How Machine Learning can help banks become customer-centric rock stars

Efma feature

06 June 2018

Rudradeb Mitra, Senior Advisor at Efma, outlines what steps banks should take if they want to use machine learning to attract, retain and better serve their customers.

“I think the most important thing is to have empathy and really listen to your customers. I cannot emphasize enough how important it is to not simply talk as if you are customer centric. You need to live and breathe this every day.” Ben Chisell, product director at Starling Bank.

rudradeb_art.png (62 KB)
Figure 1: World Fintech Forum, 2018 report from Capgemini and LinkedIn, which was published in collaboration with EFMA.

This article answers the question how banks can build systems that create empathy, build trust and listens to each of their customers.

In the past, banking customer experience was mostly restricted to in-branch, from the point a customer enters a bank branch to the point she leaves it. In such a scenario, it was easier to listen to each individual customer and build relations. Today, however, most customers interact with their banks via devices, which is both an opportunity and a threat. Opportunity as this gives a direct access to customers 24x7 with a lot of personalized data but a threat if banks decide to keep on using old models of customer experience and engagement.

Additionally, most of the banking solutions were built to make people more rational but that did not create empathy. Instead banking systems should be built to comport with, rather than confound, the actual psychology of decision making.

In today’s digital era, banks need to build new customer-centric models to comfort actual psychology of decision making and offer personalized communications, interactions, and acquisitions. Such models will ensure that banks can bring every individual customer back to the forefront of their operations.

But how can these models be built? Machine Learning comes to the rescue.

Unlearning and relearning machine learning 

Machine learning is perhaps the most hyped and misused term in the tech world (yes, even more than blockchain). Many people still relate machine learning to the one of the following: recognizing objects (human faces, cats, and dogs) in pictures; chatbots; humanoid robots; and even doomsday but few have a real understanding of machine learning.

So what is Machine learning?

art2.png (1.05 MB)

According to the definition in “machine learning algorithms can be used to learn patterns from data without explicitly being programmed.” In other words, machine learning is best suited to extrapolate information and patterns from past data.

Why is it useful to learn patterns from past data?

If we look at ourselves, we can see that we repeat a lot of our behaviors. There are patterns in many aspects of our lives, including: 

•   What we buy and when we buy it

•   What food we like and when we like it

•   What things make us feel good and engaged

•   What things make us feel bad and take away our energy

•   What music we like.

artr3.png (332 KB)

So if we can identify use cases where humans (or machines) have repeated patterns, machine learning algorithms can be used to predict future behaviors with reasonable probability.

What are the banking use cases?

There are three categories of banking use cases including:

1. Known use cases with known solutions: both the cases and solution are known, but machine learning can be used to improve upon existing solutions. For example, it could be used for risk assessment.

2. Known use cases with unknown solutions: the use cases are known, but the solution is not – for example how to provide a loan to people without a bank account

3. Unknown use cases with unknown solutions: these are new use cases that have never existed before and therefore the solutions don’t exist either. Such use cases arise due to change in environment. For example, in the past banks didn’t have to worry about acquiring customers online, but now that digital banks have opened, it’s a problem that all banks have to deal with.

Humans aren’t good at solving any of the above use cases because our minds are not suitable for statistical thinking and analyzing patterns in huge data sets. However, this is what we’re intuitively trying to do – all the best salespeople, marketers and even songwriters try to find patterns to repeat.

Using machine learning to learn patterns in customer experience

“The future of business is Consumer to Business, not Business to consumer. Businesses must customize for consumers.” - Jack Ma, Executive Chairman Alibaba.

Let us look at a couple of use cases related to customer experience and see how machine learning can help. We’ll look at both ‘known-unknown’ and ‘unknown-unknown’ use cases.

Use case one: Predicting customer needs and creating customized communications

Imagine that you are a customer who is thinking of buying a car and around the same time, you receive a communication from your bank informing that you are entitled to excellent loan terms if you take out a car loan in the next five days. Bingo! Won’t that make you feel like your bank values and cares for you?

artr4.png (67 KB)

Can banks build such a system? Yes, with the help of machine learning. To build this type of machine learning system, I recommend banks follow the below steps:

1. Intuitive thinking: marrying behavioral science with data science

In this step, banks need to think about how a human would solve the same problem. If banks want to build a personal touch, they need to know some information about the customer. Often, people of similar ages, genders and socio-economic status display similar behavioral patterns, so banks can use this information to create a model that can help us predict the future behaviors of other people in the same group. This is the basis of building a potential machine learning system. art5.png (88 KB)

Banks could focus on many customer ‘features’, but for the sake of this article, we will focus on personal data (age, sex), address and area code, and financial data (monthly income, average monthly expenditure and so on.

2. Collecting data

In this particular use case, banks should already have access to the data mentioned above.

3. Selecting the algorithm

If banks had to model a series of actions, recurrent neural networks (RNNs) seem to be the obvious candidate. One of the appeals of RNNs is that they can be used to connect previous information to the present task. Long short-term memory (LSTM), for example, is a class of RNN that can be trained with past sales data and used to build an approximation model of the process. Another simpler algorithm that can be used is Word2Vec, which is a simple neural network with one hidden layer that can be used to model semantic relations between various parameters. For example, it can produce a model for the following situation: If person X (with a certain persona) spends US$200 and earned US$500, then he is likely to do activity Y this month.art6.png (43 KB)
4. Development

Tensorflow has libraries for both LSTM and Word2Vec, so I’d recommend that banks use these libraries to train the neural network with their data.

Use case two: New customer acquisition

Can banks use machine learning to identify potential new customers before they reach out? To build a system for this, banks should follow these steps:

art7.png (36 KB)

1. Intuitive thinking

As in the previous use case, we need to think how a human would solve the same problem. The best salespeople identify new prospects based on their similarities to previous prospects. In his book New Sales Simplified, consultant, coach and author Mike Weinberg advises that banks look at the following criteria for existing customers when trying to find new prospects:

· Who are your best customers?

· Why did they become customers?

· Why are they still buying from you?

· Why do prospects choose you over other similar products and banks?

2. Collecting data

Banks can translate the above set of questions into data that can be understood by machines: 

· Who is your best customer: age, location, income data

· Why they became customers: data about location, the degree of reference and product features

· Why they still buy from the bank: data about customer service, location, and product features

· Why they choose the bank over others: data about the degree of reference, product features and location.

This data should be publicly available for any business customers, but it needs to be collected for individual consumers. This article gives a good overview of the best practices for receiving user data. The main points are that banks should:

a. Incentivize users: banks must create incentives that encourage people to share their data by providing them with a better experience or way to save their money.  For example, when I was part of the team that built the Road Skippers mobile service, we gamified the system by giving users incentives like free movie tickets and coffee. This meant people were very happy to share their data. 

b. Not force users to adopt a new disruptive technology: once banks have collected users’ data, they must look to give back value to the users.

c. Not try to change behaviors: if banks are building a tool that requires users to fundamentally change their behavior to use it, that’s not good. 

3. Selecting the algorithm

Once banks have gathered the data, we need to group a new prospect into the category of previous customers. Here we can use classification algorithm in machine learning. Classification is defined as the grouping of things by shared features, characteristics and qualities. For example, the figure below shows geometric shapes being classified based on their similarities.

art8.png (43 KB)

4.  Development

Here also Tensorflow has the library for classification and it is advised to use that to train your network.


Coding is just a small part of building a machine learning system – and perhaps the least difficult part. The bigger challenges for banks are working out how to gather data, how to incentivize users, what architecture to build, and where to use machine learning. Banks must also change their mindsets, skill sets and infrastructure if they want to realize the potential of machine learning. They need to understand where it should be applied and ensure that they have trained people in place to execute it.

Rudradeb is the Efma Senior Adviser in Machine Learning. Learn more about Efma Advisory Services.

Keywords : Big data , Customer loyalty , AI/Robotics , Customer acquisition

Geography : International