NEST Addresses Dataset Distribution Shifts in Long-Term Forecasting

This paper introduces NEST, a novel framework aimed at improving the accuracy of long-term forecasting, particularly in systems characterized by dynamic multivariate time-series data. A common challenge in such scenarios is the presence of dataset-level distribution shifts, where the underlying behavior of the system changes over time, leading to different operational modes. Unlike existing methods that primarily focus on localized temporal shifts, NEST specifically addresses these broader structural changes. NEST employs a two-phase dense Mixture-of-Experts (MoE) architecture. Initially, it partitions the dataset into distinct operational regimes using unsupervised clustering within a principled moment-entropy space. This allows the model to identify and specialize in different behavioral patterns. A regime-oriented router then generates initial expert weights, which are further refined through geometric modulation. Crucially, the individual experts within NEST are not monolithic predictors but rather specialized kernels. These kernels capture regime-specific dynamics by evolving unique variate-attention patterns, enabling them to adapt effectively to the identified operational modes. Extensive evaluations on diverse benchmarks, including network traffic and physical phenomena, demonstrate that NEST consistently achieves state-of-the-art performance.