Bank Customer Churn Prediction

Project Overview

Customer churn in banking is a direct revenue loss: acquiring a new customer costs 5–7× more than retaining an existing one. This project builds a churn prediction system on 10,000 bank customers to identify high-risk customers before they leave — enabling the bank to deploy targeted retention interventions.

The dataset has an 80:20 class imbalance (8,000 non-churners : 2,000 churners). Without addressing this, a naive classifier achieves 80% accuracy simply by predicting "no churn" for everyone — useless for the actual business problem. The core engineering challenge was choosing the right imbalance-handling strategy depending on business priority: catching more churners (recall) vs avoiding false alarms (precision).

Model Comparison: SMOTE vs Class-Weight Adjustment

Approach	Precision (Churn)	Recall (Churn)	F1 (Churn)	Best For
No imbalance handling	0.61	0.48	0.54	—
SMOTE oversampling	0.73	0.66	0.69	Minimising false positives (targeted retention spend)
Class weight adjustment	0.61	0.84	0.71	Maximising recall (catching the most churners)
SMOTE + Tuned threshold	0.68	0.78	0.76	Best overall balance — recommended

The best-performing configuration combined SMOTE oversampling with threshold tuning (decision threshold lowered from 0.5 to 0.38), achieving F1 = 0.76 on the churn class — a 41% improvement over the baseline.

Key Insights

• Class imbalance is a business problem, not just a technical one — the right approach depends on whether the bank prioritises precision (targeted spend) or recall (catching all churners). There is no universally "best" answer.
• Age (35–50), account balance, and credit score below 600 are the top three churn predictors — actionable signals for proactive outreach.
• Germany-based customers churn at 2× the rate of French and Spanish customers — suggesting regional service or product gaps worth investigating.
• Customers with 1 product churn at significantly higher rates than those with 2+ products — cross-selling is a directly addressable retention lever.
• A customer flagged as high churn risk with an estimated lifetime value above £5,000 justifies a personalised retention call; below that threshold, an automated email campaign is more cost-effective.

Technical Implementation

Preprocessing & Feature Engineering:

• Handled missing values and outliers (balance had right-skewed distribution — log transform applied).
• Encoded categorical variables: Geography (one-hot), Gender (binary).
• StandardScaler applied to numerical features for consistency with downstream SMOTE.
• Created interaction feature: balance_to_salary_ratio — stronger predictor than either alone.

Modelling Approaches:

• Random Forest + SMOTE: SMOTE applied only to training set (not test — critical to avoid data leakage).
• Random Forest + Class Weight: class_weight='balanced' in sklearn, which internally applies inverse-frequency weighting.
• 5-fold stratified cross-validation on all configurations to ensure consistent evaluation across imbalanced folds.
• Hyperparameter tuning via RandomisedSearchCV: n_estimators, max_depth, min_samples_split.

Key Learnings

• SMOTE must only be applied to training data — applying it before train/test split inflates evaluation metrics because synthetic samples near test points artificially improve test performance. This is a common mistake and a red flag in DS interviews.
• Threshold tuning is underutilised — most projects stop at 0.5 threshold. Moving the threshold to 0.38 improved F1 by 7 points without changing the model at all, purely by redefining what constitutes a "positive" prediction.
• Feature importance from Random Forest is a starting point, not a conclusion — correlated features can dilute each other's importance scores. SHAP values would give a more accurate and model-agnostic attribution.

Future Work

• Add SHAP values for model interpretability — Random Forest feature importance doesn't account for feature correlation, which skews attribution for correlated predictors.
• Build a simple customer LTV model to weight churn predictions by business impact — a high-probability churn on a low-value customer is less urgent than a moderate-probability churn on a high-value one.
• Evaluate gradient boosting (XGBoost/LightGBM) — they typically outperform Random Forest on structured tabular data with class imbalance.