Stress is known to have a range of negative impacts – for example, occupational stress faced by uniformed personnel may harm their performance and mental health. The detection of stress from electroencephalography (EEG) data is of interest in stress management. In this project, we investigated different factors that affect the performance of models in predicting stress from EEG data. We analyzed a publicly available online repository of EEG recordings of 36 individuals in stressed and unstressed states, accessed through PhysioNet, extracting two features: Pearson’s correlation coefficient and mutual entropy between channels. Nine classifiers from scikit.learn were trained and evaluated under different validation conditions (k-fold without pooling, k-fold after pooling, and leave-one-out). A trend was seen in performance, which decreased in the order of validation conditions listed above. Different classifiers emerged as the most accurate under different combinations of feature and validation methods.