← Back to Sessions

Ensemble Earthquake Forecasting With a Logistic Regression Model

Session: Beyond Poisson: Seismic Hazards and Risk Assessment for the Real Earth [Poster]

Type: Poster

Date: 4/19/2021

Presentation Time: 11:30 AM Pacific

Description: 

Combining individual forecast models into an ensemble model is proving beneficial in many research areas and applications (e.g., weather/climate forecasting, medical diagnosis, computer security). An ensemble emphasizes the individual strengths of various types of models (e.g., physical and statistical models), typically yields a superior performance than the best individual model, is more flexible than a single model, and provides a more realistic uncertainty quantification.

Constructing an ensemble relates to the performance evaluation of individual models. Marzocchi et al. 2014 [doi: 10.1785/0220130219] presented an operational earthquake forecasting (OEF) experiment for Italy, which launched in 2009 according to the standards of the Collaboratory for the Study of Earthquake Predictability (CSEP). The experiment includes three statistical models and an ensemble of them, with weights assigned using a Bayesian approach. The models provide weekly earthquake forecasts for magnitudes M3.95+ and are still running as of 2021.

Here we present a different ensemble strategy using a logistic regression. Instead of weighting the forecast models according to their individual skill (as in the Bayesian ensemble), the optimal weights are those that maximize the skill of the ensemble (via goodness of fit). The multivariate logistic regression is fit between the forecast rates of the individual models and the spatio-temporal target bins (i.e., target earthquakes with M3.95+ binned to the testing region’s grid resulting in ‘0’ [no targets] or ‘1’ [one or more targets]). We evaluate the performance of the two ensemble strategies using metrics such as their likelihood ratio or the information gain compared to a reference model. We experiment with various weighting schemes to address the multi-purpose character of OEF (i.e., with a focus on spatial skill, recent seismicity, overall rate, etc.) and try to understand when and why the logistic ensemble outperforms individual models.

Presenting Author: Marcus Herrmann

Student Presenter: No

Authors