Supervised learning

A machine learning approach where the model is trained on labeled data — pairs of inputs and known correct outputs — to learn a mapping function that can predict outputs for new, unseen inputs.

Source: AIEOG AI Lexicon (Feb 2026), adapted from NIST AI 100-1

Supervised learning is the most common type of machine learning. It works by showing the model many examples where you already know the answer, so it can learn the pattern and apply it to new cases.

For example, to build a fraud detection model, you would provide thousands of historical transactions labeled as either “fraud” or “legitimate.” The model analyzes these examples, identifies patterns that distinguish fraud from legitimate activity, and then applies those patterns to score new transactions.

Supervised learning requires labeled training data — datasets where the correct answer for each example is known. This is both its strength (clear learning signal) and its limitation (labeling is expensive and labels can be biased or incorrect).

In financial services, supervised learning is used for credit scoring, fraud detection, customer churn prediction, claim prediction, and many other applications where historical labeled data is available.

Supervised learning is the workhorse of AI in financial services, but it carries governance considerations that compliance teams must address:

• Label quality. The model learns from the labels. If labels are inaccurate, biased, or inconsistent, the model will learn incorrect patterns.
• Historical bias. Labels derived from historical human decisions may embed past biases. A credit model trained on historical approvals and denials will learn any bias present in those decisions.
• Class imbalance. Many financial services problems (fraud, default) involve rare events. Supervised learning with imbalanced data requires specialized techniques.
• Generalization. A model that performs well on historical data may not generalize to future conditions.

1. Audit label quality. Review the labeling process for accuracy, consistency, and potential bias.
2. Assess for historical bias. Evaluate whether historical labels embed discriminatory patterns that the model will learn.
3. Validate on representative data. Test model performance on data that is representative of the population it will be applied to.
4. Address class imbalance. For rare-event problems, ensure the modeling approach appropriately handles imbalanced data.
5. Document the training process. Record the training data, labeling methodology, feature selection, and model architecture.
6. Monitor post-deployment. Track model accuracy on new data to detect performance degradation.

Machine learning, Unsupervised learning, Semi-supervised learning, Training data, Algorithm, Deep learning, Data quality/validity