Metal mine with fractured blocky rock mass is much different from coal mine, which shows discontinuity and irregularity, meanwhile large differences also exist in stratum structure, geological condition, ore body shape and mining methods, so the influential factors in metal mines is more complex and volatile. The research on theory and application of ground subsidence has not reached a mature stage at present. So this paper focused on the study of this issue based on neural network with its characteristic that it has fast learning speed and could approach to any non-liner mapping, which are adapted to the complex environment in metal mine. The time series prediction model was established, which is based on the measured data of the roof subsidence in the goaf of metal mine, and the tested sample data were trained and tested by many times. Finally the predicted value of the neural network was compared with the measured value by automatic optical level, which showed that the prediction model achieved good accuracy, and could be accepted in the engineering application. This method could fill the gap of incomplete monitoring data in mine ground subsidence, and provide a reference for production in the metal mine