With the growing reliance on machine-learning (ML) methods in finance, an understanding of their long-term efficacy and underlying mechanism is needed. We document the time-varying importance of different stock characteristics over an 18-year (1998-2016) out-of-sample period to determine whether ML models, when trained on a large set of firm and trading characteristics, can consistently outperform factor models. Utilizing a combination of linear and nonlinear models, we form a ML portfolio that consistently generates a significant alpha against factor models, ranging from 2.14 to 2.74% per month. We uncover patterns in characteristic dominance that alternates between arbitrage and financial constraint features. The variation correlates with the US credit cycle, and highlights a fundamental economic mechanism underlying the ML portfolio's performance. The study's impact extends to both academics and practitioners, providing insights into the economic drivers of stock returns and the practical implementation of ML methods in portfolio construction.