We combine the galform semi-analytical model of galaxy formation, which predicts the star formation and merger histories of galaxies, the grasil spectrophotometric code, which calculates the spectral energy distributions of galaxies self-consistently, including reprocessing of radiation by dust, and artificial neural networks to investigate the clustering properties of galaxies selected by their emission at submillimetre (submm) wavelengths [submm galaxies (SMGs)]. We use the Millennium Simulation to predict the spatial and angular distribution of SMGs. At redshift z= 2, we find that these galaxies are strongly clustered, with a comoving correlation length of r0= 5.6 ± 0.9 h−1 Mpc for galaxies with 850-μ m flux densities brighter than 5 mJy, in agreement with observations. We predict that at higher redshifts, these galaxies trace denser and increasingly rarer regions of the universe. We present the predicted dependence of the clustering on luminosity, submm colour and halo and stellar masses. Interestingly, we predict tight relations between the correlation length and the halo and stellar masses, independent of submm luminosity.