Several data mining techniques are currently applied in many areas with great success. Various educational systems collect and use huge amount of data on students, staff and faculty. Researchers have the potential to employ such datasets to examine students' performance over their learning time-from one semester to another or from one academic year to another. Student's key demographic characteristics, number of examination attempts per course and final grade of each course can provide the training data and infer a regression function that estimates the performance of upcoming courses. We have carried out several experiments using eight courses modeled for regression tasks and familiar data mining models (i.e. Linear Regression, Support Vector Machines, Decision Trees, M5 Rules, and k-Nearest Neighbors). The results reported that satisfactory accuracy is achieved, provided that the first semester grades are available, indicating that an early identification of students at risk of underperforming in a course can be obtained.