Defining modelling targets.

This is the second post in the series of four where we discuss data science aspects of vulnerable customers detection. Here, we focus on setting appropriate modelling targets, i.e. what exactly we are looking for in the data.

Financial firms are encouraged by the Financial Conduct Authority (FCA) to use front line communication and data analytics to understand their vulnerable customers’ needs. Firms should provide their clients with tailored support when it is necessary for achieving fair customer outcomes.

According to the FCA, vulnerable consumers are those who, due to their personal circumstances, are especially susceptible to financial harm, particularly when a firm is not acting with appropriate levels of care. For example, somebody with a serious health condition that affects her ability to manage financial matters may require tailored support to sort out credit card debt.

From a mathematical as well as a human perspective, vulnerability is a very broad concept. This makes it challenging to understand its landscape and identify vulnerable customers consistently and at scale. It is not only about crude facts, but also about the reaction of individual people to their life situations. For example, the breakdown of a relationship may be considered a huge loss to some people but a positive change for others. As covered schematically in the latest FCA draft guidance, vulnerability is spanning through the driving event of vulnerable state and its consequences, increasing the risk of corresponding financial harms. This can be represented by the following diagram:

According to the Financial Lives 2020 survey of UK population, before the COVID-19 crisis, 6% of population exhibited health-related drivers of vulnerability, 20% had a low financial capability, 21% were in low financial or emotional resilience circumstances, and 20% of the population were experiencing major life events. In each specific case, these numbers will depend on the economic situation and the characteristics of a particular customer base.

It would be difficult to detect in the data the four categories of vulnerability drivers listed above because any of them can describe very different situations. As an example, here is a more granular view of the health vulnerability driver:

Examples of vulnerability drivers. FCA draft guidance on vulnerability, July 2020.

Examples of vulnerability drivers. FCA draft guidance on vulnerability, July 2020.

One way for a firm to proactively detect vulnerability is to utilise its experience and understanding of its customer base and its business and define major drivers of vulnerability across clients and products. It can then employ analytics and modelling to find these drivers using managed customer information.

From a modelling perspective, it seems to be more straightforward to target drivers of vulnerability in the data as they relate to factual information, while vulnerability consequences reflect human reactions to personal circumstances. However, there are a couple of exceptions and nuances to consider:

  • Some people are capable of managing their finances despite being in poor health, being elderly or having experienced a major life event. While this cohort displays signs of vulnerability drivers, they experience few consequences and have a low risk of financial harm.
  • Most of the vulnerability drivers are not directly observable in the data. A family bereavement is not normally reflected in accounts, transactions or other any mainstream analytics, though subsequent changes in behavioural patterns can be potentially detected in the data. As a general principle, the less observable the driver of vulnerability is, the longer it takes to establish a state of vulnerability. In some cases, harm can be recognised only after it has happened.
  • In the case of unobservable drivers, firms can look for vulnerability consequences in the data. However, consequences may give inaccurate and biased results because they reflect behavioural response rather than crude facts. Even when relevant information can be derived from the available data, it may be against the firm’s policies to use this information for automated decision making.
  • A number of financial firms are inclined to conduct retrospective customer outcomes testing and focus on detriment that has already happened in order to improve processes for the future. Although this can be effective, firms should endeavour to identify risks of financial harms before the damage happens and act proactively to stop or reduce it.

Organisations should aim at understanding the whole chain of vulnerability drivers, consequences and associated financial harms. This would enable them to set appropriate modelling targets and to capture vulnerability sooner. A lot can be learned by looking at past cases - for example, by working complaints data the following questions can potentially recreate the chain of vulnerability events: “what happened?”, “how did it affect you?”, and “what do you want?”. Techniques such as Natural language Processing (NLP) can be used to train predictive models that will process new conversations and complaints to predict vulnerability.

When defining vulnerability detection targets, firms should assess at what level of confidence they can detect them, then use that to design the associated decision-making process and set reasonable cut-offs. Vulnerability targets are generally a composition of multiple factors and thresholds being triggered at the same time. An example of a case that could surpass a firm's vulnerability threshold could be a combination of:

  • Demographic risk (e.g. a young individual with a low level of education);
  • Low financial resilience, detected through a combination of bank account balance history and third-party data analysis;
  • A recent major life event, inferred from behavioural analysis;
  • A credit product application that could further worsen the individual's financial position.

The above combination of factors could suggest that the individual may be experiencing transient difficulties which may worsen if future customer interactions are not carefully monitored.

It is a journey of discovery for organisations to understand their most salient vulnerability chains (drivers, consequences and financial harms). The vulnerability landscape can vary drastically between businesses and customer bases – for example an insurance business may need to look for different signals in the data than a retail bank would. While the lack of a universal answer does invariably introduce challenges, it is essential that firms develop their approach to detecting customer vulnerability. Every journey starts with a first step - in this one, it is to define a data landscape, what characteristics to identify in the data, what constitutes vulnerability and where the decision boundaries should lie.

In the following parts of this blog series we will discuss existing data-driven approaches to vulnerability detection, their strengths and limitations, and potential extensions.