Chiller systems take up the major proportion of electricity used in commercial buildings. Their energy performance in terms of coefficient of performance (COP) depends on how the controllable and uncontrollable variables change. The aim of this study is to use the random forest (RF) method to measure variables importance and predict the COP. A sophisticated data trend log was carried out on an air-cooled chiller with advanced heat rejection features. The variables measured are: the flow rate of chilled water; the supply and return temperatures of chilled water; the temperature and relative humidity of outdoor air; the compressor power; the evaporating temperature; the condensing temperature; the number and speed of condenser fans staged; the temperatures of air entering and leaving the condenser. The data were logged at 5-min intervals in Aug 2015-Mar 2016. The RF models for different operating modes were validated, with a robust coefficient of determination of 80.52-96.53% for the testing data set. The chiller part load ratio, the condensing temperature, the chilled water flow rate, the heat rejection airflow rate and the wet-bulb temperature are the top five important variables in the prediction of COP. Yet they are not fully considered in typical regression models. Results of this study provide an insight into which variables are important to predict accurately the COP under different energy efficient features. The need of identifying the changing pattern of important variables is ascertained. (C) 2017 Elsevier B.V. All rights reserved.