The breeding method has been used to generate perturbations for ensemble forecasting at the National Centers for Environmental Prediction (NCEP) since December 1992. Initially, a single breeding cycle with a pair of bred forecasts was implemented, but in March 1994, the ensemble was expanded to seven independent breeding cycles on the Cray C90 supercomputer, with forecasts extended to 16 days, providing 17 independent global forecasts valid for two weeks daily. The breeding method simulates the development of growing errors in the analysis cycle, using a difference field between two nonlinear forecasts, which is carried forward and scaled down at regular intervals. Bred vectors are superpositions of leading local Lyapunov vectors (LLVs), which are important for representing analysis errors. The method ensures quasi-orthogonality among global bred vectors from independent cycles. Experimental runs with a 10-member ensemble (five independent breeding cycles) showed that the ensemble mean outperformed optimally smoothed control and randomly generated forecasts, and compared favorably with the medium-range double horizontal resolution control. A relationship between ensemble spread and forecast error was found in both spatial and temporal domains. The breeding method provides estimates of the fastest sustainable growth of perturbations, representing probable growing analysis errors. In contrast, optimal perturbations estimate vectors with the fastest transient growth in the future. Breeding is simpler and less expensive than the adjoint technique. The breeding method generates initial perturbations by adding a small arbitrary perturbation to the atmospheric analysis, integrating the model, subtracting one forecast from the other, and scaling down the difference field. This process is repeated forward in time, resulting in bred perturbations that grow fastest on the trajectory of the evolving atmosphere. Bred vectors are determined by the dynamics of the system and are not sensitive to the type of norm used for rescaling. The method is effective in capturing the leading local Lyapunov vectors, which are crucial for representing analysis errors. Multiple breeding cycles produce perturbations that are quasi-orthogonal, ensuring that the ensemble can span the space of possible fast-growing analysis errors. Regional rescaling is used to reflect geographically varying uncertainty in the analysis, with perturbation amplitudes adjusted based on the estimated uncertainty. This approach ensures that perturbations in poorly observed regions grow freely, while those in well-observed areas are scaled back to the size of the estimated analysis error. The breeding method has been shown to improve the skill of ensemble forecasts, particularly in the Southern Hemisphere and Tropics, while showing no change in the northern extratropics. The method is a useful operational forecast tool, providing a practical solution for estimating forecast probabilities through ensemble forecasting.The breeding method has been used to generate perturbations for ensemble forecasting at the National Centers for Environmental Prediction (NCEP) since December 1992. Initially, a single breeding cycle with a pair of bred forecasts was implemented, but in March 1994, the ensemble was expanded to seven independent breeding cycles on the Cray C90 supercomputer, with forecasts extended to 16 days, providing 17 independent global forecasts valid for two weeks daily. The breeding method simulates the development of growing errors in the analysis cycle, using a difference field between two nonlinear forecasts, which is carried forward and scaled down at regular intervals. Bred vectors are superpositions of leading local Lyapunov vectors (LLVs), which are important for representing analysis errors. The method ensures quasi-orthogonality among global bred vectors from independent cycles. Experimental runs with a 10-member ensemble (five independent breeding cycles) showed that the ensemble mean outperformed optimally smoothed control and randomly generated forecasts, and compared favorably with the medium-range double horizontal resolution control. A relationship between ensemble spread and forecast error was found in both spatial and temporal domains. The breeding method provides estimates of the fastest sustainable growth of perturbations, representing probable growing analysis errors. In contrast, optimal perturbations estimate vectors with the fastest transient growth in the future. Breeding is simpler and less expensive than the adjoint technique. The breeding method generates initial perturbations by adding a small arbitrary perturbation to the atmospheric analysis, integrating the model, subtracting one forecast from the other, and scaling down the difference field. This process is repeated forward in time, resulting in bred perturbations that grow fastest on the trajectory of the evolving atmosphere. Bred vectors are determined by the dynamics of the system and are not sensitive to the type of norm used for rescaling. The method is effective in capturing the leading local Lyapunov vectors, which are crucial for representing analysis errors. Multiple breeding cycles produce perturbations that are quasi-orthogonal, ensuring that the ensemble can span the space of possible fast-growing analysis errors. Regional rescaling is used to reflect geographically varying uncertainty in the analysis, with perturbation amplitudes adjusted based on the estimated uncertainty. This approach ensures that perturbations in poorly observed regions grow freely, while those in well-observed areas are scaled back to the size of the estimated analysis error. The breeding method has been shown to improve the skill of ensemble forecasts, particularly in the Southern Hemisphere and Tropics, while showing no change in the northern extratropics. The method is a useful operational forecast tool, providing a practical solution for estimating forecast probabilities through ensemble forecasting.