The ability to monitor stress levels in daily life can provide valuable information to patients and their caretakers, help identify potential stressors, determine appropriate interventions, and monitor their effectiveness. Wearable sensor technology makes it now possible to measure non-invasively a number of physiological correlates of stress, from skin conductance to heart rate variability. These measures, however, show large individual differences and are also correlated with the physical activity of the subject. In this paper, we propose two multivariate signal processing techniques to reduce the effect of both forms of interference. The first method is an unsupervised technique that removes any systematic variation that is orthogonal to the dependent variable, in this case physiological stress. In contrast, the second method is a supervised technique that first projects the data into a subspace that emphasizes these systematic variations, and then removes them from the data. The two methods were validated on an experimental dataset containing physiological recordings from multiple subjects performing physical and/or mental activities. When compared to z-score normalization, the standard method for removing individual differences, our methods can reduce stress prediction errors by as much as 50%.