<p>Forecasting earthquake sequences remains a central challenge in seismology, particularly under non-stationary conditions. While deep learning models have shown promise, their ability to generalize across time remains poorly understood. We evaluate neural and hybrid (NN + Markov) models for short-term earthquake forecasting on a regional catalog using temporally stratified cross-validation. Models are trained on earlier portions of the catalog and evaluated on future unseen events, enabling realistic assessment of temporal generalization. We find that while these models outperform a purely Markovian model on validation data, their test performance degrades substantially in the most recent quintile. A detailed attribution analysis reveals a shift in feature relevance over time, with later data exhibiting simpler, more Markov-consistent behavior. To support interpretability, we apply Integrated Gradients, a type of explainable AI (XAI) to analyze how models rely on different input features. These results highlight the risks of overfitting to early patterns in seismicity and underscore the importance of temporally realistic benchmarks. We conclude that forecasting skill is inherently time-dependent and benefits from combining physical priors with data-driven methods.</p>