Building on CSEP principles, floatCSEP tackles a long-standing challenge in the field: the original CSEP Testing Centers relied on centralized, rigid infrastructures that tightly coupled data management to local hardware, limiting reusability, scalability, and community engagement. floatCSEP fills this gap by introducing the concept of the “Floating Experiment,” where software and data artifacts are decoupled from specific physical infrastructure and encapsulated as self-contained, reproducible packages that can run on any machine with sufficient computational resources.
The application manages the complete experiment lifecycle — from model integration and catalog handling to forecast generation, evaluation, visualization, and reporting. It supports both time-invariant and time-dependent experiments and integrates external models through containerized Docker environments. Beyond supporting new official CSEP experiments, floatCSEP enables independent researchers to create new retrospective or prospective studies, benchmark novel models against established ones, and contribute to the continuous evaluation of operational earthquake forecasting systems. Together with pyCSEP, open-source forecasting models, and long-term open-science repositories, floatCSEP helps lay the foundation for robust, collaborative benchmarks in earthquake forecasting research.
]]>This article takes a critical look at the intersection of machine learning and seismology. For decades, earthquake forecasting has been dominated by classical statistical models like the Epidemic-Type Aftershock Sequence model, which explicitly capture how earthquakes trigger subsequent events. While Neural Point Processes (NPPs) have recently emerged as a flexible alternative, their evaluation has been limited by outdated and flawed benchmarks, some even suffering from data leakage and incomplete datasets.
To address this, the authors introduce EarthquakeNPP, a new benchmarking framework built on curated earthquake catalogs from California spanning 1971–2021. Crucially, the platform aligns machine learning evaluation practices with those used in seismology, incorporating both likelihood-based and generative metrics. This creates a more realistic and rigorous testing ground, enabling fair comparisons between NPPs and established models. The benchmark also includes the ETAS model as a baseline, ensuring that new approaches are judged against the current standard.
Read the article: EarthquakeNPP: A benchmark for Earthquake Forecasting with Neural Point Processes
]]>The study, led by Francesco Serafini and colleagues (J. A. Bayona, F. Silva, W. Savran, S. Stockman, P. J. Maechling, and M. J. Werner), describes 25 forecasting models developed by nine international research groups, the tools used to evaluate them, and the processes behind their production.
This database represents one of CSEP’s most extensive efforts to date, with all models operated at the Southern California Earthquake Center (SCEC) testing center in a fully prospective manner and independently from the model developers. It provides the first long-term, prospective testbed for advancing Operational Earthquake Forecasting (OEF) and earthquake predictability research worldwide and serves as a natural benchmark for future forecasting models.
All forecasts are now openly available on Zenodo and through the CSEP website. In addition, we provide on GitHub code and a tutorial to load, visualize, combine, and evaluate the forecasts in the database, along with guidance for creating new forecasts in a format suitable for comparison.
Read the article: A benchmark database of then years of prospective next-day earthquake forecasts in California
Explore the database: download from Zenodo
Explore code and tutorial: visit the GitHub page
]]>