Credit scorecards are mathematical models which attempt to provide a quantitative estimate of the probability that a customer will display a defined behavior (e.g. loan default, bankruptcy or a lower level of delinquency) with respect to their current or proposed credit position with a lender. Scorecards are built and optimized to evaluate the credit file of a homogeneous population (e.g. files with delinquencies, files that are very young, files that have very little information). Most empirically derived credit scoring systems have between 10 and 20 variables. Application scores tend to be dominated by credit bureau data which typically amounts to over 80% of the predictive power from closer to 60% in the late 1980s for UK scorecards. Indeed there has been an increasing trend to minimize applicant or non-verifiable variables from scorecards which has increased the contribution of the credit bureau data.

Credit scoring typically uses observations or data from clients who defaulted on their loans plus observations on a large number of clients who have not defaulted. Statistically, estimation techniques such as logistic regression

Logistic regression

In statistics, logistic regression is used for prediction of the probability of occurrence of an event by fitting data to a logit function logistic curve. It is a generalized linear model used for binomial regression...

 or probit

Probit

In probability theory and statistics, the probit function is the inverse cumulative distribution function , or quantile function associated with the standard normal distribution...

are used to create estimates of the probability of default for observations based on this historical data. This model can be used to predict probability of default for new clients using the same observation characteristics (e.g. age, income, house owner). The default probabilities are then scaled to a "credit score." This score ranks clients by riskiness without explicitly identifying their probability of default.

There are a number of credit scoring techniques such as: hazard rate modeling, reduced form credit models, weight of evidence models, linear or logistic regression. The primary differences involve the assumptions required about the explanatory variables and the ability to model continuous versus binary outcomes. Some of these techniques are superior to others in directly estimating the probability of default. Despite much research from academics and industry, no single technique has been proven superior for predicting default in all circumstances.