Stress has emerged as a critical global health issue, contributing to cardiovascular disorders, depression, and several other long-term illnesses. Consequently, accurate and reliable stress monitoring systems are of growing importance. In this work, we propose a stress prediction framework based on electrocardiogram (ECG) signals recorded during multiple daily activities such as sitting, walking, and jogging. Frequency-domain indicators of autonomic nervous system activity were obtained through Power Spectral Density (PSD) analysis and utilized as input for machine learning models including Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost. In addition, deep learning approaches, namely Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, were directly applied to the raw ECG signals. Our experiments highlight the effectiveness of ensemble-based classifiers, with CatBoost achieving 90% accuracy. Moreover, the LSTM model provided superior results, attaining 94% accuracy with balanced precision, recall, and F1-score, reflecting its strength in modeling temporal dependencies in ECG data. Overall, the findings suggest that integrating frequency-domain feature extraction with advanced learning algorithms enhances stress prediction and paves the way for real-time healthcare monitoring solutions.