Overview

In order to use pypfilt you will need to define the following things:

• A simulation model that describes the behaviour of the system;

  • See Creating your own model for an example of defining a simulation model.

• Prior distribution(s) and parameter bounds;

  • Specify independent priors for each parameter (see the LatinHypercube sampler, which supports all of the scipy.stats distributions and several other distributions); and/or

  • Read correlated priors for multiple parameters from an external file (see Reading prior samples from external data files).

• Some observations of the system (either real or simulated from the simulation model);

• An observation model that describes how the simulation model and the observations are related;

  • Use the Univariate base class to define observation models in terms of any univariate scipy.stats distribution.

• Simulation and particle filter settings, such as:

  • Choose a time scale (either Scalar or Datetime);

  • Particle filter settings, such as the resampling threshold and whether to use post-regularisation;

  • Select appropriate summary tables to record things of interest; and

  • A seed for the pseudo-random number generator, so that the results are reproducible.

All of this information is collected into a single TOML file, which pypfilt can then use to generate forecasts.

Note

TOML is a simple, easy to read configuration file format, similar to JSON and YAML.