Occupancy models are essential parts of the control operations of smart buildings. Developing an appropriate model to simulate and predict the occupancy presence for common buildings will allow a better optimization of energy consumption and utility bill. However, most occupancy models tested by the on-site data only concentrate on the commercial buildings. On the other hand, the residential studies simulate mostly Time Use data. The applicability of those models are largely unknown for the real-time residential environment. This study focuses on providing a unique data set of four residential houses collected from PIR sensors in the U.S. One new inhomogeneous Markov model for predictive building controls is proposed and compared to the homogeneous model. Optimized training periods for the occupancy presence prediction are decided individually from change-point analysis of historical data. Then, various scenarios that utilize 15-minute data to forecast 15-min ahead, 1-hour ahead, and 24-hour ahead occupancy presences are presented. Both room-level and house-level predictions are evaluated. The one-to-one matching between the prediction and the ground truth demonstrates the model performance. The analysis of the residential result validates the effectiveness of the proposed method and provides more realistic analysis on the occupancy forecasting of the residential environment.