At Deloitte, we have developed an innovative approach to binding traditional regression with machine learning (ML) algorithms. Our approach not only allows lenders to make smarter credit decisions through ML, delivering fast, tailored and compliant end-to-end model design, it also avoids the creation of a black box.

One organisation in France who is benefitting from our solution is the French FinTech, October.eu. October is an SME crowdfunding platform based in continental Europe, operating across France, Spain, Italy, the Netherlands and Germany. October was looking to develop an improved credit-scoring model leveraging both internal and external data along with new techniques to help better manage the underlying risk in their portfolio. This is where they relied on the expertise of the Deloitte credit risk modelling team and our model development platform, Zen Risk.

Our Zen Risk solution offers an end-to-end platform enabling continuous predictive model design and maintenance. Furthermore, the Hybrid modelling approach at the heart of Zen Risk ensures total transparency and simple outputs that regulators and businesses can understand. What makes Zen Risk different is that we use machine-learning algorithms to design a set of complementary business rules, which enhance the traditional model’s estimates. In essence, using the machine to teach a human how to build better credit models and make better credit decisions.

Through our approach, we are able to develop an efficient and fully transparent predictive algorithm in just a few weeks. Furthermore, as Zen Risk is easy to implement, October was also able to productionise the new credit-scoring model effortlessly. As a result, the underwriters started benefitting from a new credit-scoring model allowing them to accept more customer projects whilst limiting the level of credit risk taken. Therefore, quickly yielding the benefits of increasing the number of newly accepted customers, whilst reducing the default rate of the overall portfolio.

The business rules developed by Zen Risk were complementary to the logistic regression score. The machine learning models help identify non-linear relationships, which cannot be captured by the logistic regression. These additional variables, proven to be predictive of default are then incorporated as business rules which either upgrade or downgrade the estimated PD of a customer.

In this specific use case, the business rules are able to capture additional external data features such as the number of applications submitted on the lending platform, the gross equity value or the web source. The October credit team validated these business rules as intuitive drivers of credit risk.

In a recent article Patrick de Nonneville, Chief Operating Officer of October highlighted, "working with the Risk Advisory team at Deloitte has not only accelerated the delivery of the project, but especially helped apply the best practices of documentation, auditability and transparency, this last point being particularly key for the proper use of the approach by our analysts." [Translated from French]

There are tangible benefits to improving the quantification of risk and implementing better credit scoring approaches can have a direct impact on your bottom line, by improving the acceptance of customers within risk appetite and reducing misclassified customers who end in default thereby resulting in economic loss.

Using samples of client portfolios from our case studies with total loan value of $1bn, we calculated the benefit in the underlying portfolios under the two scenarios below. The results compare performance of the hybrid model developed using Zen Risk with the client’s incumbent scorecard.

Scenario 1: Business impact when the risk appetite level is constant

Our hybrid approach increased the number of accepted customers by 50 to 55% for the same risk appetite level. Taking into account the better in risk-based pricing, which leads to higher booking rates, and the reduced losses due to previously misclassified customers being rejected; this on average results in a $100m to $150m improvement in the bottom line based on the sample portfolios.

Scenario 2: Business impact when the acceptance rate is constant

Our hybrid approach decreased the number of projects outside risk appetite by 52% to 56%. Taking into account the better in risk-based pricing, which leads to higher booking rates, and the reduced losses due to previously misclassified customers being rejected; this on average results in a $90m to $115m improvement in the bottom line based on the sample portfolios.

If you would like to know more about Zen Risk and how we can help you improve the quantification of risk using machine learning please get in touch.

Additional Reading:

Blog: Smarter Better Faster: The next generation of risk modelling with Machine Learning

Blog: Regulatory Change and Investing in the Risk Function of the Future.