Extrapolating covariate for hierarchical model about kfas

[helske]

First possibility that comes to my mind would be the use of so called "seemingly unrelated time series model", where you use multivariate model for both Y and X, i.e. something like this:

model <- SSModel(log(cbind(MyY, MyX)) ~ 
  SSMtrend(2, Q = list(matrix(NA, 2, 2), matrix(NA, 2, 2))), 
  H = matrix(NA, 2, 2), data = MyData)

fit <- fitSSM(model, inits = rep(0, 9), method = "BFGS")
sim <- simulateSSM(fit$model, "obs", nsim = 100)
ts.plot(exp(sim[,1,]))

Essentially you are modelling both the observations Y and the covariate values X, instead of treating X fixed, and here the relationship between Y and X is incorporated into the covariance terms in H and Q.

[andrewbcooper]

Brilliant! It never occurred to me to incorporate the relationship between X and Y in the covariance terms. And yes, treating those future Xs as "known" in the model for Y was bothering me, but I couldn't figure out how to avoid it. This is perfect! And it takes 1/60th of the processing time! Thanks!