Extending the coverage of vulnerability detection.

This is our last blog in the series on data science aspects of vulnerability detection. We started with discussing how to set modelling targets in this problem and proceeded to advanced analytics that can analyse client communication channels to detect signs of vulnerability. In today’s post, we will discuss how to extend the coverage of vulnerability detection in order to get a better understanding of the customer vulnerability landscape.

Dealing with financial matters is part of our daily lives and this part is with us no matter whether we are enjoying stability or treading through hard times. Vulnerable consumers are defined as those who, due to their personal circumstances, are at a higher risk of experiencing financial detriment. Firms are encouraged to use front line communication and data analytics to understand the landscape of vulnerability among their customers.

There is no single definition for the many forms of hardship and their representations in firms’ managed information. One of the possible approaches is to define vulnerability as a set of indicators that can be arranged into a vulnerability chain made of vulnerable circumstances drivers, consequent human reactions and potential financial harms.

Usually, understanding the vulnerability landscape of an organisation starts with different teams working together to share the patterns that they observe (or expect to see) and the decisions that they would make in these situations. This forms the foundation for setting modelling targets along the driver-consequence-harm chain and determining appropriate points of intersection with business processes. Although organisations often lack the full picture in their data, they can search for relevant indicators along the vulnerability chains.

We have discussed that some of the most accurate data-driven techniques for detecting vulnerability rely on analysing the audio and text of actual client communication. This approach mimics traditional methods that frontline staff use to detect vulnerability and brings numerous benefits: it neatly fits into existing processes so existing policies do not need to be revised, it follows tried-and-tested routes that the firm understands, and the technology enables this to be run at scale with a high degree of consistency. The approach is also more verifiable because an appropriately trained human can credibly review an algorithm’s prediction for each case. Despite its many benefits, however, this approach is not without drawbacks. It suffers from two issues: low coverage, as most people do not regularly communicate with their financial services providers, and poor timeliness, as most communication often happens after clients have experienced difficulties.

If a firm has an established data-driven approach to analysing customer outcomes, it can combine existing and proven vulnerability information (usually termed “vulnerability flags”) with outputs from customer communication analysis. This step can give a larger training data sample for building predictive models, so long as it's possible to identify individual customers in the data and successfully merge their information from different sources.

Often, this larger sample of vulnerability flags is still insufficient in terms of representativeness and sparsity to train supervised models that can accurately find customers similar to the ones already flagged. In order to address this, firms may need to involve broader customer information and combine a rule-based decision-making layer with customer segmentation and profiling techniques.

To use a retail banking example, the scope of managed information could include general customer information, information on products, accounts behaviour, third-party data that is often used for credit risk management purposes (such as bureau scores and CAIS), customer communication history, branch visits, and information from digital journey touchpoints.

Some data sources allow to effectively capture certain vulnerability drivers such as low financial resilience, which is closely connected to creditworthiness. Here, traditional credit risk management approaches can be adapted to detect signs of vulnerability. However, this information is less useful for inferring major life events or health issues at the stage when they have not resulted in financial hardship yet. In this case, firms could employ additional data and techniques, if benefits of adding extra layers of analysis are expected to outweigh extra cost and risks of conducting such an analysis (also, if this analysis is compliant with data protection regulation and firm’s internal policies).

Firms could adopt approaches used in customer relationship management and sales in order to combine customer segmentation information with signals extracted from unstructured data such as accounts behaviour, transactional data, banking application data, and third-party sources such as open banking. Flagged vulnerability cases from earlier stages of analysis can form “seed” customer groups and be used to guide the detection process, mitigating low case volumes and data sparsity (e.g., by building pseudo-social networks). These techniques can give a better view of the current and temporal customer profiles and segment them according to the modelling targets defined by the firm.

To supplement these methods and give firms a better coverage of potential vulnerability circumstances, techniques such as life events predictive modelling can be utilised. As suggested in the second part of this series, a combination of major life events and other vulnerability risk drivers/consequences can nudge a case past the decision boundary of whether the customer requires tailored support or not.

In order to understand the customer vulnerability landscape, firms should undertake an exploratory journey and allow data and analytics to play a major role. It is a complex problem to define vulnerability and despite the scale of financial services penetration and digitalisation, the current semi-manual approach to detecting vulnerability will fail to consistently identify cases where firms should act. That said, incomplete and inaccurate data beats the status quo, especially if the alternative is inaction. Technology can help firms break the problem down and embed fairer treatment into their culture, policies and processes.