Abstract: Using the data of temperature in Beijing during the period of 1951-1990. the predictivity and reconstruction for missing records is investigated. the prediction equations are built by six statistical models:autoregression, autoregression for selecive order, stepwise regression and stepwise regression with minimum of forecast error for uni-month series, and the later two models for multi-month series. The results show that the model of stepwise regression with minimum of forecast error has the best predictivity in all of the models. The reconstruction for missing records during 1841-1950 for temperature in Beijing has been completed using the model of stepwise regression with minimum of forecast error.