The paper discusses the application of the breeding method for ensemble forecasting at the National Centers for Environmental Prediction (NCEP). The breeding method, introduced in 1992, generates perturbations to improve the accuracy of ensemble forecasts by simulating the development of growing errors in the analysis cycle. The method involves adding small perturbations to the control analysis, integrating the model, and scaling down the perturbations to maintain a consistent norm. This process is repeated over multiple cycles, with each cycle acting as a nonlinear perturbation model. The resulting bred vectors are superpositions of the leading local Lyapunov vectors (LLVs) of the atmosphere, which are estimated through the breeding method. These LLVs represent the fastest-growing perturbations and are crucial for ensemble forecasting. The paper highlights the importance of growing errors in the analysis cycle, which dominate subsequent forecast error growth. It also compares the breeding method with the adjoint (or optimal perturbations) technique used at the European Centre for Medium-Range Weather Forecasts (ECMWF). Both methods aim to estimate the subspace of fast-growing perturbations, but the breeding method is simpler and less computationally expensive. Experimental results show that the ensemble mean generated using the breeding method outperforms both optimally smoothed control forecasts and randomly generated ensemble forecasts. The ensemble spread is also found to be useful in predicting forecast error. The paper concludes that the breeding method is a valuable operational tool for ensemble forecasting, providing improved skill in medium-range forecasts.The paper discusses the application of the breeding method for ensemble forecasting at the National Centers for Environmental Prediction (NCEP). The breeding method, introduced in 1992, generates perturbations to improve the accuracy of ensemble forecasts by simulating the development of growing errors in the analysis cycle. The method involves adding small perturbations to the control analysis, integrating the model, and scaling down the perturbations to maintain a consistent norm. This process is repeated over multiple cycles, with each cycle acting as a nonlinear perturbation model. The resulting bred vectors are superpositions of the leading local Lyapunov vectors (LLVs) of the atmosphere, which are estimated through the breeding method. These LLVs represent the fastest-growing perturbations and are crucial for ensemble forecasting. The paper highlights the importance of growing errors in the analysis cycle, which dominate subsequent forecast error growth. It also compares the breeding method with the adjoint (or optimal perturbations) technique used at the European Centre for Medium-Range Weather Forecasts (ECMWF). Both methods aim to estimate the subspace of fast-growing perturbations, but the breeding method is simpler and less computationally expensive. Experimental results show that the ensemble mean generated using the breeding method outperforms both optimally smoothed control forecasts and randomly generated ensemble forecasts. The ensemble spread is also found to be useful in predicting forecast error. The paper concludes that the breeding method is a valuable operational tool for ensemble forecasting, providing improved skill in medium-range forecasts.