How to Keep Valuable Clients: A Predictive Model for Bank Customer Churn

https://github.com/d-pap/bank-churn

https://medium.com/@dpapcodes

Purpose

Banks typically spend 5 times more acquiring new customers than retaining existing ones. For this project, I analyzed a dataset of 10,000 customer records from a multinational bank to uncover what causes customers to leave and built a predictive model to identify those most at risk of leaving. This gives banks a practical tool to proactively retain customers before they churn.

Insights

My analysis revealed four clear signals of customer churn risk:

1. Women customers churn ~10% more often (possible gaps in product fit or service).
2. Older customers (51-60) churn at nearly 30% (key demographic approaching retirement).
3. Inactive customers with poor credit scores churn ~50% more frequently.
4. Customers with exactly two products churn least, highlighting a retention "sweet spot."

Potential Impact

To see how predictive modeling translates directly into savings, let's consider a simplified, realistic scenario:

Baseline. Say the bank has 50,000 customers. About 20% churn each year, so 10,000 leave. Each lost customer costs roughly $200 in yearly profit, so churn wipes out about $2 million a year.

Current approach (no model). The bank calls everyone. A phone-or-email push is estimated to cost about $20 per customer. Assuming it persuades roughly 5% of would-be churners to stay—about 500 people, we get:

• Cost: 50,000 x $20 per customer = $1 million spent
• Profit saved: 500 customers x $200 = $100,000 profit saved
• Net result: $100,000 - $1,000,000 = -$900,000 loss

Targeted outreach (using the model). My model flags the riskiest 15% of the base (7,500 people). If outreach keeps 15% of the true churners, 518 customers stay (3,450 x 15%).

• Cost: 7,500 x $20 = $150,000 spent
• Profit saved: 518 customers x $200 = $103,600 profit saved (still a slight upfront loss)
• Immediate net = $103,600 - $150,000 = -$46,400 loss

You might be saying "Derek, that's still losing money..." but we're not done yet!

Getting new customers isn't free. This is known as the Customer Acquisition Cost (CAC) and for traditional banks can be anywhere in the range of $500 to $700. So, retaining those 518 customers also avoids that cost and we can factor it in to our projections:

• Additional savings: 518 x $500 = $259,000 saved
• Total net gain: $259,000 – $46,400 = +$212,600 overall gain

Bottom line: This turned a $900,000 loss into a $200,000 gain while bothering only a fraction of the customer base.

Tools

• Python, pandas, scikit-learn, matplotlib/seaborn for visualization

Behind the Scenes

I tested several machine learning models and selected the Random Forest model because it balanced accuracy, interpretability, and business relevance.

Model performance:

• Accuracy: 85% (overall predictions correct)
• Recall (key metric here): 46% (model identifies ~half of true churners)

Why prioritize recall?

Accuracy alone is misleading: simply predicting "no churn" would yield 80% accuracy due to class imbalance (80% retained vs. 20% churned). Recall specifically measures how effectively the model identifies the customers who actually churn.

Although 46% might seem moderate, it's significantly better than random guessing (20%). Our targeted list drastically improves efficiency, making every retention call nearly three times as likely to reach a genuine churn risk (69% vs. 20%).

Want More?

If you are curious about the technical details and insights, you can view my full analysis on GitHub or my in-depth technical report on Medium.