Why are earthquakes so hard to forecast or predict?

A recent news and feature article in the science magazine Nature by news reporter Glennda Chui discusses problems with current understanding of earthquake science and the future of earthquake forecasting. 

Historically, earthquake science has been based on what is called the elastic rebound hypothesis.  Formulated by Harry Fielding Reid in the years following the great 1906 San Francisco earthquake, this idea states that an earthquake is a response to the slow buildup of  forces within the earth's crust, resulting in fracturing of the rocks and displacement (elastic rebound) of the crust.  Although the theory of plate tectonics was unknown to Reid and his colleagues in 1906, he was able to use observations of small crustal movements to deduce that the rocks along the San Andreas fault had been subjected to increasing forces in the years prior to the 1906 earthquake.  These forces led inexorably to the earthquake that produced the destruction and fire in San Francisco.

Since the time of Reid, the elastic rebound hypothesis has been modified many times but never entirely abandoned.  One of these modifications was the idea of characteristic earthquakes, which proposed that each earthquake fault ruptures repeatedly in essentially the same earthquake time after time.  In this idea, a large earthquake is then the sum of many smaller characteristic earthquakes, each occurring on its own fault. 

This idea of characteristic earthquakes was widely accepted until relatively recently, when the Parkfield earthquake ruptured a 10 km long section of the San Andreas fault near the small town of Parkfield, in central California.  Magnitude 6 earthquakes had struck this section of the fault in 1857, 1881, 1901, 1922, 1934, and 1966.  Data from seismographs, while imprecise by today's standards, suggested that these earthquakes all had the same distribution of slip and occurred regularly in time, with the only exception being the early occurrence of the 1936 earthquake. In the early 1980's the return of this earthquake was predicted for 1988 plus or minus 5 years or so.

When the Parkfield earthquake finally ruptured the fault on September 28, 2004, 16 years late, the pattern of slip was unlike the pattern seen in the 1966 event, the only previous event for which relatively good data was available.  As a result, grave doubt has been cast upon the validity of the characteristic earthquake model, and with it, the elastic rebound hypothesis. 

What will replace these simple models is not clear at this time.  But since earthquake insurance rates in California are based on models that incorporate the characteristic earthquake and elastic rebound models, these developments hold more than just academic interest.  The earthquake forecasts on the Open Hazards web site are based on a different approach, using the observation of fundamental statistical laws, called the Gutenberg-Richter and Omori laws.  As a result, they are independent of whether the characteristic earthquake model is or is not correct.