Open source · Python 3.11+ · Reproducible research

Benchmark long-range dependence with the assumptions in view.

Compare estimators on known synthetic processes, controlled contamination, and observational signals. lrdbench keeps the target estimand, uncertainty, failures, and provenance visible from run to report.

Run a first benchmark Explore the protocol

Version 2.0.0 · View source on GitHub

A benchmark run, made inspectable

01Declare the sourceGenerator or observed series

02Fit estimatorsExplicit target estimands

03Evaluate by modeMetrics, failures, provenance

One framework · Three modes

Choose the question your data can answer

Each mode has its own evaluator. Truth-based metrics belong only where a target value is declared.

01 / KNOWN TRUTH

Ground truth

Benchmark on canonical synthetic series with a declared target estimand.

Bias · MAE · RMSE · coverage

Explore ground truth

02 / CONTROLLED CHANGE

Stress test

Measure what happens when contamination challenges an estimator's assumptions.

Drift · degradation · validity

Explore stress tests

03 / NO BENCHMARK TRUTH

Observational

Inspect stability on biomedical or user-provided signals without claiming a true Hurst value.

Window · preprocessing · resampling

Explore observational data

First run

From install to a validated report

Start with the packaged smoke suite. Preview its record-by-estimator grid before running it, then validate the output directory printed by the CLI.

Follow the full quickstart →

terminal

python -m pip install "lrdbench[reports]"
lrdbench run smoke_ground_truth --dry-run
lrdbench run smoke_ground_truth
lrdbench validate-output reports/<run_id>

Find your route

Go deeper

Interpret a resultEstimands, metric meanings, and what a leaderboard can support. Compare methodsBrowse estimator families, targets, and implementation status. Reproduce the researchFollow the audited confirmation benchmark and its evidence trail. Extend lrdbenchImplement an estimator using the public contract.