The behavior and situation of many customers has changed due to COVID-19. What can lending organizations do to respond to these changes in their portfolio and still manage risk as effectively as possible?
It is clear that many businesses are impacted, positively or negatively. Where in normal situations portfolio risk development depends on continuity, profitability and entrepreneurial spirit, in crisis, risk can get accelerated by other challenges for companies. Hence, for a portfolio manager it is very important to understand the changed risk behaviors and patterns. If, for instance, risk is monitored by a (behavioral) scorecard, the risk curve still might be applicable. That is, distribution from high risk to low risk. But, within a score the good or bad odds might have changed drastically. To put it more simply: in normal times a high-risk prediction might deliver an outcome of 6%, in crisis situations the same score will deliver a risk prediction of 12%. If you are not aware of this “doubled risk” you might take wrong decisions. So, it is important to understand these changes in your portfolio and act upon it.
We see many organizations claiming to miss an accurate risk profile of their customers, mainly due to using outdated information. With fast-changing situations like COVID this doesn’t give them the best view of a customer. How can organizations improve their risk models to cope with growing risks in their portfolio and act upon these changed risk behaviors and patterns you are mentioning?
Many companies are still mainly looking at traditional data when assessing credit risk. However, other data sources, not correlated with traditional used data, can help portfolio managers to get a better risk view and with this, put measures in place like early warning activities. Traditional data is your normal application or customer data, or your existing bureau data. Very simply: data you already use in “traditional” times for “traditional” purposes. But for example, real-time data from the web can help give a more precise risk view at that specific moment in time. Web Data Insights (WDI) is a product that uses real-time and not historical data in assessing credit risk. Then machine learning and advanced data science are deployed to generate entirely new predictive data variables. This results in models that enable lending organization to add a new dimension to credit risk management and respond to fast-changing environments with creating a better risk view.
Can you explain a little bit more how this works?
Experian has developed a risk index based on publicly available web data. In the development of this index, we have found 200+ attributes that can say something about the risk development of a company within the next 12 to 18 months. Compare it as if you were to search company details via a search engine on the internet. Depending on a company’s digital footprint, you will find only a few or many search results. We take the 200+ attributes from Google Search and Google Maps and via a machine learning model we derive the results back in a risk index: Web Data Insights (WDI). As it is so called non-traditional data (that is, not used in traditional risk models) there is very little correlation with existing internal models and therefor, in combination with your in-house used score, it gives a huge uplift in the decisioning power of the combined model.
Can traditional data sources also be improved? How?
In general, the better the data quality, the better the model performance: qualitative input, is qualitative output. Improvement really depends on how good your model already is and how good your team is capable of developing, monitoring and recalibrating the model. Machine learning can already give a substantial movement in most cases.
The other thing is that you can add macro-economic data to your model, especially if risk development is very sector or regional driven. So, add this data as attribute into your model, or cross it with your existing model, and make various strategies per sector. Adding economic data can also make models more aware of external shocks that will impact the macro-economic outlook, like COVID-19.
Finally, transactional data as a new data source is also tested in models. The added value is strong in risk and affordability calculations, and future cash flow forecasting.
Non-traditional data sources enable organizations to add a new dimension to credit risk management and respond to fast-changing environments with creating a better risk view.
If companies use more different data sources to manage credit risk, isn’t there a higher risk of declining ”good” customers?
In any risk model there is the likelihood of declining potential ‘’goods’’ below cut-off, and alternatively accepting ‘’bads’’ above cut-off. If the data quality is good and the model has a good discriminatory power, the decision power of your model should be sufficient. With adding good quality additional data resources – both traditional and non-traditional – the decision power only can improve. As a result, related to the question, less ‘’goods’’ are to get declined and more ‘’bads’’ can be accepted. Additionally, better decisions can be taken in between accepting and declining, for instance upfront payments. So again, it is all about having the appropriate data sources and models in place.
Increasing the acceptance rate of customers to grow their business is another focus point for many lending organizations. But how can they do this without resulting in a higher risk of default?
Especially increasing the acceptance rate of “new to bank” customers is a challenge. In general, a lenders strategy is more balanced and product/portfolio and margin dependent. The better the decision power of a model, the better the outcomes: increased acceptance rate with same level of risk, or alternatively same acceptance rate with lower risk of default. The more data resources and better data quality, the better view you can get to assess risk and the less risk of default.
Organizations sometimes struggle with making a good risk profile of smaller companies. How can this process be improved?
WDI is fit for purposes typically in this segment. Whereas the owner of a smaller company also behaves as a consumer, in combination with credit bureau data it will help to make better decisions. As we have seen in our recent findings specifically for SME’s, this is a crucial segment to improve decision models for as they are expected to have a higher risk at default in the coming 1,5 years.
What is the go-live period for Web Data Insights? Is it easy to add to current decisioning models?
Using WDI per application is a straightforward interfacing. There is an API that calls the WDI model and an index is delivered and used in further decisioning. Only the index is passed back to the front-end and in most cased the index together with your internal model is crossed in a matrix in which you normally set your decisions: accept, refer and decline.
And what if organizations want to enrich an entire portfolio with additional data?
For that we use a batch approach where we receive a file with the existing customer records which then is enriched with our data.
What information and data do organizations receive when using batches or Web Data Insights?
When running a batch, we will need to have company indicators like name, address and KvK numbers delivered in a file that is securely transferred to our analyst team. They run the batch through the WDI model and for almost all records in the delivered file, the WDI Risk index is set for each individual record. This batch is securely transferred back for internal follow-up where the organization can use further analytics or workflow execution like contacting the high-risk customers.
To close off, what is the main advice you want to give to lending organizations for credit risk in the coming period?
We are now more than 12 months in crisis, which in theory would give enough time for risk teams to control and monitor the decision power of their existing models. I would advise them to at least do the following:
– Segment your portfolio on sector level
– Check if companies applied for governmental support
– Get all your current available “traditional” data points
– Check against performance data and connect your data points with relevant performance flags: delinquency, default and even fraud flags
And, ideally, bring in other data types such as WDI and transactional data and run a retro- and real-time analysis to directly see the added value of new data points and identify risk. Happy to support here!