Applying machine learning to cross-sectional stock return prediction requires careful consideration of modeling choices. Common approaches that fail to account for heterogeneity or imbalanced stock representation in training data can lead to suboptimal performance. I study two strategies to address these issues: training group-specific models and predicting relative returns. Both approaches yield similar economic improvements over models trained on the full cross-section of US stock returns, with value-weighted trading strategies benefiting significantly. The findings underscore the importance of aligning machine learning modeling decisions with desired economic outcomes and provide guidance for researchers and practitioners seeking robust machine learning models.