Stress is one of the most pressing problems of modern society, which directly affects a person's mental and physical health. Prolonged or uncontrolled stress is one of the main factors of cardiovascular diseases, weakened immune system and decreased productivity. In this regard, timely detection and prediction of stress is becoming an important scientific and practical task. This article discusses the problem of predicting stress based on physiological indicators. These studies were recorded using a specially designed instrument that allows you to collect parameters of heart rate (HR), respiratory rate (RR) and skin conduction (SC). Based on the received signals, an integral Stress Index was calculated, and according to the interquartile percentiles, the stress level was divided into four categories: quiet, low, normal, high. Using this data, machine learning algorithms such as KNN, Decision Tree, Random Forest and Gradient Boosting were used. Experimental results showed that the Random Forest model, which achieved the highest result (hold-out accuracy 0.900, average accuracy of 5-layer cross-validation 0.805), proved to be a stable and reliable method for predicting stress. The results obtained make it possible to carry out real-time monitoring on wearable devices and create systems for early detection and prevention of the user's psychophysiological condition.