Automatic Identification and Forecasting of Structural Unobserved Components Models with UComp

Abstract

UComp is a powerful library for building unobserved components models, useful for forecasting and other important operations, such us de-trending, cycle analysis, seasonal adjustment, signal extraction, etc. One of the most outstanding features that makes UComp unique among its class of related software implementations is that models may be built automatically by identification algorithms (three versions are available). These algorithms select the best model among many possible combinations. Another relevant feature is that it is coded in C++, opening the door to link it to different popular and widely used environments, like R, MATLAB, Octave, Python, etc. The implemented models for the components are more general than the usual ones in the field of unobserved components modeling, including different types of trend, cycle, seasonal and irregular components, input variables and outlier detection. The automatic character of the algorithms required the development of many complementary algorithms to control performance and make it applicable to as many different time series as possible. The library is open source and available in different formats in public repositories. The performance of the library is illustrated working on real data in several varied examples.

Keywords

• unobserved components models
• state space models
• kalman filter
• fixed point smoother
• maximum likelihood
• r
• matlab
• octave