Evaluations of M>=4 Earthquake Probability Forecasts for California and Western Nevada From 2005 to 2008

Since the beginning of 2005, we have been generating three kinds of 5-day earthquake probability forecasts following the occurrences of M>=4 mainshocks in California and Nevada. One forecast model assumes that all future M>=4 mainshocks in California and Nevada are random, uncorrelated events, and the forecast probabilities are based on Poisson distributions with fixed rates of mainshock occurrences. In the second model, the short-term temporal clustering of M>=4 mainshocks in California and Nevada from 1932 to 2004 is used as the basis of the short-term forecasts of future events. The third forecasts are based on a hidden Markov model analysis that uses the past earthquake history to estimate the short-term probability of future earthquakes in the region. The second model is not adaptive, while the third model is since each forecast depends on the seismic history available at that time. Separate forecasts are issued for the eastern and the western parts of the study area. Relative evaluations of the forecasts are made using the likelihood of the observed results under the three models. For the first and second models, the events are independent and the likelihoods can be multiplied. For the third model, successive events are not independent and a joint likelihood must be computed. After almost 40 forecasts since 1/1/2005, both the second and the third models have somewhat outperformed the Poisson model, and the second model has a slightly better likelihood score than the third.