Abstract

The first estimates of macro-economic statistics are often based on partial and/or alternative information. Integration of high-frequency data, of soft data or of hard constraints, adjustment for events like the Covid or the Ukrainian crisis… are recurring problems in that context. Various model-based solutions could be considered, ranging from small temporal disaggregation models to medium-sized dynamic factor models.
Mixed-frequency state space models constitute a versatile framework to formalize them. However, their complexity and the time needed to develop them is often a barrier for most statisticians, especially in the context of fast estimates of official statistics.
To simplify the use of such models, we have developed a collection of high-performance Java components that represent common state space blocks, like trend components, various (high-frequency) seasonal components, (time-varying) regression coefficients, vector autoregression, etc. Those blocks can be easily combined or modified, for instance by aggregation or by change of frequency through “cumulators”, to form more complex models.
The usual Kalman algorithms can be applied for likelihood evaluation, filtering and smoothing. The framework hides most of the technical details, including the initialization of the model, the univariate treatment of multi-variate models or the use of the correct likelihood function. Corresponding R packages are also available for end-users.
We will present the framework and two actual uses in R. The first one is a model-based approach of the Denton method, using time-varying innovation variances to tackle the Covid crisis; that solution is used in the compilation of some series of the Belgian Quarterly National Accounts. The second one is the integration of soft data coming from business surveys in a medium-sized dynamic factor model for the nowcasting of the Belgian GDP.