Kwargs for timeseries

I was looking at the stochastic volatility example. I mentioned some arguments like dims="time", which is not explained in the api doc. I wonder if there is a full list of kwargs i can refer to and some detailed usage of those? Also I wonder what init_dist is supposed to do in these models. Thanks for your help!

Dims aren’t unique to timeseries, they’re a way to organize you data and make your code more readable. See here for a full explanation.

init_dist gives the initial conditions for the recursive relationship. For example, in the stochastic volatility model, the variance of the time series follows a Gaussian random walk, so we write the model as:

\begin{align} \sigma_{t+1} &= \sigma_t + \eta_t, \quad \eta_t \sim N(0, \sigma_{\eta}) \\ r_t &\sim T(0, \sigma_t, \nu)\end{align}

This model isn’t closed because you need to tell me what \sigma_0 is. In our case it’s a parameter to estimate, so we need to specify \sigma_0 \sim f(\cdot) where f(\cdot) is the init_dist.

1 Like

Say if I want to fit a series of equity prices using pymc.EulerMaruyama. Could I Specify the initial S_0 to be some fixed value? For example the first column in data. Also, is it possible to use a named distribution as prior and sample the posterior for this initial distribution?

I’ve never personally used the EulerMaruyama API, but there’s an example notebook here that might be useful.

You can sample from a model using all the priors you declare with pm.sample_prior_predictive. See here for discussion.