WRNet is a machine-learning approach for regularized win ratio regression tailored to hierarchical composite endpoints. It optimizes feature selection and risk prediction through an elastic net-type penalty on regression coefficients (log-win ratios), making it suitable for high-dimensional data.