This paper proposes a nonparametric simulation framework to estimate poverty vulnerability under climate shocks. We formalize vulnerability estimation as an out-of-sample prediction problem and show that flexible, regularized machine learning methods for estimating the conditional mean of welfare offer a powerful alternative to conventional linear models. The framework simulates future welfare distributions using historical realizations of climate shocks and household characteristics, enabling the estimation of vulnerability measures and related functions without imposing restrictive parametric assumptions. To interpret the model and quantify heterogeneous impacts, we employ SHapley Additive exPlanations, which decompose predicted vulnerability into contributions from climate shocks and household characteristics. An application to Ecuador reveals a strong geographic concentration of vulnerability and shows that climate shocks act as localized triggers that push marginal households, particularly low-educated informal rural workers into poverty.