Predicting Bank Customer Churn Using a Random Forest Classifier With Engineered Risk Features

In the world of banking, money is not just moved; it is held in trust. But that trust is fragile. For a retail bank, the most expensive mistake is not making a bad loan, but losing a good customer. Acquiring a new person to open an account costs significantly more than keeping an existing one happy. This reality has turned customer retention into a high-stakes game of prediction. Banks now look to the past to guess the future, asking a simple but vital question: which of our customers is about to walk away? To answer this, researchers have turned to machine learning, a field where computers learn to recognize patterns in vast amounts of data without being explicitly programmed for every rule. Instead of a human analyst reading thousands of files, a computer scans for subtle signals—like a sudden drop in account activity or a specific age group—that often precede a departure. The goal is to spot these signals early enough to intervene, offering help or a better deal before the customer has already decided to leave.

A recent study by Amula Venkatesh and Dr. Indira at Chaitanya Bharathi Institute of Technology in India tackles this challenge head-on. They focused on a dataset of 10,000 customers from a European bank, a group where nearly 38 percent had already left the bank, a figure that represents a significant loss of revenue. The researchers built a digital system designed to predict who would leave next. They did not rely on a single, simple rule, such as "if the balance is low, they will leave." Instead, they used a method called a Random Forest. Imagine a forest of decision trees, where each tree looks at the customer's data from a slightly different angle. One tree might focus on age, another on how many credit cards they hold, and a third on their location. By combining the opinions of hundreds of these trees, the system reaches a consensus that is far more accurate and reliable than any single tree could provide on its own.

To make this system sharper, the researchers did not just feed it the raw numbers found in the bank's records. They created new, "engineered" clues by mixing existing facts together. They calculated how much of a customer's salary was sitting in their bank account, a ratio that reveals how dependent the person is on that specific bank. They looked at how many products a customer held relative to how long they had been a client, measuring the pace of their engagement. They even multiplied a customer's age by the number of years they had been with the bank to see if life stage played a role in their loyalty. These new combinations acted like a magnifying glass, highlighting hidden relationships that the raw data alone had missed. When the computer was trained on this enriched information, it learned to see the signs of departure with remarkable clarity.

The results of this training were striking. When tested on a group of customers the system had never seen before, it correctly identified who would stay and who would leave with 91.4 percent accuracy. More importantly, it caught 87.8 percent of the people who were actually going to leave, a crucial metric for a bank that wants to save its customers. The system performed better than older, simpler methods like basic linear models or single decision trees, and it came very close to the performance of much more complex and computationally heavy systems. The researchers found that age was the strongest single predictor of leaving, followed closely by how actively a customer engaged with their products. They also discovered that customers in Germany were leaving at nearly twice the rate of those in France or Spain, a regional difference that suggests local factors are at play.

Perhaps the most significant part of this work is not just the prediction, but the explanation. In the past, a computer might say "this customer will leave" without saying why, leaving bank staff confused and unable to act. This new system includes a feature that acts like a spotlight, showing exactly which factors pushed a specific customer toward the "at-risk" category. If a customer is flagged as high risk, the system can tell the bank manager that it is because the customer is between 51 and 60 years old and has stopped using their products, even though they still hold several accounts. This clarity allows bank staff to move beyond guessing. They can now offer targeted help, such as a financial review for someone nearing retirement or a re-engagement campaign for someone who has become inactive.

The researchers took this model out of the computer lab and built a simple, interactive dashboard that any bank employee can use. There is no need for coding knowledge. A manager can type in a customer's details, and the screen will display a probability gauge, a color-coded risk level, and a plain-language explanation of the risk factors. This tool turns a complex mathematical result into a practical guide for daily business decisions. The study suggests that by focusing on engagement rather than just ownership, and by paying attention to specific life stages and regional differences, banks can significantly improve their ability to keep customers. While the model is powerful, the researchers note it is based on a single snapshot of history and that future work could include real-time transaction data to make predictions even more timely. For now, this system offers a clear, data-driven path to understanding why customers leave and how to keep them.