Real-time crash risk prediction is an emerging modeling approach to identify crash-prone traffic conditions, which can be further used for active traffic safety management for the expressway systems. Previous studies mostly focused on the development of various crash risk prediction models based on historical data. Although the predictive accuracies of these models are reasonably good, they are practically not transferable. Both the accuracy and model transferability issues could be potentially addressed by a model updating algorithm with real-time traffic data streaming. However, this has not been fully investigated in the literature. The basic idea of this paper is therefore to use the Stochastic Gradient Descent (SGD) based updating algorithm so as to update the commonly adopted logistic regression model for crash risk prediction. Two application scenarios were designed and tested, where the crash risk prediction model could be updated in real-time or at short time cycles (i.e., daily, weekly). Historical crash data and real-time traffic data were combined and split into training and test datasets: the training dataset was formulated in a matched case-control structure, which was used to develop the crash risk prediction model and identify a prediction threshold; the test dataset was contained the entire population data and used to evaluate the performance of the model updating algorithm through the use of the area under ROC (AUC). The results proved that model updating holds the advantage of better predictive performance and adaptability for real-time crash risk prediction. In addition, SGD was found to be adaptable in updating a crash prediction model for the two mutually exclusive and collectively exhaustive scenarios: (i) when a number of cases (i.e. crashes and non-crashes) arrive together at an instant or (ii) only one case is available at a time.