Earthquake Number Forecasts Testing

TL;DR

This study analyzes earthquake number distributions in global catalogs, finding that the negative-binomial distribution better models seismicity than the Poisson, especially over small time intervals, and proposes empirical smoothing for forecast testing.

Contribution

It demonstrates the inadequacy of the Poisson model for earthquake counts and advocates for using the negative-binomial distribution with empirical smoothing in seismic forecasting.

Findings

Poisson distribution can be rejected in favor of NBD for most cases.

Skewness and kurtosis increase with smaller magnitude thresholds and time intervals.

Observed heavy tails in earthquake counts exceed NBD predictions for small intervals.

Abstract

We study the distributions of earthquake numbers in two global catalogs: Global Centroid-Moment Tensor and Preliminary Determinations of Epicenters. These distributions are required to develop the number test for forecasts of future seismic activity rate. A common assumption is that the numbers are described by the Poisson distribution. In contrast to the one-parameter Poisson distribution, the negative-binomial distribution (NBD) has two parameters. The second parameter characterizes the clustering or over-dispersion of a process. We investigate the dependence of parameters for both distributions on the catalog magnitude threshold and on temporal subdivision of catalog duration. We find that for most cases of interest the Poisson distribution can be rejected statistically at a high significance level in favor of the NBD. Therefore we investigate whether these distributions fit the observed distributions of seismicity. For this purpose we study upper statistical moments of earthquake numbers (skewness and kurtosis) and compare them to the theoretical values for both distributions. Empirical values for the skewness and the kurtosis increase for the smaller magnitude threshold and increase with even greater intensity for small temporal subdivision of catalogs. A calculation of the NBD skewness and kurtosis levels shows rapid increase of these upper moments levels. However, the observed catalog values of skewness and kurtosis are rising even faster. This means that for small time intervals the earthquake number distribution is even more heavy-tailed than the NBD predicts. Therefore for small time intervals we propose using empirical number distributions appropriately smoothed for testing forecasted earthquake numbers.