Building Time-Series Models With Known Data Structure

This post describes how Gaussian Processes can be used to build time-series models for cases where we have domain knowledge of certain properties of the underlying time-series (seasonality, differently sources of data). The data are carbon dioxide measurements from ice cores as well as atmospheric readings from Mauna Loa (a volcano in Hawaii) with which we use PyMC to:

• Fit fully Bayesian Gaussian Processes with NUTS
• Model inputs whose exact locations are uncertain (uncertainty in ‘x’)
• Design a semiparametric Gaussian process model
• Build a changepoint covariance function / kernel
• Definine a custom mean and a custom covariance function

For the full example, see:
Example: Mauna Loa CO2 continued