← Back to Sessions

Ensemble Modeling of Physics-Based and Statistical Forecasts During the 2016/17 Central Apennines (Italy) Aftershock Sequence

Session: Constructing and Testing Regional and Global Earthquake Forecasts III

Type: Oral

Date: 4/22/2021

Presentation Time: 10:15 AM Pacific

Description: 

Developing ensemble models is an increasingly popular strategy for short-term earthquake forecasting, as they allow merging different model types and better accounting for their epistemic uncertainties. Also, ensembles offer the theoretical advantage of not having to a priori select one single model when multiple competing forecasts are available, which is a desirable feature in operational applications. However, different model weighting methods exist and there is no clear consensus on the best way to combine models, so that their actual performance remains a controversial subject.

Here, we present a retrospective daily forecast experiment for the first year of the 2016/17 Central Italy seismic cascade, where nine M5+ earthquakes occurred within five months of the initial Mw 6.0 event. First, we produce a set of forecasts belonging to two different modeling approaches, that is, the physics-based Coulomb rate-and-state (CRS) and the statistical Epidemic-Type Aftershock Sequence (ETAS) models. As input/target dataset, we use a high-resolution, high-density earthquake catalog recently released for the sequence, with ~400,000 M1+ events. We quantify the absolute and relative model performance using the likelihood-based statistical tests (e.g.,S and T-test) introduced by the Collaboratory for the Study of Earthquake Predictability (CSEP). We then form ensemble models from the highest ranked CRS and ETAS realizations using the Bayesian (BMA) and Score Model Averaging (SMA) schemes. We evaluate the ensembles adopting the same CSEP scoring metrics and we benchmark them against their individual components to assess whether they provide enhanced predictive skills.

Results from the 1-year forecast show that while the SMA model is less informative than its sub-components, the Bayesian ensemble performs better than the worse individual model and comparably to the best one.

Presenting Author: Simone Mancini

Student Presenter: Yes

Authors