Credit Scorecards: How Lenders Decide

Key takeaways

• Logistic regression predicts a probability, which is mapped to a score.
• Variables are binned and weighted by predictive power, not by intuition.
• Scorecards are validated and re-built as populations shift.
• A score is one input - policy rules and serviceability sit alongside it.

Why logistic regression?

Because the outcome being predicted is binary - the account defaults or it does not - and logistic regression models the probability of a binary outcome in a form that is stable, interpretable and defensible to a regulator.

Interpretability is doing a lot of work in that sentence. A lender must be able to explain why an application was declined, both to the applicant and to a supervisor. A model whose coefficients map cleanly to points on a scorecard can do that. A model that cannot explain itself creates a compliance problem regardless of how accurate it is.

That is the main reason more complex machine learning approaches have been adopted more slowly in credit decisioning than in areas like fraud detection, where explanation obligations are different.

How is a scorecard built?

The workflow is standardised across the industry.

1. Define the outcome - typically 90 days past due within a 12 or 24 month performance window.
2. Assemble a development sample of past applications with known outcomes.
3. Bin each variable into ranges and calculate weight of evidence for each bin - how strongly that band separates good accounts from bad.
4. Select variables using information value and correlation checks, dropping weak or redundant ones.
5. Fit the regression and convert coefficients into scorecard points.
6. Validate on a holdout sample and measure discrimination.
7. Set the cut-off - the score at which applications are declined, which is a business decision about risk appetite, not a statistical one.

Step seven is where two lenders with identical models produce different answers on the same application. The cut-off is appetite.

What typically goes into an application scorecard?

Predictive power, not fairness intuition, determines what earns a place - subject to legal limits on what may be used.

Note what is not there. Discrimination law constrains the use of protected attributes, and models are tested to check that permitted variables are not acting as proxies for prohibited ones. Enquiry frequency is the variable applicants most often damage without realising - many applications in a short window is itself a predictive signal.

Frequently asked questions

What is a credit scorecard?

A statistical model that converts application and credit file information into a score representing the probability of default. Points are assigned to bands of each variable based on how strongly they predicted default in historical data.

Why was I declined when my credit score is good?

A score is one input among several. Lenders also apply policy rules - minimum income, maximum LVR, acceptable loan purpose, employment type - and a separate serviceability assessment. Failing any of those can decline an application with a strong score.

Do lenders use AI to assess loan applications?

Some use machine learning for parts of the process, but credit decisioning has adopted it more cautiously than other areas because lenders must be able to explain decisions. Logistic regression scorecards remain widespread precisely because they are interpretable.

Does applying to several lenders hurt my score?

Multiple applications in a short period are recorded as enquiries on your credit file, and enquiry frequency is commonly a predictive variable in scorecards. This is a practical reason to have a broker assess your position before lodging applications.

Related reading

• Credit Scores in Australia: Equifax, Experian and illion
• PD, LGD and EAD: The Three Numbers That Price Your Loan
• Why Home Loan Applications Get Declined: The Top 12 Reasons

Sources

• Moneysmart - credit scores and credit reports — ASIC
• Privacy (Credit Reporting) Code — OAIC