Reader question: Ordinary bootstrap resampling treats observations as independent. What if returns, volatility, losses or forecast errors are serially dependent and the resampling itself destroys the structure we are trying to measure?

The Politis–Romano stationary bootstrap preserves dependence by resampling random-length blocks from the original time series. Block lengths follow a geometric distribution, and when a block ends the algorithm jumps to a new random starting point. The resulting bootstrap sequence is stationary under the construction and preserves short-range dependence without forcing every block to have the same length.

This article owns one precise computational job: stationary block resampling for dependent financial time series using geometrically distributed block lengths. It does not own iid bootstrap, moving-block bootstrap, circular block bootstrap, subsampling in general, or every bootstrap-based test.

This is public mathematical and computational education. It is not financial advice, and bootstrap confidence intervals are not automatically valid when the underlying time series is nonstationary or structurally broken.

1. Why iid resampling fails for dependent data

Suppose a return series has volatility clustering. A large absolute return today increases the probability of another large absolute return tomorrow.

An iid bootstrap draws observations independently from the empirical distribution. It preserves the marginal histogram but destroys serial dependence.

If the statistic depends on persistence, drawdown path, overlapping returns, long-run variance, autocorrelation or forecast dynamics, iid resampling can give badly misleading uncertainty estimates.

2. The block-bootstrap idea

Instead of resampling single observations, resample sequences:

(X_i,X_i+1,…,X_i+ℓ−1).

Within each block, the original local time order is preserved.

The central tuning question is block length: blocks must be long enough to preserve meaningful dependence but short enough to generate many approximately independent recombinations.

3. The stationary-bootstrap transition rule

Let the original sample be X₁,…,X_n.

Choose a parameter p with:

0 < p ≤ 1.

To generate each next bootstrap observation:

with probability 1−p, continue to the next observation in the original series;
with probability p, start a new block at a randomly selected original index.

When the original index reaches n, wrap around circularly to 1.

4. Geometric block lengths

Under this rule, block length L has geometric distribution:

P(L=ℓ)=p(1−p)^ℓ−1, ℓ=1,2,…

and:

E[L]=1/p.

The expected block length is therefore the natural tuning parameter.

5. Why random block lengths help

Fixed-length block methods create artificial breakpoints at regular distances.

The stationary bootstrap randomizes those breakpoints. Because every bootstrap position has the same restart mechanism, the resampled sequence is stationary under the construction when the underlying setup is stationary.

The random length also avoids making inference depend too mechanically on one chosen phase alignment.

6. A minimal algorithm

Choose expected block length ℓ̄.
Set p=1/ℓ̄.
Choose a starting original index uniformly from 1,…,n.
Copy that observation into the bootstrap series.
Draw U~Uniform(0,1).
If U<p, choose a fresh random index.
Otherwise advance to the next original index, wrapping n→1.
Repeat until n bootstrap observations are produced.
Compute the target statistic.
Repeat the entire resampling B times.

7. What dependence is actually preserved?

Within a block, the original local dependence is preserved exactly. Across block boundaries, it is broken.

The probability that two bootstrap observations k steps apart remain in the same block is approximately:

(1−p)^k.

Thus the method preserves short-lag dependence strongly when expected blocks are long, while progressively weakening very long-lag dependence.

8. Block length is the central bias–variance choice

If blocks are too short:

serial dependence is broken too aggressively;
long-run variance is understated;
path statistics can look too independent.

If blocks are too long:

there are fewer effectively independent recombinations;
bootstrap variability can become noisy;
resamples may resemble the original series too closely.

A strong implementation therefore reports block-length sensitivity rather than hiding one default.

9. Automatic block-length selection

Politis and White developed automatic block-length selection procedures, later corrected and refined in the literature, to choose block sizes from the estimated dependence structure rather than pure convention.

Automatic selection is useful, but it does not remove the need to inspect whether the estimated dependence regime is stable.

10. Applications in finance

confidence intervals for mean returns under weak dependence;
Sharpe-ratio uncertainty;
forecast-comparison statistics;
long-run variance estimation;
strategy data-snooping tests;
risk measures based on dependent paths;
drawdown or persistence statistics;
residual bootstraps for time-series models.

The technique appears naturally beside White Reality Check / Hansen SPA, where dependent resampling can be crucial to preserving time-series structure in performance differentials.

11. Stationarity is not a small assumption

The method is designed for stationary weakly dependent time series under appropriate regularity conditions.

If the data contain a permanent regime break, a 2008 crisis period, a structural market redesign and a calm modern period, random blocks can splice fundamentally different regimes together as though they belonged to one stable process.

Wintour House test: before polishing the resampling, decide whether the series deserves to be treated as one stationary edition at all.

12. Structural breaks

Suppose volatility permanently doubles halfway through the sample.

A stationary bootstrap can combine blocks from both regimes and create synthetic paths that jump repeatedly between old and new volatility states.

That may not approximate any realistic data-generating process.

Falsifier: run change-point or regime diagnostics first. If the sample is structurally segmented, use regime-aware resampling or estimate within stable windows.

13. Long memory

If dependence decays very slowly, practical finite block lengths can truncate important structure.

The stationary bootstrap is not a universal solution for genuine long-memory processes.

Falsifier: inspect autocorrelation across scales and compare inference under increasing block lengths or methods designed for long memory.

14. Multivariate financial series

When resampling several assets jointly, all variables must use the same block indices if cross-sectional dependence is to be preserved.

Resampling each asset independently would preserve each marginal series while destroying contemporaneous correlation.

Falsifier: compare original and bootstrap cross-correlation/covariance matrices.

15. Missing observations and calendars

A financial panel with holidays, delistings or asynchronous dates needs a canonical calendar before block resampling.

Blocks preserve whatever data plumbing they are given. They do not repair inconsistent calendars.

16. Inputs and outputs

Inputs can include:

ordered time series or multivariate matrix;
expected block length;
number of bootstrap replications B;
random seed;
target statistic;
stationarity/regime diagnostics;
calendar/missing-data policy.

Outputs can include:

bootstrap statistic distribution;
confidence intervals;
bootstrap p-values;
block-start frequency;
realized block-length distribution;
sensitivity across ℓ̄;
dependence-preservation diagnostics;
effective number of blocks.

17. Evidence polarity

Evidence for confidence includes:

bootstrap autocorrelation resembles original short-lag dependence;
results stabilize across a reasonable block-length band;
simulation coverage is close to nominal;
cross-sectional dependence is preserved in multivariate resampling;
structural-break tests do not contradict the stationarity assumption;
B is large enough for stable tail quantiles/p-values.

Evidence against confidence includes:

inference changes dramatically with ℓ̄;
long-lag dependence is materially truncated;
sample contains clear breaks;
bootstrap paths mix impossible regimes;
multivariate correlations collapse because series were resampled independently;
too few effective blocks create unstable bootstrap tails.

18. Counterexample: block length 1

If p=1, expected block length is one and the stationary bootstrap collapses to iid resampling.

Falsifier: unit-test this limiting case. If the algorithm does not reproduce iid bootstrap behavior at p=1, the block transition logic is wrong.

19. Counterexample: enormous block length

If p is extremely small, a bootstrap sample may contain only one or two long blocks.

The resample becomes a near-rotation of the original data and delivers very little new recombination.

Falsifier: monitor the distribution of the number of blocks per bootstrap replication.

20. Counterexample: regime-switching volatility

A single stationary bootstrap may splice calm and crisis blocks in unrealistic alternation.

Falsifier: compare regime-conditioned bootstrap distributions with the global stationary-bootstrap result.

21. Counterexample: tail-risk statistic

A 99.9% tail statistic requires many bootstrap replications and enough genuinely distinct extreme-event configurations.

Block resampling cannot manufacture tail information that the sample never contained.

Falsifier: report the number of tail-driving observations and bootstrap contribution concentration.

22. Stationary bootstrap versus moving-block bootstrap

Moving-block bootstrap uses fixed-length overlapping blocks. The stationary bootstrap uses random geometric lengths and circular continuation.

Both preserve local dependence; the stationary bootstrap removes deterministic block boundaries and has a stationary resampling law under the construction.

23. Stationary bootstrap versus circular block bootstrap

Circular block bootstrap wraps the original sample but usually retains fixed block length. Stationary bootstrap also wraps, but restarts occur randomly according to p.

24. Stationary bootstrap versus Newey–West

Newey–West HAC estimates an asymptotic covariance analytically.

Stationary bootstrap approximates the sampling distribution by dependent resampling. It can support nonlinear statistics for which an analytic HAC formula is inconvenient.

26. What would falsify confidence?

Confidence should be withdrawn if reasonable block lengths give incompatible conclusions; if structural breaks violate the stationary target; if simulation coverage fails; if dependence diagnostics show material structure beyond the effective block horizon; or if bootstrap tail results are driven by a tiny number of original observations.

27. Verification and update triggers

Preserve the exact sample, ordering, block parameter p, automatic-selection method if used, random seeds, replication count, multivariate indexing and target statistic.

Revalidate when market regime changes, sample length changes materially, dependence persistence changes, the statistic changes, the series becomes multivariate, or bootstrap software changes implementation conventions.

Research anchors

Dimitris N. Politis and Joseph P. Romano, The Stationary Bootstrap, Journal of the American Statistical Association, 1994.
Dimitris Politis and Halbert White, Automatic Block-Length Selection for the Dependent Bootstrap, Econometric Reviews.

Educational boundary: block bootstrap preserves dependence only to the extent that the original sample is a stable representation of the process. Resampling cannot repair a structural regime error.

Small Group Tutorials

How Politis–Romano Stationary-Bootstrap Algorithms Resample Dependent Financial Time Series: Geometric Blocks, Dependence Preservation, Block-Length Choice and Regime Failure