KernelICA for Identification of Structural Innovations in Multivariate Time Series Models

Details

Supervisors

Hafner, Christian

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

This master's thesis explores how Kernel Independent Component Analysis (KernelICA) can solve the identification problem in Structural Vector Autoregressive models. Identifying structural shocks is essential for accurately interpreting and forecasting multivariate time series data. We start by discussing the basics of SVAR models and the difficulties in identifying these shocks. Then, we introduce KernelICA as a new method of identification and compare it to traditional Independent Component Analysis techniques. Using Monte Carlo simulations, we test how well KernelICA identifies independent shocks and compare its performance with existing methods. An empirical analysis using data from Blanchard and Perotti shows KernelICA's practical use. Our results indicate that KernelICA is a strong alternative for identifying structural innovations in SVAR models.