Thermal loads are an important source of flexibility at a residential customer level. The uncertain economic value of residential demand response (DR), and the rising customer data privacy concerns, require non-intrusive and economical approaches to harness flexibility. Baselines are essential for evaluating DR activations, however, the frequent use of flexibility makes them less accurate. In this paper, we first propose a baseline estimation method based solely on aggregated behind the meter data, which does not require additional knowledge of the portfolio's parameters. It is suited for frequent DR activations, and relies on a combination of linear interpolation, forward-backward autoregression and load decomposition. The method is then used to evaluate DR activations, in order to construct, and continuously update, a model for the response and the rebound behavior of the loads. A portfolio of 138 real residential customers equipped with electric heaters, and a large number of DR experiments, were used to verify the proposed approach. The response model, fitted with the experimental results, shows a strong dependency of the load reduction potential on time of day and ambient temperature, with a maximum load reduction equal to 1.2 kW per household. Validation results confirm that the fitted model can be used to estimate the response with a good accuracy. Finally, a model to describe and shape the rebound behavior of the loads is proposed and validated with real experiments.