What state-space representation of VARMA is commonly used for fitting?

Is Kalman filter + MLE approach used for fitting VARMA model as a common practice?

Does the choice of which state-space representation to use affect fitting performance?


I'm new to time series modeling.

I'm trying to use a VARMA model to fit some multivariate data. (for now I use artificial data generated using dse)

n = 1000
AR <- array(c(1, .5, .3, 0, .2, .1, 0, .2, .05, 1, .5, .3), c(3,2,2))
MA <- array(c(1, .2, 0, .1, 0, 0, 1, .3), c(2,2,2))
arma <- ARMA(A=AR, B=MA, C=NULL)
result = outputData(simulate(arma, sampleT=n))

I've tried to fit the model by following steps:

  1. Convert the model to state-space model using Harvey's approach.
  2. Use FKF to kickoff a MLE optim.

The 1st part was pretty natural, I just replace each $\phi$, $\theta$ with their matrix counterparts.

However the 2nd part was tough. FKF's likelihood calculation will 'stall' as soon as Ft become non-positive-semidefinite. (I assume Ft is a covariance matrix)

After that, with doubt in whether my conversion is correct, I tried dse's toSS function, which by it's document:

If the order of the AR polynomial equals or exceeds the MA polynomial (and the input polynomial) then the model is converted by state augmentation.

I have no idea what state augmentation is, and the output matrices looks very different.

I'm seeking for answers and advices, thanks in advance!


Your Answer

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

Browse other questions tagged or ask your own question.