Unable to replicate "Stochastic Volatility model" in PyMC v5.20

[abdullahau]

I have been trying to run this notebook with the following packages:

Python implementation: CPython
Python version       : 3.12.8
IPython version      : 8.31.0

pytensor: 2.26.4
aeppl   : not installed
xarray  : 2025.1.2

numpy     : 1.26.4
arviz     : 0.20.0
matplotlib: 3.10.0
pandas    : 2.2.3
pymc      : 5.20.0

Running the function model_to_graphviz, I notice that the model is set-up exactly as intended. However, this part of the code returns extremely different results:

with stochastic_vol_model:
    idata = pm.sample_prior_predictive(500, random_seed=rng)

prior_predictive = az.extract(idata, group="prior_predictive")

Note that similar to the notebook, I have set the random seed state as rng = np.random.RandomState(1234)

Additionally, while trying to run this bit of code with the RandomState:

with stochastic_vol_model:
    idata.extend(pm.sample(random_seed=rng))

posterior = az.extract(idata)
posterior["exp_volatility"] = np.exp(posterior["volatility"])

I get the error:

TypeError: Cannot create a random Generator from a RandomStream object. Please provide a random seed, BitGenerator or Generator instead.

While setting the random generator as rng = np.random.default_rng(1234), I am unable to produce the same results.

[ricardoV94]

Yes we should use generator not randomstate.

If I remember correctly, the "results" of that notebook are very dependent on the seed, you can try a different value, but they are not very stable. And if they depend on the seed they are not very meaningful either.