Can an earthquake be predicted?

Long-term seismic hazard maps state the probability of ground shaking of a given intensity over a given period, and they are built from mapped faults, measured strain and the statistics of past events.

They are good enough to write building codes against, which is the use that saves the most lives. In January 2010 a magnitude 7.0 earthquake caused immense loss of life in Haiti, where large earthquakes were anticipated and buildings were not made to withstand them. Six weeks later a magnitude 8.8 struck Chile, which enforces high seismic safety standards, and caused substantially less damage and loss of life for its size.

The maps have critics as well. Stein, Geller and Liu point to Wenchuan in 2008 and Tohoku in 2011, both in areas the maps showed as relatively safe, and argue that the maps’ uncertainties should be stated plainly and the maps tested against simple alternatives.

A prediction, in the sense that matters, specifies place, time and magnitude narrowly enough that it could fail. No method has produced those reliably.

In 1997 a paper in Science put the case that earthquakes cannot be predicted in principle: any small earthquake has some chance of cascading into a large one, and whether it does depends on unmeasurably fine details of conditions in the Earth’s interior. Strainmeters near several large earthquakes, L’Aquila among them, recorded no change before the rupture, which fits the hypothesis that large and small earthquakes start alike and that size is settled as a rupture spreads and stops.

That argument has not been refuted and it has not been proved. The US Geological Survey says it does not expect to know how to predict earthquakes in the foreseeable future, and an international commission in 2011 called earthquake predictability poorly understood.

Early warning systems work on a different principle. They detect the first, faster, less damaging waves at instruments near the epicentre and send an electronic signal ahead of the slower, more damaging ones.

That buys seconds to tens of seconds, which is enough to slow trains, bring lifts under control and let people take cover. Close to the epicentre the shaking can arrive with the warning or before it. It is not prediction: the earthquake has already begun.

After a moderate earthquake the probability of a larger one nearby rises for days, by a factor that can be estimated from aftershock statistics. Operational forecasting turns that into an official, updated number.

The problem is what to do with a probability that goes from one in ten thousand to one in a hundred. It is a large relative change and a small absolute one. At a 2014 workshop on exactly this, some argued that such forecasts are useless in practice, because acting on them means a false alarm nearly every time, and others that even small probabilities can nudge people toward sensible precautions.

Slow slip events, changes in seismic wave speed and patterns of small quakes are all under investigation, and some have statistical support in some regions.

In 2023 Bletery and Nocquet stacked the GPS records from before 90 large earthquakes and found about two hours of accelerating slip before the ruptures. Others found that the signal vanishes once noise shared across a network is removed; the authors reply that the correction can remove a real signal too, and call the question unsettled.

Pattern-recognition methods tested prospectively have beaten chance, but only by a factor of two to four, with alarms lasting months or years over areas as large as Switzerland, which leaves them little use as predictions. For many proposed precursors, a 2011 review found, the rates of false alarms and missed earthquakes are unknown, so whether the signal means anything cannot be judged.