Block bootstrap is a resampling method for time series.

Block bootstrap is a resampling method for time series. Instead of drawing individual observations independently, it draws contiguous blocks of observations and stitches those blocks together into a new sample. The purpose is to preserve short-range serial dependence that ordinary i.i.d. bootstrap breaks by construction.

This distinction matters in financial data because returns, signals, and strategy equity changes are often not independent across time. Volatility clustering, regime persistence, signal decay, and position overlap can all create dependence between nearby observations. When a resampling method ignores that structure, the synthetic series can differ materially from the dependence properties of the original series.

The mechanism is simple. Standard bootstrap samples single observations with replacement. That preserves the marginal distribution reasonably well, but it destroys the temporal ordering that carries autocorrelation and related dependence. Block bootstrap changes the sampling unit from a single observation to a short run of consecutive observations. Because each draw keeps local ordering inside the block, the resampled series can retain some of the serial structure present in the original data.

The main design choice is block length. Short blocks preserve less dependence because the ordering is only kept over a small window. Long blocks preserve more dependence, but they also reduce the amount of recombination across the sample. In practice, block length governs the tradeoff between preserving temporal structure and generating enough variation in the resampled paths.

For strategy validation, that tradeoff affects the interpretation of performance uncertainty. If a statistic is sensitive to serial correlation, a plain i.i.d. bootstrap can understate or misstate sampling uncertainty because it treats dependent observations as if they were interchangeable. A block bootstrap is better aligned with time-series data because it resamples in a way that respects local dependence.

This is especially relevant when evaluating risk-adjusted performance metrics. Sonar’s glossary defines the deflated Sharpe ratio as a Sharpe-ratio-based measure adjusted for multiple testing and non-normality concerns in strategy evaluation. If the resampling scheme used to study a strategy breaks the dependence structure of returns, then the distribution used to assess such metrics may not reflect the behavior of the original strategy series as well as a dependence-preserving method would. That is the practical reason block bootstrap is commonly preferred over plain bootstrap for dependent financial series.

Sonar’s strategy-validation research emphasizes that rigorous validation must account for backtest overfitting and for distortions introduced by evaluation procedures. Sonar’s backtest-overfitting audit tool is designed to examine how apparent in-sample results translate out of sample under more robust validation workflows. In that context, a block bootstrap can be understood as a resampling choice that is more consistent with serially correlated strategy data than an i.i.d. bootstrap, because it retains contiguous structure rather than randomizing it away. Block bootstrap is designed to retain local serial dependence better than standard i.i.d. bootstrap because it resamples contiguous blocks instead of isolated observations.

Covered in depth in the Strategy validation & overfitting pillar hub.