The traffic safety on expressways is crucial for the efficient operation of the expressway system, and there is a close relationship between traffic states and crashes on expressways, and the occurrence of crashes may be influenced by the interaction of different combinations of traffic states upstream and downstream of the crash location. Based on the crash data and the corresponding traffic flow detector data collected on expressways in Shanghai, this study proposes a hybrid model combining a support vector machine (SVM) model with a k-means clustering algorithm to predict the likelihood of crashes. The random forest (RF) model is employed to select the important and significant variables for model construction from the data of the traffic flow 5–10 min before the crash occurred. Then, the cross-validation and transferability of different models (SVM model without variable selection, SVM model with variable selection, and hybrid SVM model with variable selection) are determined using 577 crashes and 5794 matched non-crash events. The results show that the crash prediction model along with the four most important variables selected using the RF model can obtain a satisfactory prediction performance for crashes. With the combination of the clustering algorithm and SVM model, the accuracy of the crash prediction model can be as high as 78.0%. Moreover, the results of the transferability of the three different models imply that the variable selection and clustering algorithm both have an advantage for crash prediction.