Earthquake Prediction: Can We Predict When Earthquakes Will Happen?

Few questions in earth science provoke more public fascination — and more scientific frustration — than the question of earthquake prediction. After every damaging earthquake, the same headlines appear: Why didn't scientists warn us? Can't we predict earthquakes yet? The short answer, despite more than a century of research, is no. No one has ever reliably predicted an earthquake in the way that meteorologists predict hurricanes — specifying when, where, and how big, with enough lead time for meaningful public action.

But that blunt answer obscures important nuances. While true prediction remains elusive, seismologists have developed two powerful capabilities that are often conflated with prediction: probabilistic forecasting, which estimates the likelihood of future earthquakes over years to decades, and early warning, which detects an earthquake already underway and delivers alerts seconds before damaging shaking arrives. Understanding the critical distinctions between prediction, forecasting, and early warning is essential for anyone who lives in earthquake country.

This article examines why earthquake prediction has failed, what forecasting and early warning can actually do, and where frontier research may eventually push the boundaries of what is knowable about future earthquakes.

Prediction vs. Forecasting vs. Early Warning: The Critical Distinction

The single most important concept in this entire subject is the difference between three terms that the public, media, and even some scientists use interchangeably — but which refer to fundamentally different things.

Earthquake prediction means specifying, in advance, the time, location, and magnitude of a future earthquake with enough precision to be actionable. A prediction would sound something like: "A magnitude 7.0 earthquake will strike within 50 km of Los Angeles between March 3 and March 10." No one has ever done this reliably. The USGS states plainly: "Neither the USGS nor any other scientists have ever predicted a major earthquake. We do not know how, and we do not expect to know how any time in the foreseeable future." USGS: Can You Predict Earthquakes?

Earthquake forecasting means estimating the probability of an earthquake of a given magnitude occurring in a given region over a defined time period — typically years to decades. A forecast sounds like: "There is a 72% probability of one or more magnitude 6.7 or greater earthquakes in the San Francisco Bay Area between 2014 and 2043." This is scientifically well-established and forms the basis of building codes, insurance rates, and emergency planning. What causes earthquakes

Earthquake early warning (EEW) means detecting an earthquake that has already begun and transmitting alerts to people and systems before the strongest shaking arrives at their location. An early warning sounds like: "Earthquake detected. Expect strong shaking in 15 seconds." This is operationally active in multiple countries today.

Why Earthquake Prediction Has Failed

The Nature of the Problem

Earthquakes originate from the sudden rupture of rock along faults deep in the Earth's crust. The processes that control when a fault ruptures are governed by the precise state of stress, fluid pressure, temperature, and material properties at every point along the fault — variables that cannot be measured directly at depths of 5 to 15 km where most damaging earthquakes nucleate. Even if these variables could be measured, fault systems exhibit the hallmarks of chaotic systems: small, unmeasurable differences in initial conditions can produce dramatically different outcomes.

The analogy to weather forecasting breaks down quickly. Meteorologists can observe the entire atmosphere with satellites, radar, weather balloons, and surface stations. Seismologists cannot observe fault surfaces directly. The atmosphere is a fluid that obeys well-understood equations of motion. Fault rupture involves the fracture mechanics of heterogeneous rock under enormous confining pressures — a far less tractable problem.

The Precursor Problem

Decades of research have searched for reliable precursors — observable phenomena that consistently occur before earthquakes. Proposed precursors have included changes in seismic wave velocities, radon gas emissions, groundwater levels, electromagnetic signals, foreshock sequences, and animal behavior. None has proven reliable.

The fundamental challenge is twofold. First, some proposed precursors do occasionally occur before earthquakes — but they also occur frequently without being followed by earthquakes, producing unacceptable false alarm rates. Second, many large earthquakes occur with no detectable precursors at all, producing unacceptable miss rates. A prediction method with high false alarm rates and high miss rates is operationally useless — or worse, actively harmful if it leads to costly evacuations or, eventually, to public complacency after repeated false alarms.

Foreshocks illustrate this problem perfectly. About half of all large earthquakes are preceded by smaller earthquakes in the same area in the hours to days beforehand. But at the time they occur, these smaller earthquakes are indistinguishable from ordinary seismicity. Only in hindsight, after the mainshock, can they be identified as foreshocks. The other half of large earthquakes have no detectable foreshocks at all.

The Animal Behavior Myth

The idea that animals can sense impending earthquakes is one of the most persistent myths in seismology. Accounts of unusual animal behavior before earthquakes date back to ancient Greece. Modern accounts include reports of dogs howling, fish jumping, snakes emerging from burrows, and zoo animals behaving erratically before major earthquakes.

The scientific evidence does not support this idea. A comprehensive 2018 review published in the Bulletin of the Seismological Society of America examined 180 publications reporting animal behavior anomalies before earthquakes. The authors concluded that the vast majority of reports were anecdotal, uncontrolled, and subject to severe reporting bias — people remember unusual animal behavior that preceded an earthquake and forget the many times animals behaved unusually with no earthquake following. No study has documented consistent, repeatable precursory animal behavior under controlled conditions.

Some animals may respond to the P-wave (the first, weaker seismic wave to arrive) seconds before humans feel the stronger S-wave — but this is detection of an earthquake already in progress, not prediction of a future one. It is, in essence, a biological version of early warning, not prediction.

Famous Prediction Attempts — And Why They Failed

The Haicheng "Success" (1975)

On February 4, 1975, Chinese authorities ordered the evacuation of approximately one million residents of Haicheng, Liaoning Province. Hours later, a magnitude 7.3 earthquake struck the city. The evacuation unquestionably saved thousands of lives and was hailed worldwide as the first successful earthquake prediction.

The reality is more complicated. The evacuation was based on a confluence of factors: a sequence of foreshocks that had been increasing in rate and magnitude over the preceding days, changes in groundwater levels, and reports of unusual animal behavior. Chinese seismologists had also issued a general medium-term warning for the region based on seismicity patterns. The decision to evacuate was made by local officials based on the intensifying foreshock sequence — essentially, an escalating series of increasingly alarming earthquakes that made the population and officials nervous enough to act.

The Haicheng prediction was not replicable. It depended on an unusually clear foreshock sequence — a pattern that occurs before only a minority of large earthquakes. The Chinese government's own experience demonstrated this brutally just 17 months later.

Tangshan (1976): The Unpredicted Catastrophe

On July 28, 1976, a magnitude 7.5 earthquake struck Tangshan, an industrial city of one million people approximately 150 km east of Beijing. The earthquake struck at 3:42 a.m. while most residents were sleeping. The death toll was catastrophic — the Chinese government's official figure is 242,000, though some estimates range higher. There was no foreshock sequence, no successful warning, and no evacuation. The Tangshan earthquake remains one of the deadliest earthquakes in recorded history and a sobering demonstration that the Haicheng experience could not be generalized. Understanding aftershock sequences

The Parkfield Experiment

In 1985, the USGS launched what may be the most focused earthquake prediction experiment ever attempted. The town of Parkfield, California, sits on the San Andreas Fault in a segment that had produced magnitude ~6 earthquakes in a remarkably regular pattern: 1857, 1881, 1901, 1922, 1934, and 1966 — roughly every 22 years. Scientists predicted the next Parkfield earthquake would occur by 1993, plus or minus a few years.

The USGS installed an extraordinary array of instruments: seismometers, strainmeters, creepmeters, GPS receivers, magnetometers, and water-level monitors. Parkfield became the most densely instrumented fault segment on Earth.

The earthquake came on September 28, 2004 — a magnitude 6.0 event that arrived 11 to 16 years late, depending on the prediction window used. More importantly, the dense instrument network detected no reliable precursors before the earthquake. The instruments provided extraordinary data for understanding the earthquake after it occurred, but contributed nothing to predicting it in advance. The Parkfield experiment is widely regarded as demonstrating that even under the best possible conditions, short-term earthquake prediction remains beyond current capability.

Iben Browning and New Madrid (1990)

In 1990, self-proclaimed climatologist Iben Browning predicted a major earthquake on the New Madrid Seismic Zone in the central United States on or about December 3, 1990, based on tidal forces. Browning had no credentials in seismology. His prediction was explicitly rejected by the USGS and the National Earthquake Prediction Evaluation Council. Nonetheless, the prediction generated enormous media coverage and public anxiety. Schools closed, property values dropped, and the National Guard was placed on standby.

December 3, 1990, came and went with no earthquake. The episode illustrated both the public appetite for earthquake prediction and the real economic and social harm that false predictions can cause.

The VAN Method (Greece)

Beginning in the 1980s, Greek physicists Panayiotis Varotsos, Kostas Alexopoulos, and Kostas Nomicos proposed the VAN method, which claimed to predict earthquakes based on the detection of "seismic electric signals" — transient changes in the Earth's electric field allegedly generated by stressed rock before rupture. The method generated intense controversy in the seismological community. Multiple statistical evaluations found that VAN predictions did not perform significantly better than random chance, and the method has not been adopted by any national seismological agency.

Probabilistic Seismic Hazard Assessment (PSHA)

While short-term prediction has failed, long-term probabilistic forecasting is a mature and operationally critical science. The framework is called Probabilistic Seismic Hazard Assessment, or PSHA.

How UCERF3 Works

The most sophisticated earthquake forecast in the United States is the Uniform California Earthquake Rupture Forecast, Version 3 (UCERF3), published in 2015 by the Working Group on California Earthquake Probabilities — a collaboration between the USGS, the California Geological Survey, and the Southern California Earthquake Center. USGS UCERF3

UCERF3 integrates data from geological field studies (fault slip rates, paleoseismic trench excavations), geodetic measurements (GPS-derived crustal deformation rates), and the historical and instrumental seismicity catalog. The model considers over 350,000 possible fault rupture scenarios across California's complex fault network and uses a time-dependent probability model that accounts for the time elapsed since the last major earthquake on each fault segment.

Key UCERF3 results for the 30-year period 2014–2043:

• 72% probability of one or more M6.7+ earthquakes in the San Francisco Bay Area
• 93% probability of one or more M6.7+ earthquakes in Southern California
• 7% probability of one or more M8.0+ earthquakes statewide
• The San Andreas Fault accounts for the largest share of M7.5+ earthquake probability
• Multi-fault ruptures (earthquakes that break across more than one named fault) are more common in the model than previously expected

These forecasts are not predictions — they do not specify when an earthquake will occur. But they are invaluable for building codes, emergency planning, insurance rate-setting, and public risk communication. Every building code in California incorporates PSHA results to determine the level of ground shaking that structures must be designed to resist.

PSHA Beyond California

PSHA is practiced worldwide. The USGS publishes national seismic hazard maps for the entire United States, updated most recently in 2023 (NSHM23). These maps underpin the seismic provisions of the International Building Code used across the country. Similar efforts exist in Japan (the Headquarters for Earthquake Research Promotion), New Zealand (GNS Science), Italy (INGV), and many other seismically active countries. Earthquake risk assessment tools

Earthquake Early Warning: Detecting Quakes in Progress

Earthquake early warning is the most significant operational achievement in the earthquake safety field in the past two decades. It cannot predict earthquakes, but it can provide seconds to tens of seconds of warning after an earthquake begins and before the strongest shaking arrives at a given location.

How EEW Works

The physics of EEW relies on a simple fact: electronic signals travel at the speed of light (~300,000 km/s), while seismic waves travel through rock at roughly 3–8 km/s. When a fault ruptures, seismometers near the epicenter detect the first-arriving P-wave (a compressional wave that typically causes weak, rattling shaking). Within seconds, algorithms estimate the earthquake's location, magnitude, and the expected intensity of the slower but more destructive S-wave and surface waves. Alerts are then transmitted electronically to areas that have not yet experienced strong shaking.

The warning time depends on the distance between the recipient and the epicenter. A city 100 km from the epicenter might receive 20–30 seconds of warning. A city 20 km away might receive only a few seconds — or none at all if it falls within the system's "blind zone," the area so close to the epicenter that shaking arrives before the system can process and transmit an alert.

ShakeAlert: The U.S. System

The ShakeAlert earthquake early warning system, operated by the USGS in partnership with university seismic networks, covers the states of California, Oregon, and Washington — the most seismically hazardous region of the contiguous United States. The system became publicly operational in stages: California launched statewide Wireless Emergency Alerts (WEA) integration in October 2019, Oregon in March 2021, and Washington in May 2021. ShakeAlert Earthquake Early Warning

ShakeAlert uses a network of over 1,675 seismic stations along the West Coast. When an earthquake is detected, the system can deliver alerts through multiple channels: the Wireless Emergency Alert system (the same system used for AMBER Alerts and severe weather warnings), smartphone apps (such as MyShake in California and Android's built-in earthquake alerts), and dedicated alerting systems for critical infrastructure operators. Set up earthquake alerts

The typical warning time for ShakeAlert ranges from a few seconds for areas near the epicenter to more than a minute for distant locations in large earthquakes. For the most common damaging scenario — a M6.5–7.5 earthquake on a major California fault — most populated areas would receive 5 to 30 seconds of warning. This is enough time to drop, cover, and hold on; to pull a car to the side of the road; to move away from windows; and for automated systems to slow trains, open firehouse doors, and begin shutting down sensitive industrial processes. What to do during an earthquake

The blind zone remains a significant limitation. Areas within approximately 10–20 km of the epicenter typically receive little or no warning because the shaking arrives before the system can complete its detection, estimation, and alert delivery cycle. For earthquakes directly beneath a major city — such as a rupture on the Hayward Fault beneath Oakland or a shallow thrust event beneath Los Angeles — the blind zone means the most heavily impacted areas may receive no advance warning.

Japan's JMA System: The Global Standard

Japan's earthquake early warning system, operated by the Japan Meteorological Agency (JMA), is the most advanced and most extensively deployed EEW system in the world. Japan began development in the 1990s and launched public alerts in October 2007. JMA Earthquake Early Warning

The JMA system uses a network of approximately 1,100 seismometers operated by JMA and the National Research Institute for Earth Science and Disaster Resilience (NIED), plus approximately 3,600 seismic intensity meters. When an earthquake is detected, alerts are broadcast through television and radio (with automatic audio alerts that interrupt programming), cell phone networks (all Japanese carriers support earthquake alerts), and dedicated receivers in schools, offices, factories, and transportation systems.

The system's integration into Japanese infrastructure is unmatched. The Shinkansen (bullet train) system automatically applies emergency braking when an alert is received. Factory assembly lines halt. Elevators stop at the nearest floor and open their doors. Gas valves close automatically. Schools conduct regular EEW drills, and children are trained to respond to the distinctive alert tone.

The JMA system proved its value during the March 11, 2011, Tohoku earthquake (M9.1). The system detected the earthquake and issued a public warning approximately 8 seconds after rupture began on the fault. Residents in Tokyo, approximately 373 km from the epicenter, received more than a minute of warning before strong shaking arrived. The Shinkansen system, which had trains operating at up to 270 km/h, stopped all trains before strong shaking reached the tracks. No Shinkansen trains derailed. The early warning system unquestionably saved lives, though it could not protect against the tsunami that followed.

EEW Systems Worldwide

Data sources: USGS ShakeAlert, JMA, published literature on international EEW systems

The Blind Zone Problem and Its Implications

Every EEW system has a blind zone — the region surrounding the epicenter where seismic waves arrive before the system can detect, process, and transmit an alert. The blind zone radius depends on the density of the seismic network, the speed of the processing algorithms, and the communication latency of the alert delivery channels.

For ShakeAlert, the effective blind zone radius is approximately 10–20 km for moderate earthquakes. This means that for earthquakes on faults that run directly through or beneath urban areas — such as the Hayward Fault through the East Bay, the San Andreas through the Peninsula, or the Puente Hills thrust beneath downtown Los Angeles — the people who experience the most intense shaking may receive little or no warning.

Reducing the blind zone is an active area of engineering research. Approaches include deploying denser seismic networks (reducing the distance to the nearest station), developing faster algorithms (reducing processing time), and using on-site single-station methods (which can issue very fast but less accurate alerts based on the first second or two of P-wave data detected at a single station). San Andreas Fault

Frontier Research: Can Science Do Better?

Despite the failures of prediction to date, research continues on multiple fronts.

Machine Learning and Pattern Detection

Machine learning algorithms have been applied to seismic data to search for patterns that human analysts might miss. In 2019, researchers at Los Alamos National Laboratory published results showing that a machine learning model could detect acoustic emission patterns in laboratory fault experiments that preceded fault slip. The model identified a continuous signal in the acoustic data that correlated with the time remaining before the next laboratory earthquake.

However, laboratory faults are vastly simpler than natural faults. Scaling these results to real earthquake faults — which are heterogeneous, three-dimensional, kilometers-long structures buried at depth — remains an enormous challenge. As of 2025, no machine learning model has demonstrated operationally useful earthquake prediction capability on natural faults, though several research groups continue to work on the problem.

GPS and Crustal Deformation

The Global Navigation Satellite System (GNSS), including GPS, allows scientists to measure the slow, continuous deformation of the Earth's crust with millimeter precision. By tracking how the crust deforms between earthquakes, scientists can estimate where stress is accumulating and, in principle, where the stress is closest to the threshold for failure.

In California, GPS data shows that the Pacific Plate moves roughly 50 mm/year relative to the North American Plate, with the San Andreas Fault accommodating approximately 20–28 mm/year of that motion depending on the segment. Segments of the fault that are "locked" (not creeping) are accumulating elastic strain energy that will eventually be released in earthquakes. GPS data is a critical input to models like UCERF3 but has not yet enabled short-term prediction.

Slow-Slip Events and Tremor

One of the most intriguing discoveries of the 21st century in seismology is the detection of slow-slip events (also called silent earthquakes) and associated tectonic tremor. These phenomena, first documented on the Cascadia Subduction Zone in the Pacific Northwest, involve fault slip that occurs over days to weeks rather than seconds — too slowly to generate damaging seismic waves, but detectable with GPS and sensitive seismometers. Earthquake aftershock patterns

Slow-slip events on the Cascadia Subduction Zone recur approximately every 14 months and transfer stress to the fully locked portion of the fault that is expected to produce a future megathrust earthquake. Whether slow-slip events increase or decrease the probability of a major rupture is an active area of research. Some models suggest they could eventually serve as medium-term indicators of changing stress state on the fault, though this remains speculative.

Operational Earthquake Forecasting

The most promising near-term advance is operational earthquake forecasting (OEF), which extends PSHA into shorter time frames. After a large earthquake, the probability of additional large earthquakes in the same region increases dramatically (these subsequent events may be aftershocks, or in some cases, earthquakes on adjacent faults triggered by the stress change from the first event). OEF models, such as the USGS's implementation of the Epidemic-Type Aftershock Sequence (ETAS) model, provide updated probabilistic forecasts in the hours and days after significant earthquakes.

During the 2019 Ridgecrest earthquake sequence in California, the USGS issued public aftershock forecasts for the first time, providing updated probabilities of future M5+ and M7+ earthquakes as the sequence evolved. This represents a form of short-term forecasting — not prediction, but a meaningful probabilistic statement about the near-term seismic hazard.

What You Can Do Today

The inability to predict earthquakes does not mean that people are helpless. The existing tools — probabilistic forecasting, early warning, and preparedness — are powerful when used together.

If you live in earthquake country, the USGS national seismic hazard maps and state-specific hazard assessments (such as UCERF3 for California) tell you the level of shaking your area is likely to experience over the coming decades. This information should drive your decisions about earthquake insurance, home retrofitting, and emergency supply kits.

Earthquake early warning systems provide seconds of warning — enough time to protect yourself if you know what to do. In California, ensure that Wireless Emergency Alerts are enabled on your phone. Android phones have built-in earthquake detection. The MyShake app provides ShakeAlert-powered warnings. Familiarize yourself with the alert tone and practice the protective action: drop, cover, and hold on. What to do during an earthquake

The USGS locates approximately 20,000 earthquakes per year worldwide — roughly 55 per day. Most are too small to feel. But the next damaging earthquake in your region is not a question of if but when. Prediction cannot tell you the answer. Preparation can ensure you survive it. Set up personalized earthquake alerts