This paper proposes a new thermal error modeling methodology called Clustering Regression Thermal Error Modeling which not only improves the accuracy and robustness but also saves the time and cost of gear hobbing machine thermal error model. The major heat sources causing poor machining accuracy of gear hobbing machine are investigated. Clustering analysis method is applied to reduce the number of temperature sensors. Least squares regression modeling approach is used to build thermal error model for thermal error on-line prediction of gear hobbing machine. Model performance evaluation through thermal error compensation experiments shows that the new methodology has the advantage of higher accuracy and robustness.