If you’ve attended any recent conferences or industry events, you’ve probably heard a lot about artificial intelligence, or AI. Claims about AI seem to be everywhere! None other than Sundai Pichar, Google’s CEO, has called it ‘one of the most important things humanity’s working on. It’s more profound than… electricity or fire.’ In this latest blog, Nick Sime, our Head of Modelling, explores how AI is impacting us now, with a particular focus on credit scoring.
As far as bold statements go, we’d say Pichar’s comment is right up there. But while we do tend to think AI is probably a bit overhyped at this point, there’s no denying it’s helping us achieve some fantastic things already, and has enormous potential.
Consider this. Our smartphones now understand what we mean when we talk to them and respond to our requests. Cars are driving themselves. AI-assisted radiology is helping doctors detect cancer in instances where it would have previously gone unnoticed. These are real advances in AI that have come about in a relatively short space of time, but AI is (currently) nowhere near taking over the world.
When it comes to assessing credit risk, AI also has the potential to bring about radical change. Here’s an overview of the latest advancements and how they could reshape the way we manage risk and prevent and detect fraud in the years to come.
Understanding AI: the basics
There are four main commercial areas of application of AI currently:
1. Natural language processing
This is the ability to analyse and understand text-based data, including its meaning and context.
2. Image classification
Natural language processing’s more visual sibling, this makes sense of images, graphs and other visual information.
3. Predictive modelling
AI can extend the scope of predictive modelling by bringing unstructured data – such as text and images – into scope. It can also be used to generate more accurate predictions than previous methodologies allowed. Finally, it can be useful in detecting unusual patterns of activity or unexpected behaviour.
4. Voice interfaces
Currently one of the more visible uses of AI, voice interfaces can understand text and convert speech to text. As you can probably deduce, this is what makes machine-to-human interactions possible, and it’s being deployed in customer-facing roles.
How AI Meets Credit Scoring
During our recent webinar, we asked our attendees which field of AI they thought will make the biggest impact on credit risk. And a whopping 73% told us it’s predictive modelling.
Well, we tend to agree.
AI can make it possible to identify fraudulent activity much more quickly. Which means you can tackle it at the first attempt instead of having to sit back and observe in order to establish a pattern or MO. Clearly, this has huge implications for clients, operators and regulators alike.
More to the point, AI can inject a previously unheard of level of predictive performance into credit modelling.
But how effective is AI, really?
Back in the 90s and 00s, neural networks — one of the fundamental concepts underpinning AI — were slow to learn and often performed underwhelmingly. But advances in technology, especially in the move from CPUs (Central Processing Units), to GPUs (Graphical Processing Units) and now TPU (Tensor Processing Units, which have been designed specifically to enhance AI), have led to dramatic improvements in the speed at which neural networks can learn.
And within the last few years, mathematical advances have vastly improved the way neural networks perform, leading to the deployment of ‘deep learning’ – neural networks that use many layers of neurons to achieve their goals.
So, AI is finally capable of fulfilling its potential.
Recently, we took the development sample of an application fraud model that was originally developed using traditional logistic regression and trained a neural network on the same data using our proprietary Archetype software.
A massive 10% relative performance uplift in Gini coefficient – on a model which already performed well. This type of uplift could translate into a business benefit worth millions of pounds.
Uniquely, Archetype uses mathematical means to make sure the neural networks it produces adhere to business rules. Which means it’s possible to control and explain the relationships that exist in a given credit scoring model, for each individual input variable.
This, of course, is critical in credit scoring, because:
- It allows you to understand why customers are accepted or rejected. This can come in handy when clients challenge credit decisions and is essential for good model governance
- It makes it possible to have clear oversight over credit policy. In particular, it enables you to demonstrate that you’re actually accepting those you’re supposed to be accepting
- It’s a way to show regulators that your credit decisions are robust.
Beyond credit risk
Credit scoring aside, AI is undoubtedly going to have wide-ranging implications for financial services as a whole.
We’ve already mentioned how AI can speed up fraud and money laundering detection. At the same time, natural language processing can improve decision-making outcomes significantly by bringing agent notes, client comments and other unstructured information into the scope of an automated decision.
And, of course, voice interfaces and chatbots are increasingly being deployed, hopefully (!) improving the customer’s experience while lowering overheads.
AI undoubtedly has the potential to change credit scoring — and financial services in general — for the better.
But that’s not to say it’s all roses. Or that it’s going to be an easy path.
As things stand, demand for AI talent far outstrips supply — about 300,000 professionals worldwide to fill millions of jobs. Unsurprisingly, a poll we’ve run revealed that 36% of respondents intend to seek external help to get AI-ready.
Luckily, we’re here to help. With almost 20 years of consulting experience, over 70 data scientists and deep learning specialists on our team, we’re ideally positioned to help you extract every ounce of value from AI.