Identifying at-risk students at an early stage is a challenging task for colleges and universities. In this paper, we use students' oncampus network traffic volume to construct several useful features in predicting their first semester GPA. In particular, we build proxies for their attendance, class engagement, and out-of-class study hours based on their network traffic volume. We then test how much these network-based features can increase the performance of a model with only conventional features (e.g., demographics, high school GPA, standardized test scores, etc.). We labeled students as "above median" and "below median" students based on their first term GPA. Several machine learning models were then applied, ranging from logistic regression, SVM, and random forests, to AdaBoost. The result shows that the model with network-based features consistently outperforms the ones without, in terms of accuracy, f1 score, and AUC. Given that network activity data is readily available data in most colleges and universities, this study provides practical insights on how to build more powerful models to predict student success. [For the full proceedings, see ED593090.]