How A Scorecard Is Built From Past Borrowers

A credit score is the output of a model built by observing what previous borrowers did. Understanding how that model is constructed explains most of what people find arbitrary about scoring.

The model starts with an outcome definition

Before anything is measured, the builder defines what counts as bad: typically a stated level of arrears reached within a set observation window after the account opens.

That definition is a choice. Moving the threshold or the window changes which borrowers count as bad and therefore changes which characteristics look predictive.

Every statement a scorecard makes is relative to its outcome definition. A model built to predict serious arrears is not predicting whether someone is careful with money.

Characteristics are chosen for separation

The builder examines historic files and looks for variables where the good and bad populations separate: recency of missed payments, proportion of limits used, number of recent searches.

Variables that separate strongly and remain stable over time are kept. Variables that separate weakly, or that separate only in one period, are discarded regardless of intuition.

This is why some things people expect to matter have no weight at all. They may be true of the borrower without distinguishing the two outcome groups.

Weights come from the historic population

Each retained characteristic is given a weight derived from how strongly it separated outcomes in the development sample. Points are summed to produce the score.

The weights encode the behaviour of a particular population over a particular period. Applied to a different population or a different economy, they degrade.

Lenders therefore monitor models for drift and rebuild them periodically, which is one reason a score can move when the borrower has done nothing.

The score is a rank order of risk

The output is not a percentage or a grade. It orders applicants by expected likelihood of the defined bad outcome, and the lender chooses where to cut the order.

Two lenders using the same model can accept different people, because the cut-off reflects appetite and margin rather than anything inside the score itself.

A score alone therefore never determines an outcome. It determines a position, and policy determines what happens to that position.

Models cannot see what they were not shown

A scorecard is built on accepted applicants, because only they generated outcomes. Those declined have no recorded behaviour, which biases the sample in a known direction.

Techniques exist to compensate for that gap, but the underlying limitation remains: the model learns about people the lender was already willing to lend to.

Regulatory expectations about explainability, permissible variables and testing for unfair outcomes differ between jurisdictions and have been tightening in several of them.