Closure is not default: small-business credit scoring and the use of market-area information

Sole proprietors make up about 87% of Korea's small and medium-sized enterprises, yet in credit scoring they are judged on the least information. Two boxes are empty: past repayment history and collateral. It is not that information is missing; rather, the items conventional scoring looks at do not match the records these businesses actually have.

For a long time, closure data has filled that gap. Industries with frequent closures have been classified as risky, and areas with frequent closures as markets to avoid. This essay re-examines the starting point of that practice.

1Closure and default are not the same event

An owner who recovers key money (premium for the lease) and retires, an owner who fails to renew a lease and moves to a nearby location, and an owner who shuts down because rent and labor costs became unmanageable all show up identically in public data as one line: "store count decreased." Through a lender's eyes, however, the first is a wind-down after repayment is complete, the second is a relocation while repayment continues, and only the third is close to repayment failure.

The more this distinction is skipped, the more the burden falls first on businesses with thin credit histories. Rejecting conservatively loses borrowers who would have repaid; approving loosely takes on bad loans. Yet the first loss appears in no statistic, while only the second shows up on the income statement. That is why practical incentives are structurally tilted to one side.

2What we looked at: the location, not the borrower

We shifted the unit of observation from the individual business to "market area × industry × quarter." This looks at the market conditions a borrower faces without probing further into that borrower's private life.

3Finding ①: the more clustered, the fewer closures

The indicator with the most explanatory power was not an individual store's sales but the number of same-industry stores in the same market area (log). With a standardized coefficient of −0.56, market areas with more same-industry stores showed fewer business-exit signals. This runs against the common belief that fiercer competition means more risk.

When the same industry clusters, customers choose the market area itself as their destination rather than a single store. Lease, labor and supply markets form around it too, so when one store leaves, it is replaced quickly. Conversely, where an industry is scattered, one store's exit means that demand disappears.

4Finding ②: volatility of sales matters before the level

The second factor was not the level of sales but its volatility (standardized coefficient of the coefficient of variation: +0.37). Closures appeared earlier in market areas where sales jump sharply in some quarters and leave gaps in others than in market areas where sales stay consistently low.

This parallels corporate credit-risk research, where the early warning of default was not the level of the share price but its volatility (Campbell et al., 2008). A business with steady sales can set rent and payroll to that level, but a business with large swings ends up covering the off-season with peak-season cash, a structure that is vulnerable to a single gap.

5Finding ③: industries that move together are not necessarily riskier

When closures rise in one market area, similar events sometimes follow in neighboring areas. Most of the difference across industries, however, comes not from the strength of propagation but from how many market areas the industry is spread across. Set against normal-period levels, the excess is only between −0.02 and 0.14 for all six industries.

† Industries where the normal-period mean falls within the 95% confidence interval of the co-movement index, so the excess is not statistically distinguishable. Signal rates are based on the estimation sample (117,138 observations, Q1 2020 – Q4 2025). Source: Seoul Open Data Plaza, Seoul Commercial District Analysis Service; authors' calculations.

An industry with observed co-movement is not necessarily higher risk either. Korean restaurants, with the highest index, have a signal rate of 3.5%, the lowest of the six, while pubs and bars, with the highest signal rate (12.2%), sit mid-range on the index. The size of the index does not line up with the order of risk.

It is still useful in practice because it tells you which way to check. If the same change shows up across neighboring market areas, check local conditions before the business owner's capabilities. If it shows up in only one place, check individual circumstances such as a store relocation or a lease ending.

6Combining the three indicators

Using store count, sales volatility and co-movement level together, we can to some extent flag in advance which market areas will show business-exit signals in the next quarter. When comparing randomly drawn market areas with and without a signal, the model picked the higher-risk one about three times out of four (out-of-sample AUC 0.749). Not a single line of financial information was used.

7Why this distinction is inclusive finance

Observed closures mix five different things. Public aggregate data cannot tell them apart.

A model that counts voluntary exits as defaults estimates the probability of default (PD) higher than it really is. The consequences do not stop at statistical error. Businesses that would not have defaulted are classified as defaults and either rejected or charged higher interest. A more refined estimate ends up working against the policy goal.

8What it can and cannot do

The boundary is clear. Public market-area indicators are not an individual's financial history, and they cannot be the basis for approving or rejecting an individual, or for setting rates and limits. That basis must come from verified contract and repayment records and the business owner's own explanation.

Market-area data belongs one step earlier. It is used to decide which of the records an underwriter already has to check first, and what to ask the customer. As a working procedure there are four steps: check the change in the market area, review the related existing records again, ask the business owner if something remains unexplained, and record both what was confirmed and what was not.

• Credit card companies: card loans and merchant credit for sole proprietors and small offices/home offices (SOHO). Check whether the payment terminal was replaced, when the lease expires, and whether the next month's card sales recover.
• Leasing and installment finance companies: equipment and commercial-vehicle credit for the self-employed. Check whether the equipment is actually in operation, whether delivery and sales timing match, and whether contracts continue after a relocation.
• New technology business finance companies: investment screening for local brands and franchises. Check for store relocations and business suspensions, changes in the franchise mix, and how funds are being used.

It matters that each sector checks different records. Erase those differences and merge everything into a single score, and the signal starts to replace judgment instead of supporting it.

Explainability is also a requirement. Korea's Credit Information Use and Protection Act gives individuals the right to request an explanation of automated assessment results and to object to them, and in the United States, ECOA and Regulation B require specific reasons for adverse actions. This requirement is part of why the essay kept to a simple form of estimation: which indicator acted in which direction, and by how much, becomes the statement of reasons as it stands.

9Remaining work: what has not yet been confirmed

The analysis stops at one point. Public market-area indicators show conditions in a market area, but they cannot show which businesses there actually repaid. That fact exists only in member companies' delinquency records. Only when the two datasets are combined can we confirm how many observed closures actually led to repayment failure.

The essay also set out the design conditions such a combined validation would require: limit the purpose to confirming supplementary effectiveness as an early warning; restrict the data to the minimum items needed for matching; measure separately whether the signal merely repeats what existing information already showed or actually changed the order of checks; stop if there is no added explanatory power or if disadvantages to a particular group recur; and have reviewers outside the participating companies verify the design and its interpretation.

Wrapping up: the room a distinction opens

Market areas with more same-industry stores show fewer business-exit signals. Volatility of sales signals risk ahead of its size. Industries that move with neighboring market areas are not necessarily vulnerable. None of these appear in underwriting documents, yet together they explain much of the conditions a business must cope with.

What this data cannot do is also clear. We can observe that a store disappeared, but not whether that person failed to repay a loan. If outcomes cannot be confirmed, or if it works against a particular group, stopping its use and recording that fact is also a result.

"Closure is not default. As long as this distinction holds, market-area information works not as a means of passing verdicts on borrowers but as data for understanding circumstances not yet confirmed."