Credit Scoring Comes of Age

by Cordell Wise May/June 2007
Many of today’s scoring models are optimized on various transaction types and data sources and are incorporating additional business data from emerging business data repositories. Application-specific pooled models for specialty markets have also become available. Clearly scoring models for small business decisioning are coming of age.

Credit scores have been used to assess consumer credit risk for many years. Credit grantors find scores to be useful in extending automated credit offers with the confidence that risk has been accurately gauged. Analytics and decisioning solutions help credit grantors streamline operations and cut decision turnaround time with faster application evaluation and improved customer satisfaction. In today’s credit market, where speed is a key criterion, and where as much as 80% of portfolio risk can be eliminated at account origination, the use of analytic models in a flexible decisioning system is imperative to maintain positive customer relationships. Applying such technology in a commercial setting, especially for small business credit decisioning, however has sometimes been viewed as emerging technology. In reality, credit scoring has played a major role in growing the largest small business portfolios for more than a decade. As technologies and data sources have evolved so too have the enabling analytics become quite sophisticated.

Before credit scoring, credit grantors had to evaluate each application individually, asking a series of questions to assess the creditworthiness of the applicant. It could be a long process where applicants often had to wait for days, or even weeks, before knowing whether or not they can close the lease. Credit analysts often had to use their own judgment in evaluating applicants. Leaving a margin for this type of judgment left the decision process subject to credit analysts’ biases and varying levels of experience.

Credit scoring, by contrast, gives credit analysts a way to simplify the credit origination process, and make it consistent from deal to deal and from analyst to analyst. Portfolio managers benefit by maintaining more control over the risk they’re willing to accept and more closely matching underwriting processes to portfolio objectives. By systematically quantifying, or rank ordering, the risk of each application, credit scoring speeds the decision process while simultaneously bringing greater accuracy and fairness to each decision. A credit scoring model generates a credit score for each application — an empirically derived number that analysts can easily use to make good decisions.

Additionally early credit decisioning relied almost exclusively on data provided from the applicant on the application perhaps supplemented with some references and trade credit data if it was available from one of the business bureaus.

What is Credit Scoring?
In the broadest sense, a credit score is simply a tool for assessing risk in an empirical way. Generally speaking, most credit scoring models are developed and designed to help credit grantors predict the outcome of a given transaction. The model or scorecard is composed of several questions (characteristics) about the applicant. Different answers (attributes) are rated on a point system and assigned score weights. An applicant’s score is the sum of all of his or her attribute points — the higher the score, the lower the risk.

For empirical models, the method of assigning the characteristics’ weights and point assignments is designed during model development and is usually based on statistical analysis of hundreds of variables from many thousands of historical applications for credit. The challenge is to find the unique combination of 6-15 characteristics, which can be combined to best determine risk. Exhibit 1 is an example of what a credit scoring model typically looks like.

This score is usually expressed as a three-digit number — the lower the number, the riskier the applicant.

Scores from empirical models represent the odds that an applicant will pay as promised. For example, a certain score may represent 60 to 1 odds, meaning that if 61 applicants receive this score, 60 would be expected to pay as promised and 1 would not. This odds-to-score relationship is extremely useful for lessors to make very precise and informed decisions not just about approvals and declines, but also for optimizing profitability of the deal structure.

There are many kinds of credit scoring models but most can be classified as follows:

Expert Scoring Models — Expert models are generally derived by judgmentally assigning weights to selected characteristics based on the experience and expertise of the developer. It becomes a challenge for developers to remain objective to ensure the models do not merely reproduce and codify the biases of the lessor. Although expert scoring models may be based on a high level of expertise and experience, these models are not developed empirically so no predetermined odds-to-score relationship can be established. Their assessment is essentially an “educated guess” used primarily to automate the decisioning process but can’t provide the accuracy and precision of an empirical model. They’re usually developed where sufficient data isn’t available for an empirical model.

Custom Application Models — Custom models are empirically derived statistically based tools, usually built from data for a specific population or portfolio. They are powerful but very narrowly focused models. In fact lessors can invest considerable expense in developing custom models only to find that they’re not applicable for changing environments. Custom models also require a critical mass of data — usually determined by the number of “bads” in the data sample — which most lessors are unable to assemble. Many lessors also overlook the considerable expense of implementing and maintaining custom models within their application processing systems.

Bureau-Based Models — Bureau-based models are built from broad collections of data available at various data repositories for both consumer and business. They focus on predicting delinquency on any number of tradelines reported at the repository. They are generally good measures of overall risk, but because of their broad focus, their performance can be less than optimal. In addition they’re generally limited to data from the specific data repository when it’s available.

Pooled-Data Scoring Models — These models are empirically derived from a database of pooled creditor data provided by a number of credit grantors participating in the development. Fair Isaac’s Small Business Scoring ServiceSM (SBSSSM) models are an example of pooled models. Because this data is collected on a common basis and the models focus on predicting the outcome of a specific transaction, the result is more powerful and accurate models. They also have the advantage examining data from a wide variety of applications including a spectrum of factors such as risk, industries, geographic locations and environments — a much broader and diverse sample than any one participant could assemble on their own. Furthermore models like SBSS can take advantage of multiple data sources available for decisioning.

Typically the pooled models are deployed on a computer software program, which offers the ability to process and evaluate commercial credit applications automatically. Lessors enter application information about the company, as well as all principals, into an on-screen form. Once the application is submitted for processing, the software will automatically purchase a credit report for the business and designated principals. The software evaluates the application, a business report and/or a consumer credit bureau report, through the use of an embedded pooled model, and determines the credit risk of the business applicant and states it in the form of a score. Pooled models can also be priced on a pay-as-you-go basis, which allows this technology to be accessible to small lessors or lessors who do not have a huge budget.

One of the key advances for small business credit decisioning pioneered by Fair Isaac has been the utilization of consumer information about the principal owners of the business combined with data about the business itself, including financial and application data, to produce robust, highly predictive models of small business risk. The models are able to provide superior performance because they are optimized on a pool of consumers who are small business owners, and our research shows that there are important differences. For example, research shows using a model based on small business owner data can reduce bad rates by 17% without altering approval rates. For a portfolio processing 5,000 applications a year that reduction can represent almost $500,000

Does a Score Also Tell me Which Decisions to Make?
Many lessors use a score as a primary determinant because they are very knowledgeable about scoring or because the transaction fits a certain category (e.g., under $100k), while others use it as a key determinant with other factors considered. Either way automates the process and improves the consistency and level of risk assessment. While a credit score provides a simple and easy-to-understand evaluation of an applicant’s level of risk, the score does not necessarily indicate what action should be taken or which products can be matched to the applicant. This is where a lessor’s experience and policy becomes a critical part of the decision process.

Typically, a credit score is used in conjunction with various rules representing the credit policies dictated by the lessor. One rule (or decisioning criteria) might be to decline any application if it is below the score cutoff. Another decisioning criterion might be to decline credit to any business that fails a capacity criteria like debt service ratio. Other criteria might be based on:

  • Bankruptcies on file
  • Time in business (e.g., start-ups)
  • Specific industries
  • Debt-to-income ratio
  • Lawsuits on file

The credit score can also be combined with several decisioning criteria to determine the product offered and the pricing of the product once an approval decision has been made. For example, a low-risk individual may be offered a lower interest rate versus a risky individual being offered a higher interest rate or perhaps being required to provide collateral. Exhibit 2 is one example of a product matching strategy.

In some types of loans, decisioning can be relatively simple, however, many credit grantors are finding that finer segmentation of their applicants can help them target products more accurately and result in more transactions — and more profitable transactions. This complexity of segmenting applicants into smaller groups for targeted treatment is the reason why a decisioning system is needed along with the credit score. A credit score assesses the level of risk, but the decisioning system can provide an automated way to indicate what action to take.

Weighing the Alternatives
With all the alternatives available what key factors should be considered in choosing a solution?

Performance — First and foremost solutions should be chosen on their ability to predict outcomes. While custom models generally provide the strongest prediction, pooled models can play an important role where custom models aren’t feasible.

Applicability to Population — The closer the data used for the development of the models is to the population of applicants to be evaluated the stronger the models will be. For example, we’ve discussed how using consumer analytics on small business populations can yield suboptimal results.

Applicability to Decision — Does the model address the problem you’re trying to solve? If you’re making decisions about leases, are the models designed for that purpose? Does the model offer an overall assessment of creditworthiness when what’s needed is the outcome for the specific transaction being evaluated?

Data Flexibility — With more data available about small businesses and their owners today than ever before, it only makes sense to use analytics that leverage as much as possible.

Compatibility With Decisioning & Processing Systems — Having high performance analytics is one thing. Implementing them into a technology that enables execution of high performance decisioning is another. The integration of both analytics and decisioning technology together optimizes the decision process.

Summary
Credit scoring can drive improved credit granting decisions. Predictive models are designed as a cost-effective means to assess risk for a variety of portfolios. By using credit scoring, credit grantors can confidently offer instant credit while reducing delinquency and charge-off losses, increase profitability with risk-based pricing strategies and streamline operations with faster application evaluation. The use of analytics in small business underwriting processes is beyond the early adopter stage. More and more small business credit grantors are embracing scoring and its enabling technologies as a way to compete, provide better service clients, improve consistency and risk assessment.


Cordell Wise HeadshotCordell Wise is a senior consultant for Fair Isaac’s Origination and Small Business Solutions. For the past 12 years, Wise has helped credit grantors better understand and deploy analytic solutions involving the extension of credit to small businesses. He also plays a key role in developing new innovative products for more efficient and precise origination and small business credit granting processes. Prior to joining Fair Isaac, Wise was a product manager at First Data Analytics in Omaha, NE. He holds an MBA in International Management from the Thunderbird School of Global Management and a BS in Finance from Brigham Young University.

Fair, Isaac and LiquidCredit are registered trademarks of Fair Isaac Corporation, in the United States and/or in other countries. Other product and company names herein may be trademarks of their respective owners.

Leave a comment

No tags available