This paper proposes a novel backbone architecture, the Temporal Evidence Fusion Network (TEFN), that combines accuracy and efficiency in temporal time series forecasting. TEFN captures uncertainty in both the channel and temporal dimensions of multivariate time series data by introducing a Default Probability Assignment (BPA) module based on evidence theory. We then develop a novel multi-information fusion method that effectively integrates information from both dimensions from the BPA output, thereby improving forecast accuracy. Experimental results demonstrate that TEFN achieves performance comparable to state-of-the-art methods while significantly reducing complexity and training time. Furthermore, TEFN exhibits high robustness by minimizing error variance during hyperparameter selection, and its fuzzy theory-derived BPA provides high interpretability. Therefore, TEFN is a desirable solution for temporal time series forecasting that balances accuracy, efficiency, robustness, and interpretability.
