ABSTRACT Mergers of galaxies are extremely violent events shaping their evolution. Such events are thought to trigger starbursts and, possibly, black hole accretion. Nonetheless, it is still not clear how to know the fate of a galaxy pair from the data available at a given time, limiting our ability to constrain the exact role of mergers. In this paper we use the light-cone of the horizon-agn simulation, for which we know the fate of each pair, to test three selection processes aiming at identifying true merging pairs. We find that the simplest one (selecting objects within two thresholds on projected distance d and redshift difference Δz) gives similar results than the most complex one (based on a neural network analysing d, Δz, redshift of the primary, masses/star formation rates/aspect ratio of both galaxies). Our best thresholds are $d_\mathrm{th}\sim 100\, \mathrm{kpc}$ and Δzth ∼ 10−3, in agreement with recent results.