Guiding Attention with Information Flow Dynamics for Time Series Forecasting

Effective operational planning demands knowing what drives a process, not just what correlates with it. However, standard Transformer attention is symmetric. This creates a fundamental problem for decision support systems, as key industrial and operational processes are directional. They learn spurious correlations from noisy data instead of the influential dynamic drivers needed for robust planning. This limitation directly impacts model reliability. We address this gap by integrating a neural estimator for Transfer Entropy, an explicit measure of directional information flow, directly into the attention mechanism. This entropy-guided framework learns to guide attention toward truly influential past states, moving beyond simple content-based similarity. Our main contribution is to embed this dynamic information-theoretic estimator directly into the attention computation instead of using it as a static pre-processing mask. We validated our method on widely-used, industrial-relevant forecasting benchmarks. Our model achieves highly competitive accuracy and, more importantly, demonstrates improved robustness. This work shows that encoding information-theoretic principles for directionality is a critical and effective strategy for building more reliable forecasting models.