic multiple linear regression problem...what am I missing? $\endgroup$ – redress May 24 '17 at 20:52
  • $\begingroup$ Well the thing I want to predict is the average of the next hour. Say, it is 10:00, so I can use minute measurements up to 10:45 and at 10:45 I have to make the prediction for the average of the next hour, from 11:00 to 12:00. Of course, I considered a lot of regression techniques but in my case there is not strong linear dependence/correlation this far ahead. $\endgroup$ – reicja May 24 '17 at 20:58
  • $\begingroup$ (+1) @redress You can definitely formulate this is a classic regression problem. If you do so and use overlapping observations, you'll need to take into account the correlation in your residuals due to overlapping observations (eg. Hansen-Hodrick standard errors or Newey-West or something.) $\endgroup$ – Matthew Gunn May 24 '17 at 23:00

but in my case there is not strong linear dependence/correlation this far ahead

I would consider a Recurrent Neural Network, where the shape of your input matrix is

   [number_of_observations, number_of_variables, 45]

So, after a little while I found exactly what I had in mind, which is HMM with multiple emissions per state. The problem is for me sufficiently covered in another HMM question.


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.