Known Limitations¶

This public research release is designed for reproducible benchmarking experiments, not definitive estimator rankings.

Estimators¶

No estimator is currently labelled reference-grade.
Hurst-scaling proxies and ARFIMA long-memory parameter estimates are related only under specific model assumptions; they should not be merged into a single public ranking without an explicit protocol decision.
Scale-window, bandwidth, wavelet-band, and bootstrap settings can materially change results.
Aggregation-based estimators (AbsoluteMoment, Variance, and VarianceResidual) are especially sensitive to block-size support, centring, finite-sample effects, and contamination.
Experimental wavelet estimators are included for comparison, but should not anchor public claims.
Data-driven estimators are supervised baselines whose validity depends on the declared training distribution; they are not distribution-free LRD estimators.
CNN/LSTM baselines require optional PyTorch dependencies and are not part of the lightweight RF/SVR smoke suite.

Generator checks cover ordering and sanity properties, not full distributional conformance.
Finite-sample behaviour can differ substantially from asymptotic expectations.
Contamination suites are controlled stress tests, not exhaustive models of real measurement artefacts.

Observational mode has no ground truth, so accuracy, coverage, and false-positive LRD claims are not available.
Truth-free metrics describe estimator agreement, stability, runtime, and missingness. They do not prove long-range dependence in the source data.
CSV and inline observational loaders currently support simple single-column series ingestion only.

Leaderboards are summaries of declared component metrics and weights. They are not universal estimator rankings.
Failure maps and uncertainty summaries are diagnostic artefacts. They should be read alongside raw metric exports and manifest settings.