Confronting theoretical models with the observed evolution of the galaxy population out to z=4

[abridged] We construct lightcones for the semi-analytic galaxy formation simulation of Guo et al. (2011) and make mock catalogues for comparison with deep high-redshift surveys. Photometric properties are calculated with two different stellar population synthesis codes (Bruzual & Charlot 2003; Maraston 2005) in order to study sensitivity to this aspect of the modelling. The catalogues are publicly available and include photometry for a large number of observed bands from 4000{deg}A to 6{mu}m, as well as rest-frame photometry and intrinsic properties of the galaxies. Guo et al. (2011) tuned their model to fit the low-redshift galaxy population but noted that at z > 1 it overpredicts the abundance of galaxies below the knee of the stellar mass function. Here we extend the comparison to deep galaxy counts in the B, i, J, K and IRAC 3.6{mu}m, 4.5{mu}m and 5.8{mu}m bands, to the redshift distributions of K and 5.8{mu}m selected galaxies, and to the evolution of rest-frame luminosity functions in the B and K bands. The B, i and J counts are well reproduced, but at longer wavelengths the overabundant high-redshift galaxies produce excess faint counts. The predicted redshift distributions for K and 5.8{mu}m selected samples highlight the effect of emission from thermally pulsing AGB stars. The full treatment of Maraston (2005) predicts three times as many z~2 galaxies in faint 5.8{mu}m selected samples as the model of Bruzual & Charlot (2003), whereas the two models give similar predictions for K-band selected samples. Although luminosity functions are adequately reproduced out to z~3 in rest-frame B, the same is true at rest-frame K only if TP-AGB emission is included, and then only at high luminosity. Fainter than L* the two synthesis models agree but overpredict the number of galaxies, another reflection of the overabundance of ~10^10Modot model galaxies at z > 1.