Comparative evaluation of earthquake forecasting models: An application to Italy
Authors:
Jonas R. Brehmer,
Kristof Kraus,
Tilmann Gneiting,
Marcus Herrmann,
Warner Marzocchi
Abstract:
Testing earthquake forecasts is essential to obtain scientific information on forecasting models and sufficient credibility for societal usage. We aim at enhancing the testing phase proposed by the Collaboratory for the Study of Earthquake Predictability (CSEP, Schorlemmer et al., 2018) with new statistical methods supported by mathematical theory. To demonstrate their applicability, we evaluate t…
▽ More
Testing earthquake forecasts is essential to obtain scientific information on forecasting models and sufficient credibility for societal usage. We aim at enhancing the testing phase proposed by the Collaboratory for the Study of Earthquake Predictability (CSEP, Schorlemmer et al., 2018) with new statistical methods supported by mathematical theory. To demonstrate their applicability, we evaluate three short-term forecasting models that were submitted to the CSEP Italy experiment, and two ensemble models thereof. The models produce weekly overlap** forecasts for the expected number of M4+ earthquakes in a collection of grid cells. We compare the models' forecasts using consistent scoring functions for means or expectations, which are widely used and theoretically principled tools for forecast evaluation. We further discuss and demonstrate their connection to CSEP-style earthquake likelihood model testing. Then, using tools from isotonic regression, we investigate forecast reliability and apply score decompositions in terms of calibration and discrimination. Our results show where and how models outperform their competitors and reveal a substantial lack of calibration for various models. The proposed methods also apply to full-distribution (e.g., catalog-based) forecasts, without requiring Poisson distributions or making any other type of parametric assumption.
△ Less
Submitted 17 May, 2024;
originally announced May 2024.
