Machine Learning (ML)

What it is

Machine learning is a family of techniques that fit a model to data so that it can produce useful outputs on new, unseen inputs — predictions, classifications, anomaly scores, recommendations. Unlike rule-based systems, ML does not require an analyst to encode every pattern of interest by hand; the model finds patterns in the training data and generalises from them.

Three main paradigms

• Supervised learning. The model is trained on labelled examples (e.g. transactions labelled "fraud" / "not fraud") and learns to predict the label on new data.
• Unsupervised learning. The model learns the structure of the data without labels — clustering similar transactions, detecting outliers, building behavioural baselines.
• Reinforcement learning. An agent learns by trial and error against a reward signal. Less common in production financial-defence pipelines, but useful in policy and routing contexts.

What "good" ML looks like in finance

In a regulated financial environment, ML alone is not enough — the quality bar is set by explainability, traceability and human oversight. Every score must be reducible to the inputs and the model version that produced it, every model must have a documented training data window, and every consequential decision based on a model output must have a human in the loop.