Preserving Cross-Series Dependence in Multivariate Bootstrap Samples

Summary

The document describes a block-bootstrap workflow for multivariate time series. The author estimates a block size for each series, selects the largest estimate, bootstraps one series with that block size, and uses the resulting index to reconstruct the other series. The question is whether the meboot package can replace this approach while preserving correlations across series.

The central methodological issue is dependence: resampling one series independently may not retain the timing and cross-series relationships needed for a coherent multivariate sample. The document raises that concern but contains no answer, comparison, or evidence showing how meboot behaves in a multivariate setting. It therefore serves as a useful statement of a bootstrap design problem rather than a recipe. Any proposed replacement would need to explain how indices or resampled observations are shared across series and how both within-series time structure and cross-series dependence are evaluated.

Key ideas

• The described workflow uses a common block size and bootstrap indices to reconstruct multiple series.
• A multivariate resampling method must account for cross-series dependence as well as each series’ time structure.
• Resampling a single series does not by itself establish that correlations across series are preserved.
• The document raises the meboot question but provides no solution or empirical comparison.

Tags

# Alternative to Block Bootstrap for Multivariate Time Series


# Alternative to Block Bootstrap for Multivariate Time Series












I currently use the following process for bootstrapping a multivariate time series in R:

- Determine block sizes - run the function b.star in the np package which produces a block size for each series

- Select maximum block size

- Run tsboot on any series using the selected block size

- Use index from bootstrap output to reconstruct multivariate time series

Someone suggested using the meboot package as an alternative to the block bootstrap but since I am not using the entire data set to select a block size, I am unsure of how to preserve correlations between series if I were to use the index created by running meboot on one series. If anyone has experience with meboot in a multivariate setting, I would greatly appreciate advice on the process.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.