Traditional thermal error model often shows a poor robustness and prediction accuracy when there are differences between actual condition and modeling condition. The main reasons are the limitation of modeling data and the unmodeled dynamics of the model. In order to improve the above phenomenon, a thermal error compensation modeling method based on data-driven is proposed for the computer numerical control (CNC) machine spindle. The method was utilized of data (temperature and error) collected in process of machining and a model-free adaptive control algorithm to build up model and to modify in real time, which could rapidly adapt to various operating conditions and then to improve its robustness. An experiment was develop to verify its effect on a CNC lathe. The results show that, compared to Multiple Linear Regression (MLR) model, the method improves the standard deviation, maximum residual and error squares by 41, 62 and 56 percent, respectively, with good effect in robustness and forecast. In addition, the method lay a foundation for big data to apply to thermal error compensation of the CNC machine spindle.