CREDIT SCORING PROJECT

 

Logical Designs was involved in developing the first neural network to be used for decision making in the financial services industry. This project involved the development of a network to evaluate the creditworthiness of an individual based on information contained in the loan application and credit report.

This project represents a typical application suited to neural networks. There was a substantial amount of information available in a database of loan applications. For each application, there are a large number of potential input parameters available for network training.

The problem was difficult since there was a large overlap between good and bad loans. This was known because a standard credit scoring system based on discriminant analysis had been applied to the data. This, however, gave us a performance "target". We needed to out perform the traditional method.

Before a network can be trained, a determination must be made as to what information will be presented to the network and what output is desired from the network. For each input, the proper choice of preprocessing operations had to be made. The obvious choice for the desired network output was to simply classify the loan as either good or bad.

A data set of 16,000 loan applications was supplied for training and testing the network. The decision was made to apply backpropagation learning algorithm to the problem.

The available data was split into a 10,000 application training set and a 6,000 loan application test set. A number of networks were trained on this data and the performance was evaluated on the test data and compared to the target credit scoring system.

Tests were run to determine the effects of preprocessing, size of network, desired output, and learning rates on network performance. A modification was made to the desired output of the network that reflected the value of the loan if accepted. This modification proved to provide a significant performance improvement.

The final trained network returned a profit on the test data that was 27% higher than the target credit scoring system. Over the total data set of 16,000 loans, this represented an additional $300,000 of profit.

Logical Designs can handle all aspects of the design of your neural network application. We can supply custom programming for preprocessing of data. Logical Designs can build and verify a training set from your data. We have the equipment to train and test the neural network and can also integrate the trained network into your application.