STEADY
ID. 887220
Principal investigator. Matteo Iacopini
Supervisors: * [Siem Jan Koopman](https://sjkoopman.net), Professor of Econometrics at [Vrije Universiteit Amsterdam](https://vu.nl/en/about-vu/faculties/school-of-business-and-economics) * [André Lucas](https://personal.vu.nl/a.lucas/), Professor of Financial Econometrics at [Vrije Universiteit Amsterdam](https://vu.nl/en/about-vu/faculties/school-of-business-and-economics)
Host institution. [Vrije Universiteit Amsterdam](https://vu.nl/en/about-vu/faculties/school-of-business-and-economics)
Funding. EUR 176,000 from the European Commission (EC)
Project duration. October 2020 - September 2022
Summary. Modern economic analyses require new models to study increasingly fine-grained interrelations based on increasingly complex data sources. Early dynamic economic analyses have mostly been limited to only studying univariate time series, which can be represented as a single sequence (or vector) of values. Most contemporary analyses use more complicated data with both time series and cross-sectional dimensions, such as panels of key macroeconomic figures, for many countries over time. Such data can be represented as a (2-dim) matrix. Recently, more complex data structures have rapidly emerged, requiring higher dimensional storage objects. As an example, a data set consisting of a time series (1st dimension) of the exposures of banks (2nd dim) to other banks (3rd dim) in several markets (bonds, equity; 4th dim) and for different maturities (5th dim). The storage object for such high-dimensional data sets is a generalization of a matrix, called a tensor. Models for tensor data have applications to policy-relevant questions for central banks and financial regulators, including forecasting multi-country, multi-market interest rate term structures for the evaluation of monetary policy effectiveness, and nowcasting multi-country economic activity in the heterogeneous European context. Tensor data are highly topical, however, in econometrics their use and the development of tensor models is very scant and almost exclusively limited to static tensors. The **STEADY** project fills this gap by developing novel statistical methods for time series of tensor data that account for the typical non-linear and dynamic features of economic data in a computationally feasible way. The project has two main research directions. One is the development of a general class of dynamic time-series models, which merge the linear tensor time series literature and the score-driven time-varying parameter approach based on the Generalized Autoregressive Score (GAS) model. The other contribution consists in the development of a new tensor-based compression technique for many economic time series, the tensor dynamic factor model.