A call for data-driven solutions.
This post is the first in a series of four on customer vulnerability and explores challenges and considerations in developing data-driven approaches to identify vulnerable customers.
The financial services industry has significantly improved its understanding of vulnerable customers over the past few years. These days, it is broadly accepted that tailored support may be needed to ensure fair and consistent outcomes for those in vulnerable circumstances. In consultations with the public, the Financial Conduct Authority (FCA) has developed the concept of customer vulnerability and made recommendations on how to reduce the risks of financial harms for those particularly predisposed to them.
In 2019, the FCA issued the first draft of guidance on fair treatment of vulnerable customers, calling on financial services institutions to use their data to proactively identify vulnerable individuals in their customer base. The outbreak of COVID-19 added further weight to this topic, with millions of people experiencing economic and physical constraints and at the same time facing their financial needs and commitments. This prompted the FCA to publish the COVID-19 version in July 2020 with a view to finalise the guidance in early 2021. In the updated version, the FCA clarified that firms are expected to understand their landscape of vulnerability and associated customer needs. Firms are encouraged to use available data to detect signs of vulnerability to provide clients with appropriate support.
With an increasing focus on vulnerability, firms will need novel approaches to find ways of identifying and protecting their customers. This is a complicated problem to solve – and not just because it can expose organisations to further data protection and conduct risks.
By definition, a consumer in vulnerable circumstances is somebody who, due to personal circumstances, is especially susceptible to financial harm, particularly when a firm is not acting with an appropriate level of care. From a data science perspective, identification of customers showing signs of vulnerability may sound simple at the surface, however, there are a number of factors to consider:
Vulnerability as a highly abstract modelling target
It is generally possible to set a target variable in case of event-based outcomes. For example, a loan default event happens if an individual does not repay contractual amounts. However, detecting and measuring vulnerability is not straightforward as it comes in as many forms as human hardship itself. For the consistent application of machine learning models or any other automated techniques, organisations have to define modelling targets and decision boundaries on the vulnerability scale that would depend on their own customer base and products.
Issues with data quality, breadth and coverage
Data sparsity and poor data quality can hamper the simplest of modelling efforts, let alone a problem as nuanced as predicting vulnerability. Furthermore, it may be not conducive to identifying certain types of vulnerability. In a retail banking example, while it may be possible to infer low financial resilience or life events such as job loss, it may be harder to detect circumstances such as significant health problems or substantial caring responsibilities. In addition to consolidating and maintaining their data, organisations could also consider using other data sources such as clickstream, call centre and alternative data.
Technical implementation of predictive models
Generally, only a minority of clients in vulnerable circumstances are under the radar as there are strict and highly manual ways of establishing customers’ vulnerability and not all of these cases are reflected in firms’ systems. With incomplete information and a lack of consistent approaches to verify whether a person is eligible for tailored financial support, it is hard to train an accurate predictive model. This is especially true given that life situations can change quickly, and vulnerability may happen unexpectedly or pass by as a transitional stage in one’s life.
These challenges are non-trivial and arise at the start of vulnerability landscape exploration using data analytics. The solution calls for a sensitive application of data-driven techniques with careful management of ethical and compliance risks. Successfully navigating these issues can pave the way to understanding the needs of customers in vulnerable circumstances, allowing firms to deliver good consumer outcomes for customers in vulnerable circumstances consistently and at scale.
In the following parts of this series, we will discuss implementation aspects of using data science for vulnerability detection. We will define modelling targets, discuss applicable data-driven techniques and suggest how they can be extended in order to give a better view of firms’ vulnerability landscape.