Controlled Local–Global Channel Communication for Multivariate Long-Term Forecasting

Abstract

Multivariate long-term forecasting requires deciding how much cross-channel structure can be exploited reliably, yet existing architectures fix this decision by design, from no exchange at all to unrestricted mixing. We argue that the amount should instead be selected from data, and propose CLOC (Controlled Local–Global Channel Communication), a framework in which it becomes an explicit operating point with an exact zero-communication endpoint; setting the residual scales to zero recovers the channel-independent model within the same family. Local grouped attention restricts exchange through group size g, while K learnable prototypes mediate global context and the residual scales α and β control message strength. Operating points are selected on validation data under a stated rule and fixed before testing. Across eight benchmarks under a unified 720-step lookback, validation selects qualitatively different regimes, from full suppression on the four ETT datasets to local–global communication on Electricity and Solar-Energy. CLOC attains the best or tied-best result on 50 of 64 horizon-level metrics and a mean rank of 1.45 among eleven models, against 3.73 for the next best, while dense mixing runs out of memory at 862 channels. Thus, forecasting quality depends on matching communication to the structure a dataset can reliably support, not on maximizing it.