To improve the reliability of parameter identification in the hydrologic modelling, this study combined auto regressive model with Markov chain Monte Carlo method, and proposed an auto regressive(AR)model based modified Markov Chain-Monte Carlo(AR-MCMC). The residual covariance matrix in the traditional MCMC method is modified by using the AR model. Based on a case study of snowmelt runoff model(SRM)in Xinjiang’s Tiznavu watershed, it can be found that the residual sequence of snowmelt runoff simulation has significant autocorrelation. By correcting the residual covariance matrix, the marginal likelihood of AR-MCMC is larger than that of MCMC. Based on the assessments of multiple indexes, AR-MCMC method has a better prediction interval than MCMC. In addition, when comparing the Nash coefficients in calibration and verification periods, they are 0. 86 and 0. 89 respectively for AR-MCMC, while 0. 84 and 0. 87 respectively for MCMC. Therefore, the snowmelt runoff simulation obtained by AR-MCMC method is better than that obtained by MCMC.