Disaster Probability Tracker

Statistical occurrence probabilities from historical catalogs — not predictions.

Loading catalogs


Scrub forward to widen the forecast window and watch cumulative probability grow.

—

Poisson model standard

—

Treats events as memoryless with a constant long-run rate — the standard null model for seismicity at this scale (Kagan & Jackson 1991). The probability does not depend on how long it has been quiet.

Renewal model time-dependent

—

Fits the gaps between past events (Gamma distribution) and conditions on the time elapsed since the last one — the statistically honest version of “overdue.” Note: at global scale earthquakes cluster (CV > 1), so a long quiet spell usually makes the near-term probability slightly lower, not higher. True “overdue” behavior only emerges on individual faults with quasi-periodic recurrence.

Probability vs. time

Magnitude–frequency (Gutenberg–Richter)

Epicenter map click to set a local area

Read this before quoting a number

About the model

Data. USGS ANSS ComCat (M4.5+, 2000 → Jan 2024, public domain), the ISC-GEM Global Instrumental Catalogue (1900–1999, via the Swiss Re Historical Earthquake Statistics Dataset, CC BY-SA 3.0), and — when online — a live top-up straight from the USGS feed covering everything since. Pre-1900 events (NGDC significant-earthquake database) are shown for context only and never enter the statistics.

Completeness. Rates for magnitude ≄ m are computed only over the years in which the merged catalog is complete at m: M7.5+ from 1900, M6.5+ from 1918, M5.5+ from 1964, M4.5+ from 2000 (ISC-GEM documented thresholds). This is why different magnitudes use different observation windows.

Estimators. Empirical rates carry exact Poisson confidence intervals (Garwood 1936). The Gutenberg–Richter b-value is the Aki–Utsu maximum-likelihood estimate with the Tinti–Mulargia binning correction and Shi–Bolt standard error. Where fewer than three events exist in the window, the global model extrapolates with a tapered Gutenberg–Richter law (Kagan 2002) and local areas use fixed-b scaling — both flagged in the caveats when active. The renewal view fits a Gamma distribution to inter-event times by maximum likelihood and conditions on the open interval.

Full methodology, citations, and limitations: METHODOLOGY.md.