Answers are grounded in this paper and cite their sources.
Transformer-based and multilayer-perceptron-based methods have emerged as leading approaches in time series forecasting. While transformer-based architectures demonstrate strength in modeling long-range dependencies, they encounter challenges with computational complexities and overfitting tendencies. In contrast, multilayer-perceptron-based approaches provide computational efficiency and proficiency in temporal dynamics modeling, yet they face difficulties in effectively capturing temporal patterns across complex scales. Real-world time series often exhibit diverse patterns across multiple scales, and future changes are influenced by the complex interactions of these overlapping temporal scales, necessitating high-capacity models. Drawing from the observation of multi-scale entanglement effect in time series, we introduce a lightweight orthogonal adaptive multi-scale decomposition (LOAMD) framework for time series forecasting, building upon the original adaptive multi-scale decomposition approach. Our framework decomposes time series into distinct temporal patterns at multiple scales, leveraging improved multi-scale decomposable mixing to dissect and aggregate these patterns. Additionally, LOAMD introduces lightweight expert design with orthogonal constraints and time–frequency loss functions to achieve superior parameter efficiency while maintaining forecasting accuracy. Comprehensive experiments demonstrate that our LOAMD framework not only addresses the constraints of existing approaches but also consistently delivers state-of-the-art performance across various datasets.
No takes yet. Share an insight, caveat, or question.
Jing et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: