Machine Learning Meets Panel Data: What Practitioners Need to Know

Applying machine learning models to panel data risks a significant issue known as data leakage, where future or unseen information contaminates the training process. This contamination leads to an overestimation of the model's predictive performance, creating an illusion of high accuracy. The two primary forms of leakage are temporal, where future data influences past predictions, and cross-sectional, where the same units appear in both training and test sets. To avoid this, practitioners must choose a data splitting strategy—either by unit or by time—that aligns with their specific prediction goal, such as cross-sectional imputation or sequential forecasting. An empirical analysis of U.S. county data confirms that incorrect methods yield inflated results, highlighting the real-world consequences of misallocated resources based on flawed models.