Previous research finds that machine learning methods predict short-term return variation in the cross-section of stocks, even when these methods do not impose strict economic restrictions. However, without such restrictions, the predictions from the models fail to generalize in a number of important ways, such as predicting time-series variation in market and long-short characteristic sorted portfolio returns across multiple horizons. I show this shortfall can be remedied by imposing economic restrictions in the architectural design of a neural network model and provide recommendations for using machine learning methods in asset pricing. Additionally, I shed light on the intermediate and long-run dynamics of the return forecasts generated by this model.