Earthquake science has long faced a difficult credibility test: forecasts must be made before earthquakes occur, then judged against observations that were not available when the predictions were issued. A new study published in Nature Communications reports that some earthquake forecasting models used for California show consistency with observations collected prospectively over a decade. The finding does not mean scientists can predict the exact day, location, or magnitude of the next major earthquake. Instead, it suggests that selected models can provide statistically meaningful expectations about how earthquake activity unfolds across time and space.
The study, led by J.A. Bayona, F. Serafini, F. Silva and colleagues, focuses on a central problem in modern seismology: whether earthquake forecasts remain useful when tested in real time rather than reconstructed after the fact. Retrospective analyses can be vulnerable to hindsight, because researchers may adjust a model after seeing what happened. Prospective forecasting avoids that problem. Predictions are issued in advance, preserved, and then compared with an independent sequence of observations as those observations arrive.
California provides an unusually demanding natural laboratory for this work. The state sits along the boundary between the Pacific and North American tectonic plates and contains a dense, interconnected network of faults. Its seismic catalog includes thousands of earthquakes across a wide range of magnitudes, from tiny events detected only by sensitive instruments to damaging earthquakes felt across entire regions. This abundance of data allows researchers to examine whether forecasting systems capture broad statistical features of seismicity rather than merely matching a few memorable events.
Earthquake forecasting models generally do not attempt to identify a single inevitable rupture. Instead, they estimate probabilities. A model may calculate the expected number of earthquakes above a specified magnitude within a geographic region and a defined time window. It may also assign greater or lesser likelihood to different locations based on factors such as recent earthquakes, long-term fault activity, earthquake clustering, or the distribution of accumulated seismic strain. These calculations are usually expressed through probability distributions, allowing scientists to represent uncertainty explicitly.
One important component is the magnitude-frequency relationship. In many earthquake catalogs, smaller earthquakes greatly outnumber larger ones, a pattern commonly described by the Gutenberg-Richter relation. Forecasting models use this relationship, along with information about earthquake rates and spatial patterns, to estimate how frequently events of different sizes may occur. Other models incorporate triggering effects, in which one earthquake temporarily increases the probability of additional earthquakes nearby. The challenge is determining which ingredients improve forecasts and which merely make a model appear more precise than the data justify.
Bayona and colleagues evaluated models against decadal observations collected prospectively in California. The study’s conclusion, as reflected in its title, is selective rather than universal: some forecasting models demonstrated consistency with what was subsequently observed, while others did not show the same level of agreement. This distinction is scientifically important. It indicates that earthquake forecasting should not be treated as a contest between “prediction” and “failure.” Different models can perform differently depending on how they represent earthquake clustering, background seismicity, magnitude distributions, and changing activity levels.
Consistency with observations is also not the same as perfect accuracy. A probabilistic model can be considered successful even when the precise number of earthquakes differs from its central estimate, provided the observations fall within a range that the model regarded as plausible. Researchers typically assess such performance using statistical tools that examine calibration and information gain. Calibration asks whether events assigned a given probability occur at approximately the expected frequency. Information-based measures ask whether a model provides more useful guidance than a simpler reference forecast.
The prospective, decade-scale design makes the study especially relevant to earthquake risk planning. Short-lived bursts of seismicity can make a model look impressive over a few weeks or months, while long-term tests reveal whether its assumptions remain stable. A model that performs well across a decade must contend with quiet periods, earthquake clusters, catalog changes, and the uneven distribution of seismic activity. Such testing can help identify forecasting systems suitable for applications including emergency preparedness, infrastructure planning, insurance analysis, and public risk communication.
The findings also reinforce the limits of earthquake forecasting. No statistical model can currently determine with certainty when a particular fault will rupture or guarantee that a large earthquake will occur at a specific place. California’s fault system is complex, and important physical processes remain difficult to observe directly underground. Forecasts can be affected by incomplete earthquake catalogs, uncertainty in earthquake magnitudes and locations, changes in detection capability, and assumptions about how one event influences another. For that reason, the most responsible use of forecasting models is to support decisions under uncertainty, not to issue sensational claims of imminent disaster.
The study’s broader message is that earthquake forecasting can be evaluated scientifically through transparent, prospective testing. Models that survive comparison with future observations may offer a more reliable foundation for estimating seismic risk, while models that fail can be revised or discarded. That process resembles the testing of weather and climate models: success is measured not by perfect prediction of every event, but by whether probabilistic expectations remain well calibrated over time. For earthquake-prone communities, that gradual improvement could become one of the most practical outcomes of advanced seismology—turning uncertain knowledge into better preparation without pretending that earthquakes have become fully predictable.
Subject of Research: Prospective evaluation of earthquake forecasting models and their consistency with decadal earthquake observations in California.
Article Title: Select earthquake forecasting models demonstrate consistency with prospective decadal observations in California.
Article References: Bayona, J.A., Serafini, F., Silva, F. et al. “Select earthquake forecasting models demonstrate consistency with prospective decadal observations in California.” Nature Communications (2026). https://doi.org/10.1038/s41467-026-76243-7
Image Credits: AI Generated
DOI: 10.1038/s41467-026-76243-7
Keywords: Earthquake forecasting, seismology, California earthquakes, probabilistic models, seismic risk, prospective testing, earthquake prediction, tectonic hazards.
