Bootstrap Dependence Diagnostic¶

Resampling diagnostic — instantiates Stochastic Process Modeling and Validation

Puts honest, dependence-aware error bars on a statistic by resampling the data in blocks that preserve its dependence unit rather than as if points were independent.

A Bootstrap Dependence Diagnostic quantifies the uncertainty of a statistic — or builds a null distribution for a test — by resampling the observed data many times and recomputing, but in a way that preserves the dependence structure: it draws contiguous blocks rather than individual points, so the resulting error bars are not fooled by autocorrelation. Its defining idea among the siblings is that its whole correctness hinges on keeping the dependence unit intact — the naïve independent-points bootstrap silently understates uncertainty for dependent data, and this mechanism exists to fix exactly that. Unlike the model mechanisms, it builds no process law; it wraps around an estimate to give it honest error bars.

Example¶

An operations analyst has eighteen months of a service desk's daily 95^th-percentile wait time and needs a confidence interval for the yearly average, to judge whether an SLA is genuinely being met. The days are strongly autocorrelated — busy weeks cluster — so an ordinary bootstrap that resamples individual days would treat them as independent and return an implausibly tight interval, certifying the SLA as safely met.

A moving-block bootstrap instead resamples contiguous weekly blocks, preserving the within-week dependence.^[1] The resulting interval is realistically wider, and the SLA claim that looked comfortable under the naïve interval turns out to be marginal. The block length is chosen long enough to capture the autocorrelation but short enough to leave many blocks to shuffle — the single choice on which the whole diagnostic rests.

How it works¶

Identify the dependence unit that must stay intact — a block length matched to the autocorrelation, or a natural unit such as a subject or a season.
Resample the units with replacement to build many synthetic datasets, recompute the statistic on each, and read the spread across replicates as its sampling distribution.
Turn that distribution into an answer — a confidence interval, or, resampling under a null, a hypothesis test that respects the dependence instead of assuming it away.

Its distinctive move is negative: it does not model the process, it protects an estimate from the dependence the raw data carry.

Tuning parameters¶

Block length — the core dial. Too short and you break the dependence (intervals too tight); too long and you have too few blocks (intervals noisy and too wide).
Block scheme — moving, non-overlapping, circular, or stationary (random-length) blocks. Trades bias against variance and edge effects.
Resampling unit — blocks in time versus whole clusters or subjects, matched to where the dependence actually lives.
Number of resamples — more replicates shrink the Monte-Carlo error of the interval, at compute cost.
Studentization — whether to bootstrap a pivotal (studentized) statistic for more accurate interval coverage.

When it helps, and when it misleads¶

Its strength is giving distribution-free uncertainty for statistics that have no clean analytic standard error, and — done with blocks — being one of the few honest ways to get error bars under dependence. It is an excellent reality check on a suspiciously tight interval.

Its failure modes track its one assumption. It is only as good as the preserved dependence unit: the wrong block length — especially too short — reproduces the independent-points mistake it exists to avoid, and it cannot manufacture information the sample lacks, so near the tails or with very few effective blocks it is unreliable. It also assumes the dependence is roughly stationary across the sample. The classic misuse is running an ordinary independent-points bootstrap on serially dependent data (or quietly picking a comfortably short block) to produce a reassuringly narrow interval that certifies a decision already made. The discipline that keeps it honest is to diagnose the dependence first, set the block length from the autocorrelation, and sanity-check the interval against a longer block — if it widens materially, the dependence was under-preserved.

How it implements the components¶

Bootstrap Dependence Diagnostic fills the uncertainty-under-dependence slice of the archetype's validation machinery:

stochastic_diagnostic_and_falsification_suite — it is a diagnostic: a resampling procedure that produces sampling distributions, confidence intervals, and dependence-respecting null tests.
stochastic_parameter_and_uncertainty_register — its output populates the uncertainty register with honest error bars and interval estimates for statistics and parameters.
stochastic_dependence_structure — it must faithfully preserve the dependence unit; the block scheme is a direct, operational statement about the process's dependence.

It does NOT specify or fit the process law itself — the model mechanisms such as the Autoregressive Stochastic Sequence Model do that — nor does it validate whole generated paths on path features (Held-Out Path-Feature Check).

Instantiates: Stochastic Process Modeling and Validation — supplies the dependence-aware uncertainty quantification the archetype's estimates depend on.
Consumes: any fitted model or estimator whose sampling uncertainty it quantifies — for example an Autoregressive Stochastic Sequence Model fit or a summary statistic of a raw series.
Sibling mechanisms: Held-Out Path-Feature Check · Autoregressive Stochastic Sequence Model · Markov Chain Process Model · Poisson Event-Process Model · Change-Point and Regime-Switching Model · Empirical Distribution and Increment Fit · Gaussian Process Function Model · Random-Walk and Diffusion Model · Renewal and Point-Process Model · Stochastic State-Space Model · Residual Independence and Whiteness Test · Probability Integral Transform Check · Posterior or Simulation Predictive Check · Proper Scoring Rule Comparison · Rare-Event Stress Simulation

Notes¶

The chosen block length is itself an implicit dependence model. When an explicit one already exists — an autoregression, a Markov chain — a model-based (parametric) bootstrap that simulates fresh data from the fitted model is the alternative, trading the block bootstrap's nonparametric robustness for efficiency when you trust the model.

References¶

[1] The block (moving-block) bootstrap resamples contiguous stretches of the series rather than individual observations, so that within-block dependence is carried into each synthetic dataset. Under positive autocorrelation the ordinary independent-points bootstrap underestimates variance; resampling blocks is the standard correction. ↩