AbstractΒΆ

This paper describes an object oriented approach to the estimation of time series models using state space methods and presents an implementation in the Python programming language. This approach at once allows for fast computation, a variety of out-of-the-box features, and easy extensibility. We show how to construct a custom state space model, retrieve filtered and smoothed estimates of the unobserved state, and perform parameter estimation using classical and Bayesian methods. The mapping from theory to implementation is presented explicitly and is illustrated at each step by the development of three example models: an ARMA(1,1) model, the local level model, and a simple real business cycle macroeconomic model. Finally, four fully implemented time series models are presented: SARIMAX, VARMAX, unobserved components, and dynamic factor models. These models can immediately be applied by users. [1]

[1]I thank Josef Perktold for many helpful discussions. Financial support from the Google Summer of Code program and the University of Oregon Kleinsorge Fellowship, Department of Economics, is gratefully acknowledged.