Our latest paper is live! In this work, we study how to model financial contagion over dynamic networks. When people apply for loans, banks have a pretty important decision to make—can the borrower pay it back? Traditionally, banks use credit scores to assess risk, but our new research extends our previous research by delving deeper into the relationships borrowers have with others to better understand their chances of default.

Why Networks Matter in Credit Risk

Imagine you’re considering lending money to someone, but that person is part of a group where others have also borrowed money. This idea is at the heart of our study. Rather than seeing borrowers as isolated, we treat them as part of a bigger network. Their connections to other borrowers—like being in the same neighbourhood or using the same mortgage provider—might influence their financial behaviour.

Predicting Loan Defaults Using Dynamic Networks

Borrowers can be connected in various ways and these relationships evolve over time, making them dynamic. To better capture these connections, we developed a model that combines two powerful tools: Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs). The GNNs help us map out these borrower networks, while the RNNs allow us to track how these relationships change over time. But that’s not all—we added an attention mechanism that prioritizes certain time points over others, based on their relevance to the borrower’s default risk.

What Did We Find?

We tested our model using real-world data from Freddie Mac, a major U.S. mortgage financier. The results were exciting—our model did a better job of predicting which borrowers were likely to default compared to traditional methods. It wasn’t just more accurate; it also provided a more profound understanding of why certain borrowers might struggle to repay loans.

Why This Matters

For banks and lenders, this research could change how they think about credit risk. By considering the connections between borrowers, lenders can make more informed decisions. This could even lead to more people getting approved for loans, especially those who might not have had a chance based on traditional credit scores alone. For borrowers, this kind of model could mean more opportunities. If banks can better understand the factors affecting risk, they might be more willing to take a chance on people who were previously overlooked.

What’s Next?

Our research shows that networks play a significant role in financial decisions, and there’s much more to explore. We’re excited to keep building on this work to better understand financial risk. The more we learn, the more we can help lenders and borrowers alike make informed financial decisions.

The paper is available, open to all and with CC-BY license, here. The code to replicate the paper can be found here.

Also, Juan Cristóbal Constain from Quipu created a podcast using NotebookLM from this post and the paper. Give it a listen below if you prefer!