How Nubank uses causality, machine learning and Python to support credit limit increase decisions

Written by: Nubank Editorial

During the Meetup in partnership with GruPy-SP and PyLadies, held at Nubank’s office in São Paulo, Daniela Canuta, Lead Data Scientist at Nubank, shared a high-level view of how data science techniques can support credit limit increase decisions on the credit card. The talk covered topics such as predictive modeling, causality, optimization and monitoring of machine learning systems in production.

At Nubank, credit decisions involve much more than an isolated analysis. Broadly speaking, these processes combine different signals, models and monitoring mechanisms to support choices that seek to balance customer experience, risk management and the sustainability of the operation.

This kind of challenge goes beyond predicting behavior. It also requires understanding which decisions tend to generate value consistently for customers and for the business, always with responsibility and a long-term view.

The challenge behind a credit limit increase

Granting more credit is a decision that involves multiple factors. On one side is the need to offer an experience that makes sense for the customer and follows their financial journey. On the other is the importance of maintaining a healthy operation, with attention to risk, efficiency and the regulatory context.

In practice, this means building systems capable of analyzing different dimensions of the same decision and supporting more personalized choices at scale.

The role of predictive models

An important part of this process lies in the use of predictive models, which help estimate behavior trends from signals observed over time. These models are fed by data that is processed and organized to represent relevant patterns, in work that combines technical knowledge and business understanding.

At this stage, the quality of the variables used makes a big difference. More than gathering a large volume of data, the challenge is to build useful, stable and interpretable representations to support consistent decisions.

For that reason too, the most complex solution is not always the best. Depending on the problem, simpler approaches can offer a better balance between performance, interpretability, computational cost and scalability.

More than predicting: prioritizing well

In decision systems, predicting is only part of the work. In many cases, it is also important to correctly prioritize which situations deserve more attention and which decisions tend to generate better relative results.

This requires evaluating not only the accuracy of the predictions, but also the system’s ability to support ranking, prioritization and consistency across different profiles and contexts.

When correlation is not enough

Another central point is distinguishing correlation from causality. In real problems, observing a change in behavior after a decision does not, in itself, mean that one thing caused the other.

That is why causal analysis techniques help complement predictive modeling. Instead of looking only at what tends to happen, they help investigate in which contexts a given action really makes a difference. This kind of approach contributes to decisions that are more efficient, more responsible and more aligned with the customer experience.

Models in production require continuous monitoring

Putting a model into production is just the beginning. Over time, data changes, behaviors evolve and the economic environment shifts. To keep generating value, machine learning systems need to be monitored continuously.

This includes observing indicators of performance, stability and operational quality, as well as periodically reviewing models to ensure they remain suitable for the context in which they operate.

Python at the center of this ecosystem

Throughout this entire journey, Python plays a central role. The language is widely used in analysis flows, experimentation, model development and integration with production environments, supported by tools widely used by the industry, such as PySpark, Scikit-learn, PyTorch, Airflow, Docker, Databricks and Grafana.

More than a tool, Python has become an important part of the way technical teams build, test and evolve decision systems at large scale.

Building more responsible decisions at scale

The combination of machine learning, causality, optimization and monitoring shows how data science can contribute to more robust and sustainable decision systems. In a context of millions of customers, the challenge is not only to scale technology, but to do so with responsibility, consistency and a focus on the experience of those who use the products.

As areas such as Data Science, Machine Learning Engineering and Data Engineering continue to evolve, this kind of approach tends to become increasingly relevant to the future of digital financial services.