(eng): This paper addresses the moment selection issue of the simulated method of moments, an estimation technique commonly applied to intractable agent-based models. We develop a simple machine learning extension reducing arbitrariness and automating the moment choice. Two algorithms are proposed: backward stepwise moment elimination and forward stepwise moment selection. The methodology is tested using simulations on a Markov-switching multifractal framework and two popular financial agent-based models with increasing complexity. We find that both algorithms can identify multiple moment sets that outperform all benchmark sets. Moreover, we achieve considerable in-sample estimation precision gains of up to 66 percent for agent-based models. Finally, an out-of-sample empirical exercise with S&P 500 data strongly supports the practical applicability of our methodology as the estimated models pass the validity test of overidentifying restrictions.