ecommerce-churn-ltv-models

E-Commerce Customer Churn & Lifetime Value Prediction Models

Production-grade machine learning models for customer churn risk classification (forward 90-day inactivity) and customer lifetime value (LTV) regression (forward 180-day net spend), trained on the validated UCI Online Retail II dataset.

• Kaggle Code Notebook: simonmarc/e-commerce-churn-and-lifetime-value-prediction
• Kaggle Dataset: simonmarc/ecommerce-churn-and-lifetime-value-dataset
• Interactive Space Dashboard: Kicaulah/ecommerce-churn-ltv-dashboard

1. Model Summary

Parsimony & Validation Strategy

• Models were validated using strict temporal cohort splits across monthly cutoff dates (2010-03-01 through 2011-06-01).
• Validation cutoff: 2011-02-01 to 2011-03-01.
• Out-of-time test cutoff: 2011-04-01 to 2011-06-01 (4,324 customer-snapshots).
• Random K-fold splitting was strictly avoided to prevent temporal leakage and lookahead bias.
• Parsimonious design: Logistic Regression matched or exceeded gradient boosting architectures (LightGBM, XGBoost, CatBoost) on PR-AUC while offering superior probability calibration and interpretability.

2. Business Decision Threshold Optimization

Using cost-benefit scenario analysis on test cohort ($N=4,324$): - Unit Contact Cost ($c$): £2.50 - Retention Offer / Incentive ($d$): £15.00 - Customer Lifetime Value Recovered ($v$): £45.00 - Uplift / Success Rate ($u$): 15% (Base Scenario)

The expected net business value is maximized at threshold $ au^ = 0.52$: - Customers Targeted: 2,987 (69.1%) - True Positives Targeted: 2,548 - Gross Value Rescued: £17,203.24 - Net Business Value: £9,735.74 (vs default 0.50 net value £9,715.19) - Break-Even Required Churn Risk:* $ au_{ ext{break-even}} = rac{c}{u \cdot (v - d)} = rac{2.50}{0.15 \cdot 30.00} = 0.555$

3. 2x2 Customer Value & Risk Prioritization Matrix

Combining calibrated churn probabilities with predicted 180-day LTV segments customers into four operational tiers:

                  HIGH LTV (Top 20%)             LOW LTV (Bottom 80%)
           +-----------------------------+-----------------------------+
HIGH RISK  |  PRIORITY 1: Retain Urgently|  PRIORITY 3: Low-Cost Win   |
(P > 0.52) |  85 customers (1.97%)       |  3,001 customers (69.40%)   |
           |  £150,968 Value at Risk     |  £254,168 Value at Risk     |
           |  Action: Dedicated Outreach |  Action: Automated Discounts|
           +-----------------------------+-----------------------------+
LOW RISK   |  PRIORITY 2: Nurture Loyalty|  PRIORITY 4: Lifecycle Auto |
(P <= 0.52)|  780 customers (18.04%)     |  458 customers (10.59%)     |
           |  £1,817,449 Safe Future LTV |  £83,621 Safe Future LTV    |
           |  Action: VIP Programs       |  Action: Triggered Newslettr|
           +-----------------------------+-----------------------------+

4. How to Use the Models

import joblib
import pandas as pd
from huggingface_hub import hf_hub_download

# Download models
churn_path = hf_hub_download(repo_id="Kicaulah/ecommerce-churn-ltv-models", filename="churn_model.joblib")
ltv_path = hf_hub_download(repo_id="Kicaulah/ecommerce-churn-ltv-models", filename="ltv_model.joblib")

churn_pipeline = joblib.load(churn_path)
ltv_pipeline = joblib.load(ltv_path)

# Example customer snapshot feature row
customer_sample = pd.DataFrame([{
    "recency_days": 45,
    "tenure_days": 320,
    "tenure_censored": 0,
    "frequency_lifetime": 12,
    "purchase_days_lifetime": 9,
    "frequency_365": 10,
    "frequency_180": 4,
    "frequency_90": 1,
    "monetary_365": 2840.50,
    "revenue_180": 850.20,
    "revenue_90": 180.00,
    "historical_revenue_lifetime": 3400.00,
    "avg_order_value_365": 284.05,
    "avg_units_per_order_365": 42.0,
    "avg_lines_per_order_365": 14.5,
    "avg_days_between_purchases": 28.5,
    "std_days_between_purchases": 12.1,
    "max_gap_days": 65,
    "recency_to_avg_gap_ratio": 1.58,
    "active_months_365": 8,
    "product_diversity_365": 56,
    "category_diversity_365": 4,
    "below_ref_price_share_365": 0.08,
    "discount_line_share_365": 0.05,
    "return_value_rate_365": 0.03,
    "cancellation_rate_365": 0.02,
    "revenue_trend_90": -0.35,
    "order_trend_90": -0.20,
    "is_domestic": 1,
    "preferred_category": "Home & Decor",
    "country_group": "Domestic"
}])

# Inference
churn_prob = churn_pipeline.predict_proba(customer_sample)[:, 1][0]
predicted_ltv_180 = ltv_pipeline.predict(customer_sample)[0]

print(f"Calibrated Churn Probability: {churn_prob:.1%}")
print(f"Predicted 180-Day Net LTV:     £{predicted_ltv_180:,.2f}")

5. Model Interpretability & SHAP Analysis

Feature importance analyzed using TreeSHAP and KernelSHAP: - recency_days: Positively correlated with churn probability (as days since last transaction increase, historical non-return rate escalates). - recency_to_avg_gap_ratio: Strongest non-linear signal of behavioral departure. - monetary_365 & revenue_180: Strong positive association with forward 180-day LTV.

Note on causality: All SHAP attributions represent statistical associations identified by the models under observed transaction histories, not counterfactual causal proofs.

6. Privacy & Responsible AI

• Customer Identifiers: Salted SHA-256 hashes (customer_id_hash).
• Free-Text & PII: Invoices and descriptions are strictly sanitized; zero customer names, credit card numbers, or physical addresses are ingested or stored.
• Licensing: The underlying dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).