Model Selection for Discrete Dependent Variables: Better Statistics for Better Steaks

Abstract

Little research has been conducted on evaluating out-of-sample forecasts of discrete dependent variables. This study describes the large and small sample properties of two forecast evaluation techniques for discrete dependent variables: receiver-operator curves and out-of-sample log-likelihood functions. The methods are shown to provide identical model rankings in large samples and similar rankings in small samples. The likelihood function method is better at detecting forecast accuracy in small samples. By improving forecasts of fed cattle quality grades, the forecast evaluation methods are shown to increase cattle marketing revenues by $2.59/head.

Keywords

• discrete dependent variables
• forecasting
• likelihood functions
• model selection
• out-of-sample
• quality grades
• receiver-operator curves