We present a methodology to evaluate the accuracy of baseline energy predictions. To evaluate the predictions from a computer program, the program is provided with electric load data, and additional data such as outdoor air temperature, from a "training period" of at least several months duration, and used to predict the energy use as a function of time during the subsequent "prediction period." The predicted energy use is compared to the actual energy use, and errors are summarized with several metrics, including bias and mean absolute percent error (MAPE). An important feature of this methodology is that it can be used to assess the predictive accuracy of a model even if the model itself is not provided to the evaluator, so that proprietary tools can be evaluated while protecting the developer's intellectual property. The methodology was applied to evaluate several standard statistical models using data from four hundred randomly selected commercial buildings in a large utility territory in Northern California; the result is a statistical distribution of errors for each of the models. We also demonstrate how the methodology can be used to assess the uncertainty in baseline energy predictions for a portfolio of buildings, which is an issue that is important for the design of utility programs that incentivize energy savings. The findings of this work can be used to (1) inform technology assessments for technologies that deliver operational and/or behavioral savings; and (2) determine the expected accuracy of statistical models used for automated measurement and verification (M&V) of energy savings. (C) 2015 Elsevier Ltd. All rights reserved.