Stochastic optimization is widely used in portfolio and risk management for making investment or hedging decisions. Often times, the stochasticity in these problems can only be inferred from data, thereby leading to potential suboptimal decisions. In this talk, we study a data-driven assessment of the suboptimality of a given solution, via estimating confidence bounds of its optimality gap. Our approach uses a resampling scheme that connects a stochastic optimization program to classical symmetric statistics. We demonstrate how this leads to more efficient procedure (i.e., more accurate estimation with less data) than previous approaches. We also demonstrate how using subsampling can be beneficial in saving computation costs in our context, a motivation that distinguishes from the conventional use of the latter.