Multiple hypotheses are available to explain recruitment rate. Model selection methods can be used to identify the best model that supports a particular hypothesis. However, using a single model for estimating recruitment success is often inadequate for overexploited population because of high model uncertainty. In this study, stock-recruitment data of small yellow croaker in the East China Sea collected from fishery dependent and independent surveys between 1992 and 2012 were used to examine density-dependent effects on recruitment success. Model selection methods based on frequentist (AIC, maximum adjusted R2 and P-values) and Bayesian (Bayesian model averaging, BMA) methods were applied to identify the relationship between recruitment and environment conditions. Interannual variability of the East China Sea environment was indicated by sea surface temperature ( SST) , meridional wind stress (MWS), zonal wind stress (ZWS), sea surface pressure (SPP) and runoff of Changjiang River ( RCR). Mean absolute error, mean squared predictive error and continuous ranked probability score were calculated to evaluate the predictive performance of recruitment success. The results showed that models structures were not consistent based on three kinds of model selection methods, predictive variables of models were spawning abundance and MWS by AIC, spawning abundance by P-values, spawning abundance, MWS and RCR by maximum adjusted R2. The recruitment success decreased linearly with stock abundance (P < 0.01), suggesting overcompensation effect in the recruitment success might be due to cannibalism or food competition. Meridional wind intensity showed marginally significant and positive effects on the recruitment success (P = 0.06), while runoff of Changjiang River showed a marginally negative effect (P = 0.07). Based on mean absolute error and continuous ranked probability score, predictive error associated with models obtained from BMA was the smallest amongst different approaches, while that from models selected based on the P-value of the independent variables was the highest. However, mean squared predictive error from models selected based on the maximum adjusted R2 was highest. We found that BMA method could improve the prediction of recruitment success, derive more accurate prediction interval and quantitatively evaluate model uncertainty.