Drought is a natural occurrence caused by repetitive and ephemeral that it is less rainfall than average long term and can occur in any climate. Since the drought phenomenon is stochastic and nonlinear, stochastic linear models, neural networks and hybrid models can be useful in the development forecasting results. This study models the performance of ARIMA, neural networks and hybrid models ARIMA prediction of hydrological drought in both monthly and seasonal time scale deals and SDI index is as a predictor of the watershed was selected Gamasiab in period (1353-1387). The period (1353 - 1379) used for calibration and the remaining 8 years used for verification in model. Results show that, among the three models used to predict one time step neural network models - stochastic (hybrid) models are suitable to the monthly and seasonal scales. So Kappa statistic values and the relative error of the model at monthly time scales Polchehr station (the outlet), respectively RME= 5/79  %, K= 0.565 and seasonal time scales Station Doab (the middle) RME = 22% and is K=0.232.