The aim of this study was to develop different methods for screening genetic variants by integrating genomic data， phenotypic information， and functional annotations， and to compare and evaluate the prediction accuracy of 8 genomic selection （GS） methods for complex traits. Genotype-phenotype data from 542 Beijing ducks and 2 548 Duroc pigs were analyzed using three dimensions of variant selection： 1） genomic information， including linkage disequilibrium combined with principal component analysis （LD_PCA）， other principal component variants （LD_outPCA）， and Pearson correlation coefficient analysis （LD_PCC）； 2） phenotypic information， including LD combined with genome-wide association study （LD_GWAS） and EPISNP analysis （LD_EPI）； and 3） functional annotation information， using the weight distribution method （WDM）. In total， 8 different screening methods were applied to construct genotype matrices， and the prediction accuracies of different GS models were compared. The LD_PCC method， when combined with 6 GS models （GBLUP， BayesA， BayesB， BayesC， Bayes LASSO， and RRGBLUP）， achieved the highest mean prediction accuracies for duck fossil length， pig age to 100 kg， backfat thickness， and teat number， improving by approximately 4%-13.9% compared with the initial genotype matrix， with accuracies of 0.709 3， 0.440 0， 0.497 4， and 0.440 1， respectively. The WDM method achieved the highest accuracies for backfat thickness and teat number， with improvements of about 10% compared with unweighted models. In conclusion， this study demonstrates that WDM and LD_PCC can effectively enhance the prediction accuracy of GS， providing valuable insights for further exploration of variant selection methods and their application in breeding practice.