arXiv2026
Forecast models in statistical seismology are commonly evaluated using log-likelihood (LL) scores for the full distribution $P(n)$ of earthquake counts. However, fat tails and out-of-range observations (realized counts exceed all simulations) can bias model rankings. We develop a tail-aware framework that estimates cell-wise $P(n)$ using adaptive Gaussian kernel density estimation and compares three treatments of out-of-range counts: power-law extrapolation, a fixed minimum probability, and exclusion. Using the AoyuX platform, we conduct a 25-year (2000--2023) month-by-month pseudoprospective experiment in the China Seismic Experimental Site. We compare an epidemic-type aftershock sequence model with a homogeneous background, ETAS$μ$, with a spatially heterogeneous variant, ETAS$μ(x,y)$, across six spatial resolutions and five magnitude thresholds, using 100,000 simulations per forecast window. Count distributions follow power laws, $P(n) \sim n^{-(1+a)}$, with $a \approx 1.02 \pm 0.23$, yielding a productivity exponent $α= 0.83 \pm 0.19$ for $b = 0.85 \pm 0.002$. Model rankings depend on tail treatment: power-law extrapolation is theoretically justified and empirically most robust, whereas occasional advantages of the fixed-probability approach in coarse, low-magnitude settings are attributable to noise. Cumulative information gain shows that ETAS$μ(x,y)$ outperforms ETAS$μ$ in data-rich configurations, especially for $M_t = 3.0$--$3.5$ and whole-region tests. Positive mean information gain is concentrated where earthquakes with $M \geq 5$ occurred, whereas stochastic fluctuations dominate data-poor settings. A coefficient-of-variation analysis distinguishes genuine upward trends from noise. The framework supports statistically grounded evaluation of earthquake forecasting models.