I think there’s something off with how you’re implementing the model. The outputs_info should be a scalar or starters, and it doesn’t make sense to scan in the true data as epsilon_t. epsilon should be the difference between the mean and the prediction.
I worked on it a bit and made a gist here. I thought things were going well, but they got me in the second half. Maybe more eyes on the problem will help.
I fixed it, the examples should be good to go.
EDIT 2: I added a 2nd example with a full ARMA(1,1)-GARCH(1,1) process. Scan is awesome now, all hail @ricardoV94
