A scorecard is a tool used to calculate the risk associated with a credit application. It calculates credit risk based on multiple items of information called characteristics. Characteristics can come from several sources, including the credit application and consumer and business credit reports. Each characteristic is divided into two or more possible responses known as attributes. A numerical score is associated with each attribute, so for any credit application, the numerical attribute values for all characteristics can be added together to provide a total score.
RMA and Fair, Isaac recently developed two pooled-data scorecards to meet the needs of a wide range of small business customers. RMA member banks supplied the data and credit expertise, and Fair, Isaac supplied the data analysis and modeling ability.
The pooled-data scorecards give small business lenders a cost-effective means to obtain substantial predictive power in screening applications. The objective and consistent decision-making support offered by scoring is valuable to small business lenders for a number of reasons.
The Value of Small Business Scorecards
Scoring uses the same data a loan officer uses in his or her judgmental, or nonscoring, decision process. But scoring is faster, more objective, and more consistent. With the current regulatory pressure to provide more small business loans, prospective lenders need efficient, time-saving, cost-cutting tools. Processing credit applications quickly - and with accuracy - is the key to making small business lending more efficient. With credit-scoring, a lender can increase the number of approved applications without increasing risk, time, or other resources.
A recent article in American Banker focused on the trend of using credit-scoring technology in small business lending to expedite loan processing, for marketing purposes, and to reduce costs and processing time. In the article, industry experts predicted that a retail-style loan approval process will ultimately cut credit underwriting time from an average of 12 hours to as little as 15 minutes per loan.(1)
Scoring does not require an either/or approach. Loan officers can use scoring in conjunction with traditional judgmental appraisal methods. Scoring can quickly identify and approve loans to very low-risk business applicants and decline loans to very high-risk business applicants. This process leaves loan officers with more time to use their expertise in evaluating businesses that present tougher decisions - credits with more moderate levels of risk that fall in the middle between low and high risk.
Small business lenders can also increase profitability by using scorecards to price loans. Similar to consumer lending practices, credit-scoring allows small business lenders to set interest rates and fees based on the level of risk an applicant represents. In this way, lenders can approve more loans by offering higher rates and fees to high-risk applicants and offering lower rates and fees to good, low-risk applicants.
Small business lending bridges the gap between commercial and consumer lending. Yet before scoring became available, most commercial lenders evaluated and processed small business loans the same way they handled large business loans. However, small business lending is much more like consumer lending than large commercial lending. Both small business and consumer lending generate high volumes of applications yet involve lower dollar amounts per application than commercial requests. Also, applicants expect a quick decision, and although small business lenders have traditionally taken up to two weeks to process an application, lenders are now pressured to produce a decision in one workday or less.
However, the greatest similarity between small business lending and consumer lending is that the credit-worthiness of a business is tied to the financial profile of the business' principals. In fact, the pooled-data scorecards rely as much, if not more, on information about the principals of the business as they do on the business itself. For these reasons, small business lenders can benefit from the experience of consumer lenders who have relied on credit-scoring techniques for nearly 40 years.
The Need for a Pooled-Data Model
Before there were pooled-data scorecard models, scorecard developers offered either judgmental models (built from various sources of data and the developers' professional experience) or fully customized models (built from data based on a bank's previous credit experience). Few banks were able to build custom models owing to data constraints as well as costs. As such, many lenders opted for judgmental scorecards. Judgmental scorecards provide a measure of predictive ability. They can be used as an adjunct to a bank's standard review process and have proven useful for many lenders. However, the new pooled-data models give lenders the advantages of the greater power and predictiveness found in an empirically derived customized model.
The primary development goal was to build a scorecard that strongly and consistently distinguished a lender's "good" applicants from the "bad" applicants. It is important to note that a good applicant is not simply an account that will never be delinquent, and a bad applicant is not simply an account that will be written off. Instead, the definition of each parallels the question, "If you had a crystal ball and could see the future payment behavior of any loan you provided to a business, would you still want to provide that loan?"
For example, an account that occasionally goes 30 days delinquent may still be desirable and will most likely produce greater profit for the bank through interest and fee income than an account that always pays as agreed. Likewise, an account that is repeatedly 90 days delinquent, even if it eventually pays off the loan, will still cost a lender so much in collection effort that the bank is very likely to lose money.
Once a lender has a tool in place that can statistically rank good and bad applications, the lender can:
* More actively manage the trader-writing process.
* Make decisions based on the most predictive information available.
* Determine if an applicant is a candidate for other products and services.
By using scoring to rank the applicant by risk factors, the lender can pursue different strategies, such as increasing approvals while keeping losses constant or reducing delinquencies while keeping acceptance rates constant. Of course, in today's competitive environment, most small business lenders are looking to increase approval percentages.
Development of the SBSS
The thoroughness and objectivity of the development process and the data make scorecards reliable predictors of risk. For the RMA/Fair, Isaac Small Business Scoring Service (SBSS), 17 leading small business credit grantors contributed data to form the initial data pool. These RMA member institutions represented different asset sizes and were from diverse geographic areas. The data were derived specifically from small business loan customers. Small business was defined as a firm with gross sales of less than $5 million and a total potential exposure to the credit grantor of less than $250,000.
Defining Good and Bad Applicants
To define what constituted a good and a bad applicant, an initial design meeting was held with representatives from the participating banks. The RMA bankers agreed that a bad account would be any account that had ever been 60 days or more delinquent. A good account was defined as any account that had not been 30 days delinquent more than twice during the first four years of account history.
Another purpose of the meeting was to determine what characteristics should be examined and to gauge each characteristic's predictive value. RMA Associates were able to help the model developers by contributing practical knowledge and insight on relevant characteristics and by alerting the model developers to what information would typically be available in their loan files.
The next step in building the credit-scoring model was to conduct a trial sample. The trial sample asked for data for a small number of loans from good, bad, and declined accounts. The trial served two main purposes:
1. It gave the participating banks an idea of the sampling procedure and the data that would be required.
2. It allowed the model developers insight into the size of sample the banks could reasonably provide.
After the trial was completed, a full sampling process took place that necessitated the participating banks committing extensive resources to the project. For full sampling, each bank provided data for approximately 300 accounts (100 good, 100 bad, and 100 declined) and manually precoded the relevant data on data sheets. For most banks, many credit applications had to be reviewed to find the 300 that fit the sampling criteria. The participants were also requested to attach consumer credit bureau reports for up to two principals of the business and a commercial credit report for the business itself.
The data submitted by the banks were entered into a database that included additional information from the credit reports.
Preparing the Data
The next step in the process was to take the data collected and generate the characteristics to be reviewed for the development of the scorecard. Characteristics were broken into four categories:
1. Consumer credit bureau report data.
2. Business credit bureau report data.
3. Financial ratios.
4. Credit application data.
From past experience, the model developers knew that financial ratio characteristics generated from small business data pools are surprisingly weak predictors of risk when viewed in absolute terms outside of the context of specific industries. As such, a "look up" function was created to incorporate financial ratios from RMA's Annual Statement Studies, which compared the business' financial ratios with the RMA industry standards.
Building the Scorecard
Fair, Isaac used its proprietary scorecard development technology, the InformPLUS software, to develop the pooled-data scorecards. InformPLUS is a scorecard development system that combines more than 350 software programs to address typical data problems experienced in the credit environment. It does not assume that all data are created equal. Instead, it identifies "noisy" or suspect data and can even account for missing data.
The development of the pooled-data scorecard was not a "black-box process" in which data were dumped into a model and a scorecard was pumped out. Instead, throughout the scorecard development process, developers asked the questions:
* Will adding these characteristics to the scorecard make it more predictive?
* Do the results make sense?
* Are the results believable?
Weighting the Scorecard
During the development of the scorecard, each characteristic of the principals was weighted and combined with the score based on the business' characteristics. In addition, other weighting strategies were used to account for regional differences and, in a few cases, for different sample sizes supplied by the participating banks.
Segmenting the Data
Enough data existed to produce two scorecards. Segmentation analysis was undertaken to determine the most appropriate basis on which to split the data. The goal of segmentation analysis is to divide the applicant population into distinct and dissimilar subpopulations to maximize overall predictive power. The following subpopulations were examined:
* Gross sales.
* Net sales.
* Type of business - corporation or noncorporation.
* Geographic region.
* Loan type, that is, line of credit or term loan.
* Industry groupings.
* Total current request for credit.
After analyzing the subpopulations, the model developers determined that producing two scorecards based on the total request for credit produced the most predictive set of scorecards. The cutoff point was $35,000. The smaller credit request was defined as less than $35,000 while the larger credit request was for loans of more than $35,000.
Using Reject Inference Techniques
To ensure that the scorecard included information representative of the total applicant population, the "reject inference" technique was employed. The goal was to estimate what the performance of declined applicants would have been had they been approved.
If a scorecard is based only on information from the stronger, less risky accounts, it would not sufficiently penalize the lowest quality applicants not represented in the approvals. To see the full spectrum of risk among the applicant population, performance information from the entire population should be analyzed, including declined applicants.
To maximize divergence (the statistical measure of how well the scorecard separates good applicants from bad applicants), the algorithm used for constructing the scorecards modified all score weights simultaneously. This algorithm determined which characteristics would actually be used. Characteristics highly correlated with other scorecard characteristics or contributing little to scorecard divergence were dropped.
At this point, another meeting of RMA participating banks was held for a discussion of interactive weights. The banks reviewed the scorecards and discussed either adding or removing characteristics. Because the scorecard development technology allows scorecards to be modified and rebuilt interactively, it was possible to test alternative scorecards at the meeting and to receive instant feedback from participants.
Scaling the Scorecards
The final step in the development process was to scale the scorecard weights. Scaling makes it easy for an analyst to quickly determine the relative risk of an applicant. Each scorecard was scaled so that 20 additional points doubled the odds of an account being good. For example, a business scoring 200 is twice as likely to be a good account as a business scoring 180. The scorecards were also scaled so that both scorecards offer the same probability of success at a given score.
Breaking Out Small Loan Requests
At the second meeting of the RMA participants, it was decided that the scorecard for smaller credits (less than $35,000) should be broken out into two possibilities: credit applications with financial statement data and credit applications without financial statement data. The RMA bankers decided that the option of having two small-loan scorecards would allow an institution to decide whether to request a financial statement or not.
As with any scorecard, the RMA/Fair, Isaac small business pooled-data scorecards should be validated to measure ongoing predictive strength. However, it is reasonable to assume that new scorecards will be developed before most banks can accumulate enough data to perform a validation. Fair, Isaac projects that scorecard redevelopment will take place every two years. In conjunction with future redevelopment efforts, further segmentation of the population will be explored. Additional segmentation may result in multiple scorecards in the future.
Efficiency is the key aspect in processing credit requests from small businesses. The new pooled-data scorecards are designed to provide small business loan officers with the ability to process a larger number of applications in less time and without compromising credit quality. Increased regulatory pressure for consistency and objectivity in the credit underwriting process dictates more widespread use of credit-scoring for small business customers. In addition, there are already market pressures to begin applying scoring technology in the marketing and portfolio management arenas.
As a result of the pooled-data scorecards, commercial loan officers will be able to more easily capitalize on the growing small business market, and creditworthy small business customers will be able to more quickly and easily receive the credit they need.
1 "Fast Turnaround Still Mostly a Dream, But Technology May Make It Happen," American Banker, February 13, 1995, p. 11.…