Physical activity is important for people’s health. The physical activity intervention program reported here includes daily wearing of an activity monitor to provide people with insight into their activity behavior. The activity monitor consists of a triaxial accelerometer, where measured accelerations are transformed to a physical activity level (PAL). The PAL data quantifies the level of the daily physical activity and reflects the daily energy expenditure of the wearer. In the program, coaches provide e-mail based intervention to motivate participants to increase their activity step-by-step within 12 weeks. However, a significant portion of participants (~41%) failed to complete the program. This paper examines methods to predict participants who are at risk of dropping out of the program based on a classification task. This allows for a timely delivery of tailored interventions and motivating triggers to prevent stopping of the program. In particular, this paper proposes to combine the features extracted from participants’ personal information, their behaviors during the use of the device, the observed PAL data and the features extracted from the process of predicting future PAL data to classify dropouts and non-dropouts every week. Experiment results show that a k-nearest-neighbor classifier achieved a dropout and a non-dropout prediction accuracy of 66.4 ± 13.8% and 74.1 ± 7.3%, respectively.