Skip to content
← All work

Retail Banking · Machine Learning · Risk Analytics

Predictive Churn Risk Intelligence System

Retail banks lose valuable customers before anyone in the branch network can see the warning signs in the data.

18
Engineered features
4
Models benchmarked
5
Evaluation metrics
SHAP
Explainability

Overview

A full-stack churn prediction system for retail banking that moves from raw customer data through an automated ETL pipeline to explainable risk scores, AI-generated retention strategies and a supervisory report.

Why the problem matters

Retention economics in retail banking are asymmetric: identifying an at-risk relationship early is far cheaper than reacquiring it. But a churn score on its own is not actionable — a relationship manager needs to know which customers, why, and what to do next.

Data & inputs

Customer-level retail banking records covering demographics, product holdings, balance behaviour and engagement signals, ingested through an automated ETL pipeline that handles cleaning, type validation and consistency checks before modelling.

Feature engineering

18 features were engineered from the raw customer table to capture behaviour rather than static attributes — balance dynamics, product mix, engagement recency and derived risk indicators.

Model selection

Four classification models were benchmarked side by side instead of committing to one family early.

  • Logistic Regression
  • Decision Tree
  • Random Forest
  • Gradient Boosting

Evaluation

Each model was scored across five metrics so the trade-off between catching churners and flooding the retention team with false positives stays explicit.

  • Accuracy
  • Precision
  • Recall
  • F1-Score
  • ROC-AUC

Explainability

SHAP waterfall visualizations expose the per-customer contribution of every feature to the predicted risk, so a reviewer can audit an individual decision rather than trusting an aggregate score.

Product & application layer

The model is wrapped in decision-support modules built for the people who act on the output.

  • Capital Flight Risk module for balance-weighted exposure
  • What-If Simulator for counterfactual scenario testing
  • Gemini API-generated personalized retention strategies
  • CRO-style supervisory report generator
  • Macro-Liquidity Simulator

Key technical learnings

Benchmarking across model families made the accuracy-versus-interpretability trade-off measurable, and SHAP turned an opaque ensemble into something defensible in a risk review. Wiring a generative layer on top of a statistical model required strict grounding: the language model explains and recommends, it never produces the score.

Business & decision relevance

The output is a ranked, explained and segmented view of retention risk with a suggested action per customer — the format a retention or risk function can act on directly. This is a self-directed project built on a retail banking dataset; it has not been deployed inside a bank.