Estimate Ornstein-Uhlenbeck Parameters with Euler Maruyama

Questions
4

ADVI ignore the posterior correlation, which in many cases gives sub-optimal results.
As general tips, I would suggest the following:

  1. Not using ADVI. NUTS is more or less the best thing we have for models it can be sampled. If you want a VI inference that scale a bit better FullRankADVI is the first thing to try
  2. Make sure ADVI/fullrankADVI did converge. For example, you can find in this post some discussion about making sure the fitting converge: Poor Accuracy of ADVI for Linear Regression
1 Like