Towards Generative Credit

Matt Mollison & Matt Flannery

In the world of microfinance, traditional credit assessment methods often falter. It is next to impossible for someone with no credit score and thin financial records to secure a loan through traditional means. Today, however, advances in generative AI are making the loan approval process more accurate and accessible.

By using machine learning to analyze alternative data sources, such as receipts and phone data provided by users, Branch is able to reliably predict loan repayment in markets with limited bureau coverage and an outsize segment of first-time borrowers.

These high-performing machine learning models have allowed us to automate lending decisions for millions of people. Using alternative data, we’ve made more than a billion dollars in loans — currently around 30,000 loans a day — throughout Africa and India, all while keeping delinquency and default rates low. Our advantage is in the vast quantity of data we’ve painstakingly collected over the past decade.

Now, the ability to fuse traditional machine learning with increasingly capable Large Language Models (LLMs) has the power to revolutionize microfinance, opening access to millions who have been shut out by traditional lenders.

Today: Data Labeling with LLMs

Responsible lending is paramount at Branch, considering the substantial impact on both borrowers and the company. While LLMs (like generative pretrained transformers, or GPTs) possess remarkable capabilities, they currently fall short in some areas necessary for direct underwriting. They lack domain expertise and real-time economic awareness. As generative AI continues to advance at a rapid pace, we’re working to solve these challenges.

In the past, we would use humans who write regular expressions to parse user data. For a year now, we have been using LLMs in a virtuous two-step combination: First, we use GPT-based LLMs to label data in an automated pipeline, which saves us thousands of hours and improves label accuracy.

Second, to balance efficiency and cost, we used these labeled records to fine-tune a lightweight and fast LLM (DistilBERT) to classify data in our credit risk prediction pipeline in real time. We extract information like transaction amounts and aggregate to depict user financial behavior.

The Branch underwriting model uses thousands of inputs, or features, to train credit risk models. An example of a simple feature could be “Amount spent on groceries per month” which would be gathered from commercial receipts shared by the user. This is a very important feature with a lot of predictive power.

Generating the “grocery spending amount” feature, however, is labor intensive. There are many different types of grocery stores, both online and offline. Extracting an accurate transaction amount is time consuming and riddled with complexity. Imagine doing so across India — or any of our markets — with all the languages and regional variation!

Instead, an LLM could be trained on all business data from across India, in multiple languages. Because of its ability to flexibly match patterns, the LLM is likely to outperform a team of data engineers writing regular expressions to label and categorize the data. We can effectively replace the data engineering work with a single LLM.

This is not about saving money on payroll; it’s about objectively improving the loan approval process. These LLMs can do a much better job at something that humans could never do perfectly, anyway. And it frees up our engineers to do more impactful work.

The grocery example can be extended to nearly any feature that relies upon unstructured data. How many friends do you have? Are you late on any bills? How much do you spend on fuel? Do you have steady employment? How conscientious are you? All of these rely on a somewhat subjective analysis drawn from messy data. Robots can now do this much better than people.

Integrating these features significantly improved our credit risk models. Almost instantly, this started showing up in the financials, resulting in reduced delinquency and increased growth. Since implementing LLMs, our defaults fell 25% while loans grew +300% over the same period. In lending, it’s remarkable to decouple risk and growth to this degree.

Tomorrow: Generative Credit Models

Currently, LLMs can help us analyze unstructured data about a user’s past, automating a task typically performed by humans. But what about the future? After all, LLMs are generative. They are creative. They can predict outcomes (albeit imperfectly).

Predictive Synthetic Data

With well-defined prompts and source data, LLMs can generate synthetic data: artificially generated data that can be used to train machine learning models. They can produce fake data about anything — for instance, a fake play supposedly written by Shakespeare as if he lived in the 21st century. In the case of lending, we might want to produce synthetic data that predicts a given user’s future spending, based on his or her past spending behavior.

Take bank statements. Models trained on a sequence of past bank statements can identify temporal patterns in user financial dynamics (account balances, spending habits). With knowledge of these historical patterns, they could predict what will happen next month or even next year. An LLM could generate a hypothetical future bank statement based on past spending habits.

What might a user’s financial picture look like next year? We can use that to make a more informed underwriting decision today.

Automatic Feature Discovery

Currently, machine learning lenders employ teams of data scientists who brainstorm new ideas for features with predictive power. They run experiments in the real world to test if such features help predict a given user’s repayment rate. This is an inventive process that relies on trial and error over long vintages of loans. People call this “feature discovery.”

In the future, imagine replacing this process with generative AI. You could fine-tune an LLM on user data and loan outcomes to automatically learn features that are important for evaluating creditworthiness. Then you apply advanced techniques to understand how the internal workings of the trained model identify these features from the data, making them human interpretable — potentially surfacing entirely new feature engineering ideas.

AI as Credit Officer

We’re just scratching the surface on the potential of LLMs and other generative models to extract important information from unstructured data. Given that such models can automatically discover important features and that underwriting is just math (all good lenders perform a data-driven evaluation of risk), the next logical step is to train models to directly make the loan outcome predictions, and eventually make the underwriting decisions themselves.

By incorporating knowledge of credit policy, business goals, and economic signals into this long-term vision, we can continue to sustainably address gaps in financial access for historically underserved people.

Long-Term Impact

Over the past five years, machine learning-based models have allowed lenders to distill valuable insights still overlooked by banks. This has resulted in an “emerging middle class” with access to credit. Over the next five years, generative credit models will complete the task of including the bottom billion into the banking system. Generative credit could make a generational impact. This is just the beginning.