A promising solution to reduce the testing costs of analog/RF circuits is the alternate test strategy, which permits to replace costly specification measurements by simple low-cost indirect measurements. This approach has been widely explored and demonstrated in the literature on various case studies over the past twenty years. However it is clear that the efficiency of this strategy strongly depends on the quality of the regression models used to map the indirect measurements to the device specifications. In this paper, we present a generic methodology for building efficient prediction models from a large set of indirect measurements candidates. Results are illustrated on a case study for which we have experimental test data.