Horizon-Dependent Value of Geodetic Context in Operational Earthquake Forecasting: A Leakage-Free, Multi-Region Study with Pre-Registered Negative Results
Operational earthquake forecasting (OEF) issues calibrated conditional probabilities of future seismicity; the Epidemic-Type Aftershock Sequence (ETAS) model is its de-facto benchmark, and no machine-learning temporal point process has been shown to beat a well-fit ETAS prospectively under fair testing. This preprint builds a leakage-free, multi-region OEF testbed on a global moment-magnitude catalog (M>=5, 1990-2026), scores every model with the community-standard CSEP framework (pyCSEP) in information gain per earthquake (IGPE, nats), and pre-registers ship/no-ship rules with shuffled-label negative controls. Four artifact-backed, honestly framed findings: (1) The value of a GNSS-strain geodetic context covariate, added to a Hawkes-structured neural point process, is horizon-dependent: at 1-7 days it does not beat tiled ETAS (mean IGPE -0.053 over eight weekly windows, 4/8 positive), whereas at 30 days it beats ETAS robustly (global mean IGPE +0.115 over 2166 events, 9/10 windows positive, paired-t lower bound +0.087; positive in all four high-seismicity regions). The mechanism is explicit: the calibrated model deploys a time-flat geodetic background, not a triggering model, so it wins where the background dominates the integral (30 days) and loses where ETAS triggering dominates (1-7 days). (2) A convex log-score-optimal stack of ETAS-family variants earns a small but significant global gain (+0.011 over 751 events) that does not generalize to the canonical active margins; under the pre-registered rule it is a no-ship, and a temporally adaptive variant is refuted as a multiple-comparison artifact. (3) ETAS adds skill over a stationary smoothed-seismicity Poisson null only where there is triggering to exploit, quantifying the high-versus-low-seismicity bias. (4) The binding 1-7 day consistency failure is count over-dispersion, not spatial shape: a catalog-based (branching-simulation) number test passes a window the Poisson number test rejects, and the frozen-intensity forecast systematically under-counts by about 28 percent because it omits within-window secondary triggering. Conclusion: base tiled ETAS is at or near the practical ceiling for mean-rate 1-7 day IGPE over global M>=5, and the honest place for a geodetic covariate is a longer, background-dominated outlook. This is an independent research and education tool; it is not an alarm system and must not be used for life-safety decisions. Code, configurations, provenance manifests, and the committed daily/back-analysis artifacts (MIT): https://github.com/fsantibanezleal/CAOS_SEISMIC . Static forecast viewer: https://seismic.fasl-work.com .