Port timeseries distributions to V4

[twiecki]

import pymc3 as pm
import numpy as np

returns = pd.read_csv(pm.get_data("SP500.csv"), index_col="Date")
returns["change"] = np.log(returns["Close"]).diff()
returns = returns.dropna().astype('float32')
returns.head()

with pm.Model() as model_pymc3:
    step_size = pm.Exponential("step_size", 10.)
    volatility = pm.GaussianRandomWalk("volatility", sigma=step_size, 
                                       size=returns.shape[0], 
                                       init=pm.Normal.dist(0, step_size))
    nu = pm.Exponential("nu", 0.1)
    obs = pm.StudentT(
        "obs", nu=nu, sigma=np.exp(volatility), observed=returns["change"]
    )

yields:

TypeError                                 Traceback (most recent call last)
<ipython-input-11-3f1576c8cf84> in <module>
      1 with pm.Model(check_bounds=False) as model_pymc3:
      2     step_size = pm.Exponential("step_size", 10.)
----> 3     volatility = pm.GaussianRandomWalk("volatility", sigma=step_size, 
      4                                        size=returns.shape[0],
      5                                        init=pm.Normal.dist(0, step_size))

~/projects/pymc/pymc3/distributions/distribution.py in __new__(cls, name, *args, **kwargs)
    161         transform = kwargs.pop("transform", UNSET)
    162 
--> 163         rv_out = cls.dist(*args, rng=rng, **kwargs)
    164 
    165         return model.register_rv(rv_out, name, data, total_size, dims=dims, transform=transform)

TypeError: dist() missing 1 required positional argument: 'dist_params'

[twiecki]

Seems related to pm.GaussianRandomWalk because if I comment out the init it leads to the same error.

[ricardoV94]

Is this related to check_bounds and/or V4?

[twiecki]

This is on V4, worked fine on V3.

[ricardoV94]

I could be wrong but I don't think the timeseries were refactored yet

[twiecki]

Unrelated to check_bounds. I updated the code to make it runnable.

[twiecki]

I see.