I'm taking a time series course and am learning about exchangeable time series form of dynamic linear models (DLMs). This is given by: \begin{align*} \mathbf{y}_t' &= \mathbf{F}_t'\boldsymbol{\Theta}_t + \boldsymbol{\nu}_t', \qquad \boldsymbol{\nu}_t \sim \mathcal{N}(0, v_t\mathbf{V})\\ \boldsymbol{\Theta}_t &= \mathbf{G}_t\boldsymbol{\Theta}_{t-1} + \boldsymbol{\Omega}_t, \qquad \boldsymbol{\Omega}_t \sim \mathcal{N}(0, \mathbf{W}_t, \mathbf{V}) \end{align*} $\boldsymbol{\Omega}_t$ comes from a matrix-variate normal distribution.

All parameters except $\mathbf{U}$ and $\mathbf{V}$ can be fit using forward filtering backward sampling (FFBS). In order to fit $\mathbf{U}$ and $\mathbf{V}$, we must do a Gibbs sampling step.

My question is: what is the difference between FFBS and the Kalman filter? What other ways are there to fit DLMs? For example, I have heard of the particle filter - is this something that can be used in this context? And finally, is it always possible to do Gibbs sampling (perhaps with Metropolis-Hastings if prior distributions are not conjugate) to fit a DLM?


1 Answer 1


The Kalman filter is the "forward filtering" part of FFBS, while the "backward sampling" part provides a draw from the joint distribution for the $\Theta_t$ for all $t$.

All the other ways of performing statistical parameter estimation, e.g. maximum likelihood, can be used for DLMs. Yes a particle filter could be used here, but since the model is linear and Gaussian, i.e. a DLM, you won't be gaining anything. Particle filters are better when you have non-linear or non-Gaussian models and thus cannot perform FFBS.

Yes, it is possible to perform Metropolis-within-Gibbs for any Bayesian model.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.