How Credit Scoring Models Work

Credit scoring translates a customer's risk into a number. The model looks at characteristics historically associated with non-payment and estimates the probability that a new customer behaves the same way. It is neither fortune telling nor absolute truth: it is applied statistics with a known margin of error.

Understanding how the number is produced is what lets you use a score with judgment instead of as a rubber stamp.

What the model actually estimates

Every score estimates a probability of default over a defined horizon, usually 12 months. A score corresponding to 5% probability means that out of 100 customers with that profile, roughly 5 will stop paying.

Two practical implications tend to get forgotten:

• A high score does not guarantee payment; it guarantees lower probability
• A low score is not a certainty of loss; it is higher risk, which can be offset with a smaller limit, a deposit or security

Variables that go into the model

Market models combine groups of information:

• Firmographic — years in business, size, industry, region
• Derogatory — judgments, liens, collections, amount and recency
• Behavioral — trade payment history, credit utilization, recent inquiries
• Financial — revenue and ratios, for larger companies

Recency usually outweighs amount: a $3,000 collection item from last month says more about current risk than an $80,000 judgment from four years ago.

Bureau score versus internal score

The combination is what works: bureau data for new customers, internal score for limit reviews on existing accounts. That second reading is behavior scoring, covered in behavior scoring for limit reviews.

The cutoff is your decision, not the model's

The model outputs probability; the company decides where to cut. That choice is economic, not statistical:

• Higher cutoff: less loss, fewer sales
• Lower cutoff: more sales, more loss

The optimum is where the margin generated by the last approved sale still exceeds its expected loss. That is why a business at 60% margin can approve risk that a business at 8% has to decline.

How to tell whether the model still works

A score has to be monitored, not installed and forgotten. Three basic checks:

1. Rank ordering — do lower-scored customers default more than higher-scored ones? If the ordering breaks down, the model has lost power.
2. Stability — has the score distribution of your portfolio shifted sharply? A sudden change means a different customer mix or a data source problem.
3. Calibration — does observed default match predicted default in each band?

These measures come naturally out of vintage analysis, which compares how cohorts approved in different periods perform.

Limits of scoring

• It does not catch well-built fraud: clean documents produce clean scores
• It depends on history; a new company has thin data
• It does not know commercial context (signed contract, long relationship, real security)
• It degrades over time without recalibration

What to take from this

A score is an input to the decision, not the decision. Use bureau data for who arrives and internal behavior for who stays, set the cutoff from your own margin, and monitor rank ordering and calibration — an unmonitored model ages and starts being wrong quietly.