A1163
Title: Local blockwise wild bootstrap inference for multivariate time-varying coefficient models
Authors: Isabel Casas - The University of Deusto (Spain) [presenting]
Abstract: Capturing evolving relationships in time series often requires models where coefficients change smoothly over time. While the local blockwise wild bootstrap has been shown to provide valid inference for linear time-varying coefficient models under locally stationary processes, its application to more complex multivariate structures remains limited. The theoretical framework of the local blockwise wild bootstrap is extended to a broader class of models, including time-varying vector autoregressive and seemingly unrelated regression models. By accounting for serial correlation and nonstationarity without requiring the explicit estimation of nuisance parameters, the proposed extension enables the construction of reliable confidence bands for multivariate coefficient curves. A comprehensive simulation study evaluates the performance of the method in these high-dimensional settings. Furthermore, the practical utility of the approach is demonstrated through its implementation in the R package tvReg, providing researchers with robust tools for inference in nonstationary multivariate time series.
