Does a Credit Score Predict Repayment, or Decide Who Gets the Chance to Prove Themselves?

Geoffrey Chen

When Numbers Start Making Decisions · Season One, “People on Either Side of a Threshold” · Article 6

1. What did the system know when it said no?

Imagine that someone applies for a low-limit credit card. She has a stable income, no bankruptcy and believes she has always paid on time. The application is quickly refused. The notice says only that the decision was based partly on credit reporting information and identifies the organisation that supplied the report.

When she obtains that report, she might discover three very different situations.

First, it wrongly includes a default belonging to someone else. Second, it is broadly accurate, but several late payments and recent applications have reduced a risk score. Third, she has almost no borrowing history, so the system lacks evidence that she will repay on time.

All three can produce the same result on the screen: declined.

But they represent different problems. The first is inaccurate data. The second is a prediction of the future from the past. The third is a lack of evidence caused partly by a lack of prior opportunity.

Credit scoring initially appears to do no more than estimate repayment risk. Once lenders use it to decide who receives credit, however, it is no longer merely descriptive. It helps create the reality that the next prediction will observe. A person who receives credit can build a repayment record. A person who is refused loses an opportunity to prove that they would have paid.

2. Australia has no single “official credit score”

An Australian consumer credit report can include identity information, credit accounts, credit limits, repayment history, credit enquiries, overdue accounts and defaults. A credit reporting body can derive a score or other information from those records. A lender can also use its own internal model, income information and product rules.

Different credit reporting bodies may hold different information. The Office of the Australian Information Commissioner therefore explains that a person may need to obtain a report from more than one body. OAIC: “Access your credit report”

A score displayed by one service is not a single truth adopted by the whole financial system. Organisations can use different scales, weights, time periods and definitions of risk. The same person can receive different scores in different systems.

More importantly, a credit score normally predicts the probability of some future adverse event. It does not measure honesty, human worth or complete financial capacity. It combines past account behaviour with historical outcomes among people with similar data to estimate future risk.

“Higher risk” is therefore a statement about probability, not a fact that a default has already occurred.

3. A lender cannot ask only whether the score is high enough

Consumer credit decisions do not concern risk pricing alone. Australia’s responsible-lending obligations require relevant credit licensees to make reasonable inquiries about a consumer’s financial situation, requirements and objectives, take reasonable steps to verify the financial situation, and assess whether a contract is not unsuitable. A high score does not permit a lender to ignore whether the loan would create substantial hardship. Risk scoring and affordability are not the same question. ASIC: “Responsible lending”

At least three judgements are involved:

  1. How has this person used credit in the past?
  2. How likely are people with similar records to default in the future?
  3. Is this particular loan suitable for this person’s income, expenditure, needs and objectives?

Credit reporting information contributes mainly to the first two. A responsible-lending assessment addresses the third. The final decision may also include fraud controls, security, product policy and the lender’s commercial appetite for risk.

If the only explanation is “you did not meet our criteria”, an applicant cannot tell whether the problem was inaccurate information, a predictive model, an affordability assessment or a commercial threshold. The faster a number delivers the outcome, the easier it becomes to compress the explanation.

4. How does a prediction become a feedback loop?

Using historical data to predict future outcomes is a reasonable starting point. Consistent repayment is relevant evidence of lower risk. Serious arrears should also be available to a lender considering whether to advance money.

But a past record is not produced by personal choice alone. It can reflect unstable income, illness, family violence, financial hardship arrangements, changes of residence and whether a person was previously able to obtain credit at all.

A young person or recent migrant with a thin file may not be risky; the system may simply know less about them. If a model treats “little evidence” as adverse evidence, that person will find it harder to obtain a first product. Without the product, they cannot generate the positive history the model wants to see.

Another loop can arise from enquiries. After one refusal, an applicant may submit further applications. New enquiries can enter the report, and another lender may interpret a cluster of recent applications as evidence of financial stress. The first refusal has then changed the information visible at the next decision.

This does not establish that every scoring model creates an unfair cycle. It establishes a structural point: once a prediction controls access to opportunity, it no longer sits outside the world it measures. Its decisions affect who can produce the next generation of training and assessment data.

5. What can separate people close to the threshold?

Lenders do not ordinarily publish their full models or exact cut-offs. Even if a threshold were known, two applicants separated by one score point might differ across many variables. A model may combine dozens of inputs before its output is considered alongside income, debt ratios and product policies.

Two forms of uncertainty therefore exist near a credit threshold:

  • measurement uncertainty: is the information accurate, complete and current?
  • inferential uncertainty: how well does this information predict the future of this individual?

If a default belongs to somebody else, a precisely calculated score is worthless. Even when the underlying information is correct, a model can be wrong about an individual because a probability never guarantees an individual outcome.

Part IIIA of the Privacy Act 1988 requires a credit provider to take reasonable steps to ensure that credit eligibility information is accurate, up to date, complete and relevant for the purpose for which it is used. If an application is refused wholly or partly because of relevant information supplied by a credit reporting body, the provider must give written notice within a reasonable period, state that fact and provide the body’s contact details. Federal Register of Legislation: “Privacy Act 1988, Part IIIA”

The notice does not require disclosure of an entire commercial model. It should at least tell the applicant where to begin checking the information used against them.

6. What happens when the number is wrong?

The Australian credit-reporting framework provides several specific data rights.

A credit reporting body must provide a consumer credit report free of charge once every three months. A free report is also available if a person has been refused credit within the previous 90 days or if credit-related personal information has recently been corrected. OAIC: “Access your credit report”

If a person finds an error, they can ask the relevant lender or credit reporting body to correct it. The OAIC describes this as a “no wrong door” framework: an organisation that receives a request but cannot resolve it alone should work with the other relevant participants. When information is found to be wrong, reasonable steps should ordinarily be taken to correct it within 30 days, or a longer period agreed with the person, and the person should be notified. A refusal to correct must be explained in writing, together with the available external dispute-resolution or complaint options. OAIC: “Correct your credit report”

Correction solves only the first layer. A fully accurate report can still lead to a refusal. The applicant may then need to establish whether the decision arose from affordability, internal risk policy or mishandling, and use the lender’s internal dispute-resolution process, an applicable external dispute-resolution scheme or a regulatory complaint path.

Meaningful review cannot consist solely of sending the same information through the same model again. Someone must have authority to examine the source data, the applicable rules and evidence about the individual, and to distinguish “the model predicts higher risk” from “law or product policy requires refusal”. The OAIC also explains the process for making a credit-reporting complaint. OAIC: “Make a credit reporting complaint”

Conclusion: credit scoring predicts opportunity and allocates the chance to build a record

Financial institutions cannot lend without judging risk. Ignoring repayment history would shift losses to other borrowers, depositors and the wider community. Responsible lending also requires lenders not to supply an unsuitable product to someone unable to afford it.

The question is not whether prediction should exist. It is how much decision-making authority a prediction should receive.

A credit score can be useful evidence of risk, but it should not be mistaken for a person’s financial character. It does not directly measure every pressure on income, every future change or the purpose for which credit will be used. It can also compress a thin file, earlier disadvantage and a genuine history of default into superficially similar outcomes.

The central judgement is this:

When a credit score decides who receives credit, it also decides who gets the chance to create the next good credit record.

Because prediction helps shape the future, applicants must be able to identify the source of the data, understand how the refusal relates to credit reporting information, correct mistakes and obtain meaningful human review where a score cannot resolve the individual case.

Numbers can help lenders manage risk. They must not turn “the system predicts that you may fail” into “you will never be allowed to prove the system wrong”.

Primary sources

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.