Using fertility trait data and pedigree information from Landrace, Yorkshire, and their crossbred Landrace× Yorkshire sows, three heterosis prediction models MA, MAD and MADE were constructed by stepwise adding additive effects(A), dominance effects(D) and epistatic effects(E) to a single-trait repeatability model(M). The accuracy of these models in predicting hybrid phenotypes and heterosis in crossbred offspring was compared. The results showed that the MA model, which included only additive effects, had higher prediction accuracy for phenotypes(0.290-0.720) compared to the MAD model(0.173-0.687) and the MADE model(0.055-0.365), both of which also incorporated non-additive effects. The MA model outperformed the other two models in predicting heterosis rates for 7 reproductive traits including litter weight at birth(25.41%), total number born(–25.26%), number of liveborn piglets(–26.11%), number of healthy piglets (–24.89%), number of weak piglets(–29.30%), number of stillborns(–37.25%) and number of mummified piglets (–44.88%). The MAD model was more reliable and of greater reference value in predicting heterosis rates for 4 reproductive traits including number of deformities(41.85%), number of piglets weaned(18.90%), weaning litter weight(18.29%) and 21-day litter weight(14.56%).