Thermal modelling tools have widely been used in the construction industry at the design stage, either for new build or retrofitting existing buildings, providing data for informed decision-making. The accuracy of thermal models has been subject of much research in recent decades due to the potential large difference between predicted and 'in-use' performance - the so called 'performance gap'. A number of studies suggested that better representation of building physics and operation details in thermal models can improve the accuracy of predictions. However, full-scale model calibration has always been challenging as it is difficult to measure all the necessary boundary conditions in an open environment. Thus, the Energy House facility at the University of Salford - a full-sized end terrace house constructed within an environmental chamber - presents a unique opportunity to conduct full-scale model calibration. The aim of this research is to calibrate Energy House thermal models using various full-scale measurements. The measurements used in this research include the co-heating tests for a whole house retrofit case study, and thermal resistance from window coverings and heating controls with thermostatic radiator valves (TRVs). Thermal models were created using an IESVE (Integrated Environment Solutions Virtual Environment). IESVE is a well-established dynamic thermal simulation tool widely used in analysing the dynamic response of a building based on the hourly input of weather data. The evidence from this study suggests that thermal models using measured U-values and infiltration rates do perform better than the models using calculated thermal properties and assumed infiltration rates. The research suggests that better representations of building physics help thermal models reduce the performance gap. However, discrepancies still exist due to various other underlying uncertainties which need to be considered individually with each case. In relative terms, i.e., variations in percentage, the predictions from thermal models tend to be more reliable than predicting the absolute numbers. (C) 2019 Elsevier B.V. All rights reserved.