Nowadays,　with　many　available　numerical　methods　developed　in　rock　mechanics,　researchers　have　always　focused　on　the　parameter　calibration　of　the　numerical　model　but　ignored　the　predictive　capability　of　these　methods.　A　comparative　study　of　nine　commonly　used　numerical　methods　was　performed　for　predicting　rock　failure　through　international　cooperation　organized　by　the　Discontinuous　Deformation　Analysis　(DDA)　commission　of　the　International　Society　for　Rock　Mechanics　(ISRM).　Two　steps　of　numerical　modelling　were　conducted　including　a　calibration　procedure　from　given　experimental　results　and　a　numerical　prediction　for　benchmark　tests　with　these　calibrated　parameters　for　three　types　of　rocks.　Through　the　comparison　between　different　numerical　and　experimental　results,　the　inherent　weaknesses　and　strengths　of　different　numerical　methods　in　terms　of　predicting　rock　failure　were　identified　and　analysed.　The　influence　of　human　intervention　in　terms　of　parameter　selection　is　even　more　significant　than　the　choice　of　different　numerical　methods.　Some　potential　factors　(i.e.,　different　boundary　conditions,　heterogeneity　of　rock　material,　strength　parameters,　particle　packing,　and　failure　criteria)　that　may　occur　in　numerical　and　physical　tests　were　further　discussed.　Through　the　comparison　of　different　failure　criteria,　we　found　the　selection　of　rock　failure　criterion　might　be　the　major　factor　that　affected　the　predictive　capability　of　numerical　methods,　and　the　nonlinear　failure　model　(the　Hoek-Brown　criterion)　showed　the　superiority　in　the　prediction　of　rock　fracturing　subjected　to　complex　stress　conditions.　This　comparative　work　also　enlightens　the　significance　of　a　high-quality　calibration　process　and　the　future　advancement　of　rock　failure　criteria.