MS-FLOW: A Sparse Bottleneck Framework for Reliable Multivariate Time Series Forecasting

Researchers have introduced MS-FLOW, a novel sparse-bottleneck framework designed to enhance multivariate time series forecasting. While existing methods often rely on dense cross-channel interactions to improve accuracy, these approaches can amplify spurious correlations and cause representation over-smoothing in noisy, state-dependent real-world data. MS-FLOW addresses this by modeling inter-variable interactions as capacity-limited information flow. It replaces fully connected communication with selective sparse routing, retaining only critical dependency paths under a strict communication budget. This mechanism suppresses redundant connections and prevents the propagation of false correlations. Extensive experiments conducted on twelve real-world benchmarks demonstrate that MS-FLOW achieves state-of-the-art forecasting accuracy. The framework successfully shifts the paradigm of multivariate forecasting from maximizing interaction volume to ensuring interaction effectiveness, resulting in fewer but more reliable dependencies. This advancement offers a robust solution for complex systems where accurate prediction relies on distinguishing genuine variable relationships from noise.