7
$\begingroup$

I have a dataset, which contains DateTime, target, target_type. target is basically the count of a process. target_type is binary and it says if the count was of the type, say "outflow" or of type, say "inflow". this has been recorded at periodic intervals. what I would like to do is, predict target for the next n time intervals, let's say next 5 intervals.

my question is, which approach should I look into? arima? lstm? markov-modulated Poisson process? or something else.

another thing that I'm confused is, I'm unable to figure out if I should treat this as a multivariate time series data, (i.e if I one hot encode the target type variable.

so the dataset looks like this:

enter image description here

Improve this question
$\endgroup$

1 Answer 1

Reset to default
6
$\begingroup$

You have , i.e., your time series are integer-valued, nonnegative and "mostly" zero. You may want to search for "forecasting intermittent time series" or similar.

The classical approach for point forecasts in such a case is . One alternative is a Poisson or Negative Binomial regression on whatever regressors make sense (e.g., trend, seasonal dummies, causals etc.). I have also seen Integer ARMA (INARMA) models, e.g., in a Ph.D. thesis by Mona Mohammadipour, but these are not very common.

One thing that I have not yet seen is linking multiple such time series together. In the continuous case, this would be Vector Autoregression (VAR, not to be confused with VaR, which is Value at Risk). An analogue to the integer case could conceivably be called VINAR, but as I said, I have never seen this.

Improve this answer
$\endgroup$
2
  • $\begingroup$ thanks for the suggestion! seems like intermittent time series(crostons method) is what I should work with. also, I would love to know what is your view on using a zero-inflated Poisson process/Markov modulated Poisson process with this? (ref: datalab.uci.edu/papers/event_detection_kdd06.pdf) $\endgroup$
    – ultron
    Commented Jan 25, 2018 at 4:38
  • $\begingroup$ A zero-inflated Poisson process or regression may be a possibility, depending on your data. I can't say anything about Markov modulated Poisson processes. $\endgroup$
    – Stephan Kolassa
    Commented Jan 25, 2018 at 7:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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