NVExplain Offers Interpretable Time Series Forecast Explanations.

Time series forecasting models are widely deployed in critical applications, but their "black box" nature makes their predictions difficult to interpret, especially regarding which historical factors influence specific future horizons. Existing post-hoc explanation methods often overlook the inherent temporal dependencies in time series data and fail to provide horizon-specific insights. To address this, researchers introduce NVExplain, a model-agnostic explainability framework. NVExplain explains forecasting predictions by attributing each forecast horizon to the most relevant historical lags. It conceptualizes forecasting as a latent trajectory and employs "semantic flow" to quantify how information propagates and evolves within the model's internal representations over time. By aggregating this semantic flow, NVExplain constructs a lag-horizon attribution matrix, clearly showing the temporal influence for each forecast point. To enhance human readability and temporal coherence, the framework also generates structure-preserving perturbations and fits sparse local surrogate models. Evaluations across multiple benchmark datasets demonstrate that NVExplain's semantic-flow variant achieves competitive or superior faithfulness compared to standard baselines, with significantly improved computational efficiency and robust, stable explanations.