Recommended Free Tools
Statistical modelling is the use of data and statistical methods to estimate a pattern or an outcome. In the United States, its effects reach consumers and businesses most visibly through credit decisions. A credit score is one example: it is a model output that estimates how a person is likely to handle credit, and lenders may use it when deciding whether to offer credit and on what terms. Credit scoring illustrates how a model can shape real decisions, but it is only one application. This article uses it as a concrete case and does not treat it as the whole of statistical modelling in the economy.
What statistical modelling means
A statistical model is a structured way of using past data to estimate something that is not yet known, or to describe a pattern that appears across many cases. The model takes inputs, applies a set of statistical relationships, and produces an estimate. The estimate is then used by people or systems to make a decision.
As an Amazon Associate I earn from qualifying purchases.
Two distinctions matter for anyone reading a model’s output. The first is between prediction and certainty. A model estimates likelihoods across groups of cases and cannot say with certainty what a particular person will do. The second is between a model’s output and the decision made with it. A score or estimate is one input to a business choice, and the choice also depends on policy, pricing, and the lender’s own rules.
What is a credit score?
A credit score is the output of a scoring model that uses information from a credit report to predict credit behaviour, such as the likelihood of repaying debt on time. The Consumer Financial Protection Bureau (CFPB) describes it in plain terms in its consumer guide, “What is a credit score?” (page last reviewed September 2, 2026). According to that guide, most credit scores range from 300 to 850. That is a broad statement about most scores, not a rule that applies to every scoring model or scale.
#1 Best Overall
Credit scores are used beyond lending. The CFPB identifies tenant screening and insurance as other uses of credit scores, alongside decisions about whether to extend credit and what terms to offer.
A consumer may see different numbers from different sources, and this is expected. The CFPB notes that a score can change with the model used, the data available, the loan product being considered, and the date of calculation. A single number should therefore be read as one model’s estimate at one point in time.
Rank #2
How the methods differ
Scoring and other credit models are built with different statistical techniques. A 2017 CFPB Request for Information on the use of alternative data and modelling techniques in the credit process, published in February 2017, groups the techniques discussed into two broad categories. The examples below come from that document. They are illustrations of the categories, not a complete list, and the document does not rank them by quality.
Traditional approaches
- Linear regression estimates a numerical outcome as a weighted combination of inputs.
- Logistic regression estimates the probability of a yes-or-no outcome, such as whether a loan will be repaid as agreed.
Alternative techniques
- Decision trees split the data through a sequence of yes-or-no tests.
- Random forests combine many decision trees into one estimate.
- Neural networks learn weighted connections between layers of inputs and outputs.
- Boosting builds a sequence of simple models, each correcting the errors of the one before.
More complex techniques are not automatically more accurate. Whether one method performs better than another depends on the task, the data, and how the results hold up on data the model has not seen. Those questions are covered in the sections on validation below.
Rank #3
What it means for consumers
A score can influence whether a person receives credit and what terms they are offered, including price. Because a score is an estimate, an unfavourable result is not a verdict on character. It reflects how the model weighed the information it had.
What an adverse action notice should contain
When a creditor takes adverse action on an application, such as denying it or offering less favourable terms than requested, the CFPB says the creditor must give accurate, specific reasons. This applies even when the decision was made with a complex algorithm. In Consumer Financial Protection Circular 2022-03, the CFPB states: “Whether a creditor is using a sophisticated machine learning algorithm or more conventional methods to evaluate an application, the legal requirement is the same: Creditors must be able to provide applicants against whom adverse action is taken with an accurate statement of reasons.” The circular is available on the CFPB site at Circular 2022-03.
A consumer who receives an adverse action notice can use the stated reasons to understand what drove the decision. The reasons should be specific enough to explain the outcome. A complex model does not reduce the creditor’s duty to explain it.
Free tools Windows power users keep installed
One-click scans. No signup required.
What it means for businesses
For a business, a model makes it possible to rank or evaluate large numbers of applications consistently. That efficiency comes with obligations, particularly in credit, where the rules are explicit.
Best Value
Soundness requirements for credit scoring systems
Regulation B, which implements the Equal Credit Opportunity Act, defines an empirically derived, demonstrably and statistically sound credit scoring system. Under the definition in 12 CFR § 1002.2, such a system is built on relevant empirical data and accepted statistical principles and methodology, and it is validated and periodically revalidated. Statistical soundness is therefore not established simply by choosing a sophisticated method. It depends on the data, the methodology, and the testing behind the system.
Validation and revalidation
The CFPB’s interpretive guidance says creditors are responsible for ensuring that validation and revalidation are carried out using their own data. The regulation does not set a single fixed revalidation interval that applies to every lender. Monitoring is an ongoing responsibility, and the schedule should match how quickly the lender’s data and borrowers change.
Explaining decisions
A business must be able to explain its decisions in the terms required for adverse action notices. If a model is too opaque to produce accurate, specific reasons, it creates a compliance problem regardless of its predictive strength. Businesses should test this before deployment rather than after a complaint.
How to judge a model in practice
When two or more models are compared, the question is not which one is universally best. The following axes give a more useful basis for comparison. The cited sources support validation and explanation as relevant considerations, but they do not supply comparative performance results for particular models.
Quick Recap
| Axis | Question to ask | Why it matters |
|---|---|---|
| Prediction task and target outcome | What exactly is the model estimating, and for which decision? | A model built for one outcome may not serve another. |
| Data used and its relevance | Is the input data relevant to the population and decision? | Regulation B requires relevant empirical data for credit scoring systems. |
| Predictive performance | How does the model perform on appropriate validation data? | Performance on training data alone does not show how the model will behave on new cases. |
| Stability over time | Does performance hold as conditions change? | Revalidation with the lender’s own data is the way to check this. |
| Explainability | Can the business give accurate, specific reasons for each decision? | Adverse action rules apply regardless of how complex the model is. |
Limits of the credit example
- Credit scoring is a useful illustration, but not all statistical models are credit models. Modelling is used across many sectors, and the rules discussed here apply to credit decisions.
- Not every business use of a model is covered by the cited credit rules. Whether a specific rule applies depends on the decision and the parties involved.
- A score is not fixed or universal. Different models, data, products, and dates can produce different results for the same person.
- A model’s estimate is a probability, not a certainty about any individual.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
