In regression analysis, there are two main aims: interpretation and prediction, which can be also applied in building performance analysis. Interpretation is used to understand the relationship between input parameters and building energy performance (also called sensitivity analysis), whereas prediction is used to create a reliable energy model to estimate building energy consumption. This article explores the implementation of a distribution-free bootstrap method for these two purposes. The bootstrap is a resampling method that enables assessment of the accuracy of an estimator by random sampling with replacement from an original dataset. An office building is used as a case study to demonstrate the application of this method in assessing building thermal performance. The results indicate that the probabilistic sensitivity analysis incorporating the bootstrap approach provides valuable insights into the variations in sensitivity indicators, which are not available from typical deterministic sensitivity analysis. The single point values from deterministic methods may lead to misleading prioritization of energy saving measures because they do not provide the distributions of sensitivity indicators. Information on prediction errors obtained from the bootstrap method can facilitate the selection of an appropriate building energy metamodel to more accurately predict the energy consumption of buildings, compared with the traditional one-time data splitting method (also called holdout cross-validation method), which partitions the data into a training set and a test set.