The motorcycle tilt angle variation is a crucial indicator for the motorcycle stability. However, due to the dynamic and complex nature of the environment and motorcycle mechanics, the current tilt angle is insufficient to generate a fall warning timely. To generate the fall warning in advance, it's essential to establish a predictive model for future tilt angle variations in motorcycles. Such a model could offer proactive insights for maintaining stability in tasks involving autonomous motorcycles and ensure early avoidance of erroneous driving behaviors. In this paper, we propose a predictive model for future tilt angle variations based on time series feature engineering. A simulation environment is construcapted to generate good and bad driving data, in which a deep reinforcement learning algorithm is designed to learn the control method of two-wheel vehicles. A motorcycle with varied sensors is also designed to capture the driving data of human drivers. The generated simulation data and captured experiment data are used to verify the proposed predictive model based on the LightGBM framework. Experiment results demonstrate that this approach exhibits superior predictive accuracy and generalization capabilities compared with XGBoost, Catboost, BP, and LSTM methods.