Dynamic linear models

A dynamic linear model (DLM) is a time series model in which a ‘hidden’ model state evolves over time. The state is described as hidden, or latent, as it is never actually observed, we only ever see the data it is assumed to generate. We use these observations to learn about the hidden state, and then utilise the state to forecast future observations. The framework comes from the Bayesian forecasting tradition of West & Harrison.

Note

This page is a stub. A fuller treatment of the modelling framework will follow.

Components

DLMAX builds models by composing components:

• Local level — a random-walk mean.
• Local trend — level plus a slope that may itself drift.
• Fourier / seasonal — periodic structure.
• Regression — external predictors.

Components are added together to form the model’s state structure.