No one can predict exactly when or where the next damaging earthquake will strike the San Francisco Bay Area, and no technology on the horizon is close to changing that. What scientists can do, and have done with increasing sophistication, is forecast the long-term likelihood that specific faults will produce large earthquakes within a given time window. For the Bay Area, those forecasts paint a sobering picture: the region sits atop multiple active faults, at least one of which appears overdue by historical standards. The gap between a precise prediction and a probabilistic forecast is enormous, though, and misunderstanding it shapes how millions of residents think about their risk.
Why Prediction Remains Out of Reach
An earthquake prediction, in the strict scientific sense, would specify a magnitude, a location, and a narrow time window before the event. Despite decades of effort worldwide, no reliable method exists to do this. The fundamental problem is that the crust stores stress invisibly across enormous volumes of rock, and the transition from a locked fault to a rupturing one happens in seconds. The signals that precede a quake, if they exist at all, are subtle enough to be indistinguishable from normal background noise until after the event has already begun.
Forecasting is a different enterprise. Rather than saying “a magnitude 7 will hit on Tuesday,” a forecast says something like “there is a 33% chance of a magnitude 6.7 or larger earthquake on this fault system over the next 30 years.” That kind of statement is useful for building codes, insurance, and emergency planning, even though it cannot tell you when to leave the house. The Bay Area’s forecast probabilities come primarily from the Uniform California Earthquake Rupture Forecast (UCERF3), the most comprehensive model of California’s earthquake potential.
What UCERF3 Says About Bay Area Faults
UCERF3 is the authoritative model that scientists and policymakers rely on for estimating earthquake likelihood across California. One of its biggest innovations over earlier versions was dropping the assumption that earthquakes stay neatly confined to a single fault segment. Instead, the model allows ruptures to cascade across multiple connected faults simultaneously. This change had a striking practical consequence: compared to the previous forecast, the likelihood of moderate earthquakes (around magnitude 6.5 to 7.5) went down, while the likelihood of larger events went up.
1USGS Publications Warehouse. UCERF3: A new earthquake forecast for California’s complex fault systemThe model solves for earthquake rates across all of California’s faults at once, using a massive computational approach that requires supercomputers to run through more than 1,400 alternative scenarios accounting for different sources of uncertainty.
2PubMed Central. Uniform California earthquake rupture forecast, version 3 (UCERF3): the time-independent modelFor Bay Area residents, the practical upshot is that several fault systems carry meaningful probability of producing a damaging earthquake within a human lifetime. The Hayward, San Andreas, Calaveras, and Rodgers Creek faults all run through or near densely populated areas, and UCERF3 treats them not as isolated threats but as an interconnected system where stress on one fault influences behavior on its neighbors.
The Hayward Fault and Its Alarming Clock
Among Bay Area faults, the Hayward Fault gets particular attention because it runs directly beneath some of the most densely populated cities in the East Bay, including Oakland, Berkeley, Hayward, and Fremont, and because its historical record suggests a pattern that is unusually regular for an earthquake fault.
Paleoseismic trenching, where geologists dig across the fault to read layers of disrupted sediment, has revealed a remarkably long earthquake history on the southern Hayward Fault. Researchers have identified twelve large earthquakes going back to roughly 91 A.D., with the most recent being the well-documented magnitude 6.8 event in 1868. The average time between those twelve quakes is about 161 years, with a standard error of just ten years. The sequence appears fairly regular, with a coefficient of variation of 0.40, which is low enough to suggest the fault behaves more like a ticking clock than a random process.
3Bulletin of the Seismological Society of America. Evidence for a Twelfth Large Earthquake on the Southern Hayward Fault in the Past 1900 YearsIndependent modeling of the earthquake series on the southern Hayward Fault confirms this pattern. Analysis using time-dependent statistical distributions found that a renewal model with a recurrence interval of about 210 years fits the observed record far better than any random model, by a factor of five. In statistical terms, the Hayward Fault’s large earthquakes are quasi-periodic, consistent with a process where stress builds steadily and releases in somewhat regular cycles.
4Geophysical Research Letters. Earthquake recurrence on the south Hayward fault is most consistent with a time dependent, renewal processThe 1868 earthquake was over 155 years ago. Whether you use the 161-year average from the paleoseismic record or the 210-year renewal interval from statistical modeling, the fault is well within the window where the next large event could be expected. The paleoseismic study estimates a 30-year probability of about 29% for the southern Hayward Fault alone.
5Bulletin of the Seismological Society of America. Evidence for a Twelfth Large Earthquake on the Southern Hayward Fault in the Past 1900 YearsLocked Patches and Creeping Sections
Not every part of the Hayward and Calaveras fault system behaves the same way. Some sections creep slowly and continuously, releasing stress without producing large earthquakes. Other sections are locked, accumulating stress that will eventually be released in a sudden rupture. The pattern of locked versus creeping sections is one of the most important factors controlling how big a future earthquake could be.
GPS and satellite radar measurements over two decades have allowed scientists to map this patchwork in detail. On the Hayward Fault, locked patches near San Leandro have the potential to produce roughly a magnitude 6.8 event, while a locked section south of Fremont could generate about a magnitude 6.5. On the northern Calaveras Fault near Dublin, a locked patch could produce around a magnitude 6.6. If ruptures cascade across multiple segments, the numbers climb: the Hayward Fault rupturing from Berkeley to the Calaveras junction could produce about a magnitude 6.9, and a joint rupture of the Hayward and central Calaveras faults could reach roughly magnitude 7.1.
6Journal of Geophysical Research: Solid Earth. Interseismic coupling and refined earthquake potential on the Hayward‐Calaveras fault zoneDynamic rupture modeling adds another layer of complexity. Simulations suggest that when a rupture starts in a locked section, the creeping portions of the fault can either stop it, limiting the event to a high-magnitude-6 earthquake, or the rupture can punch through the creeping zones and cascade further, producing a magnitude-7 event. Which outcome occurs depends on exactly where the rupture nucleates and how the fault geometry channels the energy.
7Journal of Geophysical Research: Solid Earth. A Geology and Geodesy Based Model of Dynamic Earthquake Rupture on the Rodgers Creek‐Hayward‐Calaveras Fault System, CaliforniaThe 1906 Stress Shadow
One reason the Bay Area went relatively quiet through much of the twentieth century, despite sitting on some of the most active faults in the country, traces back to the great 1906 San Francisco earthquake. That magnitude 7.8 rupture on the San Andreas Fault released enormous amounts of stress, and in doing so it actually reduced stress on neighboring faults across the region. This phenomenon, called a stress shadow, effectively suppressed large earthquakes on other Bay Area faults for at least 75 years.
Three-dimensional modeling of this stress transfer has shown that the 1906 shadow reduced 30-year earthquake probabilities on nearby fault segments by roughly 7 to 12 percent compared to what they would have been without it. The combined probability of a large earthquake on the Bay Area’s non-San Andreas faults dropped from a range of 65 to 77 percent (without the shadow) to 53 to 70 percent (with it).
8Journal of Geophysical Research: Solid Earth. Post‐1906 stress recovery of the San Andreas fault system calculated from three‐dimensional finite element analysisThat shadow has been fading. More than a century of tectonic loading has refilled much of the stress deficit that 1906 created. The Bay Area is now in a period where the inherited protection from the last great earthquake has largely worn off, which is part of why current probability estimates are as high as they are.
The Seismic Gap Debate
The idea that a fault segment is “overdue” rests on what seismologists call the seismic gap hypothesis: the notion that hazard increases with time since the last large earthquake. This sounds intuitive, and for individual well-characterized faults like the southern Hayward, it has some empirical backing. But globally, the seismic gap hypothesis has a troubled track record.
A landmark test published in 1991 examined over 40 large earthquakes that occurred in the decade after the hypothesis was formalized. The data showed that places identified as high-probability gaps did not fill with earthquakes at the predicted rate. In fact, recently active segments were more likely to produce another large event than the long-quiet gaps were.
9Journal of Geophysical Research: Solid Earth. Seismic Gap Hypothesis: Ten years afterA more specific global forecast based on seismic gap theory was tested using earthquake data from 1989 to 1994. It predicted that about nine zones worldwide would be filled by their characteristic earthquakes in five years. Only two were. The discrepancy was too large to attribute to chance at the 99% confidence level.
10Journal of Geophysical Research: Solid Earth. New seismic gap hypothesis: Five years afterMore recent work using earthquake simulations with realistic statistics has found that the hypothesis remains difficult to validate, partly because historical earthquake catalogs are simply too short. A fault might have a genuine recurrence pattern, but centuries or millennia of data are needed to distinguish that pattern from random clustering, and we rarely have that much data.
11Journal of Geophysical Research: Solid Earth. Testing of the Seismic Gap Hypothesis in a Model With Realistic Earthquake StatisticsFor the Hayward Fault specifically, the paleoseismic record is unusually long and detailed, which is why researchers express more confidence in time-dependent models there than they would for most faults worldwide. But the broader lesson stands: “overdue” is a slippery concept, and a fault that has been quiet does not always mean a fault that is about to rupture.
Early Warning Is Not Prediction
The closest thing to a real-time earthquake alert that Bay Area residents can access is ShakeAlert, the earthquake early warning system now operational across California, Oregon, and Washington. ShakeAlert does not predict earthquakes. It detects them after they have begun and races the seismic waves to give people a few seconds of notice before strong shaking arrives.
The system works because the fastest seismic waves (P-waves) travel ahead of the slower, more destructive waves (S-waves and surface waves). By detecting P-waves at stations close to the epicenter, the system can estimate the earthquake’s size and location and send alerts to people further away before the damaging shaking reaches them. Alerts go out through the Wireless Emergency Alert system and apps on phones. The warning window ranges from essentially nothing near the epicenter to tens of seconds at greater distances.
12Seismological Research Letters. Earthquake Early Warning ShakeAlert System: Testing and Certification PlatformA parallel effort has explored whether smartphones themselves could serve as earthquake sensors. The MyShake project turned personal phones into a citizen science seismic network: over 300,000 people worldwide downloaded the app, which uses the phone’s built-in accelerometer to detect earthquake shaking. The system showed it could detect, locate, and estimate earthquake magnitude roughly 5 to 7 seconds after the quake began, and deliver alerts in an additional 1 to 5 seconds.
13Pure and Applied Geophysics. The MyShake Platform: A Global Vision for Earthquake Early WarningSmartphone-based networks are not a replacement for traditional seismometer arrays, which are far more sensitive and precise. But they could fill gaps in regions where traditional infrastructure is sparse, and they demonstrated that magnitude 5 earthquakes could be recorded at distances of 10 kilometers or less using phone accelerometers.
14PubMed Central. MyShake: A smartphone seismic network for earthquake early warning and beyondOne complication with early warning is the risk of false alerts. Research on public reactions to unnecessary earthquake alerts found that people rated false earthquake warnings as somewhat less legitimate and were slightly less tolerant of them compared to false missile alerts. The concern is that repeated false alarms erode trust, and trust is essential for a system that demands immediate action from millions of people.
15Scientific Reports. Monitoring public reaction to an unnecessary earthquake early warning alertThe Search for Precursors
Scientists have spent decades looking for reliable signals that precede earthquakes. The most frequently studied candidates include changes in groundwater levels, radon gas emissions from the ground, and unusual animal behavior. None of these has proven reliable enough for operational prediction, though some show tantalizing patterns when examined after the fact.
Radon gas, which can seep out of the ground when rocks are stressed or fractured, has been proposed as a short-term precursor. A global review compiled 91 reported radon anomalies linked to earthquakes across multiple countries, correlating properties like the size of the anomaly with the earthquake’s magnitude and distance.
16Journal of Geophysical Research: Solid Earth. Radon content of groundwater as an earthquake precursor: Evaluation of worldwide data and physical basis The challenge is that radon fluctuations also occur for many reasons unrelated to earthquakes, including rainfall, barometric pressure changes, and seasonal variation. Separating a genuine pre-seismic signal from this noise has proven extremely difficult in real time.
Groundwater level changes are another candidate. Monitoring wells near the epicenter of the devastating 2023 double earthquake in Turkey detected anomalous groundwater changes roughly 255 minutes before the quakes struck.
17Environmental Research Communications. Groundwater level anomalies as precursors to the february 6 double earthquake in Türkiye That is an intriguing result, but a single case study from one earthquake sequence does not establish a general method. Most earthquakes have no documented groundwater precursor at all, and where anomalies have been found, they are usually identified retrospectively rather than in real time.
Animal behavior is the most publicly captivating proposed precursor, and also the weakest scientifically. A systematic review of 180 publications on abnormal animal behavior before earthquakes found over 700 claimed precursory observations across more than 130 species. But only 14 of those records were actual time series rather than anecdotal reports, and even the longest time series covered just one year. The review also noted that the timing pattern of reported animal precursors closely mirrors the timing of foreshocks, suggesting that many animals were simply reacting to small earthquakes that humans did not notice, not sensing some mysterious pre-earthquake signal.
18Bulletin of the Seismological Society of America. Can Animals Predict Earthquakes?A more rigorous study did find something interesting. Researchers fitted bio-logging tags to farm animals (cows, dogs, and sheep) near the epicenter of the 2016 magnitude 6.6 Norcia earthquake in Italy and monitored their activity continuously for months. They detected consistent anticipatory behavior 1 to 20 hours before earthquakes, but only when the animals were inside a stable, not when they were outdoors on pasture. The anticipation time correlated with distance from the earthquake.
19Ethology. Potential short‐term earthquake forecasting by farm animal monitoring Even this well-designed study, though, is a long way from a practical prediction system. The biological mechanism remains unclear, and no one has replicated the results in a way that could be deployed operationally.
Foreshocks and Aftershock Forecasting
One area where short-term forecasting has made genuine progress is in estimating the probability that a moderate earthquake will be followed by a larger one. This matters because roughly half of the largest earthquakes in some catalogs were preceded by smaller foreshocks, though identifying a foreshock in real time is almost impossible since it only becomes a “foreshock” in hindsight, after a bigger event follows it.
During the 2019 Ridgecrest earthquake sequence in Southern California, the UCERF3-ETAS model was used operationally for the first time to answer a question that comes up after every felt earthquake: what are the chances this was actually a foreshock to something bigger? The model combines the long-term fault-based probabilities of UCERF3 with a statistical framework for how earthquakes cluster in time and space.
20Seismological Research Letters. Operational Earthquake Forecasting during the 2019 Ridgecrest, California, Earthquake Sequence with the UCERF3-ETAS ModelResearch into how foreshocks are identified has revealed that the methods used to classify them can substantially change the results. Different identification techniques, including fixed time windows, nearest-neighbor clustering, and epidemic-type aftershock sequence models, yield varying conclusions about how common foreshock activity is. Some methods suggest that many California foreshocks are associated with aseismic processes deep in the fault zone, but this finding is sensitive to the detection capability of the local seismic network.
21Geophysical Research Letters. Does Foreshock Identification Depend on Seismic Monitoring Capability?Machine Learning and the Laboratory Gap
Machine learning has become a frequent headline in earthquake science, and for good reason: algorithms can find subtle patterns in massive datasets that humans would never spot. But the current state of the field is more modest than the coverage suggests. Most of the work involves laboratory experiments, not real faults.
In a typical setup, researchers run controlled stick-slip experiments where a block of material slides along a surface under stress, mimicking in miniature what happens on a fault. Acoustic signals recorded during these experiments are fed to neural networks or other algorithms, which learn to estimate when the next slip event will occur. Recent work has used physics-informed neural networks that incorporate friction laws directly into the model architecture, improving the ability to extract meaningful physical parameters from the acoustic data.
22Scientific Reports. Physics informed neural network can retrieve rate and state friction parameters from acoustic monitoring of laboratory stick-slip experimentsThe catch is obvious: a tabletop experiment with a controlled block of granular material is not a real earthquake fault buried kilometers underground, with heterogeneous rock, fluids, variable temperatures, and geometry that no one can directly observe. The leap from predicting lab slip events to predicting natural earthquakes remains enormous. Machine learning is likely to improve earthquake forecasting incrementally, particularly in areas like aftershock pattern recognition and real-time data processing. But the fantasy of an AI that tells you to evacuate before the shaking starts is not supported by anything in the current literature.
Human-Caused Earthquakes Near the Bay Area
Not all earthquakes in the broader Bay Area region are purely tectonic. The Geysers, a large geothermal energy field in the mountains north of Santa Rosa, has been a well-studied source of induced seismicity for decades. Fluid injection into hot underground rock to generate steam for electricity produces thousands of small earthquakes each year.
Studies of microseismicity at The Geysers have found that about 72% of the induced earthquakes occur on faults oriented in a way that makes them prone to failure under the existing stress field. But a meaningful fraction occur on faults that would not normally be expected to slip, and these events cluster near injection wells during periods of high injection rates.
23Journal of Geophysical Research: Solid Earth. Impact of fluid injection on fracture reactivation at The Geysers geothermal fieldThe upside, if there is one, is that induced seismicity at The Geysers appears somewhat predictable and potentially controllable. Research has suggested that intelligent management of injection rates could limit the larger seismic events.
24International Journal of Rock Mechanics and Mining Sciences. The impact of injection on seismicity at The Geysers, California Geothermal Field The earthquakes produced are generally small, rarely above magnitude 4 to 5, and The Geysers is remote enough that direct damage to urban areas is minimal. But the field offers a real-world laboratory for understanding how fluids interact with faults, which has broader implications for any region where injection activities occur near active fault zones.
Soft-Story Buildings and What Forecasts Actually Change
If precise prediction is impossible and probabilistic forecasts are the best we have, the practical question becomes: what do you actually do with a 30-year probability? For most Bay Area residents, the answer lies in structural preparedness rather than trying to guess the date.
The Bay Area has tens of thousands of soft-story wood-frame buildings, typically older apartment buildings with open garages or storefronts on the ground floor and living space above. These structures are vulnerable to collapse in strong shaking because the ground floor lacks the rigidity to support the weight above it. San Francisco, Oakland, and Berkeley have all enacted mandatory retrofit ordinances for these buildings, driven in large part by the probability estimates from models like UCERF3. Testing of retrofit methods, including those developed under the NEES-Soft project, has validated approaches consistent with FEMA guidelines for reducing collapse risk.
25Treesearch. Seismic Risk Reduction for Soft-Story Wood-Frame Buildings: Test Results and Retrofit Recommendations from the Nees-Soft ProjectThis is where earthquake forecasting delivers its real value. The forecasts do not tell you to leave your house on a particular morning. They tell cities which building types to require retrofitting, tell engineers what level of shaking to design for, tell emergency managers where to pre-position resources, and tell homeowners whether earthquake insurance is worth the premium. The inability to predict the exact moment of the next earthquake does not mean science has nothing useful to say. It means the useful things it says are about preparation, not evacuation.