Here’s how I think we could set up the model:
- Create a new timeseries class similar to MVGuassianRandomWalk
- Modify the logp function to accommodate the bivariate posterior of Y and V. In the example of the log-vol model above, the logvol follows an AR1 process so the code will look similar to the AR1 class
- Once the logp function is defined correctly, pass in the observed return series concatenated with the latent state
Alternatively, we can take the approach outlined in this paper
In section 2.3 the authors consider the same log-vol model with correlated errors and adopt a Metropolis scheme