Credit Scoring

Credit scoring produces a score, usually between 0 and 1,000, that estimates the probability that a borrower will meet their payments. It turns scattered data (payment history, outstanding debts, length of credit history, declared income, transaction behavior) into a single, comparable number that ranks applicants from lower to higher risk. The higher the score, the lower the perceived risk of default.

In practice, a company uses it to answer three questions in seconds: do I extend credit or not?, up to what amount? and at what rate or term?. The score replaces the subjective judgment of "this person seems trustworthy" with a measurable, auditable model, fed by data that in Latin America comes from credit bureaus such as Veraz/Equifax, Nosis or the BCRA (the central bank) in Argentina, plus the company's own internal information.

Building and monitoring a scoring model is a classic case of predictive analytics on data: you have to clean the sources, calculate variables, validate the model and watch that it does not degrade over time, work that relies on a data platform like Metrix.

How a scoring model works

A credit score comes from a statistical or machine learning model trained on thousands of historical cases where it is already known who paid and who did not. The model learns which combinations of variables anticipate default (delinquency) and assigns a weight to each one. The most common variables are payment history, the level of current indebtedness, the length of their credit history, the number of recent bureau inquiries and, increasingly, alternative data such as purchasing behavior or the flow of collections and payments.

The result is a probability of default that is translated into a scaled score. On top of that score, the risk team defines cutoffs: for example, approve automatically above a certain value, reject below another, and send the gray zone in between to manual review. That way, scoring does not just predict: it operationalizes the decision and makes it consistent across branches, sales reps and channels.

Why it matters to the business

Scoring resolves a real tension: selling more without the portfolio becoming uncollectible. Without a model, companies tend to be too conservative (they lose sales to good payers) or too lenient (they pile up delinquency and losses). A good score lets you calibrate risk appetite precisely and move the lever according to the context: in an inflationary economy like Argentina's, tighten cutoffs in months of stress and loosen them when risk goes down.

It also automates decisions that used to take days. A lender that evaluates each application by hand can move to instant approvals for most cases, leaving human analysis only for the borderline ones. That lowers operating costs and improves the customer experience, since the customer gets an answer on the spot.

A concrete example (Latin America)

A home appliance company in Argentina sells on installments through the traditional channel. Before, the sales rep approved customers at their own discretion based on their impression of the customer, with delinquency hovering around 12%. By implementing a scoring model that cross-checks Veraz, the customer's status with the BCRA and the internal purchase history, the company defines three cutoffs: automatic approval for high scores, rejection for very low ones and referral to an analyst for the middle band. Within six months, it approves a similar volume but with lower delinquency, because it rejects the profiles that used to slip through and approves good payers it used to turn down out of prejudice. The key point: the model is transparent and reviewable, not a hunch.

Common mistakes

• Treating the score as absolute truth: it is a probability, not a certainty; there will always be good payers with low scores and vice versa.
• Not monitoring model degradation: payment behavior changes with the macroeconomy, and a model that worked well two years ago may be predicting poorly now.
• Using poor-quality data or outdated data: if the income or debt data is dirty, the score inherits that error.
• Applying fixed cutoffs without reviewing them: the optimal threshold moves with the cost of delinquency and the margin of each product.
• Ignoring bias and regulatory compliance: sensitive variables or improper proxies can lead to discriminatory decisions and legal problems.

Credit scoring vs lead scoring

Credit scoring is often confused with lead scoring in marketing and sales, because both produce a score. But they answer different questions and draw on different data.

In short, credit scoring is one of the most mature applications of predictive analytics in the financial and commercial world: it turns risk into an actionable number and lets you scale credit decisions without losing control of the portfolio.