Closed steve-the-bayesian closed 5 years ago
Should be simple enough. Just define the likelihood using the Kalman filter, and do posterior sampling via SliceSampler (or bouncy, etc).
The "Anything Samplers" approach doesn't work particularly well. The KF is slow enough that likelihood evaluation is a bottleneck.
Closing this for now. We have the models. The samplers are not particularly fast, but they're right.
Now that the state space form for ARMA and ARIMA models is clear, add them as state space models. They don't fit as bsts state models, but they can be their own thing.