NEST Improves Forecasting by Handling Dataset Distribution Shifts

Long-term forecasting in intricate systems often suffers from dataset-level distribution shifts, where varying underlying behaviors and system states drive dynamic time-series data. Current methods primarily focus on local temporal shifts, failing to explicitly model the global structural challenge where datasets are composed of distinct operational regimes. Researchers propose NEST, a specialized framework designed to model and recompose these evolving structures using a two-phase dense Mixture-of-Experts (MoE) architecture. NEST first partitions the dataset into distinct operational regimes through unsupervised clustering in a moment-entropy space. It then employs a regime-oriented router mechanism to generate initial expert weights, refined through geometric modulation to regime centroids. Crucially, individual experts in NEST function as specialized kernels, capturing regime-specific dynamics by evolving unique variate-attention patterns rather than acting as monolithic predictors. Extensive evaluations on diverse benchmarks, including network traffic and physical phenomena, demonstrate that NEST consistently achieves state-of-the-art performance.