Abstract
Multivariate time series forecasting is crucial for numerous practical applications ranging from financial markets to climate monitoring. Traditional multivariate time series forecasting methods primarily adopt a time-centric modeling paradigm, applying attention mechanisms to the temporal dimension, which presents significant limitations when handling complex dependencies between variables. To better capture inter-variable interaction patterns, this paper proposes the Variable-Centric Transformer (VCformer), which shifts the attention paradigm from time-centric to variable-centric through sequence transposition. Building upon this foundation, we further design a dual-scale architecture that simultaneously models feature representations at both the original variable level and variable group level. Combined with an adaptive variable grouping mechanism, we construct a parameter-sharing dual-path encoder and finally select the optimal feature fusion strategy through comparative experiments. Experimental results on seven benchmark datasets demonstrate that VCformer achieves comprehensive improvements in prediction accuracy compared to traditional time-centric methods, while exhibiting stronger modeling capabilities on high-dimensional data. Ablation studies and interpretability analysis further validate the effectiveness of each component.