Machine learning model

What it is

A machine learning (ML) model learns patterns from examples instead of following rules written by hand. In supervised learning, the model sees past cases with a known outcome (a loan that defaulted, a transaction confirmed as fraud) and adjusts its parameters to predict that outcome from input features. Common techniques in finance include logistic regression, decision trees, gradient boosting and neural networks. Unsupervised learning finds structure without labels, for example clustering clients by behaviour.

The model lifecycle has clear stages: define the problem, prepare data, train, validate on data the model has not seen, deploy into a process, and monitor performance over time. MLOps is the practice of running this lifecycle reliably in production.

Why it matters

ML models now support credit decisions, fraud detection, anti-money laundering alerts, pricing and customer service. The Financial Stability Board notes that wider use of AI can raise efficiency, but also concentration on a few providers, model risk and data quality concerns. In the EU, ML models used to assess the creditworthiness of individuals or to price life and health insurance count as high-risk AI systems under the AI Act. Banks using models for capital also apply supervisory model risk and validation standards.

Worked example

Illustrative numbers. A fraud model is tested on 100,000 card payments, of which 200 are fraud. It flags 500 payments. Of those, 150 are real fraud. Precision is 150 divided by 500, or 30 percent: seven in ten alerts are false alarms. Recall is 150 divided by 200, or 75 percent: one fraud in four is missed. Raising the alert threshold might cut false alarms to 200 but lower recall to 60 percent. The business chooses the threshold by weighing analyst workload against fraud losses.

Who uses it

Data scientists build, train and test models and choose features. Credit risk model developers apply ML alongside traditional scorecards and document model choices. Model validation analysts challenge the design, data and performance independently. Fraud analysts work the alerts a model produces and feed back confirmed outcomes, which become new training labels. AI transformation analysts in banking scope use cases, assess benefits and coordinate governance.

Common confusions

Machine learning and artificial intelligence are often used as synonyms. ML is the main family of techniques inside AI; rule-based systems and large language models are other examples. A second confusion is to judge a model on accuracy alone: with rare events like fraud, a model that flags nothing is 99.8 percent accurate and useless. Finally, a model that performed well at launch can degrade as behaviour changes, which is why monitoring for drift is part of the lifecycle.