Abstract: The successive one-step ahead predictions are used for the realization of multi-step ahead predictions for bilinear stochastic system (BLSS), i. e. the multi-step ahead predictions may simply be obtained via the system model recursions Furthermore, under some assumptions for the simplification of analysis, a Generalized Predictive Control (GPC) algorithm for BLSS can be obtained according to the cost function of generalized minimum variance. Numerical simulations have shown that the algorithm proposed is feasible, requires less on-line calculation, and possesses high steady state control precision. By adjusting the prediction horizon and properly setting the initial values of parameter estimation, fast starting can be gained without overshoot. The robustness against stepwise disturbance and time-delay variation is satisfactory.