Joint latent-Gaussian dynamic regressions, with block elliptical slice inference.
Gaussian prior scales and NB dispersion are fixed model specifications.
Group coefficients equal an uncertain population coefficient plus a Gaussian
deviation; group random walks can additionally share a common random walk.
Composable structural models with exact Gaussian FFBS and conjugate variance
updates. Student-t observations use Gamma precision mixtures with fixed df.
Initial state priors are independent of innovation variances; x[-1] is included
in every state draw so every transition contributes to the variance update.
Regression columns on an integer calendar: known-date interventions,
piecewise-linear trend changepoints, Fourier terms, and named columns the
caller supplies.