E-Commerce Churn & Lifetime Value Intelligence Platform

Calibrated Predictive Modeling • Multi-Cohort Temporal Validation • Cost-Sensitive Profit Optimization

🤗 Hugging Face Models 🎯 Kaggle Model Hub 📓 Kaggle Notebook
Churn Model PR-AUC
0.8185
ROC-AUC: 0.7721 • Brier: 0.1786
LTV Model Spearman Rank
0.5837
MAE: £615.11 • R²: 0.4623
Optimal Threshold τ*
0.52
Net Profit Max: £19,326.70 (Base)
Priority 1 Value at Risk
£150,968
85 Top VIP Accounts (1.97%)
Customer Behavioral Profile
Select an archetype preset or adjust individual parameters
Predictive Output & Operational Playbook
Model inference across 90-day churn & 180-day LTV
CALIBRATED CHURN RISK
80.1%
CRITICAL RISK TIER
PREDICTED 180-DAY LTV
£482.30
FORWARD 6-MONTH HORIZON
🚨 PRIORITY 1: RETAIN URGENTLY
RECOMMENDED INTERVENTION STRATEGY

High Value at Risk: Immediate personal phone outreach or dedicated account rep call. Offer customized renewal packages without blanket discount codes.

KEY BEHAVIORAL EXPLAINABILITY SIGNALS
  • Dormancy: Last purchase was 95 days ago (exceeds 60-day threshold).
  • Cadence breakdown: Current inactivity is 3.4x their average inter-purchase gap.
  • Velocity collapse: 90-day momentum dropped by 100%.
2x2 Customer Value & Risk Prioritization Matrix
Empirical cohort segmentation on test set (N=4,324 customer snapshots)
High Risk • High LTV
PRIORITY 1: Retain Urgently
85 Customers (1.97%)
Total Forward Value at Risk: £150,968.32 (Avg £1,776/cust)
Commercial Action: Dedicated account rep outreach, tailored renewal packages, white-glove concierge. High ROI target (£433% campaign ROI).
Low Risk • High LTV
PRIORITY 2: Nurture Loyalty
780 Customers (18.04%)
Safe Future Enterprise Value: £1,817,449.10 (78.81% of total)
Commercial Action: VIP loyalty club tiers, early access to new product drops. Strictly protect margins: DO NOT offer unnecessary discounts.
High Risk • Low LTV
PRIORITY 3: Low-Cost Win-Back
3,001 Customers (69.40%)
Total Value at Risk: £254,168.45 (Avg £84.69/cust)
Commercial Action: Automated triggered drip email sequences with tiered 10%-15% voucher codes. Avoid expensive manual contact methods.
Low Risk • Low LTV
PRIORITY 4: Automated Lifecycle
458 Customers (10.59%)
Safe Future Value: £83,621.19 (Avg £182.58/cust)
Commercial Action: Standard seasonal newsletters and organic catalog touchpoints. Zero manual intervention overhead required.
Campaign Economics Parameters
Interactive cost-benefit scenario modeling
Break-Even Risk Required: 48.5%
Optimal Threshold (τ*): 0.14
Projected Net Profit: £19,326.70
Expected Campaign ROI: +53.5%
Net Profit vs Decision Threshold (τ)
Dynamic curve showing financial outcomes across [0.01 - 0.99]
Top Behavioral Feature Importance
TreeSHAP & KernelSHAP mean absolute impact on predictions
Model Architecture Benchmarks
Strict out-of-time temporal cohort evaluation
Model Architecture PR-AUC ROC-AUC Brier Score Status
★ Logistic Regression (Calibrated) 0.8185 0.7721 0.1786 Selected (Production)
Random Forest Classifier 0.8089 0.7663 0.1915 Benchmark
LightGBM Classifier 0.7949 0.7445 0.1948 Benchmark
XGBoost Classifier 0.7939 0.7456 0.1913 Benchmark
CatBoost Classifier 0.7910 0.7390 0.1938 Benchmark
Why Parsimonious Models Won: On temporal distribution shifts, uncalibrated tree boosting models tended to overfit local feature intersections. Logistic Regression with Isotonic Calibration yielded superior probability reliability and zero risk of erratic probability swings.
Data Card & Python Hub Quickstart
Download models directly from Hugging Face or Kaggle with 3 lines of code
import joblib
from huggingface_hub import hf_hub_download

# Download production-ready calibrated models
churn_model_path = hf_hub_download(repo_id="Kicaulah/ecommerce-churn-ltv-models", filename="churn_model.joblib")
ltv_model_path   = hf_hub_download(repo_id="Kicaulah/ecommerce-churn-ltv-models", filename="ltv_model.joblib")

# Load scikit-learn pipelines
churn_pipeline = joblib.load(churn_model_path)
ltv_pipeline   = joblib.load(ltv_model_path)

# Predict churn risk & 180-day LTV
# p_churn = churn_pipeline.predict_proba(customer_df)[:, 1]
# y_ltv   = ltv_pipeline.predict(customer_df)
print("Models loaded successfully from Hugging Face Hub!")
Dataset Provenance

UCI Online Retail II dataset (2009–2011). 1.04M raw records cleaned to 797,815 high-fidelity transactions across 5,939 salted SHA-256 customer hashes.

License & Ethics

Licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Salted hashing eliminates re-identification. Zero PII stored or leaked.

Kaggle Registry

Official Kaggle Model Hub integration: simonmarc/ecommerce-churn-ltv-models.