Generation of runoff sequences for simulation of complex water resources systems requires the use of multivariate stochastic models. To face the many practical difficulties existing with the building of pure multivariate ARMA type of models, contemporaneous-ARMA (CARMA) models were considered in this approach, and a conceptual stochastic ARMA univariate framework for monthly streamflows was adopted as the base model for reproduction of the serial correlation of individual series. In that framework, ARMA parameters have conceptual meaning and the model residual resembles the effective rainfall process. In the reproduction of the spatial correlation of series the proposed multisite model structure takes advantage of this analogy and considers the effective rainfall as the variable of which the spatial correlation structure is analysed and reproduced, in agreement with the climatic evidence. Evaluation of the correlation structure of this process, which has finite probability to present zero values, requires the use of an appropriate technique to deal with the contemporaneous presence of zeros throughout the stations. To face this problem, we used a stochastic approach which accounts separately for the correlation of zero and nonzero values.