Even though simulation is mainly used for computer models with inexact outputs, there are direct benefits in viewing results from samples of an existing dataset as replications of a stochastic simulation. We propose building Machine Learning prediction models with the Monte Carlo approach. This allows more specific accountability for the underlying distribution of the data and the impact of uncertainty in the input data in terms of bias. We opt for nonparametric input uncertainty with multi-level bootstrapping to make the framework applicable to large datasets. The cost of Monte Carlo-based model construction is controllable with optimal designs of nested bootstrapping and integrating variance reduction strategies. The benefit is substantial in providing more robustness in the predictions. Implementation in a data-driven simulation optimization problem further indicates the superiority of the proposed method compared to the state-of-the-art methods.