How to detect 70% of churning bank customers with big data analytics and ML

A bank spends $100 or more on average to attract one customer, according to Unicom. Retaining a customer base is much cheaper, so companies not only strive to attract a new audience but also to retain the existing one. The main question is how to identify customers who are likely to leave.

The answer lies in customer data, which swirls in vast arrays within every enterprise company's processes. Eastwind analyzed the data of a Russian bank using its own ML models and identified churning customers whose decisions could still be influenced by the right NBO (Next Best Offer).
 

Key Results

 

Why to predict churn 

Predicting customer churn helps to:

● Reduce the cost of attracting new customers
● Increase the loyalty of the existing base
● Reduce financial risks since income from long-term customers is more predictable
● Increase customer LTV (Life Time Value)



The customer gradually decides to switch to another company: offers from your competitors flicker around them constantly, pushing them to take this step. Big data analytics enables you to identify when doubts first arise in the consumer. It records essential indicators of the customer's behavior that show they intend to leave but have not yet made a final decision.
 

Project format

In a pilot project with a bank whose customer base is estimated at hundreds of thousands, our client sought to understand how effectively customer churn could be predicted and whether this could be achieved with a limited data set. Eastwind specialists created a predictive model that forecasted churn and compared the analysis results with actual data. 

In this project, we used EW DataFlow, a platform that works with big data and AI/ML models. We carried out the project in four stages:

1. Setting the task: what kind of customers  we are looking for
2. Collecting information: what data we are interested in
3. Data analysis: what signs we use to determine potential churn
4. Summing up: how accurate the prediction is

Learn more about DataFlow

Stage one: setting the task

The client identified three types of churn within their bank:
 

1. Intentional: the customer decides to switch to a competitor
2. Rotational: the customer leaves for reasons beyond their control; for example, they move 
3. Forced: the bank terminates the contract with the customer

Together with the bank’s specialists, we decided to consider only the first two groups as churning customers.

Then, we determined how much time before the churn date could still be effective for customer retention measures. We found that the customer is usually already confident in their decision two months before the expected churn, making it nearly impossible to retain them. Therefore, we analyzed the customer's actions five months before the churn. Three of them — the earlier ones — were included in the machine learning model (highlighted in green in the diagram).


Events that led the customer to change banks occurred during the initial months. We needed to identify what these events were and how to track them in advance
 

Stage two: collecting information

Once we had determined what exactly constituted churn and when it needed to be predicted, we began collecting information for analysis.
 

1. The bank downloaded customers’ data from its information systems for nine months.
2. The data was depersonalized and encrypted to prevent the transfer of personal information.
3. The bank transferred data to Eastwind specialists in text files.
4. We converted all files to the required format and structured them for further work in EW DataFlow.

We have collected the history of customer activity in an anonymous form:

● General information about customers 
● General information about accounts
● Agreements with customers, including tariffs
● Account balances and movement of funds for each day
● Transfers within the bank and to external counterparties
● Transactions with bank cards
● Communications between the bank and customers
● Applications for account closures
 

Stage three: data analysis

Filtration

After collecting the data, we filtered it, retaining only what was necessary for developing the model. For example, we filtered out unreliable customers whom the bank did not plan to retain. We also excluded those with a transaction history of less than 30 days.

After filtering, several hundred bank customers remained in our database. Based on information about their activities, we began developing an algorithm to train the model.
 

Preparing the training sample

Eastwind analysts formed a training sample to forecast churn — a group of customers about whom we know whether they have churned. We visualized the activity history of each participant in the sample and compared the behavior graphs of customers who have and have not churned from the bank.
  The number of transactions of a customer planning to change the bank gradually decreased. As a rule, churning customers had no debit turnover for over a month and minimal credit turnover.


Green dots are credit transactions, and orange dots are debit transactions. Here and further on, the upper graph shows the transactions of customers who did not plan to change the bank; the lower one shows the churning customers
 

How activity decreases

If customers regularly used the bank's services, their activity line on the graph would be far from zero. The less active the customer is, the more likely they were to churn.


Activity became nearly zero for a customer who wanted to change the bank (bottom graph) Another marker of churn was that the customer hardly replenished the balance before leaving the bank.


The black line shows the balance values, and the blue line smoothes the black line. For a churning customer, account replenishments became increasingly rare and insignificant
 

Activity analysis 

In addition to studying financial activity, we examined behavior patterns in customer groups: who interacted with whom and how often. To do this, we studied their social graphs.


Blue dots on the graph indicate the bank's customers and red dots indicate competitors' customers. Lines show transactions between them

One indicator of potential churn is the lack of transactions with customers of the same bank. If a person has no interactions within your bank, but there is movement of funds with other counterparties, then the customer is prone to churn. On the graph, this may look like this:

"Our client particularly liked the idea of ​​visualizing customer interactions. According to them, the traditional format lacks clarity, so it is sometimes difficult to notice the obvious. As part of the pilot, we wrote a small program in which they can enter the contract ID and see who it interacts with."
 

Pavel Oliver, Development Director, Eastwind

Creating an ML model

Stage four: evaluation of the result

"The accuracy of the model could improve further with additional data, such as information on customer activity in online banking. We would know the emotional coloring of requests, the speed of processing applications, and the results of surveys of churning customers. In that case, we could determine the fact and the hypothetic cause for the churn."
 

Andrey Pluschenko, Head of Data Analysis Group, Eastwind

What other tasks can EW DataFlow be used for?