A new AI model called Fuzzy-MoE has been proposed for forecasting non-stationary multivariate time series. Unlike existing models that compress different variables and samples into a single mapping, Fuzzy-MoE uses a fuzzy logic-based dynamic Mixture-of-Experts approach to identify latent temporal states and select the most relevant experts for each variable. This allows for improved interpretability and captures heterogeneous temporal dynamics. Experimental results show that Fuzzy-MoE outperforms other forecasting methods on multiple benchmark datasets. --- Why it matters: Fuzzy-MoE matters because it addresses a key limitation of existing time series forecasting models: their lack of interpretability. By providing explicit IF-THEN rule-based expert selection, Fuzzy-MoE enables researchers to understand which mechanisms are activated under different conditions, making it easier to debug and improve the model. Source: https://arxiv.org/abs/2608.20761