Can Earthquakes Be Predicted?

The honest answer is no, and it is worth understanding precisely why, because the reasons are more interesting than the answer.

What "prediction" would require

A useful earthquake prediction must specify three things with enough precision to act on: location, time window, and magnitude range. It must also be accompanied by a stated probability, and it must be issued before the event.

That last requirement rules out most claimed successes, which are identified retrospectively. Retrospective identification is trivially easy — after any earthquake you can find some anomaly that preceded it — and scientifically worthless unless the method also generates specific predictions in advance, and does so without an unacceptable rate of false alarms.

No proposed method has passed this bar.

Precursors that did not survive testing

Many candidate precursors have been proposed and investigated seriously.

Radon gas emission from groundwater was studied for decades. Anomalies do occur, but not consistently before earthquakes, and they occur frequently without any earthquake following.

Ground deformation measured by GPS and satellite radar reveals strain accumulation beautifully, but strain accumulates steadily for decades before an event with no distinctive acceleration signalling imminent rupture.

Electromagnetic signals and ionospheric disturbances have generated a large and contested literature. Independent replication has been poor, and reported effects are difficult to separate from the many other sources of electromagnetic noise.

Groundwater level changes in wells do sometimes precede earthquakes, but they also respond to rainfall, atmospheric pressure and pumping.

Seismic quiescence — a decrease in small earthquakes before a large one — has been reported in some sequences and absent in others, and it is difficult to define objectively without hindsight.

The recurring pattern is the same in each case: an effect that sometimes appears before some earthquakes, appears frequently without any earthquake following, and cannot be used to make a specific advance prediction.

The Parkfield experiment

The most instructive test was conducted at Parkfield, California, on the San Andreas Fault. Parkfield had produced moderate earthquakes of around magnitude 6 at strikingly regular intervals — 1857, 1881, 1901, 1922, 1934, 1966 — averaging roughly 22 years.

In 1985 the USGS formally predicted the next Parkfield earthquake would occur before 1993, and instrumented the segment more densely than any other piece of fault on Earth in order to capture precursors.

The earthquake arrived in 2004, eleven years after the prediction window closed. And despite the extraordinary instrumentation, no clear precursory signal was detected in the extensive data collected before it.

Parkfield was not a failure — it produced enormously valuable data and it decisively answered a question. But the answer was not the one hoped for. Even a well-instrumented, apparently regular fault does not announce itself, and its recurrence is not regular enough to time.

The foreshock asymmetry

The most tantalising near-miss is the foreshock. As covered in the aftershocks guide, some large earthquakes are preceded by smaller ones on the same fault.

But the operationally relevant question is the reverse of the one usually asked. Not "what fraction of large earthquakes have foreshocks?" (perhaps 5–10%) but "given a small earthquake, what is the probability a larger one follows?" That probability is only a few percent.

Small earthquakes are extremely common. Acting on every one would mean an overwhelming preponderance of false alarms — and false alarms carry real costs, including economic disruption, erosion of public trust, and injuries during unnecessary evacuations.

What does work: probabilistic forecasting

Forecasting is a genuinely successful science and it is used everywhere.

Long-term seismic hazard forecasts estimate the probability of exceeding a given level of ground shaking at a location over a period of decades. They combine known fault locations and slip rates, historical and paleoseismic records, and ground-motion models. These forecasts are the basis for building codes, insurance pricing and infrastructure design — and they save far more lives than any prediction ever could, because they influence how millions of buildings are constructed.

Operational aftershock forecasts are issued routinely after significant earthquakes, giving the probability of further events of given magnitudes over coming days and weeks. These are well calibrated and genuinely useful for emergency management.

Time-dependent renewal models incorporate the time elapsed since a fault's last rupture. Their validity is debated but they represent a serious attempt to sharpen long-term forecasts.

What also works: early warning

Earthquake early warning is often confused with prediction. It is not prediction at all — it detects an earthquake that has already started and races an alert ahead of the damaging waves.

Because P-waves travel faster than S-waves, and because electronic signals travel far faster than either, a system can detect the rupture and warn locations further away seconds to tens of seconds before strong shaking arrives. That is enough to stop trains, halt surgery, open lift doors and let people take cover.

It offers no advance notice, but it works, and it is deployed operationally in Japan, Mexico, Taiwan and along the US West Coast.

What to do with this

Since no one can tell you when an earthquake will strike, the rational response is structural rather than predictive. If you live in a seismically active region, the actions that matter are the ones you take now: securing heavy furniture and water heaters, knowing what to do during shaking, maintaining supplies, and supporting enforcement of modern building codes.

Preparedness works precisely because it does not depend on knowing when.