Select earthquake forecasting models matched what happened in California over 11 years, according to a University of Bristol-led study in Nature Communications that checked 51,625 next-day forecasts against 597 earthquakes of magnitude 3.95 or greater. The authors say it is the strongest evidence so far that such models can describe next-day activity.
What the Study Tested
The paper, titled “Select earthquake forecasting models demonstrate consistency with prospective decadal observations in California,” evaluates 27 forecasting models run by the Collaboratory for the Study of Earthquake Predictability, known as CSEP. The main analysis focuses on 21 models that change from day to day and forecast earthquakes of magnitude 3.95 or greater. It compares them with a benchmark model that does not change over time.
The test ran from August 2007 to August 2018, the paper says, and the models produced 51,625 forecasts. Over that period 597 earthquakes of magnitude 3.95 or greater and depth of 30 kilometers or less struck the test region. That is about 54 a year, or roughly one every 6.7 days, a calculation from the paper’s counts and the 11-year span.
The tests were prospective, which means the models issued forecasts before the earthquakes happened and researchers compared them with later observations. The paper applied separate tests for the number of earthquakes, their locations and their magnitudes.
Authors named on the journal page include José A. Bayona, Francesco Serafini, Fábio Silva, Pablo Iturrieta, William H. Savran, Marcus Herrmann, Warner Marzocchi, Philip J. Maechling and Maximilian J. Werner, and the page truncates the list. The paper points readers to the pyCSEP software toolkit, a companion data publication in Scientific Data from 2025 and the U.S.
Geological Survey’s Quaternary Fault and Fold Database. The paper uses four families of checks. Number tests ask whether a model forecast the right total count of earthquakes. Spatial tests ask whether it put them in the right places, and the authors use calibration plots and Kolmogorov-Smirnov statistics for that purpose. A magnitude test asks whether it forecast the right mix of sizes. A fourth measure, information gain per activated cell, compares each model with the benchmark.
Which Models Held Up
Results depended on the test. On the number of earthquakes, the paper reports that under a standard Poisson test only two models, ETAS-PPE-M2.95 and STEPJAVA, were consistent with observations. Under a second test that allows for clustering, a negative binomial test, every model but one passed, which is about 95%.
The authors explain the gap in the limitations they list. The Poisson test assumes counts that are not clustered or skewed, which fits poorly with earthquakes that come in sequences.
On location, 15 of the 21 models were consistent, or about 71%. On magnitude, all models passed the standard magnitude test, and all but STEPJAVA passed a second statistical check. Several newer models, including ETAS-HW-KERNEL-M2, ETAS-HW-KERNEL-M3 and the ETAS-PPE models, gained more information than the benchmark in later periods, the paper says.
Where the Models Fell Short
The paper lists weaknesses. Models generally underestimated the number of earthquakes on days with large earthquakes and their aftershocks. After the El Mayor-Cucapah earthquake sequence, daily forecasts ran too low for two days. One model, KJ-SMOOTHING-V0, underestimated counts, and another, KJ-SMOOTHING-V1, overestimated them, and the authors recommend a code review. A third, PPE-BACKGROUND-M2, overestimated the long-term background rate by about 44%.
Some models produced forecasts that were too smooth across space and others too localized. The authors did not update model parameters during the evaluation, and the models ran over different, staggered periods, which makes comparisons across models uneven. The main analysis covers magnitude 3.95 and above only, and a subset at magnitude 4.95 appears in the supplementary material.
What the Lead Author Said
Dr. José A. Bayona, an Early Career Leverhulme Research Fellow in the School of Earth Sciences at Bristol and the lead author, said in the university’s announcement that “No scientific method can currently predict exactly when and where” the next earthquake will happen. He said the results give strong evidence for using the select models in public operational forecasting systems.
The announcement describes probabilistic forecasts as tools for estimating where, when and how large earthquakes are more likely over a set period. It says they do not predict an exact time or place. Bayona said the authors provide a recipe of elements to consider when building forecasts that match the observed number, location and size of earthquakes. The authors also released an open-source software package with code, data and documentation so others can repeat the analysis.
The announcement mentions recent earthquakes in Colombia, Japan and Venezuela as reminders of why preparedness matters. It says government agencies, civil protection authorities and scientific organizations developing operational forecasting systems may find the work useful.
Publication Date in Two Sources
The sources disagree on when the paper appeared. The university’s announcement, posted on Newswise, says the study was published on Oct. 15, 2026. The journal’s own page lists the paper as published on Aug. 6, 2026, as volume 17, article 9444.
How This Experiment Differs from Earlier Ones
The announcement says earlier CSEP experiments covered Japan, New Zealand and Italy, and that the California experiment stands out for the number of models, its length and the amount of prospective earthquake data. It also says the models used in some countries’ operational forecasting can stay consistent with observations over long tests.
The Leverhulme Trust’s Early Career Research program, the U.S. Geological Survey and the Statewide California Earthquake Center funded the work, according to the announcement. The announcement does not list limitations or future research plans. Those appear in the paper, and none of them extends the results beyond California or beyond next-day forecasts.
