Among the methods applicable to time series forecasting the autoregressive models stand out. This is explained by their easy mathematical tractability allied with good results. The adjustment of the free parameters is done by an analytic solution - Yule-Walker equations -, which allows to find the global optimum of their cost function. However, the prediction of up to one step ahead in a recursive way violates the global optimization, once suboptimum points may achieve better performances. This work performs the streamflow series forecasting from hydroelectric plants, for a twelve steps ahead horizon. The calculation of the coefficients is done by the artificial immune systems and LMS algorithm. The parameters found were different from those achieved by the Yule-Walker equations, which means that the optimization processes converge to distinct solutions. The computational results indicate that other optimization methodologies should be considered to increase the performance of autoregressive models in multistep prediction.