RiskQuest is a Dutch financial consultancy firm that specialises in risk modelling. RiskQuest creates tailor-made models for large banks, insurers and other financial institutions in the Netherlands. We had the pleasure of speaking with Sven de Man and Emma Immink, with Emma being a former VSAE member. Sven and Emma gave us an introduction to the practical application of what is possible with a degree in Econometrics or Actuarial Science.
Emma and Sven, could you tell us something about yourselves?
Emma: I am Emma, I live in Amsterdam and I have studied at the UvA as well. I graduated two and a half years ago. I studied Econometrics and did a minor in programming. I was really interested in Data Science, so I did my Master in Econometrics. However, I combined it with the VU for the data science track. When I finished, I was still contemplating my options, so I checked the vacancies on the VSAE website. The website gave me a very good overview of the possibilities. I was looking to be a technical consultant. I wanted a job I can program in and that involves higher thinking than regular consulting. I joined RiskQuest and I liked it straight away because it is a relatively small company, but we work with prestigious clients and we can really make models. Before, I had an internship at a company that creates models for healthcare companies. I really liked modelling, so I wanted to work at a company that makes models and RiskQuest seemed like a perfect fit. I’ve been here for almost two years now.
Sven: I am Sven, I’m 37 years old and live in Amsterdam, I have a PhD in experimental physics. After I got my PhD in 2010, I switched to consulting. For a year, I worked for a firm in risk management. As I was still very much interested in the academic world, I went back to university to do a postdoc. I did the postdoc for almost 5 years part-time. I had a small research group in quantum optics. That was half-time and the other half I worked as a freelance financial risk manager, working for ABN AMRO and Rabobank mostly. Then in 2017, I joined RiskQuest. I am a partner and work mainly in credit risk and counter financial crime.
At what point did you realise you wanted to go into risk modelling? That is a big step from experimental physics.
Sven: After my PhD, I knew I didn’t want to make a career in academia and was considering multiple possibilities. In between, I was in training for air traffic controller which was also very different from experimental physics. However, that didn’t work out and I rolled into consulting. I always enjoyed the probability theory and statistics part. In physics education this is a small part, only statistical physics. For example, you don’t get stochastic differential equations. I actually liked this kind of mathematics so that’s where my personal interest laid.
On a more practical note, what do you actually do on a day-to-day basis?
Sven: I mainly run a couple of projects. I typically spend 20% of my week on RiskQuest related tasks, for example recruitment or internal discussions about strategy. The rest of my time I try to split equally between my projects. At the moment, I spend approximately 40% of my time on countering financial crime and 40% on credit risk. My day-to-day tasks include discussing on a high level with the members of the teams. What are the results of certain analyses? What are the implications for the business? Should we go a certain direction or not? What is our opinion with respect to model validation? These are mostly the high-level project decisions from a content perspective.
So you are a lot less involved in the technical aspect?
Sven: Yes, exactly. What I really like is sitting next to someone who is programming.
Emma: That would be me.
Sven: the data scientists perform the analyses and when everything comes together we can think about the results and put them in the proper business perspective.
And Emma, what about your day-to-day?
Emma: My day is a bit different. Most of the time I only work on one project, ABN AMRO right now. Before corona, I went to their office four days a week and I liked to go to our own RiskQuest office on Fridays since it is nice to work together with your RiskQuest colleagues. I think my split between programming and operational tasks is 50-50. Some weeks it’s 70-30, i.e. 70% programming and 30% in meetings, but most of the time it’s 50-50.
How do you actually model and find financial crime?
Emma: Right now, we fight financial crime by monitoring all the transactions of the client. You can open up your bank’s app and see all your transactions within the bank. We try to monitor all of these transactions. Of course, you can do that in several ways. For instance, you can aggregate them on a monthly basis and create features such as the sum of incoming transactions, the amount of cash available on the bank account or money transfers towards high-risk companies. Once you have created those features on an aggregated level, you can add historical cases of fraud or money laundering, the so-called target value. Based on this data, we fit a model that correctly predicts the target value historically. This model will then be used to try to predict future fraud cases. On some projects, historical labels are sparse. Then we could create an unsupervised learning model that tries to predict abnormal behaviour, instead of the target value. Fighting financial crime is mostly machine learning since there is so much data available.
Was that a difficult transition for you? In econometrics we mostly use regressions.
Emma: I think most colleagues of RiskQuest are working with credit risk models and those rarely use machine learning. However, if you are working on the financial crime projects, that may be a transition. I believe it is something you can easily learn, with a background in Econometrics, the task will stay the same: creating a model that solves a complex issue.
What’s the difference between credit risk and counter financial crime?
Sven: There are many differences, but there are also similarities. You are basically trying to predict something based on data in a bank. But as Emma was saying, this overall level of ‘you have a problem and you think analytically’ is essentially the same. But the tools you use are different. In financial crime, it involves a lot of big data, feature engineering, more data science type of work. Credit risk is more traditional, the tools we use are logistic regressions, Tobit regressions or linear models. If it gets advanced, you might use panel data. There is a lot of emphasis on model testing and checking your model assumptions, whether they hold or not. Therefore, it is more statistical in a traditional sense, while counter financial crime involves much more data science and machine learning. For example, you have to check your convergence and check your hyperparameters. It is much more technical, while credit risk is more traditional.
When we work with large banks, credit risk is heavily regulated, while in counter financial crime, you can think more freely on how to attack the problem. In credit risk, there is a lot of regulation from the European Central Bank (ECB) that we have to follow regarding the methodology and how to prepare the data. In loss given default models, for example, if you have a client that defaults (the clients doesn’t pay back the loan for a certain amount of time or is very unlikely to do so), then cures (so it gets back to fully paying), but then defaults again within nine months, you have to merge these default events. You have to pretend that it’s one event instead of two. However, for the probability of default model you should not merge these events. There are many small rules and articles you should consider when building the credit risk models.
Emma: A bit like restrictions you get in operations research.
RiskQuest is working on a new product, RiskNavigator. What is that about?
Sven: Consider a financial institution that for example issues loans. When a new client comes in, the company has to decide whether to grant the loan or not. The financial institution can choose to accept or can prefer not to have this client on her books. It is always a trade-off. The client will bring in interest, but maybe the risk of this client is bigger than the compensation. Then the company does not want to grant the loan to this client. So that is something the financial institution has to decide. This is quite hard as there is often not much information available.
For example, what is typically done when a client comes to the bank and asks for a loan, the bank needs to evaluate the financial situation of the client. The bank will request the annual report from this client from last year. An analyst then needs to evaluate the annual report and decide based on the balance sheet and the P/L whether to trust these figures and accept the risk of this client, or not. This kind of information, however, is typically quite static. These annual reports are months old, they are sensitive to fraud and they can be manipulated. So it is a tricky decision.
RiskNavigator is a tool which, instead of using the more static annual account data, focusses on transactional data of the client. In order to do this, the bank will ask permission to the client to use their transactional data. This is allowed under the Payment Services Directive 2 (PSD2) regulation, which financing providers can use as of January 2018.
The client can give permission to a one-time share of his transactional data of the last 18 months. We then enrich this data, for example by categorising it. All the transactions will be categorised: groceries, car, mortgage, etc. By doing this, we obtain summary statistics that are very much up-to-date. For example, for a restaurant we can see the incoming card and cash payments, the monthly rent and the monthly salaries. Based on this data, we can quickly analyse if it is a healthy company or not. So, the RiskNavigator tool will provide these summary statistics together with a verdict on whether to provide the loan or not.
This sounds very different from your usual consulting business. Are you moving more into products or is this just a one-off?
Sven: I think it is a general trend. You have generic consulting services that, for example, the accounting firms, like the Big 4, would do. We have a very niche consulting service business model, as we consult on risk models. We see that the whole economy is changing into products and subscriptions. Clients prefer to pay for a product and not worry about it anymore. That’s one reason and the other reason is that you only have to maintain and improve one model. We have built mortgage models at many different clients. The models are slightly different at every bank because the data is a bit different and the portfolio is a bit different. But in the end, it is the same type of model. It would be beneficial to have one model that you can constantly improve upon.
What does the first year at RiskQuest look like, either for a new consultant or a part-time student?
Sven: Students usually combine working on their thesis with an internship. Two days a week they work on their thesis where one of our colleagues will function as the thesis supervisor, together with a professor from the university. The other three days a week they work on an internal assignment such as helping in the development of a tool like RiskNavigator. There are a couple of students working with consultants on developing this tool. The other option is to help out on an assignment for a client together with a consultant. A lot of working students stay at the company and become consultants. Then they go on a client assignment as soon as possible. In the end that is the quickest way to learn how the job works.
Emma: I think it’s really interesting for the student because we offer 2-3 days a week working on the Master’s thesis and 2-3 days of work at our company. The subject is basically the same so you can work on a project and write your thesis about it. I think that is very interesting.
Do you have any last remarks?
Sven: What really defines RiskQuest as a company is this mixture of consulting and also very technical work. We make models, where we focus on both the business implications, but also on the statistics. For example, we can have a meeting with twenty people talking about Kalman filtering. That is special in the industry. In a lot of bigger firms, consulting is much more generic. We also do the generic consulting around it, of course, we have to explain our model to stakeholders at the client. However, the core work and the core discussions we have are technical. Therefore, I think it is a very special company. If you like mathematics and you like programming, it is a place where you can really do intellectually challenging work.
Emma: It is pretty fun as well, that helps.