Block Bootstrapping for Synthetic Return Series in Stress Testing

Summary

The document asks how to create synthetic return series for sensitivity analysis when historical data are limited and the true distribution of market returns is uncertain. The proposed method pools daily returns from multiple instruments within one asset class, samples returns with replacement, and recombines them into new series. Restricting the pool to a single class is intended to preserve broad asset characteristics.

The accepted response recommends investigating block bootstrapping as a way to retain some serial dependence that ordinary independent resampling can erase. The exchange gives no implementation details, comparison, or performance evidence, so it does not establish which block size or sampling design is appropriate. Synthetic series remain dependent on the historical pool and resampling assumptions; they cannot represent market outcomes absent from that sample. The discussion is a concise methodological pointer rather than a complete stress-testing framework.

Key ideas

• Independent resampling of individual returns can disrupt serial patterns in the original data.
• Block bootstrapping resamples groups of adjacent observations to preserve some time dependence.
• The proposed sampling pool uses instruments from the same asset class to maintain broad comparability.
• Synthetic returns inherit limitations from the historical data and the chosen resampling method.

Tags

# Generating Return Streams for stress testing


# Generating Return Streams for stress testing












There is never enough market data for testing. And sampling from user defined distribution is a hotly debated subject as which distribution does the market really go with?

There are many ways to generate synthetic data series for sensitivity testing but the methods should be sound. Does the follow method warrant further investigation? Extract daily returns from n different instruments. Mix and match them from sampling with replacement. After each new series, mix them again before generating another one. To avoid major differences in the underlining characteristics, the pool of return should come from one asset class. For example, synthetic data from stocks should come from the pool of equities.

Does this make sense? Or is it too naive?

Thanks,

## Answer by pat (score 1, accepted)

https://quant.stackexchange.com/a/4257

You can look into block bootstrapping as one alternative to mitigate loss of any serial dependency effects.

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.