Data-driven techniques for vulnerability detection.

This is the third post in the series of four where we discuss data science aspects of vulnerability detection. In the second part, we focused on setting appropriate modelling targets. In this post, we will discuss approaches to vulnerability detection that target different links in the driver-consequence-harm chain of vulnerability.

Financial services play an important role in people’s lives and many interact with them on a daily basis. However, certain life events can compromise financial stability and sound decision-making. Vulnerability describes situations when customers are at higher risk of financial harms due to their personal circumstances and may need tailored support from their financial services providers.

The Financial Conduct Authority (FCA) notes that vulnerability is not a discrete event but rather a process that is unfolding in time. It starts with vulnerability drivers that can be associated with health conditions, major life events, or low financial resilience/capability. Drivers may cause vulnerability consequences such as high-stress levels, mental pre-occupation or lack of perspective in understanding implications of financial decisions. These difficulties complicate managing one’s financial matters and increase the risk of poor and unfair financial outcomes. For example, people may be sold inappropriate products because they do not realise the implications of entering into long-term financial commitments.

Organisations generally detect drivers and trigger behaviours (consequences) during direct customer communication (for example, during a bank branch office visit, a telephone call or messages exchange). Asking general questions allows the frontline staff to understand the customer better. Further questions clarify the situation, how the customer perceives it and how it affects her capability to manage finances.  This interaction allows to triage the situation and determine whether to refer the customer to a dedicated support team.

Usually, advanced data analytics follows the human path in approaching vulnerability assessment. Analysis of audio recordings (e.g., phone calls) or text (e.g., chat logs or complaints history) can capture factual and sentiment/emotional information on personal circumstances, which is then verified by a human reviewer. As a result, it may be possible to detect one or more links in the driver-consequence-harm chain or even several links of the same type. For example, poor health is often associated with low financial resilience, which translates into two vulnerability drivers at the same time.

TrueVoice, Deloitte’s speech analytics solution, adopts this approach by transcribing phone calls and analysing their verbal content and non-verbal cues to detect vulnerability. This is achieved by combining a data-driven (bottom-up) approach which makes use of supervised machine learning techniques with a rule-based (top-down) methodology developed in collaboration with our risk and regulatory expertise.

Text and audio analytics are powerful techniques for vulnerability detection because they cover communication channels where customers discuss their concerns directly. Analysing actual customer engagement rather than MI collected about the customer increases the accuracy of any vulnerability detection methods, especially if text and audio analysis can be combined with extra available information about the customers.

This approach allows valuable information to be collected about situations where a firm can improve itself in order to serve its customers better. It also helps to remediate issues faster and benefits the customer directly by reviewing past communication and following up on cases have been mishandled in any significant way.

The main limitation of this approach is that these are point-in-time interactions that cover a limited number of customers, meaning that it is hardly able to provide a measure of vulnerability across the customer base. Voice and text analytics cover only the portion of the customer base that interacts with the bank using these channels, i.e., phoning the call centre, using webchat and e-mail and attending recorded in-branch meetings. To be honest, few of us regularly call and text our banks or raise complaints. Also, most customer journeys are designed to convey information to the bank prior to a sale, or when the customer has experienced a poor outcome, further reducing the coverage of these analytic techniques.

At the same time, people are increasingly switching to digital channels and manage their finance through internet banking accounts or banking applications where their user-app interaction becomes the main communication with the services provider. This means that customer engagement is becoming increasingly digital, leading to the creation of a new channel that can be richly analysed.

Using broader customer information in conjunction with analysis of different communication channels can allow firms to get a better view of their customer base, detect more diverse signs of vulnerability and to do it earlier. 

In the final part of this blog post series, we will suggest how vulnerable customers detection can be extended to provide a better view of the vulnerability landscape for the whole customer base.