The Failures of Marginal Influence-Based Attribution Methods for Global Time Series Explanations

Attribution methods for time series models have a fundamental flaw: The paper proves that popular explainability methods like SHAP fail not because they produce single numbers instead of richer explanations, but because they incorrectly compute which past events actually influenced a model's decision by conflating direct causes with indirect ones when data is autocorrelated (i.e., when values are correlated with their own past values).

Current explanations may mislead users about what the model actually learned: The authors introduce "DAG-faithfulness" as a standard—explanations should accurately reflect the temporal cause-and-effect relationships the model has learned. They find that standard methods, including SHAP and recent time-series-specific variants, fail this test, meaning users could misunderstand which historical events truly drove the model's predictions.

The problem is widespread across the field: Rather than being isolated to one method, this computational mismatch affects "the dominant" explainability approaches for time series, suggesting that many deployed systems using these methods may be providing unreliable explanations to stakeholders.

This matters for high-stakes decisions: For applications like financial forecasting, healthcare monitoring, or infrastructure management where time series models inform critical decisions, incorrect explanations could lead organizations to trust the wrong signals or make decisions based on misunderstood model logic.