How De Volksbank took its credit risk model to a higher level
The Credit Risk Modeling department of De Volksbank faced a challenge: how to implement new regulations, including the new definition of default that form the basis of credit risk models. Having to execute the PD modeling process and the LGD modeling process simultaneously convinced the bank they needed help. “That help came in the form of someone who was ‘one of the few’ able to do both LGD and PD components.”
De Volksbank, which is the parent company of SNS, ASN Bank, RegioBank and BLG Wonen, is the Netherlands’ fourth largest bank. It offers its 3.2 million customers mortgages, checking and savings accounts, and, albeit to a lesser extent, insurance and investment products too. A key difference with other major Dutch banks is the social roots of its four subsidiaries. These typically focus on the Dutch consumer society, independent entrepreneurs and small- and medium-sized enterprises. SNS, for example, helps people with their financial affairs, ASN Bank finances sustainable projects, RegioBank stimulates local entrepreneurs and BLG Wonen makes it possible for first-time homebuyers to get a mortgage.
From framework to validation
As a retail bank, De Volksbank mainly serves private individuals and savings and checking account customers. It therefore doesn’t have an overly large business portfolio. “During the past year, a small number of our savings and mortgage customers have felt the financial stress imposed by the corona crisis,” says Guido van Baal, who heads De Volksbank’s Credit Risk Modeling department. “De Volksbank has several measures at hand that can be deployed to help these customers. We developed these measures a while ago so that we could help customers during periods of financial hardship. To date, we’ve been able to support more than 1,500 customers. And last year we ran several projects to develop models with regulatory deadlines, all of which we complied with.”
“A significant modification to an IRB model must be submitted to the ECB for approval”
Van Baal’s team comprises 17 employees, which includes modelers, project managers, and data specialists. About four years ago, the team split from the then-larger department it formed together with Market Risk Modeling. The Credit Risk Modelling department develops models for both the mortgage and corporate portfolios, such as IRB (internal ratings-based) models and IFRS (International Financial Reporting Standards) models.
“Developing a credit-risk model starts with building a framework,” explains Van Baal. “This is a process-based, step-by-step plan documenting what needs to be done. The departure point for all this is the relevant legislation and regulations. Furthermore, we also collate all manner of data. Only after the process has been thought through and we are in possession of all the data, the modeling actually can begin. This starts with the analysis of the data, which we do by means of, among other things, holding expert sessions to test whether the behavior we observe in the data corresponds with expectations. We then test various model methodologies to establish their relevant limitations, after which we opt for the most ‘fit-for-purpose’ methodology.
As soon as the appropriate model methodology has been chosen, we select the risk indicators so that we can arrive at a prototype model. We then calibrate this model so that model prediction and realization are a good match. Everything is then tested and documented. The final internal step is to send the model to an independent internal department for validation. This determines whether or not we proceed with the model and request external approval. A significant modification to an IRB model must also be submitted to the ECB for approval, and only when ECB approval has been received the model can be used.”
Down two model routes
In 2018, the team embarked on a model-development route for a new credit risk model, which was then submitted to the regulator for review at the end of the same year. Given that this model did not yet fully comply with the new regulations, the regulator decided not to pursue the approval process.
“We had to go down two model-development routes simultaneously, so we decided to split our department into two teams”
“We were also mandatorily obliged to adapt our credit risk model in accordance with the new definition of default, which defines when a client is in default on a loan,” adds Van Baal. “That had to be done before 30 June 2020, which meant we had to go down two model-development routes simultaneously. It’s for that reason that we decided to split our department into two teams: one to redesign the existing model to include the new definition of default, and the other to focus on a completely new model that incorporated the regulator’s findings (and also including the new definition of default). The problem, however, was that we didn’t have enough capacity to carry out two such substantial projects at the same time. We started by engaging several freelancers to carry out a specific part of the assignment, but we were also looking for a specialist who was familiar with several aspects of credit risk models. And that’s why we came to Zanders. John de Kroon had just finished a project with another bank in which the new default definition had also been implemented.”
From LGD to PD
To facilitate the development of the models, Van Baal was looking for someone experienced in the area of loss given default (LGD) models. “But no sooner had John started with us than all manner of things changed, necessitating some changes in the project. What we wanted was someone who was also familiar with probability of default (PD) models. Fortunately, John turned out to be ‘one of the few’ who was also able to do the PD component. It was great to be able to solve the problem like that.”
So, after starting the development of the LGD model in May 2020, the assignment changed for Zanders, which had first helped make the data accessible for the PD part before moving on to the development of the actual PD model. It is an IRB model, which is used to calculate capital requirements. The model was delivered for internal validation in April 2021.
“This PD model is to be used to calculate our risk-weighted assets,” continues Van Baal. “That’s why the ECB, in its supervisory capacity, has to approve these models. They can only be used if the ECB is convinced that they comply with the regulations, are prudent enough, and that all the necessary steps have been correctly followed. It is an intensive process and we expect the model to be evaluated at the end of Q2, 2022. Subject to the final decision of the ECB, it should then become clear from which quarter the model can be used. By then, the underlying data platform and model engine will also be ready and we’ll be able to make monthly prognoses for every client in our portfolio.”
At the business end
According to Zanders consultant John de Kroon an additional component has also been incorporated into the PD model. “De Volksbank will also be able to use the model at the business end to assess new applications. Based on the magnitude of the loan being applied for and the characteristics of the applicant, the model can generate a creditworthiness estimate. In combination with human input from the credit assessor, the bank can then ascertain whether or not to accept a credit application. In this way, the model also supports the business end.”
“In combination with human input from the credit assessor, the model enables the bank to decide whether to accept a credit application”
“In addition to being a regulatory requirement, credit risk hedging is also, of course, a bank requirement,” assures Van Baal. “If you have a model that can predict risk, you use it for reporting and carrying out capital calculations. But it can also be used to support a customer-selection process. In this respect, in addition to being a regulatory capital model, this also makes it a client-acceptance model. In other words, the model isn’t just used at the back end, but at the front too – at the business end. It’s a new way of working for us; in the past we used a separate scorecard for that.”
The power of monitoring
The model uses both macroeconomic data, such as unemployment rates, as well as customer and loan specific characteristics. “You can see the repercussions of the past year, for example, during which the corona crisis played havoc with the economy,” continues Van Baal. “That said, a model is always one step behind reality, so you always have to monitor closely – and there lies the power of good model management. The model must work over several years. To realize an average value that’s as accurate as possible, the peaks and troughs must compensate for one another.”
Van Baal insists that both these model-development routes are part of a broader improvement plan. “In addition to modifying the models, we also have a new policy, which has laid a better foundation on which to build our models.”