Machine learning algorithms have shown to be remarkably successful tools for predicting asset returns. However, the underlying economic mechanisms behind their performance remain unclear. This paper proposes a model-based dynamic arbitrage trading strategy that combines economic and statistical nonstationarity to demystify this black box. In predicting stock returns based on 153 firm characteristics (anomalies), our strategy ranks anomalies similarly to neural networks in the cross-section. Overall, it accounts for approximately 87.9 bps monthly alphas of the high-minus-low portfolios selected by neural networks in the time series. When unpublished anomalies and microcap stocks are excluded from trading, this strategy can fully explain the performance of neural networks. Our results reveal three economic sources of neural-network performance: a time varying strategy analogous to dynamic arbitrage, a tendency to weight portfolios on unpublished anomalies, and exposure to microcaps.