Revolut Built a Foundation Model for Money (3 minute read)
Revolut replaced six separate machine learning systems with a single foundation model trained on 24 billion banking events, achieving massive performance gains in credit scoring, fraud detection, and marketing.
What: Revolut trained a unified foundation model on 24 billion banking transactions across 111 countries that delivers 130% better credit scoring, 65% better fraud recall, and 79% higher marketing engagement compared to their previous six separate ML systems.
Why it matters: This signals a fundamental shift in financial services from traditional feature-engineering ML approaches to foundation model infrastructure, potentially creating a new competitive moat where the model itself becomes the core intellectual property rather than just a tool. The approach could reshape how neobanks and crypto-native fintechs compete on underwriting and risk assessment.
Deep dive
- Revolut consolidated six separate machine learning systems into one foundation model trained on 24 billion banking events spanning 111 countries
- The performance improvements are substantial: credit scoring improved by 130%, fraud recall by 65%, and marketing engagement by 79%
- This represents a strategic shift where the trained model itself becomes the primary intellectual property asset, rather than just supporting traditional banking operations
- Financial services are transitioning from feature-level machine learning (where engineers manually design features) to foundation-model-level infrastructure (where models learn representations from massive datasets)
- The competitive implication is that whichever bank builds the most comprehensive foundation model next could capture billions in value through superior risk assessment and customer engagement
- This approach mirrors developments in other AI domains where foundation models trained on massive datasets outperform narrowly-trained specialized models
- The scale of training data (24 billion events across 111 countries) provides a data moat that smaller competitors would struggle to replicate
- Direct implications for crypto-native fintechs and neobanks who will need similar infrastructure to compete on underwriting quality and fraud prevention
Decoder
- Foundation model: A large AI model trained on vast amounts of data that can be adapted for multiple downstream tasks, rather than separate models trained for each specific task
- Fraud recall: The percentage of actual fraudulent transactions that the system successfully identifies (higher recall means catching more fraud)
- Credit scoring: Automated assessment of how likely a borrower is to repay a loan, traditionally based on credit history and financial behavior
- Underwriting: The process of evaluating risk when deciding whether to extend credit or insurance to a customer
Original article
Revolut just moved the IP of banking into a model.
Trained on 24 billion banking events in 111 countries.
One foundation model replacing six separate ML systems.
- Credit scoring: +130%
- Fraud recall: +65%
- Marketing engagement: +79%
The model is the new moat.