<p>This paper presents a novel multi-neural network architecture based on the NARMAX framework for modeling and forecasting complex dynamical systems with multiple interdependent state variables. In the proposed model, each state variable is predicted by an independent neural network, all of which share a common input vector that represents the current state of the system. This architecture ensures consistency, scalability, and interpretability in high-dimensional nonlinear systems. The theoretical foundation of the model is supported by an extension of the universal approximation theorem to the vectorial setting. Empirical validation is carried out using two real-world case studies: (i) multi-step forecasting of stock prices from several Brazilian companies, and (ii) prediction of river flow dynamics in the Furnas Basin (Rio Grande, Brazil). In both applications, the multi-neural NARMAX model outperformed standard single-network and statistical baselines in terms of accuracy and robustness. These results demonstrate the potential of the proposed approach for a wide range of coupled forecasting tasks in finance and environmental sciences.</p>