According to research conducted by Herbert L. Jensen, Ph.D., an Ernst & Young Research Fellow at California State University Fullerton, "building a neural network capable of analyzing the credit worthiness of loan applicants is quite practical and can be done quite easily."
The credit scoring neural network was trained on no more than 100 loan applications yet achieved a 75-80% success rate. One day's work by an operator familiar with the BrainMaker software package was required to build, train and test the credit scoring neural network. Except for showing a greater bias towards approving weak loan applications, the neural network's loan classification rate was identical to that achieved using a commercial credit scoring scheme.
The input data for the credit scoring with Brainmaker neural network software study consisted of information typically found on loan applications. The outcomes of those loans were classified as either delinquent, charged-off, or paid-off. The actual outputs from the network were 0 to 1 ratings for the three alternatives.
Once the network was built, it was subjected to two training trials. In the first trial, the data was arranged in random order and the first 75 applications were used to train the network. The remaining 50 applications were then evaluated using the trained network. The network misclassified 10 of the 50 applications in the sample for an 80% success rate. In short, the network favored approving loan applications. More traditional and much more costly, credit scoring method used by 82% of all banks, resulted in a 74% success rate. The credit scoring method proved to be more conservative than the neural network in granting credit
In the second trial, the data was rearranged in different random order and the first 100 applications were used to train the network. The remaining 25 applications were then evaluated using the trained network. The network misclassified 6 of the 25 applications in the sample for a 76% success rate. Classifications of good loans as bad and of bad loans as good were equal at 12% each. The credit scoring method for this sample of 25 applications, also misclassified 6 of the 25 applications.
INPUTS
Own/Rent your home
Years with Employer
Credit Cards
Store Account
Bank Account
Occupation
Previous Account
Credit Bureau
OUTPUT
Credit Score