AME-TS: Anchored Mixture-of-Experts for Time Series Forecasting

Time series forecasting models are increasingly scaled through large Transformer backbones, yet most existing approaches process all series through a shared dense computation path despite substantial heterogeneity in temporal structure. Mixture-of-Experts (MoE) offers a natural alternative by enabling conditional computation, but standard MoE routing leaves expert specialization weakly identified and often unstable during downstream adaptation. We propose AME-TS, a regime-aware MoE framework for time series forecasting that aligns expert routing with interpretable temporal structure. AME-TS first uses a lightweight regime predictor to estimate series level descriptors, including forecastability, seasonality, trend, and sparsity, and maps them to a soft structural prior over experts. This series-level prior guides token-level routing during training through a training-only prior-alignment loss, encouraging structure-aligned specialization. On the GIFT-Eval benchmark, AME-TS achieves state-of-the-art or competitive performance across model scales while using substantially fewer active parameters than recent time series foundation models. We further show that AME-TS learns more interpretable routing geometry and substantially more stable expert specialization than standard MoE during fine-tuning on M5. These results suggest that structure-aware routing is an effective and reliable way to realize the benefits of sparse expert models for time series forecasting.