Who Wanted the Loan—and Who Got It?
For Chile's guaranteed-credit program, the crucial question is not simply how many firms borrowed. It is whether firms with different pre-pandemic repayment prospects entered the application pipeline at different rates, and whether banks filtered those applicants differently once they did. This chapter follows that chain from a modeled risk proxy to application, conditional approval, and final use. The distinction matters because a loan portfolio can contain more risky borrowers even when banks are actively rejecting a larger share of risky applications.
Risk was reconstructed from the information available before the shock
The study begins with a practical problem: “riskier firm” is not a fact that appears fully formed in the administrative data. The authors estimate a model of default using information from before the pandemic. A firm counts as having defaulted when a loan is more than 90 days past due. The model predicts that outcome using 2018 characteristics and then applies the estimated relationship to 2019 information to construct a 2020 ex ante default probability—the risk assessment that could have been formed before the pandemic's consequences were known.
The predictors combine the firm's real economic position with its financial history. They include net worth, value added per worker as a measure of productivity, age, wage bill as a measure of labor intensity, sales as a measure of size, outstanding debt, and the spread between the firm's loan rate and a risk-free rate. Industry and municipality effects account for differences associated with where the firm operates and what it produces. The point is not that any single variable identifies a future defaulter. The point is to approximate the information set that distinguishes a firm with stronger repayment capacity from one already carrying more financial risk. Each variable describes a different route to repayment. Net worth provides a buffer, productivity indicates how effectively the firm turns inputs into value, and age can capture accumulated operating experience. Sales describe scale rather than safety by themselves. Debt and loan spread bring in the financial side: a firm may have substantial activity and still face a heavier servicing burden or a lender's prior assessment of risk. The model combines these signals instead of treating any one of them as a verdict.
This also explains why industry and municipality effects are useful without being a complete explanation. Firms in the same sector can face similar demand conditions, and firms in the same municipality can share local disruptions or credit environments. Holding those differences constant makes the comparison less dependent on obvious location or sector composition. It does not eliminate unobserved differences among firms, nor does it reproduce the private information a bank may have from a long relationship with a borrower. The measure is a common research yardstick, not a claim that the researchers see everything a lender sees.
That timing gives the measure both value and a hard boundary. It is useful for asking whether risk visible before COVID-19 was related to participation in the new credit program. It is not a measure of the pandemic shock itself. A firm could have a sound pre-pandemic balance sheet and then lose sales because a lockdown prevents customers from reaching it. Another firm could be less exposed to the local operating restriction but already have high debt and a weak repayment record. Treating those two dimensions as one would confuse vulnerability to the shock with the probability of default. Consider the implication for a bank looking at two applications. The first firm may need cash because an external restriction has interrupted otherwise productive activity; the second may need cash because its finances were already strained. Both can report a fall in revenue, but the first fact points to a temporary operating shock while the second points to a repayment problem that the crisis may have intensified. A pre-pandemic risk proxy cannot settle the distinction on its own, and a pandemic exposure measure cannot replace it. The value of the study's design is that it keeps the two signals available for separate parts of the analysis.
The model performs reasonably well as a prediction exercise. In the main eligible-firm sample, the ex ante measure explains 75 percent of the variance in later ex post defaults. That is strong enough to support comparisons across risk groups, but it is not perfect knowledge. The pandemic brought uncertainty about its duration and severity that could not be learned from 2019 observations. In addition, firms without prior loans lack the financial variables—such as outstanding debt and loan spread—that make the risk estimate more precise. For those firms, the prediction uses real variables alone; their predicted 2020 default probability is 10.7 percent, roughly two percentage points higher than for firms with previous loans.
A useful reading rule follows. The study's risk measure answers, “How risky did this firm look using information available before the pandemic?” It does not answer, “How much did the pandemic damage this firm?” Nor does it observe the bank's private decision rule directly. It gives the researchers a common comparison across firms, after which applications and approvals reveal how demand and supply respond to that comparison.
The distinction also guards against hindsight. If later default is used to label a firm as risky, the analysis would allow the outcome to define the characteristic it is supposed to explain. By constructing risk before the 2020 program unfolded, the study can ask whether the program attracted firms that had looked riskier in advance. The later default data are then used as an outcome for checking the proxy, not as the original source of the selection claim.
The application and approval pipeline tells a different story
The financial data make it possible to split “obtaining a loan” into two decisions. First, an eligible firm decides whether to apply. This is the demand side: it reflects the firm's need for liquidity, the attractiveness of the subsidized terms, and its expectations about being approved. Second, conditional on receiving an application, the bank decides whether to approve it. This is the supply side: it reflects the bank's evaluation of repayment risk and the loss it would retain under the partial guarantee. Final use is the equilibrium result after both decisions, and it is not equivalent to either one alone.
The study finds that riskier firms were more likely to obtain a guaranteed loan among comparable eligible firms. Moving from the 25th to the 75th percentile of the risk distribution was associated with a 3.4-percentage-point increase in the likelihood of using the credit program, about a 7 percent increase relative to the average likelihood of use. The result is consistent with the incentive created by cheap guaranteed borrowing: firms that would face a higher private cost of credit have more reason to seek the subsidized loan.
The application data show that this pattern was not produced by banks blindly approving every request. About 36 percent of eligible firms applied. Of 79,319 applications, banks approved loans for 60,770 firms, a 77 percent approval rate. Conditional on applying, riskier firms were less likely to be approved. In other words, the risk gradient points in opposite directions at the two stages: risk raises the probability of applying but lowers the probability of approval once the application is in front of the bank. The two margins should not be confused. Entry into the program is an extensive-margin decision—whether a firm becomes a user at all. The amount borrowed after approval is an intensive-margin decision, which the later debt analysis examines. Even at this stage, the distinction matters because a high-risk firm can be more likely to cross the entry threshold without receiving the largest possible loan or the most favorable treatment from the bank. A count of users and a dollar allocation answer different questions.
Why, then, were riskier firms still more likely to use the program? Because the increase in demand was stronger than the reduction in approval probability. A risky firm may value a low-cost guaranteed loan enough to apply even when the bank is cautious. A safer firm may not apply because it does not need the liquidity, while the data do not make every non-application's reason observable. The users observed at the end are therefore selected by the difference between the firm's willingness to seek the loan and the bank's willingness to provide it. A high approval rate among applicants does not show that the whole eligible population faced equally favorable treatment; it describes only the pool that crossed the application threshold.
This sequence is more informative than a single participation statistic because it shows where selection occurs. If only final use were observed, a researcher could not tell whether risky firms borrowed more because they demanded more credit, because banks supplied it more readily, or because both forces operated together. Here, the application and approval records show that both forces mattered, with demand dominating in the observed equilibrium.
The result should not be turned into a universal claim that public guarantees always channel credit to risky borrowers. The estimates are conditional on the program's eligibility rules and on the industries, municipalities, and other characteristics included in the analysis. The program had already excluded some of the riskiest firms through its past-due-payment requirement. The findings describe how risk operated inside that gate, not what would happen if every firm in the economy—including firms already in serious default—could apply.
There is also a statistical selection problem in the approval stage. If the researchers compare only applicants, they are comparing a selected group that is already more inclined to seek credit. A riskier applicant may look very different from the lower-risk firms who never applied, but those non-applicants do not appear in the approval regression. The paper addresses this with a two-step Heckman correction. It first estimates the probability of applying using both applicants and non-applicants, then uses the resulting selection adjustment in the approval equation estimated among applicants.
That correction reduces the absolute risk coefficient in the approval equation from 26.4 percent to 12.3 percent. The direction of the relationship remains: banks are less likely to approve riskier applicants. But its magnitude is smaller once the composition of the applicant pool is taken into account. The uncorrected result overstated bank responsiveness because the bank was being evaluated on a group that was disproportionately risky to begin with. This is a useful lesson for interpreting lending data: conditioning on a decision can make the next decision appear more sensitive than it really is. The correction does not claim that selection into applying has been made random, and it does not reveal every reason a firm stayed away from the program. It adjusts the approval comparison for the modeled tendency to enter the applicant pool. The remaining coefficient is therefore best read as evidence of conditional risk sensitivity under the study's assumptions, not as a complete structural description of bank underwriting.
A second caution concerns what a rejection means. A bank can reject an application because expected default is too high, because the requested amount is too large for the firm's capacity, or because other information makes the loan unattractive under the program's terms. The observed data identify a relationship between the study's risk proxy and approval, but they do not turn every individual rejection into a publicly observable explanation. The strength of the result is at the pattern level: riskier applicants were less likely to pass the approval stage, while riskier eligible firms were more likely to seek the program.
The evidence consequently supports a layered conclusion. The credit program's low price made it attractive to riskier firms, and their demand helped expand the program's reach. Banks did not respond by treating risk as irrelevant; conditional approval fell with the ex ante risk measure. Yet screening was not strong enough to reverse the demand pattern, so riskier firms remained more likely to use guaranteed credit. The program was neither a pure demand story nor a pure supply story. It was a market outcome produced by their interaction.
The boundaries of the comparison matter just as much as the direction of the estimates. The risk proxy is most informative for firms with prior borrowing histories, while firms without prior loans are evaluated with fewer financial inputs. The sample is based on eligible formal firms, not an unrestricted census of all enterprises. The study's application and approval results are model-based comparisons rather than a randomized experiment in which firms were assigned different risk levels or banks were assigned different information. They tell us how observed risk and program decisions moved together under Chile's rules.
That leaves a separate question about the nature of the shock. Pre-pandemic risk helps explain who wanted credit and how banks screened applications, but it does not explain why a municipality entering lockdown would make employment protection more attractive. A firm can be financially risky, operationally disrupted, both, or neither. The next step is to examine how sales declines and staggered lockdowns separated those conditions—and why employment protection became a different kind of information in the credit decision.