Centrality-Augmented Time Series Forecasting

초록

Time-series forecasting often struggles to capture complex patternsand inter-series interactions, especially when global structuralrelationships among multivariate series are left implicit. This paperproposes a network-feature-augmented forecasting framework thatconstructs separate date-node networks based on calendarco-occurrence, price-regime similarity, and shared event labels,extracts degree, closeness, and betweenness centrality features fromeach network, and injects them as static covariates into a deepforecasting backbone. The centrality features are weighted through alightweight gating module and provided to the temporal model so thatpredictions can leverage global relational context in addition to localtemporal dynamics. To prevent data leakage, all networks areconstructed using historical training data only and kept static; theextracted centralities remain fixed during inference. Experiments onthe M5 dataset show that the proposed approach consistently improvesmean absolute error compared with representative baselines fromregression, deep learning, and Transformer families, while root meansquared error remains competitive. The model maintains a 7.5-16.6%MAE advantage over the original Temporal Fusion Transformer acrossprediction horizons from 20 to 100 steps. Ablation results confirm thatall three centralities contribute, with betweenness providing the largestmarginal benefit. Overall, integrating graph-derived global descriptorsas static covariates improves accuracy and stability for long-horizonmultivariate forecasting

키워드