Since it was introduced in the latter half of the twentieth century, there’s been a sense – until recently – that Artificial Intelligence (AI) has not quite lived up to its own hype. That’s changed in the last few years as advances in processing power, modelling techniques and a proliferation of data have transformed the field. AI is everywhere.

But until now it has been nothing more than a buzzword for the risk industry as regulatory transparency has prevented its use in credit scoring. Martin Benson explores the issue and outlines a genuine shift in approach that is set to revolutionise the credit risk industry.

AI was the big news of 2017. From self-driving cars and personal assistants on our smartphones, to tumour detection and machine translation, AI is being used to automate previously manual tasks and generate deeper insight from data. But, decades ago the credit industry was a pioneer of exactly that approach – using stats and data to automate decisions.

As far back as the 1980s, the industry was moving away from manually reviewing every single application to reach a judgement-based decision, to a process that used statistical techniques and data to automate those decisions to a greater or lesser extent. The benefit of this was more consistent and more accurate decision-making compared to manual underwriting, and at a lower cost, leading to more profitable lending.

The techniques used in credit scoring have been much simpler than the neural network based approaches that are driving the recent progress in AI, though, and in fact, have not moved on significantly since they were adopted some 30 years ago. Neural networks have been known and available throughout but they have not been widely adopted for credit scoring, despite having the potential to deliver superior results. There are three main reasons for this:

1. ‘Black box’ problem

Credit scoring systems need to be interpretable and predictable. You need to be able to look at the input data and understand why the model generated a particular decision. But neural networks produce ‘black box’ models that are so complex that it is essentially impossible to understand how decisions are reached. This not acceptable in a heavily-regulated industry like credit scoring, where the decisions made have significant implications for both lenders and the borrowers they serve.

2. Modest gains

Historic experience was that the performance gains of neural networks compared to linear models were modest. With hindsight, we now see that methodological weaknesses were a key contributing factor to that. It simply wasn’t possible to create powerful deep neural networks without techniques that were invented only in the last few years.

3. Lack of computational power

Until very recently, the processing power available was simply not up to the job of training deep neural networks to consistently perform the modelling required. It was just too slow to be useful.

However, there have been some key advances in recent times that mean deep neural networks now can revolutionise the credit scoring field.

A new dawn for AI?
The first step-change is processing power. Computing power has marched on according to Moore’s law, which says that processing power will double roughly every 18 months. The shift from CPUs (Central Processing Units) to GPUs (Graphical Processing Units) has also super-charged AI’s modelling capabilities. It’s now possible to train deep neural networks more than a thousand times faster than 10 years ago.

Furthermore, hardware designed specifically to accelerate deep learning is being brought to market (for instance the Tensor Processing Units that Google has developed), and will drive further significant gains.

The second big development that has really moved things forward is the algorithmic improvements that have been made in the last few years, such as ‘dropout’ and ‘rectified linear units’. These mean it’s now possible to train better, deeper neural networks than ever before, which has radically improved efficiency and significantly improved the performance of deep neural networks compared to linear models.

Solving the final task in the AI puzzle
The one area that was still causing hesitancy until now was the ‘black box’ issue. Without more certainty in interpreting, predicting and understanding decisions made by deep neural network models, their relevance to the credit scoring field was always going to be limited.

Standard AI approaches simply cannot work, as it is necessary to constrain the AI in order to generate the requisite interpretability and predictability and overcome the concerns over transparency that have until now hindered its use in credit scoring. Jaywing have developed an innovative, patent-pending approach to solving the remaining blocker to the adoption of AI in credit scoring – a significant advancement in the credit risk marketplace.

Our approach to creating deep neural networks and ensuring they behave in an intuitive and understandable way is to allow the business user to impose constraints on specific fields within the model so that the output adheres to certain business rules. For example, you might expect credit risk to increase as salary levels fall, or to decrease as disposable income and affordability improves, and the model needs to reflect these expectations. By guaranteeing that the model outputs behave in this way – in all instances – our methodology ensures that the models obey common-sense relationships and can be understood by the business and signed off through normal governance routes. No more black boxes to worry about.

Commentators in the field have often discussed the challenge of solving this one remaining issue that has prevented the mass adoption of AI in credit scoring. Producing neural network models that behave in this way is far from trivial and, we are pleased to be able make our unique, patent-pending solution broadly available to the industry through our new AI-based predictive modelling product, Archetype.

Furthermore, this development opens up the real possibility of having automatically-generated, constantly learning models that update themselves as new data becomes available – and without ever needing to worry that they will develop unwanted behaviours. This allows for far greater performance as models can be redeveloped at will, without the need to offset the typical associated costs of redevelopment.

We believe this kind of innovation is a significant advancement in realising the potential of AI-built models in highly regulated industries.

Martin Benson, Head of AI Consulting, Jaywing