Case Study

Case study: the need for data science in banks:

• A bank offered personal loans to 800,000 customers with savings or credit card accounts over five years (2013–2018). While the program aimed to benefit customers and increase bank revenue, it ultimately resulted in losses.

• The loan repayments were set up as monthly installments, determined by the loan amount, tenure, and interest rate. If a customer’s account didn’t have enough funds to cover an installment, the loan became “delinquent” (missed payment). Approximately 10% of customers missed their payments, leading to dissatisfaction due to hefty penalty fees.

• The bank's risk management system didn’t account for the likelihood of customers missing payments, which negatively impacted both the loan program and overall customer satisfaction. To address this issue, the bank’s management tasked the data science team with identifying customers likely to miss payments in the next three months. By doing so, the bank could proactively call these customers and remind them of their payment, testing whether such reminders could reduce missed payments.

❑ Objective: Identify the 30,000 customers most likely to miss loan payments in the next three months to prioritize telephone reminders.

A few ways to approach this problem are:

❑ Approach 1: Account Balance Threshold

• Method: Identify customers whose account balance falls below a specific amount, such as 2–3 times their monthly installment.

• Pros: Simple and quick to implement.

• Cons: Doesn’t account for account balance fluctuations (e.g., customers waiting for their salary but typically paying on time).

❑ Approach 2: Rule-Based Analysis

• Method: Analyze historical data of customers who missed payments and create rules based on common characteristics (e.g., customers aged 30–40 with an income below $2,500 are more likely to miss payments).

• Pros: Provides insights into customer behavior.

• Cons: Rules may not account for all variations and might miss some patterns.

❑ Approach 3: Machine Learning

• Method: Use historical data to build a machine-learning model that predicts the likelihood of missing payments for each customer.

• Pros: Generates a specific score for each customer and adapts to changing patterns over time.

• Cons: Requires more data and resources for implementation.

❑ Steps to Implement a Machine Learning Approach:

1

❑ Step 1: Data Collection and Preparation:

  • Gather data about customers who missed payments vs. those who didn’t since 2013.
  • Collect relevant details such as:
    • Account balance history.
    • Demographics (e.g., age, income).
    • Loan details (e.g., tenure, amount, interest rate).
  • Clean and preprocess the data for accuracy and consistency.
2

❑ Step 2: Modeling

  • Supervised Learning: Use labeled data (e.g., delinquent = Yes/No) to train predictive models.
  • Unsupervised Learning: Explore patterns without predefined targets to find useful relationships in the data.
  • Apply algorithms such as regression or decision trees to generate predictions for each customer.
3

❑ Step 3: Model Evaluation

  • Evaluate how well the model predicts delinquency by testing it on new data.
  • Metrics to check:
    • Accuracy: How often the model predicts correctly.
    • Precision: How many of the predicted "delinquents" are actual delinquents?
4

• Step 4: Experimentation

  • Test the model in real-life scenarios:
    • Call some customers flagged as likely to miss payments and track their response.
    • Compare their payment behavior with those who weren’t called.
  • Questions to answer:
    • Does calling reduce missed payments?
    • Are there customers who still miss payments despite reminders?

There are several approaches to performing an exploratory data analysis:

• Exploratory Data Analysis (EDA)

• To better understand the data:

• Tables:

Summarize key statistics like average balance, income, and payment history.

• Visualization:

Use graphs (e.g., bar charts, histograms) to identify trends.

• Correlations:

Analyze relationships between variables (e.g., income and delinquency).

❑ Final Recommendations:

  • ▪ Implement a machine learning model for precise predictions.
  • ▪ Use the model to generate a ranked list of customers with the highest risk of missing payments.
  • ▪ Conduct monthly calls targeting the top 30,000 high-risk customers.
  • ▪ Regularly update the model with new data to improve predictions and adapt to changing customer behaviors.
By integrating data science into its operations, the bank can reduce missed payments, improve customer satisfaction, and optimize its loan program. This approach also demonstrates the powerful role of data science in solving real-world financial challenges.