Predictive Analytics is the use of historical data, statistical methods and machine learning to estimate the probability of a future event, such as a borrower defaulting, a customer leaving or a card payment proving fraudulent. Models train on past cases with known outcomes, are tested on unseen data, and are judged on ranking power, measured by Gini or AUC, and on calibration against actual outcomes. The output is a score attached to each customer or transaction that a rule or a person acts on. Fintechs use it for credit scoring, fraud checks before authorisation, churn prediction and payout forecasting. The EU AI Act treats creditworthiness models as high-risk, and models degrade as behaviour changes, so they need retraining.
How does predictive analytics work
A model is trained on past cases where the outcome is known: for example, 200,000 loans, each flagged for whether it defaulted within 12 months. Inputs might include income, existing debt, account age and repayment history, and in alternative data models, patterns in bank-account transactions. The model is then tested on loans it has not seen. Performance is measured for ranking power, with metrics such as the Gini coefficient or AUC, and for calibration: whether a group given a predicted 5% default rate actually defaults at about 5%.
Models degrade as customer behaviour and economic conditions change, so lenders and payment firms monitor the gap between predicted and actual outcomes and retrain on a schedule.
Predictive analytics in fintech: examples
Credit scoring models estimate probability of default and feed loan pricing and limits. Fraud detection models score card and account-to-account payments in milliseconds, before authorisation. Churn models flag customers likely to close accounts or cancel subscriptions. Treasury teams forecast daily payout volumes so they can pre-fund settlement accounts without leaving excess cash idle.
Illustrative churn example: a neobank with 100,000 active users loses 2% of them a month, or 2,000 users. A model flags the 5,000 users most likely to leave, and 800 of those do leave that month. The model has found 40% of all churners by contacting 5% of users. A retention offer costing €10 per flagged user (€50,000) pays off only if it keeps enough of those 800 customers to exceed that cost in lifetime value.
How the EU AI Act treats predictive models
The AI Act classifies AI systems used to evaluate the creditworthiness of individuals or establish their credit score as high-risk, with an explicit exclusion for systems used to detect financial fraud. High-risk systems need risk management, data governance, logging, human oversight and a conformity assessment. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, moved the application date for these stand-alone high-risk systems from 2 August 2026 to 2 December 2027.
Data protection law already applies today. In SCHUFA (C-634/21, 7 December 2023) the Court of Justice held that producing a credit score is itself an automated decision under Article 22 GDPR where a lender relies heavily on it, so the scoring provider must meet Article 22 conditions as well as the lender.
For a buyer of a lender or payments business, predictive models are part of the asset being acquired: due diligence should cover model documentation, validation reports, rights to the training data, and whether credit models will meet AI Act obligations by December 2027.