Software Design for Dynamic Linear Modeling

摘要

Dynamic Linear Models (DLM) are used to represent time-varying data. The applications range from mostly data driven to settings in which detailed physical information is incorporated. A key feature of these models is the separation of the description of the underlying state and how observations or measurements are related to the underlying state. This paper describes a design for software that represents a fairly broad portion of the models in the family of DLMs. The software design closely follows the mathematical structure of DLMs. The design has been implemented in Lisp-Stat; and takes advantage of the object model, GUI support, and nonlinear optimization routines that are part of the Lisp-Stat system. DLM-specific features of the software include support for estimation, smoothing and forecasting. Nuisance parameters can be estimated in the software via maximum likelihood. Additionally, models are implemented that include support for both equispaced and non-equispaced time steps. This software has been applied in various settings, including forecasting short-term power loads for the Bonneville Power Administration and modeling time-varying gas retention in the Hanford nuclear waste tanks.