What's the Score?

Abstract

Traditionalists don't like it, but credit scoring-increasingly sophisticated-often proves more accurate than human judgment for certain loans At the same time that delinquency rates and bankruptcy rates are climbing, competition for the credit consumer is also toughening. That combination places a high demand on the ability of credit vendors to be able to identify new customers without exposing themselves to unacceptable levels of risk. The tools requisite for finer discrimination, fortunately, have been developing concurrently. Statisticians have developed ever more sophisticated mathematical modeling techniques, especially in the last half-dozen years, which are allowing faster and more sensitive readings of a credit applicant's financial pulse. Although credit scoring (credit risk prediction using numerical formulas) has been around for more than a decade, the degree to which lenders are relying on it for their deci sion-making has been increasing sharply in recent years. Where early credit scoring methods were generally limited to revolving credit, the range of products and questions assessed by current techniques is broad and expanding. According to Peter E DiGiammarino, vice-president of American Management Systems, Inc. (AMS), of Arlington, Virginia, and general manager of AMS's Credit Management Systems Group, Only recently have we found statistically valid models for many other credit products, including instalment loans, mortgage loans, and auto loans. In addition, models can be fi ne-tuned once they are in operation. Judgmental vs. automated. Traditionalists may shake their heads and imagine some noble resistance to relinquishing their right to exercise time-tested judgmental evaluation powers. But in fact, banks are not looking for a black box that takes human faculties out of the loop. Rather, they are looking for reduced delinquency rates and higher levels of control of their credit policies, and more precise feedback on what is occurring within their operations. The extent to which lenders are locking in their decisions to the information provided by credit scoring methods varies widely. Anthony Scotto, vice-president for consumer credit at Chase Manhattan Bank, Garden City, New York, states that although Chase lenders . do not use credit scoring, for the last several months they have subscribed to TRW's Gold Report, an on-line enhancement of the TRW credit report. The report supplies a single number Derived Credit Risk Score that signifies risk of delinquency or bankruptcy in the next 8 to 12 months, and up to four single-letter codes identifying major factors determining the scoreWe call for that report on every account and use it as part of our judgmental evaluation. I think it is a helpful tool. If nothing else it certainly guides the analyst to a correct decision says Scotto. David Harris, vice-president, NCNB National Bank, Greensboro, N.C., reports that the changeover to greater reliance on credit scoring has been gradual. have been using credit scoring in some form for eight years, but I don't think very effectively People did too much second-guessing of the models. Recently we've been moving in the direction of relying more heavily on the score. We try to let the score be 80-90% of our decisions. Loan officers can over our the score, but they have to get approval from their supervisors to do it. Harris says that although some of the more experienced officers had some resistance, eventually they have come to see scoring as a helpful tool. One reason why is because NCNB has monitored the track record of override decisions, and in their instalment loan portfolio the 30-day-plus delinquency rate for overrides has been about seven times higher than for decisions based strictly on the cut-off levels. Developing a scorecard. Typically, according to AMS's Peter DiGiammarino, development of a credit scoring model costs between $50,000 and $100,000. …

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 19 citations exist upstream.