Call Credit

Executive Summary

Given the recent economic turmoil resulting from the banking crisis, consumer credit lending decisions have come to the fore of the consciousness of both lenders and the public and companies are looking carefully at approaches to credit rating and where improvements may be made. For companies who work within or with credit scoring, some questions remain key for each new approach suggested. Those questions include:

• Does this approach offer anything new?
• If it does, is that 'new' thing of value?
• What is that value?

Logical Glue believes that an automated system which can analyse and predict an individual's attitude to credit and credit obligations can assist in producing better lending decisions. In short, Logical Glue propose that learning and understanding those key characteristics of individuals that are highly suggestive of a good or bad credit loan can enable accurate predictions of the outcome of a loan to each individual profile and thereby mitigate the lenders exposure to bad loans.

However, those key questions must be addressed in order to understand if and how such an approach does indeed add value to the mature lending decision market. In order to do this, Callcredit kindly worked closely with Logical Glue on analysing the Logical Glue system and approach in order to help us answer those questions.

Project Brief

For the project with Callcredit, we used the Logical Glue Standalone product offering. Callcredit supplied un-cleaned and unprocessed data in order for the Logical Glue approach to be evaluated in full.

The data set comprised 125,000 instances and 557 variables divided into 75% training and 25 % testing data in random. The data was highly skewed data where bad customers represented only 1.5 % of the whole data set while the rest of the data represented good customers.

The data was noisy with a significant amount being Out of Bounds or Indeterminate Data / Not Deliverable data. We could not remove these data instances as this would leave very few data instances. Hence, we have worked with these data categories.

Project Objectives

The objective of the project was to:

• Find the most influential 45 variables from the available 557 variables and identify their weights in terms of level of influence on the output.
• Produce a system to predict if the given applicant is likely to be a good or bad customer.
• Grade predictions based upon the likelihood of the given applicant becoming a good or bad customer.
• Perform the prediction based on eight categories of output, rather than two categories. Where one of the categories represented good customer categories, the remaining categories represented shades of bad customers. The challenge for Logical Glue was that the data available in these bad categories was very small compared to the data available for the good category.
• Generate customer profiles in the various categories which can be used to understand in depth the key identifiers and characteristics of each profiled category. This method of profile generation could have a secondary application to allow targeted marketing campaigns.

Results

The Logical Glue system:

• Identified the top 45 variables and their respective weights in terms of importance.
• Achieving on unseen data an uplift in predicting bad customers of 15.1 % when compared to the prediction ability of statistical regression models.
• Achieving on unseen data an uplift in predicting good customers of 4.2% when compared to the prediction ability of statistical regression models.
• An average prediction accuracy improvement of 9.65% when compared to the predictions generated by statistical regression models.
• Generation of linguistic profiles \ rules for those customers in each of the eight categories.

The accuracy of predictions generated both by the statistical regression model and the Logical Glue model using the same data were input into Gini Coefficients.

Gini Coefficients are widely used to measure the degree of concentration, or inequality, of a variable in a distribution of its elements. In the case of credit analysis, it is measuring the ability of a predictive measure to discriminate between defaulting and non-defaulting customers. A further advantage of the Gini analysis is to enable different credit evaluation techniques, either quantitative or based on qualitative rating grades, to be compared in their ability to predict default. The more predictive the model, the more the defaults cluster to the left of the chart.

Logical Glue What's New

• Improved prediction accuracy of the creditworthiness of individual applicants when compared to statistical regression models.
• Can efficiently identify credit worthiness based on an applicant's standing in real time.
• Profiles the key characteristics of the customers falling within each of the eight output categories.
• Provides clearer and meaningful linguistic explanations as to why an individual has been predicted or classified within a given category.

• Download Case Study (PDF)