Supervised Learning
Supervised learning is the form of machine learning in which a model is trained on examples that carry the correct answer — transactions labelled fraud or not, applications labelled defaulted or repaid, documents labelled by type — so that it learns the mapping from inputs to labels and can predict the label for new inputs. Classification and regression are its two tasks; the labels are its cost and its limit, since the model learns exactly what they say.
- Also known as
- Supervised machine learning, Labelled training, Classification and regression
- Used at
- fraud detection, credit scoring, chargeback prediction, document classification, churn, pricing, OCR — nearly every predictive model in a financial company; also the fine-tuning of language models on labelled examples
- Binding
- the labels and features are where regulatory obligations attach — a credit model's labels must reflect the outcome the regulator recognises, the features must not proxy protected characteristics, and the EU AI Act's data-governance duties for high-risk systems cover both
- Typical algorithms
- logistic regression, gradient-boosted trees (the workhorse for tabular data — XGBoost, LightGBM, CatBoost), random forests, neural networks, support vector machines; the choice depends on data type, size, and the explainability required
- Typical cost
- the labels — outcomes that take months to observe (a loan default), human judgements that take analysts' time (a fraud confirmation), or annotation that is bought; label quality bounds model quality
- Typical pitfalls
- labels that encode past decisions rather than truth (a credit model trained only on approved loans learns nothing about the declined — selection bias); labels that arrive late and change (a fraud confirmed a year later); leakage of the label into the features
- Parties
- data scientists, the operations and analyst teams who produce the labels, and the reviewers who ask where the labels came from
How supervised learning works in practice
A fintech has three years of card transactions, 1.2 billion rows, and for each a label: confirmed fraud (from chargebacks and investigations), or not. It builds features — amount, merchant, time, device, velocity, history — and trains a gradient-boosted model to predict the label from the features, minimising a loss on the training years and validating on a held-out period. The model learns that certain combinations predict fraud with a probability; at inference it scores each new transaction in milliseconds. Everything the model knows came from the labels: a fraud type that was never confirmed (because nobody investigated it) is invisible to the model, and a label that was wrong (a genuine dispute recorded as fraud) taught it the wrong pattern.
The label problem is the central one. Credit models see outcomes only for the loans that were approved — the declined applicants never defaulted or repaid in the data — so the model learns the approved population and must be corrected (reject inference) or it will be over-confident on the applicants it never saw. Fraud labels arrive with delay and are incomplete; a model trained on last month's labels misses last month's undetected fraud. Human labels drift with the analysts' policies. The engineering responses — label pipelines with documented sources, delays handled explicitly, audits of label quality, out-of-time validation — are as much of the work as the algorithm, which is usually a gradient-boosted tree that anyone can run.
For a fintech, supervised learning is the technology behind its decisions, and the labels are the asset: a competitor can copy the algorithm, not three years of confirmed outcomes.
Example
A lender builds a default model on 80,000 approved loans, labelled default or repaid at 12 months. Gini on the approved population: 0.60. Applied to all applicants, the model would approve 30% of those the old policy declined. The risk team runs reject inference — estimating outcomes for the historically declined using a mix of bureau performance on their other credit and the model's own scores — and retrains on the augmented set; the model's confidence on the previously declined segment falls, and the projected approval increase drops to 12%. A pilot on 3,000 applications in that segment observes a 12-month default rate of 9.1% against the model's 8.4% forecast — close enough to expand. The label pipeline, the reject-inference method and the pilot outcome go into the model card.
Common mistakes
- Training on labels that reflect past decisions rather than outcomes, and deploying on a population the labels never covered. - Ignoring label delay; last month's fraud labels are incomplete. - No audit of label quality; a wrong label is a wrong lesson. - Features that leak the label and produce impossible development scores. - Choosing a neural network for tabular data where gradient boosting is more accurate and easier to explain.