Abstract: A forecast factor set, or a certain combination of forecast factors, is crucial to forecasting accuracy, since its size determines the number of information sources for the forecasting. Considering the shortcoming in previous methods that focused on the cases of one single factor, this paper, from a holistic perspective, presents a method for selecting hydrological forecast factor set based on joint mutual information. First, we introduce the concept of mutual information and extend it to high-dimensional cases. Both conditional mutual information and joint mutual information are illustrated and calculated by the Parzen window estimation method. Then, using conditional mutual information, a maximum joint mutual information (JMI) model is constructed and solved for application of hydrological forecasting. Finally, the results of this method are tested via calculations of back propagation (BP) neural network and compared with those of previous methods in a case study of the Luning hydrological station in the Yalong River basin. This work shows that the new method can generate more appropriate inputs for hydrological forecasting models.