A successful neural network-based methodology for predicting small business loan default

Credit & Financial Management Review, Fourth Quarter 2001 by Yegorova, Irena, Andrews, Bruce H, Jensen, John B, Smoluk, Bert J, Walczak, Steven

Abstract

This study contributes to the credit risk management literature by describing a new, user-- friendly, generic neural network-based methodology for developing credit-scoring models for small businesses based on commonly available data. The methodology is used to construct and validate a model employing data from a pool of terminated small business loans made by an economic development lender based in Maine. A total of 138 variables representing loan characteristics are initially examined, and are subsequently reduced to a set of five input variables that are effective predictors of loan default. These variables, which are composed mainly of traditional financial ratios, are then used to build a probabilistic neural network model that correctly predicts the ultimate disposition of 92% of the loans in the out-of-sample testing. These results are better than those of a binary logistic regression model that correctly classified 86% of the loans.

Introduction

During the last few decades, credit-scoring models have relied on traditional statistical methods, and more recently, on neural networks (NNs) and various hybrids of statistical and neural network methods. This growing body of knowledge is quite complicated and can be confusing to a practitioner who is not a trained statistician.1 This article reports on a highly successful, but uncomplicated, application of NNs designed to predict the default of loans made to small manufacturing companies by an economic development lender based in Maine.

This article builds on the work of Yegorova et al. (2000), who successfully applied binary logistic regression to loan default prediction. Here, we use nearly the same data set of loans to illustrate a more effective, efficient, versatile and user-friendly methodology for credit scoring that handily outperforms traditional statistical methods. Our application exemplifies the usefulness of NNs for practitioners involved in loan making, mortgage and credit approval and other financial applications that attempt to reduce future loan losses.

The remainder of this article is organized as follows: Section 2 presents the main objective of the article. Section 3 discusses recent applications of NNs to classification problems in finance and then enumerates some of the advantages of NNs over statistical methodologies in the area of credit scoring. Section 4 describes the data used to build and evaluate the accuracy of our model. Section 5 discusses commonly used NNs and the training strategy often employed in their application and then describes the probabilistic neural network (PNN) in detail. Section 6 reviews the PNN-based selection method used to prescreen input variables for the final neural network model described in Section 7. In Section 8, we compare and contrast the results of our final model with those of Yegorova et al. (2000). Section 9 concludes the article and provides some direction for future research.

Motivation for the Study

The objective of this study is to build upon the credit-scoring work performed by Yegorova et al. (2000). Based on our literature review, the application of NNs to credit-scoring should result in several improvements over traditional statistical methods. First, and most important, we hope to show increased accuracy in predicting loan default. Second, we expect to see improvement in model-building efficiency, versatility and user-friendliness. Consequently, the neural network methodology that we develop here should help minimize the challenges often experienced when building quantitative credit-scoring models.

Literature Review

NNs have served as versatile tools for data analysis in a variety of complex environments. In finance, they have been successfully applied to bankruptcy and loan-default prediction and credit evaluation [see DeLurgio and Hays (2001) and Jain and Nag (1995)]. Recent technological advances have made NNs an attractive alternative to traditional statistical methods commonly applied in many financial decision-making settings, particularly credit scoring. The reasons for their success seem to revolve around NNs being more capable than traditional statistics in effectively modeling complex relationships. Dorsey et al. (1995) conducted an extensive review of literature and found that neural network models consistently improve bankruptcy forecasts. They found that classification with NNs performed better, both in-sample and out-of-sample, than discriminant analysis and logistic regression.

Practitioners in the financial services industry are also starting to benefit from the improved accuracy of neural network-based credit scoring systems. American Express and Security Pacific Bank are using neural credit scoring systems for credit card fraud detection and small business loan approval, respectively. Lloyds Bowmaker Motor Finance, which now employs a neural network credit scoring system for granting automobile financing, claims that their neural network system is 10% more accurate than the system they used previously [see West (2000)]. As mentioned in Klimasauskas (1996) and Klein and Rossin (1999), hybrid approaches, which employ both statistical and neural network tools, are also being used successfully.