In this study, we introduce a non-linear estimating approach for risk factor loading. This new estimate is based on mixed vine copula with the aim of separating upside and downside risk exposure. We provide empirical evidence of Chinese stocks that copula-based method fits better than OLS for single-factor model, then we present that adjusted estimate adapted for time-serial weights performs better when fitting factor loadings. For multi-factor model, copula-based method is also superior in explaining the changes of asset return and interpreting the influence of extreme changes from risk factors. By exploring the usage of copulas in factor model regression, we enhance the accuracy for predicting asset returns as well as extending the application of factor model during extreme events.