The sliding distance of an earthquake landslide is significantly different from that of a gravity landslide. Scientific prediction of the sliding distance of seismically-induced landslides in loess regions is an effective way to reasonably assess the risk and minimize the hazards of such landslides. Based on 400 groups of field survey data of loess landslides triggered by the 1920 Haiyuan earthquake and 67 sets of verification data, feasibility and suitability of the back propagation (BP) neural network model for predicting sliding distances of seismic landslides was demonstrated. Comparison of the results of BP neural network algorithm with those of traditional multiple linear regression and multiple nonlinear regression showed that the BP neural network was a superior predictor of real-life situations. This study can be used to predict landslide slip of loess earthquakes.