In view of the null results (so far) in the numerous channel-by-channel searches for new particles at the LHC, it becomes increasingly relevant to change perspective and attempt a more global approach to finding out where BSM physics may hide. To this end, we developed a novel statistical learning algorithm that is capable of identifying potential dispersed signals in the slew of published LHC analyses. The task of the algorithm is to build candidate "proto-models" from small excesses in the data, while at the same time remaining consistent with all other constraints. At present, this is based on the concept of simplified models, exploiting the SModelS software framework and its large database of simplified-model results from ATLAS and CMS searches for new physics.

In this talk, we explain the concept as well as technical details of the statistical learning procedure. A crucial aspect is the ability to construct a reliable likelihood in proto-model space; we discuss the various approximations which are needed depending on the information available from the experiments, and how they impact the whole procedure. Finally, we also discuss various aspects of the test statistic employed in our approach. With the current setup, the best-performing proto-model consists a top partner, a light-flavor quark partner, and a lightest neutral new particle with masses of about 1.2 TeV, 700 GeV and 160 GeV, respectively, and SUSY-like cross sections; for the SM hypothesis we find a global p-value of 0.19.