Businesses, researchers, and policymakers in the agricultural and food sector regularly make use of large public, private, and administrative datasets for prediction, including forecasting, public policy targeting, and management research. Machine learning has the potential to substantially improve prediction with these datasets. In this study we demonstrate and evaluate several machine learning models for predicting demand for new credit with the 2014 Agricultural Resource Management Survey. Many, but not all, of the machine learning models used are shown to have stronger predictive power than standard econometric approaches. We provide a cost based model evaluation approach for managers to analyze returns to machine learning methods relative to standard econometric approaches. While there are potentially significant returns to machine learning methods, research objectives and firm-level costs are important considerations that in some cases may favor standard econometric approaches.