With the growing trove of publicly available building energy data, there are now ample opportunities to apply machine learning methods for prediction of building energy performance. In this study, we test different predictive modeling approaches for estimating Energy Use Intensity (EUI) for US commercial office buildings and the individual energy end-uses of HVAC, plug loads, and lighting, based on the latest Commercial Building Energy Consumption Survey (CBECS) 2012 microdata. After preliminary statistical analysis, six regression or machine learning techniques are applied and compared for prediction performance. Among all candidates, Support Vector Machine and Random Forest demonstrate both accuracy and stability. However, machine learning algorithms are better than the linear regression only to a limited extent, with on average 10-15% lower prediction errors for Total EUI prediction. Conversely, linear regression models slightly outperform machine learning methods in estimating Plug Loads EUI. These mixed results suggest careful consideration in applying advanced predictive algorithms to the CBECS dataset. Individual variable importance was tested using Random Forest, with the top 10 predictors differing for the total and sub-system Wis. The analysis demonstrates that, for the techniques applied, the variables reported in CBECS have inadequate predictive power to map actual energy consumption. Filling information gaps in areas such as occupant behavior, power management, building thermal performance, and their interactions may help to improve predictive modeling. (C) 2017 Elsevier B.V. All rights reserved.