The paper proposes a hybrid numerical-neural-network model developed based on the simulation of unheated and uncooled indoor temperature and humidity for buildings. This model is initiated with a numerical simulation and the output is then passed to a neural network for calibration. This approach utilizes both numerical and neural network models and it can analyze the influences of specific parameters on building performances without sacrificing accuracy and generalizability. An experimental study was conducted using a simulation of unheated and uncooled indoor temperature in a sports hall with a non-operating hour ventilation rate that is about half that of the operating hour rate. Several cases were examined. The indoor temperatures simulated by the hybrid model were more accurate than predictions by the numerical model alone for all cases. Particularly, results indicate that the hybrid model can generalize about a building parameter having only a constant value in training data, which a conventional neural network model cannot do that. More importantly, the hybrid model is adaptable for other building simulations, which is the main value of this model. (C) 2014 Elsevier B.V. All rights reserved.