In view of the limitations of traditional Markov chain and its improved prediction methods which can only predict the state, in this paper we realized a critical improvement of the Markov chain forecasting method to being able to conduct numerical prediction. We did so by using weighted summation of the average value of each state multiplied by the corresponding predicted probability, on the basis of obtaining the predicted probability of each state with the traditional Markov chain forecasting method according to the characteristics of dependent stochastic variables. The data of this study were collected from Daojieba hydrological station on the Nujiang river, which is a famous international river in southwest China. We used the runoff series from 1957 to 2010 and the suspended sediment series from 1964 to 2010 for analysis, and used the runoff and suspended sediment series from 2011 to 2015 for validation. Results showed that the re-weighted Markov chain forecasting had a high accuracy in numerical prediction and could meet the demand of short-term numerical prediction in stochastic time series.