Demand Response (DR)-activated smart thermostats can be used to exploit the flexibility of residential heating and/or cooling systems. However, the acceptance/rejection of DR events depends on how occupants interact with their thermostats during the activated setpoint modulations. This interaction is mainly driven by their thermal comfort needs. Thus, understanding and modelling occupants’ comfort-driven interactions with thermostats is crucial for the design, assessment, and control of DR strategies. In this paper, we describe, calibrate, and show the in-use potentialities of a novel framework which is able to model occupants’ interactions with thermostats in residential buildings in winter. The framework includes a stochastic agent-based model of thermostat adjustments, whose dynamic thermal discomfort predictions are based on a two-node thermo-physiological model coupled with a dynamic thermal perception model. This represents a novelty with respect to the most often used static PMV/PPD model. Furthermore, the agent-based model is built on an activity and presence model and, therefore, is able to account for the diversity of the activities carried out by the occupants. User interaction data from about 9,000 connected Canadian thermostats included in the Donate Your Data (DYD) dataset are used to calibrate and establish the empirical foundation of the thermostat interaction model. Finally, we simulate typical DR-activated setpoint modulations in two residential buildings characterized by different levels of insulation and we use the framework to predict occupants’ override rates as a function of the indoor temperature and the time since the start of the DR event. The derived relationship can be directly used to inform the design and control of setpoint modulations in residential buildings.