In light of the negative effects of mental stress on health, there is growing scientific attention placed on understanding it better. This study aims to introduce a method that uses biological indicators such as skin surface temperature and skin resistance using easily accessible data to detect and predict stress levels. The publicly available dataset used in this study included the physiological signals of a group of university students whose signals were recorded by a wearable device during three exam sessions (two midterm and one final exam). Features were extracted from heart rate, electrodermal activity, and physiological and skin surface temperature and were input for the classification. In the labeling process, the stress levels were classified as high, low, or moderate. Machine learning algorithms such as Support Vector Machines (SVM) and K-Nearest Neighbors (KNN) with classifiers were trained and tested with the labeled data The stress level of a subject was predicted with the highest F1 score of up to 0.8 and accuracy of higher than 70%. Such findings highlight the important role of skin conductance and surface temperature in estimating stress levels.