Jamie Twiss is an experienced banker and data scientist working in the intersection of data science, artificial intelligence and consumer loans. He is currently the CEO of Carrington Labs, a leading provider of interpretable AI-driven credit risk scores and loan solutions. Previously, he was the chief data officer of a major Australian bank. Prior to that, he worked in various roles in banking and financial services after his career as an advisor at McKinsey & Company.
Can you explain how Carrington Labs’ AI-driven risk scoring system is different from traditional credit scoring methods?
Carrington Lab’s risk scoring method is different from the traditional credit scoring method:
Our platform uses larger datasets than previous methods. Traditional credit scores rely on outdated technology and are based on a small amount of information available in customer credit documents, mainly payment history, which gives only a limited snapshot of the individual and has no opinion in many. With the client’s consent, we obtain line project bank transaction data and use it to create more detailed and richer personal images of individuals.
We then use modern AI and machine learning techniques to transform this massive amount of data into a keen view of personal credibility, compute hundreds of individual variables, and combine them into a comprehensive holistic view. Unlike credit scores, the final score can be fully explained and used by credits, which is the mysterious black box. These scores are also tailored to lenders’ specific products and customer segments, which makes them more relevant and therefore more accurate than credit scores, a universal score trained across a wide range of products and customers.
Finally, not only can our platform assess customers’ risk more effectively than traditional scores, but also can recommend the best loan terms such as limits and durations using that score. Due to all these factors, CL risk scores are a significant advancement in providing lender insights into traditional methods.
How does your AI integrate open banking transaction data to provide a complete picture of applicant credibility? What major predictors does your AI model identify when assessing credit risk?
Our models can be trained on many different types of data, but bank transaction data is often the core. We use tens of millions of rows of transaction data to train the overall model and then use thousands of transactions for each new customer that the model scores. Open banking is usually the best way to collect this data as it provides a consistent format, good security and fast response time. We can collect it in other ways, but open banking is usually preferred.
For example, we can analyze cash withdrawal habits to see if someone frequently withdraws a large amount, always uses the same ATM, or gets cash multiple times a day. We can determine gambling activities by looking for frequent transactions on the betting platform. We can see how quickly someone spends after receiving the money, or if they adjust their spending if they start running low. We also mark unexpected financial patterns that may indicate risky mindsets or behaviors, such as frequent speeding tickets.
Our model trains about 50,000 possible variables and is actively used in a typical risk model. This data-driven approach helps lenders make more accurate loan decisions and tailor the loan to each applicant’s unique risk profile. It is important to note that the data we identify and analyze are anonymous and therefore we do not process personally identifiable information (PII).
How does Carrington Labs ensure that its AI model has no gender, racial or socioeconomic bias in loan decisions, and what steps have you taken in your credit risk assessment to mitigate algorithmic bias?
Carrington Lab’s model is much less likely than the objectivity of traditional methods (not involving human “gut sensations”) and the broad range of data we use to create models.
We have three pillars of anti-bias approach: First, we will never let protected class data (race, gender, etc.) anywhere near the model creation process. If you don’t even give us data, we prefer it (unless you want us to use it for bias testing; see below). Second, our model is completely explainable, so we review each feature used in each model to understand potential biases, proxy variables, or other problems. The lender also has access to the feature list and can make his own comments. Third, if the lender chooses to provide us with protection-level data for the test (only; away from training), we will conduct statistical tests on the model output to determine approval rates and limitations and ensure that changes in each category are clearly driven by explainable and reasonable factors.
As a result, Carrington Labs’ model’s higher predictive power and the ability to fine-tune restrictions based on risk make it easier for lenders to approve more applicants on smaller restrictions and then increase them with good repayment behavior over time, thus enabling wider financial inclusion.
How do you ensure that AI-driven credit risk assessments are interpretable and transparent to lenders and regulators?
When we use AI in multiple steps in the model creation process, the model itself, the actual logic used to calculate customer scores – based on predictable and controllable mathematical and statistical information. A lender or regulator can view every feature in the model to make sure they are satisfied with each feature, and we can also provide a score breakdown for our customers and map it into bad action codes if needed.
How does your AI model help democratize lending and expand financial inclusion for underserved populations?
Many people are more credible than what traditional credit scores suggest. The traditional credit scoring method does not include millions of people who are not suitable for traditional credit models. Our AI-driven approach can help lenders recognize these borrowers and expand access to fair and responsible credit without increasing risk.
To give an example of an underserved audience, think of an immigrant who has recently moved to a new country. They may be financially responsible, diligent and diligent, but they may also lack a traditional credit history. Since the credit bureaus have never heard of them, they lack the ability to prove that the person is credible, which in turn makes lenders reluctant to offer them loan opportunities.
These non-traditional transaction data points are key to an accurate assessment of the credit risk scores that are not familiar to credit bureaus. They may lack traditional credit history, or have credit history that may seem risky for lenders without proper backgrounds, but we have the ability to show lenders that these people are creditworthy and stable by leveraging a large amount of financial data. In fact, based on a sample of anonymous data, our platform is up to 250% accurate, less risky borrowers than traditional credit scores identifying limited credit information, which is why lenders lenders expand their borrower base and ultimately increase loan approval.
Furthermore, since many lenders only have an approximate sense of risk to individual clients, it is difficult for them to fine-tune an offer to reflect the client’s personal situation, often offering them more than they can afford, less than what they need, or (most often) reject them altogether. The ability to accurately set loan restrictions has a particularly powerful impact on enabling lenders to enter the financial system by showing good repayment behavior, from where they can improve their borrowing capacity by showing good repayment behavior, which is the first time they have had the opportunity to show that they can work responsibly with debt.
What role do regulators play in shaping the way AI-driven lending solutions are developed and deployed?
Regulators are an important part of embedding AI into financial services and the broader economy. Where and how AI is used, boundaries will achieve faster growth and new use cases, and we support various processes of ongoing legal and regulatory responsibilities.
Overall, we believe that the AI tools used in loans should be subject to the same type of oversight and scrutiny as other tools – they should be able to prove that they are dealing with customers fairly and make the banking system safer, not more risky. Our solution can clearly demonstrate both.
Can you tell us more about Carrington Labs’ recent selection of MasterCard Start Path program? How will this accelerate your U.S. expansion?
We are pleased to work with MasterCard in our U.S. and global expansion plans. They have unparalleled experience in providing financial solutions to banks and other lenders around the world, and it has been very helpful as we increase our engagement with potential U.S. clients. We hope both parties will provide advice, introductions and solutions through MasterCard, and Carrington Labs provides high-value services to MasterCard customers.
More than 4 million loans have been issued before your consumer-facing brand. What insights have you gained from this experience and how do they shape the AI model of Carrington Labs?
Through this experience, we learned how to build models quickly and effectively, thanks to our access to their excellent R&D labs and some massive amounts of data. If we have an idea about model frameworks, architectures, code, etc. We can try it before sales. The sharp drop in pre-sales default rates is also a good case study to show how the model works.
Overall, it was a very inspiring experience because our employees have a large stake in the company. We use Carrington Labs’ models every day to lend our own money, so it focuses on what makes sure these models work!
How do you view AI development in the loan space in the next decade?
Once the industry completely shifts to the large-scale risk model that Carrington Labs leverages over the next decade, loans will change significantly. This will be – these models are more efficient. It’s like the role of electricity in manufacturing; it’s a game-changer, and everyone goes on shifts or exits.
Big data models can be built manually (I used to be myself, but the process takes months or even years, and it is also very expensive to provide the best results. Or, you can automate the construction of the model. With AI, you can automate more time, and at higher time and quality of execution, you can also save thousands of custom features and provide you with impossible time to generate thousands of custom features.
The key is knowing how to do this correctly – if you just throw a bunch of stuff on LLM, you’re going to be in a huge mess and hit your budget.
Thanks for your excellent interview, readers who hope to learn more should visit Carrington Labs.
