Fifty-year seismicity analysis of the East Anatolian Fault System (1976–2026): leakage-safe probabilistic forecasting for seismic-risk resilience
Year: 2026
Type: preprint
Abstract: Background The East Anatolian Fault System (EAFS) has experienced pronounced changes in seismicity rate, catalog sensitivity, and sequence behavior over five decades. We aimed to characterize this 1976–2026 record and test whether historical seismicity supports reproducible short-term fault-trace probability updates relevant to seismic-risk preparedness. Methods USGS and AFAD catalogs were merged with source provenance, conservative duplicate matching, pre-test magnitude harmonization, and time-varying completeness control. The dense EAF-corridor catalog contained 73 532 events; 5 026 events met the common Mc = 3.4 modeling threshold. These events were associated computationally with 157 source-traceable active-fault geometries, yielding 1 509 869 fault-trace-day observations. The primary outcome was at least one M ≥ 4.0 event associated with a trace during the following seven days. Training ended in 2019, model selection used 2020–2022, 2023 was locked for testing, and 2024–29 July 2026 was retained as a forward holdout. Formal inference used non-overlapping seven-day origins and 28-day moving-block bootstrap confidence intervals. Results The validation-selected History+ETAS logistic forecast achieved PR-AUC values of 0.3479 in 2023 and 0.03261 in the forward holdout, compared with 0.0388 and 0.01127 for a Bayesian historical fault-rate reference. Paired PR-AUC gains were 0.314 (95% CI 0.157–0.443) and 0.0201 (0.00888–0.03844), respectively. However, prespecified incremental tests did not support unique added value from the frozen ETAS parametrization, graph/stress proxies, boosted trees, or deep learning. Conclusions A source-auditable 50-year EAFS benchmark indicates that recent seismic history contains reproducible short-term information beyond a static fault-rate reference, including after the exceptional 2023 sequence. More complex model components should not be assigned physical or operational significance until they demonstrate independent prospective skill. The framework supports cautious probabilistic risk communication rather than deterministic earthquake prediction.
Source: Research Square
Author: Mehmet Tevfik AĞDAŞ
Institution: Munzur University
Language: English
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204259193