Stress can cause mental disorders such as depression and anxiety disorders. To detect such mental disorders at an early stage, it is necessary to detect stress accurately. One of the effective methods for this purpose is observing changes in biological signals caused by sensory stimuli such as video presentation. This study aims to identify effective video stimuli for stress estimation. We hypothesize that the emotional state evoked by the video stimuli influences the accuracy of stress estimation. To test this hypothesis, we utilized an open video dataset consisting of 444 responses on an emotion scale (valence and arousal) as emotional stimuli. Ninety videos were divided into emotion subsets based on the emotion scale for each video, and biological signals were measured when each video was presented to the subjects. Machine learning models were constructed for each subset, and the prediction errors were compared. The results showed that the prediction error was lower for the high valence and high arousal subsets than for the others. These results suggest that high-valence or high-arousal videos effectively estimate stress.