The nested simulation approach is one of the most widely used methods for the valuation and risk management of variable annuity (VA) portfolios. Due to the complexity of the simulation algorithms, large size and inhomogeneity of a VA portfolio, running the nested simulation for the entire portfolio is often computationally expensive and sometimes prohibitive. In this paper, we propose a hybrid regression-based nested simulation algorithm that can be used to effectively calculate/approximate the quantities of interest (e.g. predictive distribution, required capitals, etc.) for a large VA portfolio. An additive model based smoothing approach is proposed to reduce the number of inner-loops and the number of outer-loops for simulation. Further, we adopt a finite population estimation framework to reduce the number of policies in use for calculation. Numerical studies show our proposed method performs well in accurately approximating the distribution of the total future liability.