ABSTRACT This article investigates optimal model averaging for regression kink models with heteroscedasticity. When all candidate models are misspecified, we establish the asymptotic optimality of the corresponding model averaging estimator in the sense of minimizing the squared prediction loss. When correct models exist, we demonstrate that the sum of weights assigned to the correct models converges to one in probability as the sample size increases. Simulation experiments reveal the superiority of our method over other commonly used model averaging methods. Furthermore, we provide an empirical illustration using a dataset of baseball pitcher salaries.