Predicting Delayed Train Trajectories on the Dutch Railway Network: Explainable AI Evaluation of Topological, Operational and Weather Features with Tree Based Ensemble Methods

Explainability matters more than raw accuracy: The researchers chose interpretable tree-based models over more complex "black box" deep learning systems, allowing them to understand why predictions fail—a crucial consideration for public services like railways where decisions affect millions of passengers.

Models work well in the short term but fail over time: The study found that train delay predictions performed well when tested on data from the same month, but accuracy systematically declined when predicting delays in future months, revealing a critical limitation that opaque AI systems might hide.

Environmental factors are unstable predictors: Weather and seasonal variations caused prediction performance to degrade over longer timeframes, suggesting that simply adding more data features (like weather conditions) cannot solve fundamental forecasting challenges without accounting for how these factors change seasonally.

Current AI approaches have inherent limits for long-term planning: The research demonstrates that richer datasets alone cannot overcome the constraints of predicting train delays far into the future, indicating that railway operators need fundamentally different AI architectures rather than incremental improvements to existing models.

Responsible AI requires acknowledging uncertainty: By transparently documenting where and why their model fails, the researchers provide actionable insights for practitioners—showing that admitting limitations is more valuable for public infrastructure than claiming high accuracy without understanding the underlying causes.