Abstract: Due to the non-linearity feature of data series of traffic flow,a nonparametric predicting method based on the theory of decision tree is proposed to overcome the deficiencies of current short-term forecasting methods of traffic flow.Using time series lags,the data series of traffic flow are converted into a format which can be recognized by a non-parameter model.Considering the long-term co-integration relationship between traffic flow parameters,combinational vectors of volume and speed lags are established,providing a basis for this forecasting model.Then a decision tree model based on classification and regression tree (CART)is developed for the prediction for the parameters of short-term traffic flow.Based on actual data of traffic flow,the performance of this CART decision tree model is evaluated under different types of road and speed intervals.The results show that the CART decision tree model is superior to the general time se-ries model in terms of the prediction accuracy.In addition,the accuracy of speed prediction is generally higher than that of volume prediction.The mean average percentage error (MAPE)of speed prediction is less than 13% while the MAPE of volume prediction is 30%.Besides,this model can perform better when it is constructed with the data from weekdays or weekends separately.The evaluation of performance under different speed intervals indicates that the CART decision tree model presents high accuracy and stability in medium intervals speed for all types of road.