A machine learning-based method is proposed for predicting transmission efficiency in high-speed gear reducers. An extreme gradient boosting (XGBoost) model, trained on fundamental friction test data, is employed to estimate the gear friction coefficient and meshing power losses, showing strong alignment with experimental results. Additionally, oil churning power losses in the rotating components are evaluated through computational fluid dynamics (CFD) simulations, incorporating the sliding mesh technique. The influence of input speed and torque on various power losses, efficiency losses, and overall transmission efficiency is investigated. As a result, meshing efficiency losses decrease significantly at low rotational speeds, leading to a positive correlation between transmission efficiency and speed. However, at higher rotational speeds, oil churning losses increase sharply, resulting in a negative correlation between transmission efficiency and speed. Notably, the ratio of churning power loss to total power loss at 15000 rpm is 215 times higher than at 1000 rpm, highlighting the critical importance of optimizing lubrication strategies and housing design to enhance the transmission efficiency of high-speed reducers.