Aimed at the poor performance of short term prediction of navigation satellite clock error, a method is proposed for prediction of satellite clock error based on the least square support vector machine (LS-SVM) and artificial fish-swarm algorithm (AFSA). To avoid the man-made blindness and enhance the efficiency of online forecasting, penalty parameter and kernel bandwidth parameter of LS-SVM are optimized by artificial fish-swarm algorithm with a rather good ability of global optimization based on AFSA model. The clock data of four typical GPS satellites are chosen and respectively used in three models to forecast short term clock error. The results show that the accuracy of LS-SVM based on model is superior to the other models, especially in the field of rubidium clock; the error is less than 0.5 ns, and running time is in 5 minutes. The work provides a new way for short term prediction of satellite clock error.