When dealing with a large customer base, such as in the case of banks and insurance companies, it is not easy to understand how to serve them best. Each customer has their own problems, needs, interests, and sensitivities.
But how can marketing strategies or entire business plans be developed for hundreds of thousands of customers?
The solution lies in using “marketing personas” (or “customer personas”), which help understand customers by identifying their needs, relevant characteristics, and recurring behaviors.
Marketing personas are customer profiles that represent broad segments of the customer base. This simplifies the reality and helps formulate effective business strategies.
Segmenting customers solely based on income or, at most, adding age as a dimension is somewhat limiting, isn’t it? It’s obvious that a customer is more complex than their bank account and date of birth.
So, how can we construct these marketing personas in the financial services sector to simplify processes while still being representative of customers?
Building marketing personas for the financial services sector is different from doing so for consumer goods or any other sector: although the subject remains the same, it is the change in perspective that matters.
Imagining a customer abstractly as a cylinder, one company may want to view it from the side, perceiving a rectangle, while another company may analyze it from above, seeing a circle. In other words, the angle of approach is crucial.
For example, let’s consider a person who is accustomed to mountain running. Their characteristics would be defined differently when seen by brands in the outdoor industry such as Salomon or Garmin, compared to the characteristics that define them as a customer of their bank. Yes, there may be some intersection in certain demographic and economic aspects, but not much else. Considering the right perspective means that the right data is needed.
Specifically, it requires highly relevant information for a bank or insurance company (which is typically not found on Instagram). And then, this information needs to be selected and combined through “feature selection” and “feature engineering” to ensure the perspective is accurate.
With the advent of Machine Learning or AI, marketing personas can be identified and monitored based on data. Unsupervised Machine Learning, particularly clustering, is the typical tool for this task. The idea is to explore customer data, searching for similarities and differences, and grouping similar customers together. Simple, right?
In unsupervised Machine Learning, there is no underlying truth in the form of a response variable Y that allows estimating the relationship Y = f(X). Only X variables are available. However, we still need to estimate f(X), which represents the function that segments customers and identifies marketing personas. The algorithm must find structures and information in the data, improvising.
In Machine Learning in general, there are two main sources of knowledge:
When labeled data is lacking (as in the case of clustering), only the second source remains, which relies on specific Machine Learning and business knowledge possessed by the team.
There are numerous choices to make, and this is not the place for an extensive discussion on algorithm selection, parameter and hyperparameter estimation, and so on. However, I want to give you an idea (albeit simplified) of the variety of methods available and the vast array of errors that can arise from naive clustering. The following image illustrates what happens when some of the available algorithms (columns) are applied to simple two-dimensional artificial datasets, known as “benchmarks” (rows). The different clusters are differentiated by colors. The lesson: it is easy to make mistakes, and quite significant ones at that.
The absence of the “underlying truth” (Y variables) also means that verifying the validity of the results is not straightforward. You might say, “Just look at the data.” Well, the problem is that, for example, at Virtual B, we estimate financial marketing personas based on hundreds of variables. And we humans have a limitation: we can visualize things in a 3D space at best, while here we have an N-dimensional space, with N >> 3.
So, how can we assess the quality of our customer base segmentation when representing it in a 100-dimensional world using only 2D or 3D graphs?
To gain an understanding of the reasonableness of the results through visual inspection of the data (which is quite useful since the human brain is capable of great things), dimensionality reduction methods are necessary. These involve clever mathematical transformations such as rotations and compressions, allowing the original data to be represented in a reduced space. Data with 50, 100, 1000 dimensions (or more) are “squeezed” and forcefully projected into 2D or 3D.
There are many techniques to accomplish this, such as factor analysis, t-SNE, UMAP, and PACMAP. Unfortunately, these complex transformations can lead to the appearance of distinct aggregations and clusters when they do not actually exist. In other words, we enter potentially psychedelic topological spaces for a Sapiens accustomed to living in 3D.
Categorical variables are labels, not numbers. For example, profession, gender, residence, and so on. Let’s observe what happens when we simulate two completely independent variables that do not form any clusters, but:
Looking at the graph below, you can see that the categorical variable of male/female acts as an attractor, grouping the other variable. It appears that there are two clusters (the two lines), or two personas. Unfortunately, they are not meaningful. To avoid statistical hallucinations, clustering techniques suitable for categorical variables are required, such as k-prototypes.
I could keep on listing points of consideration, but I will stop here, hoping that I have been able to make the discussion as clear as possible.
As we have seen, simply relying on machine learning tools is not enough to enhance our understanding of customers. If you choose to do so, always seek companies that specifically operate in the financial sector and understand its systemic and organizational challenges. This is because clustering systems offer great knowledge, but only when used correctly. Otherwise, they will be superficial technicalities that do not yield real results.
This is what sets Wealthype apart from other companies in the Italian market. The services and features we have been offering for years to wealth management professionals, as pioneers in Italy, are born from a deep understanding of the financial services world, its challenges, and its untapped potential within the territory.
If you are interested in this topic, please contact us through the link below and discover how to apply Machine Learning principles to your business processes.