Precursory Accelerating Moment Release: An Artifact of Data-Selection?

Accelerating Moment Release (AMR), a power-law increase in seismicity prior to a large earthquake, has been identified after-the-fact in many recent studies and is sometimes used to attempt to forecast future earthquakes. We ask whether the observations of AMR are real and potentially predictive, or if apparent AMR could be an artifact due to data-selection. The risk of discovering fictitious AMR exists because the time period, area, and in some cases magnitude range analyzed before each mainshock are often optimized to produce the strongest AMR signal. Searching a range of selection criteria could produce a large probability that some data set will be captured that, by pure chance, satisfies the AMR criteria. We test this hypothesis by using the procedure outlined in Bowman et al. (1998) to compare AMR found in actual seismicity with that found in simulated catalogs in which no underlying seismicity acceleration actually occurs. We generate three types of test catalogs using different approaches to simulating the seismicity of California and Nevada. These catalogs range from purely random earthquake times and locations to more complex catalogs based on an ETAS (epidemic-type aftershock sequence) model that includes both realistic spatial-temporal earthquake clustering and spatial distribution of seismicity based on the real fault network. For all of the simulated data sets, we observe apparent AMR before 91-93% (±7%) of mainshocks M≥6.0, despite the fact that no actual AMR is present in these catalogs. Using the same approach on the real California and Nevada catalog, we find apparent AMR before only 90% (±8%) of the M≥6.0 earthquakes. If at least 10 pre-mainshock events are required and mainshocks restricted to M≥6.5, the AMR model performs better for the data than for the simulations, but not significantly better at the 95% confidence level. These tests use a circular region around the mainshock. We demonstrate that using a region based on Coulomb stress loading of the mainshock does not significantly improve the performance of the AMR model. We also investigate the proposed scaling relationship between mainshock magnitude and the size of the optimal AMR region. We find that the proposed scaling cannot be strictly applied. If the time period and area are fixed to values derived from proposed scaling relationships, AMR is found for only 13% of the southern California M≥6.5 mainshocks. The apparent scaling relationship also appears to be an artifact of data selection, in this case the use of a minimum magnitude that scales with mainshock magnitude, usually 2 magnitude units below. For a larger mainshock, there will be fewer events within 2 magnitude units, and hence a larger area and/or a longer time period may be needed to accumulate enough events to observe significant AMR. We find no scaling between mainshock magnitude and the size of the optimal AMR region when we use a fixed minimum magnitude. These results suggest that the observed AMR in the California and Nevada dataset is actually the result of data-selection and does not represent a real, precursory process.