When Attention Does Not Explain the Peak: Temporal Reference vs. Forecast Output in Attention-Based Time-Series Forecasting
The study investigates if attention mechanisms correctly identify when a forecast peak occurs by comparing attention weights to actual forecast timing. It found a dissociation: while the model's internal attention reference ($\Psi_{out}$) is structured and reproducible, it does not align with the true or predicted peak horizons. This suggests that the structure of attention is an internal reference, not necessarily a selector for the forecast peak.