M${}^2$TSF: LLM-empowered time series forecasting via multi-stream multi-scale modeling

Time series forecasting (TSF) demands capturing multi-scale temporal dynamics that range from short-term fluctuations to long-term evolving trends. Recent advances have witnessed the growing success of Large Language Models (LLMs) in TSF, particularly in capturing multi-scale temporal dynamics. However, existing methods remain constrained by a single-stream architecture that concatenates multi-scale features into a unified sequence, inevitably entangling heterogeneous temporal patterns and introducing cross-scale noise interference. To address this, we propose ${\mathrm{M}}^2\mathrm{TSF}$, an LLM-empowered **T**ime **S**eries **F**orecasting framework via **M**ulti-stream **M**ulti-scale modeling, which assigns each temporal scale an independent processing pathway within the LLM. Specifically, a multi-scale tokenization module is designed to obtain the multi-scale representations. Then, a constrained intermediate interaction (CII) module is designed to enable controlled and progressive cross-scale information exchange aligned with the hierarchical semantics of LLM layers, integrating fine-grained features at lower layers and high-level global semantics at deeper layers. In addition, a pattern-aware fusion (PAF) module is introduced to emphasize informative streams while mitigating interference from irrelevant ones via a hyperedge-centric mixture-of-experts (HMoE) mechanism. Extensive experiments on 17 real-world benchmarks across diverse domains demonstrate that ${\mathrm{M}}^2\mathrm{TSF}$ achieves state-of-the-art performance. Code is available at: https://anonymous.4open.science/r/M2TSF.