In the last decade the interest in adaptive models for non-stationary environments has gained momentum within the research community due to an increasing number of application scenarios generating non-stationary data streams. In this context the literature has been specially rich in terms of ensemble techniques, which in their majority have focused on taking advantage of past information in the form of already trained predictive models and other alternatives alike. This manuscript elaborates on a rather different approach, which hinges on extracting the essential predictive information of past trained models and determining therefrom the best candidates (intelligent sample matchmaking) for training the predictive model of the current data batch. This novel perspective is of inherent utility for data streams characterized by short-length unbalanced data batches, situation where the so-called trade-off between plasticity and stability must be carefully met. The approach is evaluated on a synthetic data set that simulates a non-stationary environment with recurrently changing concept drift. The proposed approach is shown to perform competitively when adapting to a sudden and recurrent change with respect to the state of the art, but without storing all the past trained models and by lessening its computational complexity in terms of model evaluations. These promising results motivate future research aimed at validating the proposed strategy on other scenarios under concept drift, such as those characterized by semi-supervised data streams.