Crisp: A Spectral-Based Interaction Strategy for Multivariate Time Series Forecasting

Multivariate time series (MTS) forecasting critically relies on effectively modeling inter-variable dependencies. However, existing paradigms often face an inherent trade-off: channel-isolation strategies may lead to information fragmentation in strongly coupled systems, while channelinteraction methods can introduce spurious dependencies among irrelevant variables. To address this challenge, we propose Coherent Resonance Interaction with Spectral Priors (Crisp), a novel framework built on the principle that effective information exchange should occur only among variables exhibiting compatible oscillatory patterns. Specifically, Crisp derives spectral priors in the frequency domain to construct dynamic resonance topologies. Through a differentiable, adaptive, and strictly sparse blocking mechanism, Crisp explicitly sets the attention weights of spectrally inconsistent neighbors to zero, thereby suppressing spurious interactions. Furthermore, we introduce a spectral-gated feature filtering module that refines variable representations according to their intrinsic spectral characteristics. Extensive experiments demonstrate that Crisp achieves the best or highly competitive performance across most settings. The code of Crisp is available at GitHub.