How Nubank models risk for smarter, scalable credit limit increases

At our latest Nu DS & MLE Meetup, senior data scientists Jacob and José walked us through the complex and ever-evolving world of credit limit management. And, of course, how data science helps us make smarter, scalable, and sustainable decisions for our customers across Latin America.

With over 122 million clients and counting, offering the right credit limit is both a business opportunity and a responsibility. Too low, and we risk frustrating customers. Too high, and we may expose them, and ourselves, to unnecessary financial risk. So how do we strike the right balance? Let’s find out!

A risk modeling challenge, reframed

The heart of the presentation was our risk modeling strategy, specifically for deciding when and how to increase credit limits for existing customers. José began by framing the business challenge: managing credit risk while promoting financial health and ensuring our credit card products remain useful and profitable.

To do this, the team focuses on a crucial metric: the probability of default, or the likelihood that a customer will not pay their bill within a defined time window (typically 60–180 days).

But simply predicting if someone will default isn’t enough, we also care about when. Enter survival analysis: a more nuanced approach that models time-to-default using survival curves, allowing us to capture not just the risk, but its evolution over time.

Combining classic techniques with modern flexibility

Jacob then dived into the technical details of our modeling approach. While traditional non-parametric and parametric models offer strong baselines, each has its limitations. Non-parametric models, for example, don’t scale well and can overfit when too many variables are involved. Parametric models, while interpretable, often rely on strong (and sometimes incorrect) assumptions about the data distribution.

To address these issues, the team applies a “divide and conquer” strategy: first building a robust risk ranking model to capture signal strength, and then using that signal to guide the fitting of more precise risk curves.

This two-step process allows for more frequent updates to the survival curve calibration, while keeping the core ranking model stable. The result is a scalable and modular framework that adapts well to different countries and evolving customer behaviors.

Smart engineering for a dynamic environment

Of course, great models are only part of the equation. José closed the talk by emphasizing the importance of monitoring and operational rigor. As customer behavior shifts, especially in less mature markets like Mexico or Colombia, the team keeps a close eye on performance indicators, concept drift, and feature stability. This includes managing the impact of product launches, macroeconomic shifts, and changes to external bureau data.

The team also highlighted their investment in strong engineering foundations: reusable feature stores, CI/CD pipelines, and automated alerting for underperformance or anomalies.

And they’re not stopping there. The team is now experimenting with foundational models trained on trillions of transaction records to explore new frontiers in customer behavior modeling.

Financial health at scale

This session was a deep dive into how Nubank makes data-driven credit decisions that scale—not just in terms of data or technology, but in impact. By continuously refining our models, infrastructure, and strategies, we’re able to offer meaningful credit while promoting financial inclusion and stability.

In Jacob’s words: “We aim for simple but robust methodologies—never static, always improving.”

At Nubank, smart credit limit management is a commitment to scalable risk assessment, financial inclusion, and long-term customer health. By combining robust data science with flexible engineering practices, we’re able to deliver credit decisions that are not only precise and cost-effective, but also aligned with the financial well-being of over 122 million customers across Latin America.

Want to learn more about how we’re using data to drive impact at scale? Stay tuned to the Building Nubank blog for more deep dives from our engineering and data science teams.