Bank Term Deposit Predictions Dataset

Group Members

• Nidhi
• Jeetesh
• Ramish
• Fatima
• Yukang Lin

Overview

This project focuses on predicting customer subscription to term deposits based on historical marketing campaign data. By leveraging machine learning techniques, we aim to identify potential subscribers and optimize marketing efforts.

Dataset

Data Source

This dataset is sourced from Kaggle: Bank Term Deposit Predictions.

Description:

The dataset contains records of a bank's marketing campaign interactions with customers.

• Training set: 45,211 instances
• Test set: 4,521 instances
• Features: Age, job, marital status, education, balance, contact type, previous campaign outcomes, etc.
• Target Variable: y (Indicates whether the customer subscribed to a term deposit: yes or no)

Dataset Description

Below is a description of each field in the dataset:

Field	Description	Type
age	The age of the customer.	Numerical
job	The occupation/employment status of the customer.	Categorical
marital	The marital status of the customer.	Categorical
education	The education level attained by the customer.	Categorical
default	Whether the customer has credit in default.	Categorical
balance	The balance in the customer's account.	Numerical
housing	Whether the customer has a housing loan.	Categorical
contact	Type of communication used to contact customers (phone, cellular, etc.).	Categorical
day	Day of the month when customers were last contacted.	Numerical
duration	Duration (in seconds) of the last contact with customers during the previous campaign.	Numerical
pdays	The number of days passed by after contact from the previous campaign.	Numerical
poutcome	Outcome from the previous marketing campaign.	Categorical

Business Questions

1. What factors influence a customer's decision to invest in a term deposit?
2. Can the bank's operational status be inferred from customer usage records?
3. Can we predict term deposit subscription likelihood and optimize marketing efforts?
4. Can we classify customers into potential vs. non-potential subscribers?

Challenges

• Class imbalance: More negative instances (no) than positive (yes)
• Handling missing values and categorical variables
• Feature selection and importance analysis

Data Preprocessing

Handling Missing Values:

• No missing values detected
• Replaced unknown values in education column with the most frequent value (secondary)
• Merged month and day columns to create a proper date format

Exploratory Data Analysis (EDA)

Findings:

• Customers with higher balances and longer call durations were more likely to subscribe.
• Students and retirees showed higher subscription rates.
• Higher education levels correlated with increased subscriptions.
• Clients with no default records, no housing, or personal loans had a greater likelihood of subscribing.
• The month of contact significantly affected the subscription rate.

Model Development

Machine Learning Models Used:

1. Random Forest Classifier
   • Provided baseline feature importance.
   • Identified balance, duration, and contact type as key features.
2. Neural Networks
   • Captured complex, non-linear interactions between features.
   • Identified education and job type as additional significant predictors.
3. Logistic Regression
   • Achieved 89.88% accuracy after feature selection.
   • Used for baseline comparisons.
4. XGBoost Classifier
   • Achieved 90.55% accuracy after hyperparameter tuning.
   • Strong feature importance insights.
5. Decision Tree
   • Tuned via GridSearchCV (optimal depth: 5 layers).
   • Confusion matrix and ROC-AUC analysis.
6. Adaptive Boosting (AdaBoost)
   • Improved classification balance.
   • Increased True Positive Rate while controlling False Negatives.

Feature Importance Analysis

• Key Features:
  • balance: Higher balance → Higher subscription likelihood
  • duration: Longer call duration → Increased success rate
  • month: Certain months (Aug, Sep) showed higher engagement
  • job & education: Higher education and job types mattered
  • previous campaign outcome: Customers with previous successful interactions were more likely to subscribe

Model Performance Comparison

Model	Accuracy
Logistic Regression	89.88%
XGBoost	90.55%
Decision Tree	89.01%
AdaBoost	Improved TPR & FPR

Insights & Recommendations

• Targeted Marketing: Focus on high-balance customers with longer call engagement.
• Optimal Timing: Execute campaigns in August and September for maximum conversion.
• Customer Segmentation: Classify customers based on past interaction success rates.
• Campaign Optimization: Reduce excessive follow-ups to avoid customer fatigue.

Conclusion

• Predictive modeling improves marketing efficiency by identifying potential subscribers.
• Feature engineering plays a crucial role in enhancing model accuracy.
• Advanced techniques like XGBoost and AdaBoost offer superior classification performance.

Future Enhancements

• Further Hyperparameter Tuning for improved model performance.
• Customer Profiling & Segmentation using clustering techniques.
• A/B Testing for Campaign Strategies to refine marketing effectiveness.

Thank You! 🚀