Nubank
transactional foundation model for credit scoring
In 2025, Nubank put nuFormer into production, a transactional foundation model that improves credit scoring AuC by +1.25% relative to LightGBM, reduces churn by 4.4%, and earns it its largest quarterly credit card market share gain in Brazil in ten quarters.
Key points
- Transactional foundation model nuFormer for credit scoring.
- Custom self-supervised transformer, embeddings fused with tabular variables (PyTorch, Databricks).
- +1.25% relative AuC vs LightGBM, -4.4% churn, largest market share gain in 10 quarters.
- Evidence B, confirmed status.
Objective
Approve credit for profiles that classic scoring rejects, without degrading portfolio quality, and gain credit card market share by making better risk decisions at scale.
The deployment
Nubank built nuFormer, a self-supervised foundation model trained on its transaction data. The model produces embeddings of the customer's transactional behavior, which it fuses with classic tabular variables to score credit risk and feed origination and limit-increase decisions. On comparable segments, the fused version delivered a relative AuC improvement of +1.25% over the industry reference (LightGBM), and the model in production reduced customer churn by 4.4% in relative terms against the reference model. Nubank attributes to nuFormer its largest quarterly credit card market share gain in Brazil in ten quarters, in the fourth quarter of 2025. The model is operational in Brazil and is being deployed toward personal loans and the Mexico and Colombia markets.
Results Proof B
The model gains (AuC, churn) come from research published by Nubank and its production deployment; the market share gain appears in the fourth quarter 2025 results. Concordant sources (official Nubank page, third-party analysis repeating the same figures). Attribution of the market share gain to nuFormer is carried by Nubank, so kept as such.
How it works
Documented architectureThe stack in detail
- llm nuFormer Self-supervised transformer foundation model trained in-house on transaction sequences, producing customer behavior embeddings fused with tabular variables.
- llm LightGBM Gradient boosting serving as the industry scoring baseline, beaten by +1.25% relative AuC by the fused model.
- outil PyTorch Training framework for the foundation model.
- plateforme Databricks Large-scale data processing platform for training and scoring.
- infra PySpark (Apache Spark) Transactional data preparation pipelines.
How it runs, concretely
For ops teams-
1Pre-training data team
nuFormer learns self-supervised on transaction sequences to produce customer behavior embeddings.
-
2Fusion and scoring AI
The embeddings are combined with classic tabular variables to score credit risk.
-
3Credit decision risk team
The score feeds origination, the limit, and its increase, under risk-policy guardrails.
-
4Production monitoring data team
More than a thousand monitoring indicators are watched to detect drift and trigger a retraining.
Transaction behavior and observed repayment. If the transactional data degrades or default labels arrive late, risk calibration becomes less reliable.
How your customers perceive this type of use
Sourced studiesC'est la famille la moins acceptee : 68% des Americains jugent inacceptable un score financier personnel calcule par algorithme et 67% l'analyse video automatisee d'entretiens d'embauche (Pew Research, 2018). La demande d'explication et de recours est massive : 83% veulent savoir quelles donnees l'IA utilise et 91% veulent pouvoir corriger des donnees erronees (Consumer Reports, 2024). A l'echelle mondiale, seuls 46% se disent prets a faire confiance aux systemes d'IA et 70% jugent une regulation necessaire (KPMG / Universite de Melbourne, 2025).
Acceptance conditions
- Transparence sur les donnees utilisees : 83% des Americains la reclament (Consumer Reports 2024)
- Droit de correction des donnees erronees : 91% le demandent (Consumer Reports 2024)
- Explication de la logique de decision : 44% des consommateurs sont plus enclins a utiliser un agent IA si sa logique est clairement expliquee (Salesforce 2024)
- L'acceptabilite depend du contexte de la decision : 50% des Americains jugent equitable un score de risque criminel pour la liberation conditionnelle, contre 32% pour un score financier applique aux consommateurs (Pew Research 2018)
Red lines
- La decision opaque et sans recours sur l'emploi, le credit ou le logement : 45% tres mal a l'aise pour l'embauche, 39% pour le pret, 39% pour le logement (Consumer Reports 2024)
- Le scoring des personnes a partir de donnees comportementales : 68% le jugent inacceptable pour les offres financieres (Pew Research 2018)
Sources: Pew Research Center 2018 · Consumer Reports 2024 · KPMG / Universite de Melbourne 2025 · Salesforce 2024
How to replicate
Inference, not sourcedData prerequisites
- A large volume of first-party transaction sequences
- Labeled default and repayment history
- Large-scale processing infrastructure
Org prerequisites
- A data science team able to train a foundation model
- A credit risk and compliance function
- Model governance (AI Act high-risk)
Possible stack
- Custom foundation model (the Nubank path)
- LightGBM-type gradient boosting as a baseline
- PyTorch, Databricks, Spark
The plan, step by step
- Step 1Establish a robust scoring baseline (gradient boosting on tabular variables) and document its AuC per segment.Deliverable: Documented baseline with reference metrics.
- Step 2Pre-train a self-supervised model on the first-party transaction sequences.Deliverable: Transactional embeddings evaluated offline.
- Step 3Fuse embeddings and tabular variables, then measure the marginal AuC gain against the baseline on comparable segments.Deliverable: Fused model beating the baseline, with a quantified gain.
- Step 4Pass risk and compliance validation: explainability, human oversight, high-risk system governance (AI Act, GDPR art. 22).Deliverable: Approved model governance file.
- Step 5Deploy progressively on origination and limit decisions, with drift monitoring and triggerable retraining.Deliverable: Model in production with monitoring indicators and a fallback plan.
First step: Establish a robust scoring baseline (gradient boosting) and measure the marginal gain of transactional embeddings before industrializing.
The same pattern, proven in other industries
Sources
- S1 Nu's AI-powered credit model extends access while maintaining portfolio quality Interested party archive pending
- S2 Nubank's AI Model Rewrites Credit Underwriting (WhiteSight) Secondary archive pending
- S3 How Nubank uses causality, machine learning and Python to support credit limit increase decisions Interested party archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.