Classification of Binary Data with one Low-Frequency Class using Logistic Discrimination an.pdf

Classification of Binary Data with one Low-Frequency Class using Logistic Discrimination and Feedforward Networks Daniel Enache Visiting Scholar Information Systems Department, MEC 9-181 New York University, STERN School of Business 44 West 4 th Street, New York, NY 10012 denache@stern.nyu.edu Working Paper RC-0001 ? Copyright 1998 by Daniel Enache Daniel Enache (1998), Classification of Binary Data with one Low-Frequency..., Working Paper RC-0001. 1 1 Introduction and Notation The sample used in the credit risk model in Arminger, Enache, and Bonne (1997) is not repre- sentative for the population. The probability p of a good credit (class 1) in the population lies between 93% and 99%. If a representative sample had been taken to estimate a statistical model, like an artificial neural network or a logistic discrimination model, it is very likely, that the bad credits (class 0) are treated as an estimation error, because the probability of this class is too small (around 5%). Therefore, a sample had to be taken, where the proportion s of good credits is around 50%. The error introduced by sampling is usually called sample selection bias. In cases, where p is too big (class 1 is a frequent event) or too small (class 1 is a rare event), the non-representative sample should be used for estimation, despite of the sample selection bias. The goal of this paper is to find a transformation, which can be applied after parameter estima- tion, to correct this sample selection bias. To simplify the task, the focus will be set on feedfor- ward networks with the cumulative density function of the standardized logistic distribution )exp(1 )exp( )( x x xF + = (1) as activation function of the output unit. 2 Solutions First, the solution for the logistic discrimination is developed. It can be seen, that the developed solution is equal to the transformation Anderson (1972, 1982) suggests. Then the application for certain types of feedforward networks is described. 2.1 Logis