•Discussed cross-building energy prediction in a limited-data context.•Compared different methods in a same case via a competition.•Identified hybrid strategies for hybrid building energy prediction models.•Discussed the shortcomings and suggestions for the data preparation process.•Highlighted the importance of data selection for accurate cross-building prediction. With the evolution of new energy and carbon trading systems, it is important to accurately predict building energy consumption to help energy arrangements. Additionally, the widespread use of smart meters has introduced a new data context for building energy prediction. Building energy prediction techniques need improvement but the ideas of various new prediction methods are still on the way and have not yet been compared and tested side-by-side in the reported studies. Thus, we held a competition called ‘Energy Detective’. To investigate the status quo of the current prediction techniques, we designed a representative prediction case: cross-building prediction with limited physical parameters and historical data. A total of 195 participants formed 89 teams to participate in the competition. This paper describes the models presented in the competition. By analysing the methods and results, we identified strategies for the future development of energy prediction in hybrid modelling and data-driven modelling. For hybrid modelling, we discuss the basic strategies for hybrid models and suggest that more hybrid models can be developed by combining a wide variety of individual models in sequence or parallel or via feedback methods to achieve accurate and interpretable models. For data-driven modelling, we analyse and discuss the areas of improvement for the current data-driven workflow and suggest that processes other than model application are also important and should be carefully considered. Considering the increasing amount of data available for prediction, we discuss the shortcomings and suggestions for improving the current data preparation process. We recommend comprehensive consideration of the anomaly types in data pre-processing and a focus on feature engineering for higher accuracy and model interpretability, while emphasising the vital role of data selection in cross-building energy prediction.