Neural Earthquake Forecasting with Minimal Information: Limits, Interpretability, and the Role of Markov Structure

Deep learning models for earthquake forecasting show significant performance degradation over time: While neural network models outperformed simpler Markovian baselines during validation, their accuracy "degraded substantially in the most recent quintile" of test data, demonstrating that past performance does not guarantee future reliability in this domain.

Models may overfit to outdated patterns rather than capture fundamental earthquake physics: The research found that "feature relevance shifted over time," with later earthquake data showing "simpler, more Markov-consistent behavior," suggesting models trained on earlier data may learn spurious patterns that don't persist—a critical concern for real-world deployment.

Temporal realism in testing is essential for responsible AI evaluation: The authors emphasize the importance of "temporally realistic benchmarks" and "temporally stratified cross-validation" rather than random data splits, highlighting that standard machine learning evaluation practices can mask serious generalization failures in time-dependent domains.

Explainability tools are necessary but not sufficient for trustworthiness: The paper applies Integrated Gradients to interpret model decisions, but this transparency alone doesn't solve the underlying problem of performance decay, suggesting that interpretability must be paired with rigorous temporal validation to build genuinely reliable systems.

Hybrid approaches combining physics and data-driven methods show promise: The research concludes that forecasting "benefits from combining physical priors with data-driven methods," suggesting that responsible AI in high-stakes domains like seismology requires integrating domain expertise rather than relying on pure machine learning.