CSEP 2.0 Developments

CSEP Working Group Home Page

Introduction

Retrospective experiments of earthquake forecasts represent an important aspect of the CSEP software infrastructure. They provide significant insight into the performance of earthquake forecasting models and allow scientists to evaluate the performance of forecasts on historic earthquake sequences. The CSEP1 system is designed to support retrospective forecast experiments in addition to prospective forecasting. CSEP operates entirely in serial; meaning that both retrospective and prospective processing capabilities must happen sequentially. This sequential behavior is enforced within CSEP using locks that are placed on active directories during computation. While the serial nature of the system is not strictly enforced, working around the CSEPLocks require non-elegant hacks that violate the core concept of forecast groups within CSEP. This document explores the possibility of repurposing the current CSEP framework into a module to exclusively facilitate retrospective forecasting experiments.

Guiding Principals

1. Controlled Environment
2. Transparency
3. Comparability
4. Reproducibility

The locks are used to prevent data contamination, and to prevent overwriting data by a different process. This is required because CSEP uses the same directory for all forecast models before moving forecast files to archive directories