A motion analysis system analyzing a prescribed motion performed by a person using a prescribed tool, with a prescribed portion of the person or portion of the tool in the motion being taken as a measurement portion, a predictive model generated by learning based on learning data including an image frame of a motion video, wherein the person performing the motion, and an inter-frame differential frame indicating a difference in pixel values between frames of each pixel of the image and an adjacent frames, which is adjacent to the image frame, in the motion video; generates an inter-frame differential frame of a given analysis object video; and predicts by using the predictive model a position of the measurement portion in an image frame of the analysis object video on the basis of the image frame of the analysis object video and the inter-frame differential frame of the analysis object video.