A new method in selecting the inputs of the neural networks for multivariate time series is proposed. The input and output time series are analyzed and suitable mathematical models are built in the input-output model parametric representation. The inputs to the best input-output models are chosen as the inputs to the neural network model. The estimates from different models are then combined to obtain an unbiased estimate of the output series. The resulting combined neural network model is used in forecasting the bench marking gas furnace multivariate time series. The new developed procedure outperformed previously used forecasting techniques.