Neural Negative Binomial Regression for Weekly Seismicity Forecasting: Per-Cell Dispersion Estimation and Tail Risk Assessment
Researchers introduce EarthquakeNet, a neural architecture that learns per-location overdispersion parameters for seismic forecasting rather than assuming a global statistical model. The work demonstrates that standard Poisson assumptions fail dramatically on real seismic data (p < 10^-179) and proposes learned spatial embeddings to capture localized variance patterns. This represents a methodological shift in domain-specific forecasting: moving from hand-tuned statistical assumptions to neural-learned heterogeneous parameters, a pattern increasingly relevant across scientific computing and risk modeling where one-size-fits-all distributional assumptions break down.
Modelwire context
ExplainerThe paper's core contribution isn't just that Poisson fails for earthquakes (that's been known), but that learning spatial embeddings to capture location-specific variance patterns outperforms fitting a single global overdispersion parameter. This shifts the burden from domain experts hand-tuning statistical assumptions to neural networks discovering heterogeneous parameters from data.
This follows the same architectural logic as the GNSS positioning work from earlier this week, which applied activation functions to classical weighted least squares to learn domain-specific error patterns rather than relying on purely geometric models. Both papers treat a classical engineering problem as a learned optimization task where neural primitives absorb noise patterns that fixed statistical assumptions miss. The seismic work goes further by making the learned component explicitly spatial (per-cell embeddings) rather than just signal-level, suggesting a maturing pattern: when one-size-fits-all distributional assumptions fail catastrophically, learn the heterogeneity.
If EarthquakeNet's per-cell dispersion estimates correlate with known geological features (fault zones, plate boundaries, crustal composition) without explicit geological inputs, that confirms the model is learning meaningful spatial structure rather than overfitting. If the approach transfers to other rare-event forecasting domains (volcanic activity, landslides) with minimal retraining, that signals genuine methodological portability.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsEarthquakeNet · Central Asia seismic dataset
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.