Temporally Coherent Counterfactual Explanations for Time Series Forecasting

Counterfactual explanations provide interpretability for machine learning models by identifying minimal, actionable changes to an input that would alter a model’s prediction. Most prior work focuses on tabular settings, where features are treated as independent within a static instance, making counterfactual generation relatively tractable due to limited structural constraints. In contrast, counterfactual explanations for time-series forecasting are substantially more challenging. Time series exhibit temporal dependencies, seasonality, and structured evolution. Time-series counterfactuals must therefore satisfy predictive validity while remaining sparse and temporally coherent, as outcomes depend on the evolution of features over time. Existing time-series approaches either rely on input-space optimization, which often introduces unrealistic temporal artifacts, or operate in flat latent spaces that lack mechanisms to enforce temporal coherence. Consequently, generating counterfactuals that are both valid and temporally coherent remains an open challenge. In this paper, we propose TempACE (Temporal Attention-Guided Counterfactual Explainer), a framework for counterfactual generation within a hierarchically structured latent space that explicitly encodes temporal dependencies. TempACE employs a ladder variational autoencoder with Temporal Convolutional Attention (TCA), capturing multi-scale temporal structure, while attention-guided optimization constrains edits to influential time steps on the data manifold. Experimental results demonstrate that TempACE outperforms state-of-the-art input- and latent-space baselines.