Based on the samples of typhoons in July,August and September from 1960 to 2003 over the South China Sea,multivariate regression prediction models of typhoon track for July,August and September developed by using the condition number method in selecting predictors from numerical prediction products and predictors used in the CLIPER model of typhoon track,and prediction experiments indicate that the average error of the three models is 153.9 km,which is smaller than those of objective prediction models currently used in China and western countries.Comparison experiments of new regression models with stepwise regression models and CLIPER models are performed under the conditions of identical modeling and independent samples and identical or almost identical predictor mumber,and results show that new regression models are superior to the stepwise regression models in prediction accuracy,and the forecast skill of new regression models relative to the CLIPER models is positive.