Supervisory predictive control of heat pumps in residential buildings can improve comfort, reduce operating costs, and reduce emissions. However, all the experimental demonstrations so far in residential buildings have considered a “sensible” formulation, where the effect of dehumidification is assumed to be constant during the day. In this work, we demonstrate the use of a model-free machine learning method to predict the indoor wet-bulb temperature, and subsequently the sensible heat ratio. This paper presents the formulations for these “latent” and “sensible” models. The two modes of operation are tested in on-off operation for 4 weeks in a detached single-family home in West Lafayette, Indiana. This paper finds that when the real electrical power is predicted with significant uncertainty, the two models perform very similarly. The energy savings achieved by both models were comparable and in the range of 7 – 21% (95% confidence interval), with a mean of 14%. However, the “latent” formulation better predicted the electrical power, which is expected to be significant for formulations that concern the magnitude (constraints to power, penalizing peak power) rather than the direction, i.e., minimizing the total energy.