We aimed to investigate the performance of three deregression methods (VanRaden, VR; Wiggans, WG; and Garrick, GR) of cows’ and bulls’ breeding values to be used as pseudophenotypes in the genomic evaluation of test‐day dairy production traits. Three scenarios were considered within each deregression method: (i) including only animals with reliability of estimated breeding value (RELEBV ) higher than the average of parent reliability (RELPA ) in the training and validation populations; (ii) including only animals with RELEBV higher than 0.50 in the training and RELEBV higher than RELPA in the validation population; and (iii) including only animals with RELEBV higher than 0.50 in both training and validation populations. Individual random regression coefficients of lactation curves were predicted using the genomic best linear unbiased prediction (GBLUP), considering either unweighted or weighted residual variances based on effective records contributions. In summary, VR and WG deregression methods seemed more appropriate for genomic prediction of test‐day traits without need for weighting in the genomic analysis, unless large differences in RELEBV between training population animals exist.