Abstract
Time series forecasting is often influenced by heterogeneous contextual data, whose integration is essential for accurate prediction. Although such data can be temporally aligned with a time series, their structural heterogeneity remains, limiting the forecasting performance of state-of-the-art Temporal Fusion Transformer (TFT) model. To tackle this problem, we propose a Mixture-of-Experts enhanced Temporal Fusion Transformer model, termed MoE-TFT, which explicitly models heterogeneity that is present in temporal contextual data even after alignment. Specifically, we extend TFT by specialized expert subnetworks together with a learnable gating mechanism that dynamically routes feature representations to the most suitable experts. Experiments on real-world datasets show that MoE-TFT consistently outperforms TFT and other baseline models in terms of quantile risk.