A regional bank wanted to reduce attrition to retain millions of dollars that flee when customers close their accounts. Could machine learning predict which checking accounts are likely to close so that they could change the outcome? Read their story.

Challenge

With the banking industry in flux, disruptive competitors grabbing market share, and customers raising the bar on what kind of experiences they expect, banks must find ways to attract and retain customers on a level never known before. Today’s banking customers grow impatient more quickly than in the past, and if they are unhappy with an experience, their loyalty is fleeting.

One constant is that each year most banks lose about 10% of their account deposits due to customers closing their accounts, according to CDC/NCHS National Vital Statistics. Of that segment, 50% leave because they are dissatisfied with the bank’s service, fees, rates, products, or lack of convenience. The other 50% leave due to events that the bank cannot control, such as death, divorce, or displacement.

For example, if a bank handles $700 million to a billion in deposits annually, nearly $100 million in capital walks out the door each year due to customers closing their accounts.

A regional bank saw an opportunity to reduce attrition in this area. This long-time client wanted to be able to predict which checking accounts were likely to close within the next 90 days so they could take action to retain the customer. They knew machine learning could provide them with that data, but they had never leveraged ML before.

Machine learning (ML) is a data science technique that analyzes massive quantities of data, especially historical, to discover trends and insights and rapidly predict future behaviors and outcomes. The technique lets the data learn from itself, free of human bias or the need for explicit instruction.

Traditional analytics tools don’t have the capability to rapidly uncover patterns when there are billions of datapoints to be analyzed, nor can humans identify patterns in such large quantities of data, not to mention in real time. This is the value of machine learning in enabling your data to be a market differentiator, and that’s why this bank wanted to explore a proof of concept through a Fusion Alliance Machine Learning Jumpstart.

on-demand webinar

How to increase your customer wallet share using machine learning

Get Access

Solution

Our team had more than one goal when we began this jumpstart. Foremost, we wanted to develop a machine learning model for this bank to reduce deposit attrition. We also wanted to support the education of the client’s team on the key elements of machine learning, such as the process for training models and the process for generating predictions. This would occur organically as we worked side by side and guided them through the journey.

The bank was interested in machine learning because it decreases the risk and expense of traditional analytics by allowing the data to speak for itself. They additionally wanted to understand the key metrics for evaluating the ML models.

Innovation

  • Use case identification. Prior to beginning our technical work, we explored a variety of use cases in a workshop with the bank’s business and technical stakeholders. Our team rated the potential use cases on different criteria such as how complex it would be to develop the model, what data was available, and the value impact to the bank. Together we agreed upon the deposit attrition proof of concept, deciding it would drive maximum predictive value with minimal risk. That would be the basis for the succeeding steps.

  • Data processing. We performed an inventory of existing data, sourced the data, cleaned it, and loaded it in the target on-premises environment where the models would be developed. We provided the option of loading the models in the cloud to enable additional ML models and more complex computations.

  • ML model development. Within three weeks, we began to engineer the machine learning models, choosing the subset of data most relevant to the question, “Which checking accounts are likely close in the next 90 days?” We selected the ML algorithms, then trained and tuned the ML model. We then met with the bank’s stakeholders to present metrics to measure the model’s success, focusing on KPIs.

  • Model insights integration. In this step, accounts at high risk of closing are referred to the bank’s retention team, and Fusion helps expose these insights so that the bank can take action. The end-to-end process to generate daily predictions using real-time data can be accomplished in less than an hour using the bank’s current infrastructure.

As a next step, optimization can occur. In this phase, model efficacy is captured, and the bank can optimize and expand the use case or use the model as a template that can be modified for one of the other vetted use cases.

The entire proof of concept took eight weeks, and the bank is now in possession of machine learning models that can be implemented in marketing campaigns in the next phase.

Outcomes

With machine learning models in place, the bank’s immediate goal is to develop a deeper understanding of the technology. In the next phase, they will  automate the entire marketing pipeline and feed predictions to Salesforce or another enterprise platform.

Currently, the models are set to identify customers who will leave within 90 days at 83% precision. In addition, the bank’s team now understands all the key elements of machine learning and has a meaningful list of indicators of attrition.

Jumpstart defined attributes of accounts likely to close
ML models predict customers likely to leave with 83% precision
Enabled opportunities to increase capital