Time Series Explainability Methods Flawed by Autocorrelation.

Explainability methods for time series models typically assign a single scalar score to quantify the direct influence of a feature at a specific timestamp. However, this research highlights a critical limitation: these methods, which often rely on marginal conditioning or off-manifold gradients, fail to distinguish between direct temporal dependencies and those mediated through other variables when autocorrelation is present in the data. This conflation leads to misleading explanations of how features impact model predictions over time. The paper introduces a new concept called DAG-faithfulness. An explanation is considered DAG-faithful if the temporal dependency graph it implicitly encodes is Markov-equivalent to the temporal directed acyclic graph (DAG) that the model has learned. Crucially, the study proves that widely used attribution methods, including SHAP, are generally not DAG-faithful. Furthermore, recent extensions designed specifically for time series data inherit this same computational limitation, meaning they also struggle to accurately represent the true causal or temporal relationships within the model. This suggests a need for new approaches that can correctly disentangle direct from mediated influences in autocorrelated time series data.