<p>Load spectrum is the objective basis for product reliability design and reliability testing. Aiming at the high sensitivity problem of threshold selection in the peak over threshold model when extrapolating the time domain load, a random forest model load time domain extrapolation method based on entropy weight-Topsis is proposed. Firstly, the weights of each POT model threshold evaluation index are objectively assigned by the entropy weight method. Secondly, a comprehensive evaluation is carried out by using the Topsis method with multiple indexes to preliminarily screen the optimal thresholds by calculating the relative proximity of the candidate thresholds to the positive–negative ideal solutions, and the relative proximity is nonlinearly calibrated using the random forest model to optimize the selection of thresholds. Thus, the load extrapolation method of establishing a generalized Pareto distribution (GPD) model based on the optimal threshold is proposed, and the technique is utilized to prepare the dynamic cutting force load spectrum of the servo tool turret for constant speed cutting conditions. Example analysis shows that the load data extrapolated by this method has a broader distribution range in the high amplitude region, the number of cycles increases significantly, and the range of extreme value action is extended. At the same time, the cumulative frequency curve and the pseudo-damage value indicate that the method has not changed the original load distribution characteristics, and the extrapolation process increases part of the large-amplitude signal and the pseudo-damage part increases, which is in line with the actual operation of the servo tool turret.</p>