AI, a helping hand for managing credit risk

The pandemic’s impact on the world economy has been wide-ranging. But how has it affected credit risk? A determining factor for the profitability of any lender, the management of credit risk could soon involve the use of Artificial Intelligence (AI) to achieve objectives that were previously out of reach, provided that the regulations allow its deployment. Julien Valletoux, director of risk, permanent monitoring and compliance at Crédit Immobilier de France, and William Violet, CEO of Homiwoo, analyse the situation.

What’s your view of the mortgage market since the start of the pandemic?

Julien Vailletoux: Getting a mortgage is becoming harder because the conditions for approval in France (a maximum level of indebtedness set at 35% of income, and mortgage periods limited to 25 years) are proving to be a real constraint. These new requirements will inevitably lead to a reduction in households’ ability to borrow, which will make it ever more difficult for them to become homeowners. In addition, individuals with healthy finances who could have borrowed over a very long period will also have their applications turned down. By imposing strict criteria, the section of society on modest incomes will be excluded from property ownership and will be driven to the rental market, even though mortgage experts can identify long-terms solutions despite these criteria.

William Violet: Overall, with these new rules, people’s appetite for spending, which is one of the components of our algorithms, will be reduced. And you can anticipate two consequences of this drop in households’ ability to borrow: a fall in demand in certain areas, due to a reduction in the number of people in a position to buy a property, and greater stability in house prices. Over the past year, we have seen a significant rise in the time taken to conclude a sale. In the major cities, prices stabilised at first, before returning to a more normal rate of growth. After the Fed’s announcement about its planned rate rises by the end of 2023, we can a whole series of rate increases to start soon. The cumulative effect of the pandemic, the rate rises, the shorter borrowing periods and the lowering of household debt ratios all point to a lasting reduction in households’ ability to obtain financing.

JV: In March 2020, it was difficult to predict whether we were heading for a major crisis or a non-event, in terms of the economy. To manage the credit risks, we had to develop a series of different scenarios. From mid-March, we were facing requests for the deferment of mortgage payments, although in truth these were being sought as a precautionary measure, and significant increase in arrears. In March-April 2020, we saw a 50% increase in the number of people going into arrears for the first time. However, by May, both these indicators had returned to their normal levels. Paradoxically, since then, the number of first-time arrears has been at its lowest level, while the overpayments have been more numerous than expected. Meanwhile, the number of cases of being handled by the indebtedness committee of the Banque de France has fallen. It’s a paradox, but the Covid-19 crisis seems to have improved household finances overall.

How has the pandemic affected credit risk management at CIF?

JV: Very quickly, it became clear that there was a greater need for agility and flexibility in our procedures for risk monitoring. The traditional indicators were no longer sufficient when it came to identifying the potential difficulties that could be faced by our portfolios. So, we developed new indicators for detecting the various sectors, types of financing, and properties that were most at risk of a default. We introduced a more detailed segmentation of the arrears which, for example, enabled us to identify the financing difficulties of managed residences.

Have you had to rethink your customer relationship management or your monitoring?

JV: The pandemic led to a major improvement in our risk management methods. The need to introduce new indicators meant that we developed analytical and automated tools (mainly for data visualisation). We also began to integrate innovative algorithms that use AI, enabling us to better detect the fainter signals coming from the market. We wanted to identify the profiles of at-risk clients in a more granular and more agile way, so that we could arbitrate between the security of the debt financing and the settling of unpaid arrears. Once this had been carried out, we could contact the clients concerned and find a solution. And throughout this process, we were able to maintain as far as possible both the CIF’s policy on risk, and our usual approach to handling the situations faced by our customers, separate from the pandemic. Some banks opted for the mass deferment of debt repayments. By contrast, we were preferred to analyse each situation individually, to make sure that deferment was indeed the best solution, given the customer’s difficulties. When those difficulties were due to the circumstances of the time, we chose to stagger the payments over a short period, of between three and 12 months.

What did AI bring to your understanding and optimisation of risk management?

JV: Compared to traditional methods, AI provided benefits in terms of the speed and the usefulness of our risk analysis. When it comes to risk management, it seems essential to me to combine AI with human intervention. With AI, you can identify irregularities in behaviour more quickly, but you cannot do without the analysis by experts of the results produced by AI. This complementary approach makes it possible to create strategies more quickly and in a more structured way, while also making them more transparent and homogenous.

WV: AI is a decision-making tool that aims to provide valuable, reliable information. Using AI means handling risk with greater accuracy, and being ahead of the curve rather than behind it. As for property valuations, it gives you greater geographic granularity, right down to the precise address. Prices in different areas in the same district do not change in the same way, and analysis by address, and property by property, makes a huge difference in terms of accuracy. With AI, you can instantly find out the true state of the market, which again makes for better risk management. AI models are also able to analyse complex data and identify weaker market signals, compared to simple models that miss out on this wealth of information. AI models provide a targeted, differentiated response, while simple models provide a uniform response.

The level of detail provided by AI also makes for better ‘loan to value’ calculations, and therefore provides the promise of better performance in terms of risk detection. 

Lastly, it provides much greater transparency. While simple models offer an easy understanding of a situation, they are not very reliable. So there needs to be a way of mathematically auditing AI models, so that people feel more comfortable about using them.

How can AI optimise risk management in the period between the end of a crisis and the return to normal life?

 JV: You have to wonder whether the crisis is really here, or whether it is still to come. Unemployment and bankruptcies are both on the way down, and the government has compensated for the low level of household consumption. With the help of AI, we can certainly identify the sub-groups that will struggle in a few months’ time, in order to strengthen the support we provide. Our aim is for our customers to keep up their payments, or find ways of resolving their problems, and not for the situation to end up as a legal dispute. AI helps us to take action to prevent this. It enables you to achieve a new objective – of becoming an ‘augmented analyst.’ It’s wrong to only look at sub-groups and linear relationships. Understanding a trend needs an all-round perspective. The whole is more than the sum of its parts. AI helps us to understand this reality in all its complexity. On the other hand, regulations should not be allowed to hold back these technological tools on the basis that there is no deterministic formula involved On the contrary, regulations should be embracing them.

WV: AI is clearly an unrivalled tool for managing risk on a large scale. It’s regrettable that AI hasn’t so far been taken into account by France’s regulator, notably in the rewriting of Article 3 of Regulation 99-10 on mortgage companies. This covers the analysis of the property market using a fixed formula that enables you update the value of a property, based on its original price. While its laudable to have an audit trail, it doesn’t make it possible to manage the actual risk. In fact, when it comes to the operational reality, an audit trail is actually counter-productive, because although it may be reassuring – it’s not very effective. Simple models based on indices are structurally ineffective, because there is not sufficient granularity.

What’s more, reality is regularly disrupted by crises. Simple models soon find themselves out-of-step with the new reality, while AI models are able to learn and to adapt to the crisis – and they are more resilient as a result. There is no doubt that algorithms using AI methods give more predictive and more reliable results.

The proposed legislation would also require the updated price to be no higher than the market value or the mortgage value. Achieving this objective would involve applying a safety margin or a significant discount. And the less accurate the model, the bigger the discount would be. This creates a ‘double penalty’ for banking institutions that use simple models, as it would require an over-provision against default, compared to the real risk, as a result of less accurate information. In turn, it means that banks would have less liquidity and would therefore be less competitive. In its current form, this approach therefore excludes all the benefits that AI could bring to our understanding of the market!

Today, not using AI means depriving banking institutions of the tools needed to deal with future crises and reducing the competitiveness of our country!