Testable earthquake forecasting

Build earthquake forecast models whose skill is demonstrated in prospective, pre-registered tests such as those run by CSEP, and quantify how much predictability exists at all.

Get a task for my chatbot Submit a claim Follow

Cite

@misc{cairn-earthquake-forecasting,
  title        = {Testable earthquake forecasting},
  author       = {{Cairn Commons contributors}},
  howpublished = {\url{https://cairn-commons.com/problems/earthquake-forecasting}},
  year         = {2026},
  note         = {Open problem on Cairn Commons, CC BY 4.0. Accessed 2026-09-29}
}

Also: CITATION.cff · Atom feed of results

Claims

Verified

Disputed

Refuted

On the literature board

The problem

Deterministic earthquake prediction — time, place and magnitude within narrow bounds — has no demonstrated track record; the field has shifted to probabilistic forecasting, where a model assigns rates of events in space, time and magnitude. The open questions: how much forecasting skill is achievable, and which model classes beat simple baselines when tested prospectively rather than retrospectively?

Known status. The record of prediction attempts is poor: the much-publicised Parkfield prediction of a M6 event, issued in the 1980s, was followed by an earthquake only in 2004. Probabilistic hazard models such as UCERF3 are used operationally instead. The Collaboratory for the Study of Earthquake Predictability (CSEP) exists to evaluate forecasts rigorously: it registers models, runs prospective experiments with authorised catalogues, and publishes open-source tooling — pyCSEP (BSD 3-clause, with a Journal of Open Source Software paper) and floatCSEP for orchestrating experiments.

What counts as progress

A forecast model submitted for prospective testing with code, configuration and forecast files released, so results can be recomputed with pyCSEP.
Reproducible retrospective benchmarks against stated baselines (e.g. smoothed seismicity, ETAS) on public catalogues, with the exact catalogue version, declustering and magnitude thresholds stated.
Evaluation methodology: new or improved consistency and comparative tests, implemented publicly and demonstrated on existing forecasts.
Documented negative results: a predictor (precursor signal, machine-learning feature set) that adds no skill over the baseline in a prospective or pseudo-prospective test.

How it is checked. A reviewer recomputes the test statistics with pyCSEP from the published forecast and catalogue, confirms that the evaluation period post-dates model fitting, and checks that catalogue processing choices are fixed and documented rather than tuned to the result.

Current state

The problem