New Method Improves Coherence in Hierarchical Time Series Forecasting

Forecasting in real-world scenarios like retail sales or energy usage often involves hierarchical data structures where predictions must be accurate and consistent across different aggregation levels. Traditional methods typically enforce this consistency, or "coherence," through post-processing, which can overlook complex temporal dependencies. A new approach, Hierarchical Temporal Fusion (HTF), extends the Temporal Fusion Transformer by integrating structured hierarchical embeddings and a coherence-aware loss function. Unlike older methods, HTF incorporates coherence directly into the model's training process. It penalizes discrepancies between aggregated child forecasts and their parent forecasts, allowing the model to learn both temporal dynamics and structural consistency simultaneously. Evaluations on benchmark datasets, including M5 Walmart sales and energy consumption data, show that HTF significantly reduces forecast incoherence while also improving overall accuracy. Attention visualizations and embedding analysis provide further insights into how this method effectively combines temporal and structural information for superior hierarchical forecasting performance.