Abstract
Abstract Macro-scale earthquake forecasting on spatially discretized grids presents a severe class imbalance problem, where rare seismic events are embedded within long periods of inactivity. In Central Asia, forecasting \((M \ge 3.0)\) earthquakes on a weekly \((1^\circ \times 1^\circ)\) grid produces approximately 96% zero inflation and an event prevalence of about 4.5%, yielding a constant Precision--Recall Area Under the Curve (PR-AUC) baseline of 0.045. This study investigates whether predictive performance under such extreme imbalance is governed primarily by model architecture or by structured feature design and baseline calibration. We introduce a framework that integrates tectonic regime--conditioned normalization, Omori energy decay proxies, structural fault descriptors, and Log-Odds baseline initialization that encodes historical cell-specific event rates directly into the learning objective. Across six heterogeneous architectures evaluated under a strict chronological split, performance converges to PR-AUC \((\approx 0.451)\) and Receiver Operating Characteristic Area Under the Curve (ROC-AUC) \((\approx 0.844)\), representing a tenfold improvement over the constant baseline without synthetic resampling. The results indicate that under extreme low-prevalence conditions, calibrated baselines and physically structured feature spaces dominate architectural complexity, suggesting the existence of a macro-scale predictability ceiling in seismic forecasting.