Forecasting future trends in complex systems requires the integration of diverse data sources beyond traditional numerical time series. However, effectively fusing fundamentally heterogeneous data, where external signals, such as textual reports, are independently sourced rather than derived from the series itself, remains a significant challenge. Current multi-modal frameworks often rely on tightly coupled architectures that require complex joint optimization of internal representations, or utilize static fusion mechanisms that fail to adapt to shifting data relevance. We propose GMM-TS, a modular gating architecture inspired by Mixture-of-Experts, designed for universal heterogeneous integration. Unlike monolithic models, GMM-TS operates directly in the target prediction space, enabling the seamless integration of disparate model families, such as Large Language Models and specialized numerical forecasters, without internal architectural modification. Furthermore, GMM-TS replaces traditional fixed-weight interpolation with a Transformer-based gating mechanism that computes dynamic, per-time-step expert weights. This enables the framework to adaptively prioritize specific modalities as temporal context evolves, providing both superior predictive accuracy and inherent interpretability by revealing the relative importance of each data source across different time horizons and events. Extensive evaluations show that GMM-TS consistently outperforms state-of-the-art baselines across nine domains, multiple forecast horizons, and various expert configurations. We also include, for the first time, to the best of our knowledge, the option to integrate more than two experts. Our framework is efficient, extensible, and interpretable, offering a modality-agnostic solution for real-world forecasting in multi-source data environments. Code is available at https: //github. com/NVIDIA/gmmtsₗib.
Razmadze et al. (Tue,) studied this question.