In the present study the performance of several machine learning regression techniques for the analysis of displacement data acquired in dynamic identification shaking table tests to assess a building’s prototype state of damage were compared. The displacement data of a typical Italian historic masonry building prototype were recorded by a passive 3D motion capture system. These recorded displacement data were analyzed by calculating a widely accepted Damage Index (DI) based on the decay of the first modal frequency of the tested structure. To such purpose, a conventional Frequency Response Function (FRF) algorithm for modal analysis was adopted. Then, several regressors, commonly used in machine learning applications, were applied to cepstral coefficients extracted from time-series segmentation of the 3D motion markers signals to predict the DI values. To assess the quality of the tested regressors, Coefficient of determination (R2) and Relative Percent Deviation (RPD) parameters were evaluated. Different lengths of time-series segments were also explored and the results were statistically compared to highlight the differences between regressors. Excellent results were obtained in terms of R2 (0.98) and RPD (6.87) with a weak dependence on the time partitioning performed on the data with optimal results for 4 or more seconds. Moreover, Stacked Regressor, Decision Tree Bagging Regressor and XGB Regressor were found as the best overall regressors. The results obtained demonstrated that machine learning methods can be effectively utilized to predict the level of damage of historic masonry buildings with very high accuracy.