Machine Learning
Machine learning (ML) is the discipline of building systems that learn to perform a task from data — finding the patterns that connect inputs to outcomes and using them to predict, classify, rank or detect — rather than being programmed with explicit rules for the task. Supervised learning learns from labelled examples, unsupervised learning finds structure without labels, reinforcement learning learns from rewards. It is the working core of fintech's fraud, credit and identity systems.
- Also known as
- ML, Statistical learning, Predictive modelling, Machine-learning models
- Used at
- fraud detection, credit scoring, transaction monitoring, churn and lifetime-value prediction, pricing, collections, document classification, identity verification, recommendation, and, through deep learning, language and vision; in trading, signal generation and execution
- Binding
- model-risk management guidance (the Federal Reserve's SR 11-7, the PRA's SS1/23, the EBA's expectations) governs development, validation and monitoring; consumer-credit and anti-discrimination law govern what features and outcomes are permitted; the EU AI Act classes credit scoring, insurance pricing and some employment and identity uses as high-risk, with data-governance, documentation and human-oversight obligations; data protection governs the training data
- Typical algorithms
- linear and logistic regression; gradient-boosted trees (XGBoost, LightGBM, CatBoost — the workhorses for tabular data); random forests; support vector machines; k-means and other clustering; neural networks; and the transformers behind language and vision. For tabular financial data gradient boosting usually wins; deep learning wins on text, images, sequences and scale
- Typical lifecycle
- define the target and the population; assemble and clean data; engineer features; split into train, validation and test; train and tune; evaluate on held-out and out-of-time data; validate independently; deploy; monitor for drift and performance; retrain. Most of the work is data and evaluation; the algorithm is a small part
- Parties
- data scientists and ML engineers, the data engineers who build pipelines, model-risk validators, compliance and legal, the business owners whose decisions the model makes, and the regulators who examine the model file
How machine learning works in practice
A lender wants to predict which applicants will default. It assembles 200,000 past applications with their twelve-month outcomes, builds 300 features from the application, bureau and transaction data as of the decision date, and trains a gradient-boosted model on 70% of them, tuning on 15% and testing once on the remaining 15%, with the test period after the training period. The model reaches a Gini of 0.62 against the incumbent scorecard's 0.54. Validation by an independent team checks the data lineage, the feature definitions, the absence of leakage, the performance by segment, the fairness metrics across protected groups, and the stability of the model on the most recent quarter. Deployment puts it behind an API scoring applications in 50 milliseconds, with reason codes from SHAP for every decline, monitoring of score distributions and outcomes, and a retraining schedule. The model card records all of it, and the regulator reads the card.
That lifecycle — not the algorithm — is what machine learning in a regulated business is. The algorithm is a library call; the data, the target definition, the validation, the fairness analysis, the monitoring and the documentation are the work and the risk. The failures are almost always there: labels that encoded past decisions, features that leaked the outcome, a random split that hid drift, a model that was never retrained as the population changed, a proxy for a protected characteristic nobody probed. Deep learning extended machine learning to language, images and long sequences, and rented general-purpose models made those available without training; but the fintech's own decisions about customers still run on models trained on its own data, and those are the ones the regulator examines. For a buyer, a target's machine-learning capability is its data, its lifecycle discipline and its model files — not its algorithms, which are the same everywhere.
For a fintech, machine learning is the technology behind its decisions and the governance obligation that comes with them; for an acquirer, the models are worth what the data and the validation behind them prove.
Example
A payments company's fraud model, trained in 2022 and never retrained, sees its detection rate fall from 74% to 58% over two years as fraud patterns change; its false-positive rate rises as the customer base shifts. An audit finds no drift monitoring, no retraining schedule, no out-of-time validation in the original file, and a feature — "days since account opened" — whose distribution has shifted so far the model's thresholds are wrong. The rebuild: a monthly retraining pipeline on the trailing 18 months, drift monitoring on the top 30 features and on the score distribution, out-of-time validation as the acceptance criterion, a champion–challenger process, and a model card regenerated on each release. Detection returns to 76%, false positives fall by a third, and the model-risk file now shows a lifecycle rather than a moment. The company's next acquirer reads the pipeline, not the algorithm.
Common mistakes
- Treating the algorithm as the work; the data, validation and monitoring are. - Models deployed once and never retrained as the world moves. - Random splits that hide drift and leakage; validate out of time. - Features that proxy protected characteristics, found by the regulator first. - Deep learning on tabular data where gradient boosting is more accurate and easier to explain.