💳 RetainCred

AI-Powered Credit Card Churn Prediction, Segmentation & Personalized Offer Engine

🔗 Live App: retaincred-app

---

Overview

RetainCred is an end-to-end machine learning product that predicts credit-card customer churn, segments users by behaviour, and generates personalized retention offers using AI.

The system combines:

• Predictive modelling
• Customer segmentation
• Recommendation engine
• LLM-generated marketing messages
• Full deployment pipeline

---

Problem Statement

Customer churn is one of the biggest revenue losses for credit card companies.

Instead of reacting after churn happens, RetainCred answers:

• Which customers are likely to churn?
• What type of users are they?
• What retention offer should we give?
• How do we communicate it personally?

---

System Architecture

User Input → Feature Engineering → Churn Model (XGBoost)
                                 → Segmentation (KMeans)
                                 → Offer Recommendation
                                 → LLM Message Generation
                                 → Streamlit App

---

Dataset

Kaggle Credit Card Customers Dataset ~10,000 customers with demographic, behavioural and transaction data.

---

Machine Learning Pipeline

1️. Churn Prediction Model

Models evaluated:

• Logistic Regression
• Random Forest
• XGBoost (final model)

Cross-validation results:

Model	ROC-AUC	Recall
Logistic Regression	0.93	0.63
Random Forest	0.99	0.81
XGBoost	0.993	0.86

Final performance:

• ROC-AUC: 0.993
• F1 Score: 0.91
• Recall (Churn detection): 86%

Threshold tuning performed to optimise recall–precision trade-off.

---

2️. Customer Segmentation

KMeans clustering using behavioural features:

Features used:

• Transaction count & amount
• Credit utilisation
• Relationship count
• Engagement metrics
• Derived features (avg spend/txn, credit used ratio)

Segments discovered:

• High Value Spenders
• Credit Revolvers
• Low Engagement Users

---

3️. Risk Tier Classification

Churn probability is converted into:

• Low Risk
• Medium Risk
• High Risk

This allows business-friendly decision making.

---

4️. Offer Recommendation Engine

Offers generated based on:

• Risk tier
• Spending behaviour
• Engagement level
• Credit utilisation

Examples:

• Fee waivers for high-risk users
• Cashback offers for active users
• Reward boosters for loyal customers

---

5️. LLM Message Generation

HuggingFace Inference API (FLAN-T5) generates personalised messages:

Example output:

“We value your loyalty! Enjoy bonus reward points on your next purchases.”

---

Production Engineering

Backend

• FastAPI inference service
• Modular pipeline architecture
• Separate preprocessing pipelines
• Saved model artifacts & scalers

Frontend

• Streamlit interactive dashboard
• Real-time predictions

DevOps

• Docker containerisation
• GitHub Actions CI pipeline
• Environment variable secret management
• Cloud deployment (Streamlit Community Cloud)

---

Run locally with Docker

docker build -t retaincred .
docker run -e HF_TOKEN=your_token -p 8501:8501 retaincred

---