The focus of the project is forecasting of time series in high-dimensional settings. This situation is increasingly common and requires new approaches as many common methods do not work well when dimensions, both the number of time points and the number of variables, are high.
In this project, we will study the performance of a non-linear factor model that we are developing as a way of attacking the dimensionality issue. Furthermore, we will extend newly-proposed Bayesian methods for forecasting with large multivariate time series models to the case where the data is sampled on different frequencies. In our work, we consider monthly and quarterly data on more than 130 macroeconomic time series for more than 50 years. The final aim of the project is to compare our extensions to existing methods for the high-dimensional setting.