Citing a paper does not mean this website has been scientifically validated.
The model is experimental. Past tests do not prove future accuracy.
We tested 1-, 7- and 30-day earthquake counts at magnitude 2, 3 and 4 or greater. We also tried fitting ETAS to the previous three years, updating its parameters once a year. All comparisons use the same dates and recorded events.
The starting date is uncertain. Our original M2+ screen selected 2014. Changing the test settings selects 2022 or 2025; the strictest setting finds no passing period. This does not tell us exactly how many earthquakes are missing. It means that “complete since 2014” would be too strong a claim.
Lower scores mean less error across the possible earthquake counts. These are error scores, not earthquake magnitudes or probabilities. Each row is a separate comparison. “Recent average” uses the previous year.
Swipe sideways to see the rest of the table.
| Period / magnitude | Tests | Fixed ETAS | Updated ETAS | Recent average |
|---|---|---|---|---|
| 1 days · M2+ | 94 | 0.55 | 0.52 | 0.67 |
| 1 days · M3+ | 94 | 0.13 | 0.13 | 0.15 |
| 1 days · M4+ | 94 | 0.01 | 0.01 | 0.01 |
| 7 days · M2+ | 94 | 3.46 | 2.39 | 2.84 |
| 7 days · M3+ | 94 | 0.58 | 0.51 | 0.6 |
| 7 days · M4+ | 94 | 0.14 | 0.14 | 0.14 |
| 30 days · M2+ | 94 | 20.65 | 12.67 | 13.44 |
| 30 days · M3+ | 94 | 3.05 | 2.74 | 3.09 |
| 30 days · M4+ | 94 | 0.47 | 0.47 | 0.48 |
The catalog was revised later, and we had already examined some of these periods. These are research comparisons, not independent proof of future accuracy. We have not selected a winning model or promoted these candidates to the live model.
Keeping only ML magnitudes would remove 444 of 6844 M2+ records since 2014. We retain the reported scales and do not invent conversions.
These tests cannot tell when or where a large earthquake will happen. They do not estimate M7 probability or building safety.
Seif et al. (2017) · Wiemer & Wyss (2000) · Kumazawa & Ogata (2013) · CSEP
Our rolling stationary refit is a simpler control, not a reproduction of the paper’s penalized nonstationary ETAS. Parameter profiles refit the other parameters, but are not calibrated confidence intervals.
All test results (JSON) · Catalog checks (JSON) · Sequence checks (JSON) · Earlier-event checks (JSON) · Parameter sensitivity (JSON)
We compared four approaches in 95 seven-day windows from 2019 to 2026. Issue dates are 30 days apart. This uses a later-revised catalog, not forecasts saved before the earthquakes.
| Method | Score |
|---|---|
| Full-history average | 4.8 |
| Past-year average | 2.85 |
| Past-year variable counts | 2.82 |
| Offline ETAS candidate | 3.36 |
The recent-rate controls scored better than the new ETAS candidate here. The fit still hit a parameter bound. We have not replaced the live model or established future reliability.
Parameters frozen using 2014–2018; 200 simulated futures per result. Regional M2+ counts only. Conditional uncertainty; no M7 probability, building-safety or fault-rupture inference. Historical cutoff choice, native magnitude differences and limited simulations remain limitations.
Download research data (JSON) · Download calibration checks (JSON) · Download catalog audit (JSON)Hainzl, S.; Kumazawa, T.; Ogata, Y. · 2024 · Geophysical Journal International · 236, 1609–1620
Examines the effect of missed detections on fitting in a Türkiye sequence. This version does not implement ETASI detection corrections; the study motivates our warning about changing detection in the historical archive.
Ogata, Y. · 1988 · Journal of the American Statistical Association · 83, 9–27
Basis for the ETAS conditional intensity and likelihood approach. Our implementation is a temporal adaptation with a constrained parameter search, not a reproduction of every method in the paper.
Utsu, T.; Ogata, Y.; Matsu’ura, R. S. · 1995 · Journal of Physics of the Earth · 43, 1–33
Review of the Omori–Utsu law describing aftershock decay. Our triggering kernel uses this functional form.
Reasenberg, P. A.; Jones, L. M. · 1989 · Science · 243, 1173–1176
Foundational probabilistic aftershock work. California coefficients are not copied into Marmara; parameters here are fitted from regional observations.
Page, M. T.; van der Elst, N.; Hardebeck, J. L.; Felzer, K.; Michael, A. J. · 2016 · Bulletin of the Seismological Society of America · 106, 2290–2301
Motivates caution about regional transfer and missing observations. This version does not implement the paper’s full Bayesian and time-varying detection model; that remains an explicit limitation.
Wiemer, S.; Wyss, M. · 2000 · Bulletin of the Seismological Society of America · 90, 859–869
Explains why one magnitude threshold does not guarantee uniform reporting through time. An annual M2 screening supports a 2014 start; this does not prove all earthquakes were detected.
Schorlemmer, D.; Gerstenberger, M. C.; Wiemer, S.; Jackson, D. D.; Rhoades, D. A. · 2007 · Seismological Research Letters · 78, 17–29
Motivates predefined evaluation of forecasts. The retrospective count tests here are not a full RELM/CSEP validation exercise.
Styron, R.; Pagani, M. · 2020 · Earthquake Spectra
Published source of the map’s traces. Original selected geometry and provenance are retained; this does not support parcel-level safety assessments.
Seif et al. · 2017 · Journal of Geophysical Research: Solid Earth
Research reference for magnitude limits and modeling assumptions. Missing aftershocks and magnitude cutoffs can bias fitted parameters; its results have not been reproduced here.
Kumazawa, T.; Ogata, Y. · 2013 · Journal of Geophysical Research: Solid Earth
Studies time-varying seismicity rates. The live model does not implement this method.
Mizrahi, L.; Nandan, S.; Wiemer, S. · 2021 · Journal of Geophysical Research: Solid Earth
Addresses incomplete detection in earthquake forecasting. This correction is not implemented in the live model.
Mizrahi et al. · 2024 · Reviews of Geophysics
Guides evaluation and communication of limitations; it does not validate our Marmara model.
Serafini et al. · 2025 · Scientific Data
An example of evaluating forecasts recorded before events occur. California results do not establish skill in Marmara.
Han; Mizrahi; Wiemer · 2026 · Journal of Geophysical Research: Solid Earth
Studies directional aftershock distributions. This site does not calculate directional forecasts or fault-rupture probabilities.
Mignan, A.; Woessner, J. · 2012 · CORSSA
Methods for studying detection of small events across time and place. We do not assume every M2 event was recorded.
Collaboratory for the Study of Earthquake Predictability · 2026 · Software documentation
Evaluation reference. Standard independent-Poisson tests are not applied directly to clustered ETAS counts.
Swiss Seismological Service · 2026 · Research software · MIT license
Research reference for method comparison. The live site does not run this package.