This section provides background information in areas that were central to the preparation of this report: (1) credit-risk evaluation systems, (2) the emergence of credit scoring, (3) the credit-reporting agencies, (4) the content of credit records, (5) the development and estimation of credit-scoring models, (6) generic credit history scores, and (7) the current uses of credit scores.
The ability to quantify credit risk--the risk that a borrower will not pay back a loan as agreed--is central to the core aspects of lending: soliciting accounts, extending credit, pricing (that is, setting the interest rate or fees or other terms), and managing existing credit accounts. As noted earlier, systems in which the credit decision is made manually by a loan officer or other person are referred to here as judgmental systems; those in which the credit decision is made mechanically on the basis of a statistical model are commonly termed credit-scoring systems. Although these systems differ in how the credit decision is made, they can rely on similar information in reaching the decision. For example, both judgmental and credit-scoring systems ordinarily consider individuals' past experiences with credit as reflected in the credit records maintained by credit-reporting agencies. Moreover, a factor considered in many judgmental systems is a statistically derived credit score.8
Both judgmental systems and credit-scoring systems assume that past experience can be used to predict future performance, but not with certainty: Even the best-rated loans might suffer default, and even the worst-rated loans might be repaid as agreed. Rather, the basic goal of any credit-risk evaluation system is simply to differentiate loans that are more likely to be repaid from those that are less likely to be repaid.
Assessments of credit risk have been conducted as long as credit has been offered: Lenders collect information that they believe is relevant to the question of whether a loan will be repaid, and the summary of that information determines whether to make the loan. Whereas judgmental assessments generally rely on less standardized information that may be subjectively evaluated, statistically based procedures draw on types of information that will be similar for all borrowers and evaluate the data through a mathematical process that yields a numerical score.
Judgmental systems continue to be the only practical approach for the few large loans (in relation to the large number of smaller consumer loans) a lender will make to larger businesses. In such multimillion-dollar agreements, the specific attributes of the loans and the circumstances of the borrowers tend to be unique and often highly complex and thus unsuitable for a standardized system. However, judgmental systems--which entail detailed attention to each case--are expensive for lenders to apply to the vast size of their consumer lending portfolio, particularly different types of revolving credit and personal installment loans.
If applied to the same loan application, the judgmental and statistical methods of credit-risk assessment will not always produce the same predictions of repayment likelihood or result in the same decision of whether to lend. Part of the reason is that in judgmental systems, evaluation criteria are often set up as distinct "hurdles" such as a maximum debt-to-income ratio or minimum loan size; as soon as an application is confronted by a hurdle it cannot surmount, it may be rejected without ever being tested against other hurdles. In contrast, in credit-scoring systems, shortfalls or weaknesses in one area may be offset by strength in one or more other areas.
The potentially inconsistent treatment of information is another reason that a judgmental system may reach an outcome that differs from a statistically based decision. Judgmental systems rely on the experiences of individual loan officers to discern the factors that will be good predictors of loan repayment and to identify the tradeoffs among those factors. Differences in loan officers' experiences may lead them to consider different factors and make different tradeoffs among factors.
In evaluating information, statistical systems rely on automated statistical procedures, not on the experience and judgment of loan officers. The statistical procedures consider many credit-related factors simultaneously, statistically identify the relative ability of these factors to measure risk, and assign corresponding weights to each factor. Unlike judgmental systems, credit-scoring systems are consistent in their treatment of information; different outcomes arise entirely from differences in the underlying information and not from the inconsistent treatment of information from case to case.
Credit-scoring systems generally involve significant fixed costs to develop, but their "operating" cost is extremely low--that is, it costs a lender little more to apply the system to a few million cases than it does to a few hundred. This low "marginal" cost-- or the highly "scalable" nature--of the credit-scoring system greatly enhances the lending process by allowing lenders to compete for a wider range of customers and by making their management of existing account relationships more efficient.
Credit scoring first emerged in the late 1950s to support lending decisions by the credit departments of large retail stores and finance companies.9 By the end of the 1970s, most of the nation's largest commercial banks, finance companies, and credit card issuers used credit-scoring systems. Over these two decades, the primary use of credit scoring was in evaluating new applications for credit, and creditors used their own experience and data, sometimes with the aid of consultants, to develop the models. Although often available at the time from local credit bureaus (today more commonly referred to as credit-reporting agencies), credit history records were limited in scope and relatively expensive to access. Thus, lenders essentially had no practical way of obtaining the complete credit histories of noncustomers and so could not effectively target them for solicitations on the basis of credit history.
By the late 1980s much had changed. Creditors were no longer restricted to the credit histories of their own customers and credit applicants. Rather, lenders could purchase the generic credit history scores of individuals who were not their account holders and, with that data, market consumer credit products tailored to various credit scores to appropriate potential borrowers.
The use of credit scoring then spread to additional loan products including home mortgage and small-business lending. Scoring technologies also were applied in new ways, such as in assessments by institutions of whether to purchase individual loans or pools of loans backing securities.10 Finally, credit-scoring technologies were developed to focus on outcomes beyond credit-risk assessment to include, for example, account profitability and various aspects of account management.
As the use of credit scoring was growing, so was the demand for consumer credit and the number of credit instruments offered to finance such activities. Since the early 1900s, merchants have been offering installment credit to allow customers to stretch out their payments for the purchase of furniture, major appliances, and other large durable goods. Charge cards, such as those offered by oil companies and large retailers, first emerged in the 1950s, but in most instances full payment were expected within the billing cycle. In the 1960s, retailers began converting their charge cards into credit cards, a credit instrument that allowed the consumer to extend payments over a long period.
Generic revolving credit, that is, a re-usable credit account not tied to a specific retailer, dates to the 1950s with the emergence of the first bankcards, but it begin to flourish with the introduction of credit cards carrying the Visa and MasterCard logos; its usage more than doubled over the 1970s, with much of that growth taking the place of small installment loans.11 The substitution accelerated in the 1980s and 1990s as credit cards--some tied to home equity lines of credit--became widely accepted for the purchase of larger durable goods and as a ready source of funds through cash advance features.
The development of statistical methods to evaluate credit risk was necessary for the emergence of large-scale open-ended consumer lending, that is, the extension of very large numbers of relatively small loans, each of which has only a small expected return to the lender. In all likelihood, making such loans at the rates they are offered today would not have been possible had it not been for the advances in credit scoring, which have dramatically reduced the cost of offering such credit. Likewise, in the home mortgage market, the application of credit-scoring technologies in the 1990s lowered the costs of both underwriting and funding and promoted greater competition as lenders extended their reach far beyond their traditional branch office locations.
Borrowers with poor payment histories have incentives both to seek out new sources of credit and to withhold information about their credit histories. In the latter part of the nineteenth century, private-sector firms arose to share credit information among lenders and others who were allowed to subscribe to their service. These firms, known today as credit-reporting agencies, do not make credit decisions; rather, they collect, standardize, and disseminate to their subscribers information on a wide range of consumer activity by individuals over time. The activity covers loans, leases, non-credit-related bills, and money-related public records such as court-ordered collections and bankruptcy.12 The agencies also record, and report, the requests for such information that have come from their subscribers, which include not only lenders but employers and others with a legally sanctioned interest in the information.
Credit-reporting agencies, historically referred to as "credit bureaus," were initially established by localized retail establishments and personal finance companies to share information on their customers.13 In 1906, the bureaus established a trade association, the Associated Credit Bureaus, Inc. (ACB), to facilitate the sharing of credit-related information across the country. The membership of the ACB grew substantially, as did the number of individuals covered. However, as late as the 1960s, technological limitations restricted the coverage of even the largest credit bureaus to only a few cities.
As retail establishments sought to serve customers beyond the reach of their local outlets and as consumers became more mobile, the demand intensified for the credit bureaus to efficiently obtain comprehensive information on consumers in many different markets. At the same time, commercial banks, particularly those involved in regional or national credit card lending, had a growing need to gather information about prospective customers in geographically dispersed markets. Technological advances ultimately enabled the bureaus and banks to meet their needs. Those advances also encouraged consolidation among credit bureaus as the smaller entities found the costs of adopting the new technologies prohibitive.
As improved technology reduced costs and increased capabilities over the late 1970s and 1980s, the current national system of gathering and reporting credit-related information emerged. Today the credit-reporting industry is dominated by three national credit-reporting agencies, although the industry still includes a number of smaller firms with only local or regional scope.
The three national credit-reporting agencies--Equifax, Experian, and TransUnion LLC (TransUnion)--seek to collect comprehensive information on all lending to individuals in the United States; as a consequence, the information maintained by each agency is vast.14 Each of these national credit-reporting agencies has records on perhaps as many as 1.5 billion credit accounts held by approximately 225 million individuals. Together, the three national agencies generate more than 1 billion credit reports each year. The vast majority of these reports are provided to creditors, employers, and insurers and individuals have also long been able to purchase a copy of their own report. To improve consumer awareness and understanding of the information included in credit records and to help individuals identify potential errors in their reports, a 2003 amendment to the 1970 Fair Credit Reporting Act (FCRA) provides that individuals may obtain a copy of their credit report free of charge from each of the credit-reporting agencies once a year.15
Credit records a wealth of information about the credit-related experiences of individuals (indeed, all the information needed to construct a comprehensive credit history score; however, they include limited information about individuals apart from name, date of birth, Social Security number, and current and previous home addresses. In particular, credit records do not identify the race, ethnicity, sex, national origin, marital status, or religion. Credit scores are not maintained as part of credit records but rather calculated upon request using the information in the credit records. (A credit score may also be based on additional information not maintained in credit records.) There is a time dimension to a credit record. The credit-reporting agencies can produce a report that shows what an individual's credit record included at any point in time.
Credit records contain information from four broad sources: (1) creditors and some other entities such as utility companies and medical facilities, who report detailed information on the status of current and past loans, leases, and non-credit-related bills such as utility and medical bills (each such loan, lease, and bill is referred to here as a credit account); (2) monetary-related legal records of bankruptcy, foreclosure, tax liens (local, state, or federal), garnishments, and other civil judgments (these records are referred to here as public records); (3) collection agencies, who report on actions associated with delinquent credit accounts and unpaid non-credit-related bills (the credit accounts and bills being handled by collection agencies are referred to here as collection agency accounts); and (4) the credit-reporting agencies' record of inquiries about an individual's credit record made by creditors and others legally entitled to the information.16
Credit accounts constitute the bulk of the information in the typical individual's credit record, and thus the information on credit accounts represents most of the information maintained by the agencies. Credit-account records include the following details about each account: the date it was established, closed (if applicable), last reported on by the creditor, and last used; type of account, such as revolving, installment, or home mortgage; current balance owed; highest balance owed; credit limits (if applicable); and payment performance, such as the extent to which payments are, or have been, in arrears.
The information available on public records, collection agency accounts, and creditor inquiries is significantly less detailed than the data covering credit accounts. In the public records and collection accounts, only the amount of money involved, the type of creditor, and the date last reported are generally available. Entries for inquiries show only the type of inquirer and the date of the inquiry. Inquiry information is retained for up to 24 months; information from public records is retained longer, generally seven or ten years depending on the type of information. Information on credit accounts has no legally mandated time limits except for those that relate to adverse information such as records of delinquency or default.
Credit-reporting agencies collect information from more than 30,000 sources, primarily creditors, governmental entities (mostly courts at the state and local level), collection agencies, and third-party intermediaries. Generally the agencies collect data from each source every month, and they typically update their records within one to seven days of receiving new information. According to the Consumer Data Industry Association (CDIA), credit-reporting agencies receive more than 4.5 billion items of information each month.17
No law requires creditors or others to report data to the agencies. However, although participation in the credit-reporting process is voluntary, entities that do report to the agencies, and the agencies themselves, are subject to rules and regulations governing credit reporting. Access to credit-related information held by a credit-reporting agency and maintenance of each credit report held by the agencies is governed by conditions spelled out in the FCRA.18
The information provided to the credit-reporting agencies has expanded and become much more comprehensive over time.19 However, not all creditors report to the agencies, and not all always report or provide updates on all requested items.20 For these reasons the information on an individual is not always complete. Moreover, reporters do not always report to each of the national credit-reporting agencies, and if they do, they may not report the same information or at the same time to each agency. As a consequence, the information on an individual may differ across the agencies.
The information on an individual may also differ across agencies because each applies its own rules in determining how to assign reported information to a given individual. Such rules are necessary because reporters are not always able to provide a Social Security number when furnishing information or the reported number may be wrong. Also, individuals may have accounts under different names (because of marriage or variations in the use of a middle name or initial) or different addresses (because of changes in residence).21
Fundamental to any underwriting process (that is, the process of evaluating the credit risk of a prospective borrower) is the accuracy and completeness of the information considered. Numerous studies have reviewed the degree to which credit report information is accurate and complete and the implications of data limitations for credit availability and pricing. These studies have reached quite different conclusions.22
Inaccurate data may cause some consumers to pay more, or less, for credit than is warranted by their true circumstances. For the full benefits of the credit-reporting system to be realized, credit records must be reasonably complete and accurate. Yet, under the country's voluntary system of credit reporting, complete information is not always reported to the credit-reporting system. Moreover, data accuracy is an issue under any credit-reporting system. The accuracy of the data affects both credit scoring and judgmental evaluations because both techniques rely on the quality of the information included in credit reports. Judgmental underwriting, which requires a loan officer's individual attention to an application, provides an opportunity to identify inaccuracies that credit scoring does not.
Despite the importance of accurate and complete credit reports, the subject is beyond the scope of this study. However, section 319 of the Fact Act directs the FTC to conduct ongoing studies of the quality of the data in credit reports and report its findings to the Congress.
Developing an effective credit-scoring model is complex, time consuming, and costly. By contemporary standards, early credit-scoring models were built on less robust databases and often focused on information derived from applications, but advances in computing power, access to more-comprehensive credit history information, and improved empirical methods have made credit-scoring models more sophisticated and effective. This section provides a general description of the development of credit-scoring models. A detailed description of the specific, generic credit history scoring model developed for this study is presented in a later section.
Development of a credit-scoring model begins with the collection of data on a sample of individuals and accounts that is broadly representative of the accounts whose performance is to be predicted. Typically, the sample of credit records drawn for estimation is a stratified random sample that includes a larger representation of credit accounts with specific characteristics, such as elevated delinquencies rates, to ensure the model predicts well for each segment of the population. The data must include the outcome of interest--typically, whether the borrower defaulted on a loan--as well as information that may be used to predict the outcome of interest, such as data contained in credit records or data collected as part of the loan application process. The predictive information typically includes the data contained in loan applications and thus antedates the outcomes. When complete, the model can be applied to the data in a new application for credit to generate a quantitative score--the credit score; in most systems the highest possible number represents the greatest certainty that the account holder will pay as agreed.
For the predictive information (termed the "explanatory variables") in a loan-default model to be useful in determining whether a borrower will repay as agreed, the data must include a fairly large number of each type of outcome (termed the "dependent variable")--both defaults and proper repayments. Most accounts are in good standing (such an account is commonly referred to as a "good"); thus, the challenge most often is to acquire a data set that has a substantial number of defaults ("bads"). A traditional rule of thumb for loan-default models is that the sample must include at least 1,500 bads although some use fewer.23
After assembling the sample of data (for example, credit records), the model builder creates explanatory or predictive variables from the data, often referred to as characteristics. Characteristics then are the key inputs of the model used to generate credit scores. Although credit records can be used to create hundreds of characteristics, only those proven statistically to be the best predictors of future credit performance are included in the final model.
The specific characteristics and the weights assigned to each can vary according to the purpose of the model. For example, to support the evaluation of specific loan products, such as home mortgages or automobile loans, a model will typically include characteristics (for example, loan-to-value and debt-to-income ratios) derived from loan applications, as well as information drawn from records of the credit-reporting agencies.
More generally, characteristics representing two types of data are typically used to develop credit-scoring models: continuous data and data that can take only a limited set of values. For characteristics that represent continuous data, such as outstanding balances or the degree of credit utilization (outstanding balance divided by the maximum amount the individual is authorized to borrow), the model builder generally simplifies the data by defining ranges that differentiate meaningfully among different levels of risk. For example, credit utilization might be represented by ranges such as above 90 percent, between 50 percent and 90 percent, and below 50 percent.24 The options are by definition more limited for characteristics that can take only a limited number of values, such as "yes" or "no" (for example, for the characteristic that represents whether or not an individual has an entry for a public records).
Finally, each value of each characteristic--including each range for a continuous characteristic--is assigned a specific point count, and the credit score for any given individual is equal to the sum of these point counts over all characteristics considered in the model. The point counts and selection of the specific characteristics used in the model are derived from a statistical analysis of the relationship between characteristics at an initial point in time and credit performance over a subsequent period. The statistical model typically used in predicting loan performance takes the form of a so-called logistic regression, in which the dependent variable is the logarithm of the odds ("log-odds") of the probability of default versus nondefault. Specifically, the log-odds is the logarithm of the ratio of the number of "good" accounts to the number of "bad" accounts in the estimating sample.
The model estimation undertaken to identify and assign weight to each characteristic to reflect their relative importance in determining borrower performance is generally done using multivariate techniques. Because the characteristics that bear on credit risk are likely to be correlated with each other, the weights assigned in a multivariate analysis are likely to differ from the weights that would be assigned if each characteristic was used to predict performance in isolation. It also may be the case that characteristics which are highly predictive when considered in isolation may contribute little in a multivariate framework. The converse can also be true. A characteristic can have a significant role in a multivariate model even when it does not exhibit strong predictive power in a univariate setting. A tendency for a high degree of correlation among credit risk characteristics is one reason that scoring models ordinarily include only a relatively small number of distinct characteristics. According to industry sources, a typical credit-scoring model will include eight to fifteen characteristics.
An important stage of model development involves validation of its predictive accuracy through a series of statistical tests. One common validation method is to establish a "hold-out" sample (a portion of the original sample not used to estimate the model) to test how well the estimated model predicts the outcome of interest. Two of the most widely used statistical measures of accuracy are the Kolmogorov-Smirnov (KS) test statistic and the divergence statistic (refer to box "The Kolmogorov-Smirnov and Divergence Statistics").
These sorts of statistical measures are used not only to determine the overall effectiveness of a model but also to help determine the number of characteristics to include in the model. Typically, the final choice involves a tradeoff between the additional effect of a characteristic on the model's predictive accuracy and a desire to keep the complexity of the model manageable. The hold-out sample is useful in deciding the issue. Testing the model against the hold-out sample reveals whether each characteristic included in the model is predictive using data not used to construct the model. Characteristics that do not prove predictive for the hold-out sample would likely be dropped from the final model.
The final stage of model development typically involves translating, or "normalizing," the raw statistical output, which is typically a log-odds prediction, into an easily understood score. Such normalizations must preserve the relative order among individuals.
The Kolmogorov-Smirnov and Divergence Statistics
The Kolmogorov-Smirnov (KS) test statistic is the maximum, across all credit-score values, of the difference in the cumulative proportions (in percentage points) of goods and bads. A zero value for the KS statistic means that the two credit-score distributions are the same and indicates that the credit score fails to differentiate between defaulters and nondefaulters; a value equal to 100 indicates that the credit score perfectly differentiates defaulters from nondefaulters. The KS statistic for a given credit-scoring system is the maximum vertical distance between the two curves for that system.
Whereas the KS statistic describes the ability of a credit-scoring model to differentiate goods from bads at a single point, the divergence statistic compares how the entire distributions of defaulters and nondefaulters differ. The divergence statistic is calculated as the square of the difference of the mean of the goods and the mean of the bads, divided by the average variance of the score distributions. When the model performs poorly, so that the average credit score of bads is not much different from the average score of goods, the divergence statistic will be close to zero. As the model's performance improves, increasing the difference between the mean scores of bads and goods, the divergence statistic increases. The larger the divergence statistic, the greater the predictive power of the model.
Under the Federal Reserve's Regulation B (Equal Credit Opportunity), a credit-scoring system that considers age must be empirically based, must be demonstrably and statistically sound, and cannot use "prohibited" information, which is information about an individual the use of which by creditors is prohibited by the Equal Credit Opportunity Act.25 Prohibited information includes of race, ethnicity, national origin, religion, sex, and marital status. Certain information, such as age, receipt of child-support, and receipt of income from public assistance can be used, but only in restricted ways.
Creditors also exclude from their credit-scoring systems still other information available to them. Such information consists mostly of certain inquiries made to the credit-reporting agencies to check on the status of an individual's credit record. These inquiries consist of those made by consumers to check on their own credit reports; by employers or insurance companies; and by lenders either considering extending an unsolicited credit offer or checking for changes in the credit circumstances of their existing customers. However, inquiries made by creditors evaluating credit applications from the individual may be included in credit-scoring systems because they are consistently found to be predictive of future performance.
A concern has been raised in recent years about the possible adverse effect on credit scores of multiple inquiries stemming from credit shopping. From a credit-risk perspective, multiple inquiries arising from shopping for a specific loan for a specific purpose are not as significant as those arising from simply trying to obtain as much credit as possible. In an attempt to implement this distinction, generic credit history scoring models now customarily attempt to consolidate into one inquiry those that are similar (typically, from the same type of lender or for the same type of loan) and made over, say, a rolling two-week period.26
A new type of credit score emerged at the end of the 1980s--one based entirely on the information included in the credit records maintained by credit-reporting agencies: a generic credit history score. Previously, most credit-scoring models were custom models developed with information specific to an individual lender and product. The demand for credit scores that could be used to acquire new customers for a variety of loan products stimulated the development of generic credit history scores. Developing the models for such a score became affordable only when computer technology and the structure of the credit-reporting agency industry had sufficiently evolved.
Over time the lending industry and firms that support their activities have developed a great many versions of a generic credit history score. The first two widely available scores were the MDS Bankruptcy Score introduced in 1987 and produced by Management Decision Systems, Inc., and the FICO Prescore, developed by Fair Isaac Corporation (Fair Isaac).27 The FICO Prescore scores were used in underwriting new credit card accounts. TransUnion was the first credit-reporting agency to offer a credit history based score with an online, real-time credit report in 1987.
The use of generic credit history scores expanded over time to a wider array of loan products and uses. In the mid-1990s, Fannie Mae and Freddie Mac recommended the use of both FICO scores and the MDS Bankruptcy Score for the underwriting of the home mortgage loans they purchased. According to Fair Isaac, FICO scores are involved each year in more than 10 billion credit decisions of all types. Fair Isaac also estimates that FICO scores are involved in more than 75 percent of all mortgage originations (refer to box "FICO Scores").28
The FICO score, like most other generic credit scores, ranks consumers by the likelihood that they will become seriously delinquent on any of their credit accounts in the near future (typically over the next 18 to 24 months). The most commonly used FICO score ranges in value from 300 to 850 (the higher the number, the lower the credit risk). Each of the three national credit-reporting agencies calculates a FICO score, to the extent possible, for each individual in its records. In doing so, each agency uses models developed by Fair Isaac specifically for that agency and with that agency's data. Upon request by a creditor or others, the agencies calculate an individual's FICO score using the most up-to-date information in each individual's credit record.
Fair Isaac has developed generic credit history scoring models that focus on different populations. Versions of the models are used for varying purposes, such as for underwriting automobile credit and credit cards. Two of these versions (which Fair Isaac calls the Classic FICO score and the NextGen FICO score) generate ratings on the basis of data drawn from the general public. A third model, designed for use with individuals who have little or no credit history in the files of the three national credit-reporting agencies, generates a rating called the Expansion score.
Each of those three credit-scoring models is calibrated separately for several subpopulations; each group has one or more distinguishing characteristics in common (a technique discussed in more detail later in the main text). The model for the Classic FICO score has ten variations (called "scorecards" by Fair Isaac); the NextGen model has eighteen scorecards. The selection of scorecards is analytically driven to more effectively predict risks in certain key subpopulations, such as those that have severe derogatory information in their records. Compared with the Classic FICO score, the NextGen model seeks to better distinguish individuals who are likely to perform well (or worse) on multiple credit obligations. The NextGen model also focuses on individuals with credit records that evidence little use of credit or that contain only limited information (individuals for whom the conventional FICO model often cannot generate a score at all). Fair Isaac estimates that the NextGen model increases the proportion of such individuals who are scorable, principally those with little credit experience, by about 2 percent. The firm also reports that, in tests, the NextGen scores substantially outperform the Classic scores.*
Each of the credit-reporting agencies offers Fair Isaac credit scores to lending institutions and the broader public under a unique name, in part to reflect the fact that the model created to generate the score was calibrated from the agency's own particular data. The Classic FICO score, for example, is called the Beacon score at Equifax; the Experian/Fair Isaac Risk Model score at Experian; and the FICO Risk score, Classic (formerly the Empirica score) at TransUnion. The NextGen FICO score is known as Pinnacle at Equifax; the Experian/Fair Isaac Advanced Risk Score at Experian; and FICO Risk Score, NextGen, at TransUnion.____________________
* Matthew Hubbard and Steve Gregg (2001), "NextGen FICO Scores: More Predictive Power in Account Management," a Fair Isaac Paper (September), www.fairisaac.com. Return to text
Because each national credit-reporting agency uses a Fair Isaac model developed specifically for that agency and its data, the models differ to a certain degree. In addition, information on an individual may differ across the three agencies. Hence, an individual credit score may differ across the three agencies.
Besides the FICO score, each of the three national credit-reporting agencies makes available a generic credit history score derived from its own models. Recently, a new generic credit history score named the VantageScore became available to the marketplace. The VantageScore was developed by VantageScore Solutions LLC, a joint venture by Equifax, Experian, and TransUnion to create a measure of credit risk that scores individuals consistently across all three companies.29 The VantageScore applies a single credit-scoring model to the data at each of the national credit-reporting agencies to ensure that the only reason that the credit score for an individual might vary across the three agencies would be differences in the data maintained by these firms.30 The VantageScore ranges in value from 501 to 990, with lower scores representing greater credit risk. As with the FICO models, the algorithm used to generate the VantageScore involves multiple scorecards.
Proprietary models can be developed and used by individual lenders instead of, or in addition to, the generic scoring systems described above. Little information is publicly available about proprietary credit-scoring models; however, they may supplement credit history information with information beyond that included in credit records. Although the various credit history scoring models differ in their scoring ranges, in their estimation samples, and in their methods of measuring performance, they all rely exclusively on credit-record data from the national credit-reporting agencies.
The characteristics created for a generic credit history scoring model tend to be similar across such models. These characteristics are generally of five broad types: (1) payment history, (2) indebtedness, (3) length of credit history, (4) types of credit used, and (5) acquisition of new credit.31 These five types are not of equal importance in determining credit scores. For example, for the general population, Fair Isaac reports that payment history characteristics are the most important type, accounting for about 35 percent of the FICO score's predictive accuracy; consumer indebtedness accounts for about 30 percent; length of credit history, 15 percent; and types of credit used and acquisition of new credit, each about 10 percent. These proportions may vary for particular subgroups of individuals, such as those with only a short history of credit use.
Payment history. In general, the most important characteristics considered in credit-risk evaluation are those that relate to an individual's history of repaying credit and any evidence of money-related public actions or non-credit-related collections. The essential issue captured by payment history is timely repayment. Specific measures include the frequency of delinquencies, the severity of delinquencies, their age and dollar amount, and how recently they occurred. Repayment performance is evaluated on the full range of accounts that an individual holds, distinguishing among accounts by type (revolving, installment, mortgage, and others) and source (banking institution, finance company, retailers, and others). In general, an individual whose credit record includes a major-derogatory account, collection account, or public record will find qualifying for new credit difficult, may face higher interest rates for the credit received, or may be limited in further borrowing on existing revolving accounts.32
Indebtedness. When evaluating credit history, creditors also consider the type and amount of debt an individual has and the proportion of available credit in use (credit utilization). For revolving accounts, credit utilization is measured as the outstanding balance divided by the credit limit, which is the maximum amount the individual is authorized to borrow on the account. For mortgage and installment accounts, credit utilization is generally measured as the unpaid proportion of the original loan amount. High rates of credit utilization may reflect a financial setback, such as a loss of income or an inability to manage debt, and thus are generally viewed as an additional risk in credit evaluations.
Length of credit history. The age of credit accounts is relevant to an evaluation of credit quality because it provides information on the extent of experience an individual has had with credit. New accounts may convey little information other than that the consumer had a very recent need for additional credit and was approved for credit.
Types of credit used. The use of many or all of the several types of credit accounts (revolving, retail, automobile, and mortgage) by an individual, together with how recently they have been used, has been found to have a bearing on credit risk.
Acquisition of new credit. Searching for new credit, as well as obtaining it, provides information about credit risk. A relatively large number of new accounts or efforts to obtain loans as indicated by recent inquiries from creditors tend to indicate elevated risk.33 For example, the recent opening of a relatively large number of accounts may signal that an individual is becoming overextended.
Like characteristics, the estimation process used in the development of generic credit history scoring models is similar across such models. The goal of the estimation process is to choose the characteristics that best predict borrower performance and assign weights to them to reflect their relative importance.
Typically, the estimation process uses a representative sample of individuals available at two points in time separated by 18 to 24 months. Performance of the borrower is measured by delinquencies or defaults that take place in that period. The predictive characteristics are calculated entirely from the initial sample.
Although a generic credit history score can be estimated over the entire sample, experience has shown that the predictive accuracy of the model may be improved by first segmenting the sample of individuals into distinct subpopulations (scorecards) for purposes of estimation. A separate model is then estimated for each scorecard. The predictive characteristics and their weights will generally differ across scorecards given the differences in the information in the credit records for each subpopulation. The final choice of characteristics for each scorecard is guided by, among other things, the marginal predictiveness of each characteristic and whether the implied statistical relationship between the values of the characteristic and performance is reasonable.34
Generic credit history score models, like other types of credit-scoring systems, need to be periodically re-estimated to reflect changing conditions in credit markets, although the models have been found to be robust over differing economic conditions. There is no formal timetable for re-estimation, but typically it is undertaken every couple of years. Periods that have witnessed substantial volatility or notable changes in the credit environment warrant more frequent re-estimation than other periods. The FICO score models developed using data from each of the credit-reporting agencies are not re-estimated at the same time.
The final credit score for an individual normalizes the results from each scorecard to a common scale representing a prediction of future performance. When models are updated through re-estimation, typically the credit scores are normalized in a way that aligns with a risk-to-score relationship observed at a given point in time.
In most credit-scoring systems, a higher credit score represents a lower degree of estimated credit risk.35 Each lender determines, on the basis of its own business strategy, which credit scores represent an acceptable degree of credit risk or at which points in the continuum of scores it will establish different interest rates.
As noted, FICO scores are the most widely used generic credit history score. According to Fair Isaac, nearly 60 percent of individuals with credit records that are scorable have FICO scores of 700 or more; about 15 percent of individuals have FICO scores below 600 (table 1). The median FICO score for the population of scorable individuals is about 720.
Fair Isaac's analysis of the relationship between payment performance on loans and FICO credit scores finds that individuals with low credit scores are much more likely to experience a serious delinquency or default than individuals with higher scores (table 2). For example, for new accounts extended to individuals with FICO scores below 520, about 40 percent subsequently experienced a "bad" (a delinquency of at least ninety days or other serious derogatory such as bankruptcy), compared with a "bad" rate of less than 1 percent for accounts extended to individuals with FICO scores of 760 or more. Moreover, according to Fair Isaac, default rates on new credit decrease consistently with increasing credit scores.
Not surprisingly, then, many lenders underwrite and set the price (interest rate or fees) on their loans according to risk as estimated by credit score. Both previous research and fair lending reviews conducted by banking institution supervisory agencies indicate that, all else being equal, individuals with lower credit scores or weaker credit histories are more likely to have their applications for credit denied.36
The connection between loan price and credit score is not rigid or uniform. For most lending, the credit history score is only one of several factors used to assess credit risk, and creditors differ in their willingness to bear risk. Consequently, no universally established credit-score threshold exists to define acceptable risk, and no universally established correspondence exists to link a specific score to a specific loan price. Nevertheless, information on the relationship between credit scores and interest rates made available by Fair Isaac shows that better credit scores are associated with lower interest rates on credit (table 3). This relationship is routinely observed in the "rate sheets" used by loan officers when establishing the interest rate on new credit.37
Industry estimates suggest that between 35 million and 50 million individuals either do not have a credit record at a national credit-reporting agency (so-called no-file individuals) or have a record with too little credit experience to reliably calculate a traditional generic credit history score (so-called thin-file individuals).38 Individuals lacking robust credit records disproportionately include young adults and students; recent immigrants; recently divorced or widowed individuals; and those who do not use much credit or rely primarily on non-mainstream sources of financing, such as pawn shops or payday lenders.39
The inability to calculate credit scores for some individuals may limit their access to credit. For example, they may not be included in solicitations for credit that rely only on credit-reporting agency records. And because creditors may not be able to reliably gauge the credit risk posed by individuals lacking a credit score or because they do not wish to spend the time or money required to gather additional information, such individuals may find it more difficult obtain credit or receive it on the best terms available. Creditors are, however, more likely to expend extra time or money to gather information about the credit-related experiences of individuals applying without a credit score for larger loans, such as home mortgages or automobiles.
Given the tens of millions of individuals without a credit score, the credit industry has an incentive to develop cost-effective techniques and sources of information to determine which individuals present a profitable lending opportunity.40 In response, alternatives to the traditional credit history score have emerged recently.41 These alternative credit scores are based on information gathered to supplement the data in traditional credit-reporting agency records, including information related to deposit account records, payday loans, purchase payment plans (rent-to-own transactions), rent and utility payments, and regular child-care payments.42 These expanded data are used to develop a more informative credit record for individuals that may be used to generate nontraditional credit scores. These alternative scores may be used by creditors to underwrite and establish the terms of loans and for marketing purposes. By expanding the information available to judge credit risk, alternative credit-scoring systems and the expanded data upon which they are built allow creditors to better assess credit risk and offer credit to more consumers and on terms more consistent with the risks they pose.
Creditors vary greatly in their use of credit scores for credit evaluation and pricing. Even for a given creditor, the use of credit scoring may differ markedly across loan products: The weight accorded the score in judging creditworthiness may vary, and for some products a specific score may be established to define unacceptable risk.
Perhaps more common for loan underwriting, however, is to associate particular score ranges with particular interest rates. Creditors often distribute rate sheets to underwriters to specify the interest rates corresponding to various credit-score levels. The rate sheets are sometimes rendered as a grid, with each cell representing a combination of a credit-score level and the level of another key underwriting factor, such as the loan-to-value ratio. In this type of underwriting structure, the creditor is defining the tradeoffs between changes in credit-score level and offsetting changes in the other factor that will maintain an essentially unchanged credit risk and, consequently, unchanged pricing.
Over time, however, credit scores have increasingly been applied to other aspects of the lending process, including prescreening and account marketing, loan pricing, account management and loan servicing, fraud detection, estimating loss in the event of default, and estimating account profitability.
Another aspect of account management is the servicing of delinquent loans. Borrowers tend to react differently to the various options used by lenders to recover delinquent loan payments. For example, a reminder that a payment due date was missed will be appreciated by some account holders but will antagonize others. The costs of the various recovery options--ranging from letters and telephone calls to legal action--also vary greatly. Credit and behavioral scoring are used to predict the actions that are likely to have the highest return net of expenses.47 Perhaps most important, lenders and loan servicers have found that credit scoring can be used to target delinquent borrowers for early intervention to help avoid default and minimize losses.48 Credit scores are also used for monitoring and auditing purposes in the context of account management.
A credit-scoring model developed for one purpose (for example, to answer the question, What is the likelihood of default?) may be ineffective when used to answer a different question. Moreover, a credit-scoring system generally applies only to borrowers who are similar to the group of borrowers used in developing the scoring system. Thus, to use scoring methods to answer a different question or to ask the same question but for a different group of borrowers generally requires gathering new data and developing an entirely new scoring model.