STATSML 604: Car Loan Propensity Prediction using Logistic Regression

Overview

Predicting whether a customer is likely to take a car loan can significantly improve a bank’s ability to design targeted campaigns, manage credit risk, and optimize resource allocation.

Stage 1: Awareness

• The customer realizes a need — perhaps their current car is unreliable, or their lifestyle has changed (family, job move, etc.).

• Signals:

  • Search behavior (Google/search logs)

  • Browsing car loan info on bank websites

  • Interactions with car dealerships or vehicle-related services

Stage 2: Research

• They begin comparing options, calculating EMIs, and checking eligibility.

• Signals:

  • Clicks on car loan calculators

  • App logins increase

  • Engaging with bank agents or chatbot loan FAQs

  • Increasing balance inquiries

Stage 3: Financial Readiness

• Evaluating “Can I afford it?”

• Signals:

  • Growth in monthly net income

  • Stable or rising credit score

  • High cash reserves (in savings + checking)

  • Existing auto loans closed or paid off

Stage 4: Application Intent

• They are ready to apply.

• Signals:

  • Clicking on “Apply Now” or loan inquiry forms

  • Uploading documents to the portal

In this tutorial, we build a simple logistic regression model to classify a customer's car loan propensity focusing on profile and bank transaction behavior.

Use Cases

• Targeted Campaigns: Focus offers on high-propensity segments

• Loan Eligibility Filtering: Pre-qualify candidates automatically

• Customer Risk Profiling: Understand financial behavior deep.

Prerequisites

PREP 400: DBVisualizer SQL Editor Setup for Data DistillerSTATSML 600: Data Distiller Advanced Statistics & Machine Learning Models

Goals

Build and deploy a classification model that predicts if a customer is likely to opt for a car loan, based on:

• Demographic and behavioral data

• Financial account balances

• Derived income and loan features

Sample Dataset

Fields & Description

Field Name	Data Type	Description
customer_id	STRING	Unique identifier
age	INT	Customer age
gender	STRING	Male / Female / Other
marital_status	STRING	Married / Single / Divorced
employment_status	STRING	Employed / Self-employed / Retired etc.
annual_income	DECIMAL	Yearly income
credit_score	INT	Credit bureau score
checking_balance	DECIMAL	Balance in checking account
savings_balance	DECIMAL	Balance in savings account
monthly_debit_volume	DECIMAL	Monthly average spending
monthly_credit_volume	DECIMAL	Monthly average income
loan_history	ARRAY<STRUCT>	Past loans (type, amount, status)
existing_auto_loan	BOOLEAN	Existing car loan
owns_vehicle	BOOLEAN	Whether customer owns a car
propensity_car_loan	FLOAT	Target: Likelihood (0-1) of taking a loan

If you were to go and explore the sample data, it should look like this:

A simple data exploration query with DB Visualizer

Derived Features

These features are engineered to improve model performance:

Feature	Formula / Logic
savings_to_income_ratio	savings_balance / annual_income
debt_to_income_ratio	(monthly_debit_volume * 12) / annual_income
avg_monthly_net_income	monthly_credit_volume - monthly_debit_volume
loan_count	COUNT(loan_history)
previous_auto_loans	COUNT WHERE loan_type = 'Auto'
good_credit_flag	credit_score >= 700
high_cash_reserve_flag	checking_balance + savings_balance > 10000

Select
customer_id,
cast(age as int) age,
gender,
marital_status,
employment_status,
cast(annual_income as decimal(18,2)) annual_income,
cast(credit_score as Int) credit_score,
cast(checking_balance as decimal(18,2))checking_balance,
cast(savings_balance as decimal(18,2)) savings_balance,
cast(monthly_debit_volume as decimal(18,2)) monthly_debit_volume,
cast(monthly_credit_volume as decimal(18,2)) monthly_credit_volume,
--Derived Features
ROUND(savings_balance / annual_income, 3) AS savings_to_income_ratio,
ROUND((monthly_debit_volume * 12) / annual_income, 3) AS debt_to_income_ratio,
(monthly_credit_volume - monthly_debit_volume) AS avg_monthly_net_income,
CASE WHEN credit_score >= 700 THEN TRUE ELSE FALSE END AS good_credit_flag,
CASE WHEN (checking_balance + savings_balance) > 10000 THEN TRUE ELSE FALSE END AS high_cash_reserve_flag
from customer_bank_data;

This should look like the following:

Derived features along with the base features

Auto-Labeling of Training Data

Feature	Weight	Explanation
Not having an auto loan	0.30	More likely to consider buying
Doesn’t own a vehicle	0.20	May need a car, hence loan
Good credit score	0.20	More eligible for credit
Income > $60K	0.10	Likely to get approved
Checking balance > $2K	0.10	Has funds for down payment
Net income > $2K	0.10	Better repayment capacity

Auto-labeling of training data

Model Definition

DROP MODEL IF EXISTS customer_propensity_using_LogisticRegression;

CREATE MODEL customer_propensity_using_LogisticRegression
TRANSFORM (
  vector_assembler(array(
     age,
     annual_income,
     credit_score,
     checking_balance,
     savings_balance,
     monthly_debit_volume,
     monthly_credit_volume,
     existing_auto_loan,
     owns_vehicle,
     savings_to_income_ratio,
     debt_to_income_ratio,
     avg_monthly_net_income,
     good_credit_flag,
     high_cash_reserve_flag
  )) features
)
OPTIONS (
  MODEL_TYPE = 'logistic_reg',
  LABEL = 'propensity'
)
AS
SELECT
  *,
  CASE WHEN propensity_car_loan > 0.5 THEN 1 ELSE 0 END AS propensity
FROM vw_customer_bank_profile_data_train
ORDER BY RANDOM()
LIMIT 50000;

Model Evaluation

SELECT * FROM model_evaluate(
  customer_propensity_using_LogisticRegression, 
  1, 
  (
    SELECT
      *,
      CASE WHEN propensity_car_loan > 0.5 THEN 1 ELSE 0 END AS propensity
    FROM vw_customer_bank_profile_data_train
  )
);

Results

The query should return the folllowing:

Query results on the evaluation

Metric	Value
AUC ROC	0.9362
Accuracy	0.9361
Precision	0.9367
Recall	0.9372

Predict on New Customers

SELECT * FROM model_predict(
  customer_propensity_using_LogisticRegression, 
  1,
  (
    SELECT
      *,
      CASE WHEN propensity_car_loan > 0.5 THEN 1 ELSE 0 END AS propensity
    FROM vw_customer_bank_profile_data_predict
  )
);

The query looks like the foilowing:

Predictions on the same dataset