Jamie Twiss is an experienced banker and a data scientist who works at the intersection of data science, artificial intelligence, and consumer lending. He currently serves as the Chief Executive Officer of Carrington Labs, a leading provider of explainable AI-powered credit risk scoring and lending solutions. Previously, he was the Chief Data Officer at a major Australian bank. Before that, he worked in a variety of roles across banking and financial services after beginning his career as a consultant with McKinsey & Company.
Can you explain how Carrington Labs’ AI-powered risk scoring system differs from traditional credit scoring methods?
Carrington Labs’ approach to risk scoring differs from traditional credit scoring methods in several ways:
Our platform uses a much larger dataset than previous methods. Traditional credit scores rely on outdated technology and are based on the small amount of information available in a customer’s credit file, mostly payment histories, which only give a limited snapshot of an individual, and no view at all of many people. With customer consent, we take line-item bank transaction data and use it to create a far more detailed and richer picture of an individual.
We then use modern AI and machine-learning techniques to turn these large volumes of data into a sharp point of view on the creditworthiness of an individual, calculating hundreds of individual variables and combining them into a comprehensive overall view. The resulting scores are fully explainable and transparent to the lender using them, unlike credit scores, which are mysterious black boxes. These scores are also tailored to a lender’s specific product and customer segment, which makes them more relevant and therefore accurate than a credit score, which is a generic score trained across a wide range of products and customers.
Finally, our platform can not only assess the risk of a customer more effectively than a traditional score, but it can use that score to recommend the optimal lending terms such as limit and duration. As a result of all these factors, CL risk scoring is a significant advancement upon the insights that traditional methods give lenders.
How does your AI integrate open banking transaction data to provide a fuller picture of an applicant’s creditworthiness? And what are some of the key predictors that your AI models identify when assessing credit risk?
Our models can be trained on many different types of data, but bank transaction data is usually at the core. We use tens of millions of lines 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 generally the best way to collect this data, as it provides a consistent format, good security, and fast response times. We can collect it through other means, but Open Banking is usually preferred.
For example, we can analyze cash withdrawal habits to see if someone frequently withdraws large amounts, if they always use the same ATM, or if they take out cash multiple times a day. We can identify gambling activity by looking for frequent transactions on betting platforms. We can look at how quickly someone spends money after receiving it, or whether they adjust their spending if they start to run low. We also flag unexpected financial patterns that can indicate risky mindsets or behaviors, like frequent speeding tickets.
Our models are trained on around 50,000 possible variables, with about 400 actively used in a typical risk model. This data-driven approach helps lenders make more precise lending decisions and tailor loans to each applicant’s unique risk profile. It’s important to note that the data we identify and analyze is anonymous, so we don’t deal with personally identifiable information (PII).
How does Carrington Labs ensure that its AI models are free from gender, ethnic, or socio-economic bias in lending decisions, and what steps have you taken to mitigate algorithmic bias in your credit risk assessments?
Carrington Labs’ models are significantly less likely to be biased than traditional approaches due to their objectivity (no human “gut feel” involved) and the wide range of data we use to create models.
We have three pillars to our anti-bias approach: First, we never let protected-class data (race, gender, etc.) anywhere near the model-creation process. We prefer it if you don’t even give us that data (unless you want us to use it for bias testing; see below). Second, our models are fully explainable, so we review every feature used in each model for potential bias, proxy variables, or other problems. Lenders also have access to the list of features and can conduct their own reviews. Third, if the lender chooses to provide us with protected-class data for testing (only; kept far away from training), we will conduct statistical tests on model outputs to determine approval rates and limits and ensure variation across classes is clearly driven by explainable and reasonable factors.
As a result, the higher predictive power of Carrington Labs’ models and the ability to fine-tune limits based on risk makes it much easier for lenders to approve more applicants on smaller limits and then increase them over time with good repayment behavior which enables broader financial inclusion.
How do you ensure that your AI-driven credit risk assessments are explainable and transparent to lenders and regulators?
While we use AI in a number of steps in the model-creation process, the models themselves, the actual logic used to calculate a customer score—are based on predictable and controllable mathematics and statistics. A lender or regulator can review every feature in the model to ensure they are comfortable with each one, and we can also provide a breakdown of a customer’s score and map it back to an adverse-action code if desired.
How do your AI models help democratize lending and expand financial inclusion for underserved populations?
Many people are more creditworthy than their traditional credit scores suggest. Legacy credit scoring methods exclude millions of people who don’t fit into traditional credit models. Our AI-powered approach helps lenders recognize these borrowers, expanding access to fair and responsible credit without increasing risk.
To give one example of someone who falls into an underserved audience, think about an immigrant who just recently moved to a new country. They might be financially responsible, hard-working, and industrious, but they might also lack a traditional credit history. Because the credit bureau has never heard of them, they lack the capability to prove that this person is creditworthy, which in turn makes lenders reluctant to present them with loan opportunities.
Those non-traditional transaction data points are the key to building an accurate assessment of credit risk scores for people that credit bureaus aren’t familiar with. They might lack a traditional credit history or have a credit history that might seem risky to lenders without proper context, but we have the ability to show lenders that these people are creditworthy and stable by leveraging a larger quantity of their financial data. In fact, our platform is up to 250% more accurate, based on a sample set of anonymized data, at identifying low-risk borrowers with limited credit information than traditional credit scores, and that’s what empowers lenders to expand their base of borrowers and ultimately increase loan approvals.
In addition, because many lenders have only an approximate sense of an individual customer’s risk, they struggle to fine-tune an offer to reflect a customer’s individual circumstances, frequently either offering them more than they can afford, lending them less than they need, or (most frequently of all) turning them down altogether. The ability to set lending limits precisely has a particularly strong effect on enabling lenders to bring new borrowers into the financial system, from where they can increase their borrowing capacity by showing good repayment behavior—giving them that first chance to show that they can work responsibly with debt.
What role do regulatory bodies play in shaping the way AI-powered lending solutions are developed and deployed?
Regulators are an essential part of embedding AI in financial services and in the wider economy. Clear boundaries on where and how AI can be used will enable faster growth and new use cases, and we’re supportive of the various processes underway to create legal and regulatory accountability.
As a general principle, we believe that AI tools used in lending should be subjected to the same kinds of oversight and scrutiny as other tools—they should be able to demonstrate that they are treating customers fairly, and that they are making the banking system safer, not riskier. Our solution can clearly demonstrate both.
Can you tell us more about Carrington Labs’ recent selection into the Mastercard Start Path Program? How will this accelerate your US expansion?
We’re delighted to be working with Mastercard on our US and global expansion plans. They have unparalleled experience in delivering financial solutions to banks and other lenders around the world and have already been extremely helpful as we increase our engagement with prospective US clients. We expect both parties to benefit, with Mastercard offering advice, introductions, and possibly elements of our solution, while Carrington Labs provides a high-value service to Mastercard clients.
Beforepay, your consumer-facing brand, has issued over 4 million loans. What insights have you gained from this experience, and how have they shaped Carrington Labs’ AI models?
Through this experience, we learned how to build models quickly and effectively thanks to the access Beforepay gave us to their great R&D lab and some tremendously large volumes of data. If we have an idea for a model framework, architecture, code, etc. we can try it out in Beforepay first. The precipitous decline in Beforepay’s default rate is also a great case study in showing how well the model works.
It’s been a very motivating experience in general, as our employees have a big stake in the company. We’re using Carrington Labs’ models every day to lend out our own money, so it focuses the mind on making sure those models work!
How do you see AI evolving in the lending space over the next decade?
Lending is going to change massively once the industry fully moves over to the kinds of big-data-powered risk models that Carrington Labs is leveraging over the next decade. And it will—those models are just so much more effective. It’s like the role of electricity in manufacturing; it’s a game-changer and everyone will either make the shift or exit.
Big-data models can either be built by hand (which I used to do myself, but this process takes months or even years while also being hugely expensive and incapable of providing the best outcome. Or you can automate the model-building. With AI, you can automate far more of it at higher quality while also saving time and doing things that would be impossible if you were building by hand, like generating thousands of custom features for a mid-sized lender.
The key is knowing how to do it correctly—if you just throw a bunch of stuff at an LLM, you’ll get a giant mess and blow through your budget.
Thank you for the great interview, readers who wish to learn more should visit Carrington Labs.
Credit: Source link
