Keywords: 
• forest trees /
• phenological forecasting /
• linear autoregression models

Abstract: The linear autoregression model with input variables is a comprehensive forecasting model which is superior to the conventional model for phenological forecast. It belongs to the dynamic random difference equation in structure and combines the merits of both the linear autoregression model and linear multivariant regression model. As the new model has its lag ,the prediction value is not only related to current input ( long-term weather forecast results but also is affected by historical input and hysteresis. This means that the long-term weather forecast results do not exert much influence on the prediction value, hence increase the prediction precision. The revision of model paramaters is made by means of recursion least squares and the parematers are constantly revised along with the increase in the number of predictions, approximating the prediction value to the actual value. Results from error contrast on Prunus yedoensis,P . serrulata and Robinia pseudoacacia show that the error is kept within one or two days if the new forecast method is used, while the error could vary from one to eleven days in other methods applied. This means that the new method could better comform to dynamic variation in forecasting tree phenology.