Churn Propensity

Churn is defined by a prediction horizon—a specific future time window (e.g., 30, 60, or 90 days). The model must distinguish between a customer who will churn within this window versus one who will churn later. This requires time-aware feature engineering, such as recency of last purchase, frequency decay rates, and trend-based aggregations over rolling windows. A common pitfall is defining the window too narrowly, which misses early warning signals, or too broadly, which dilutes precision.

Churn events are typically rare, often representing 2-10% of the customer base in a given period. This severe class imbalance renders standard accuracy metrics useless. Effective models employ specialized techniques:

• SMOTE (Synthetic Minority Over-sampling Technique) to generate synthetic churn examples
• Cost-sensitive learning that penalizes false negatives more heavily
• Anomaly detection framing where churn is treated as a deviation from normal behavior
• Evaluation relies on Precision-Recall AUC and F1-score rather than raw accuracy.

Raw transactional data is insufficient. Predictive power comes from engineered behavioral signals:

• Recency, Frequency, Monetary (RFM) metrics decomposed over time
• Engagement velocity: rate of change in login frequency or session duration
• Complaint-to-transaction ratio and support ticket sentiment scores
• Product usage entropy: diversity of features used, indicating stickiness
• Social graph decay: loss of connections in collaborative or multi-user products These features capture the trajectory of a relationship, not just its current state.

A churn prediction without a reason is operationally useless. Business teams need to know why a customer is likely to leave to design an effective intervention. This demands intrinsically interpretable models (logistic regression, decision trees) or post-hoc explanation frameworks:

• SHAP (SHapley Additive exPlanations) values to quantify each feature's contribution to an individual prediction
• LIME (Local Interpretable Model-agnostic Explanations) for local surrogate models
• Partial dependence plots to visualize the directional impact of key drivers like price sensitivity or support latency

A high correlation does not imply a high-leverage intervention point. A customer exhibiting low login frequency may be a churn risk, but forcing them to log in does not prevent churn. Advanced churn models integrate uplift modeling to distinguish:

• Sleeping dogs: customers who would not churn but might be triggered to leave by an intervention
• Persuadables: customers who will only stay if treated
• Lost causes: customers who will churn regardless
• Sure things: customers who will stay regardless This causal framing ensures retention budgets are allocated efficiently.

Churn patterns are non-stationary. A model trained on pre-pandemic behavior will fail in a recession. Continuous monitoring is essential:

• Data drift detection: tracking shifts in feature distributions (e.g., average order value declining across the population)
• Prediction drift: monitoring the distribution of churn probability scores over time
• Performance decay: regularly backtesting against realized churn events using a delayed ground truth window
• Automated retraining pipelines triggered by drift thresholds prevent silent model degradation.

The foundational statistical technique that estimates the probability of a specific binary event—in this case, churn—occurring within a defined future window. A churn propensity model is a specific instance of a propensity score.

• Input Features: Aggregated behavioral data (recency, frequency, monetary value), engagement metrics, and demographic attributes.
• Output: A continuous probability score between 0 and 1, often calibrated using Platt scaling or isotonic regression.
• Common Algorithms: Logistic regression, gradient-boosted trees (XGBoost, LightGBM), and deep neural networks for high-cardinality categorical features.

A causal machine learning technique that goes beyond predicting churn to estimate the incremental impact of a retention intervention. It isolates customers into four quadrants:

• Persuadables: Customers who will stay only if treated. These are the optimal targets.
• Sure Things: Customers who will stay regardless. Treating them wastes resources.
• Lost Causes: Customers who will churn regardless. Treatment is ineffective.
• Sleeping Dogs: Customers who will churn because of the treatment. Avoid contacting them.

This directly addresses the fundamental business question: 'Will this discount actually prevent this specific customer from leaving?'

A forward-looking metric representing the total net profit attributed to the entire future relationship with a customer. Churn propensity is a critical input to CLV calculation.

• Relationship: CLV = (Average Transaction Value × Purchase Frequency × Margin) ÷ Churn Rate. A high churn propensity directly diminishes CLV.
• Action Threshold: Retention offers should cost less than the expected CLV loss prevented. If a customer's CLV is $500 and their churn probability is 80%, the expected loss is $400, setting a ceiling on intervention cost.
• Segmentation: High-CLV, high-churn-propensity customers form the most critical intervention segment.

The degradation of a churn model's predictive performance over time due to changing statistical relationships in the underlying data. This is the primary operational risk for deployed churn systems.

• Real Drift: The actual relationship between features and churn changes. Example: A global pandemic makes historical engagement patterns irrelevant.
• Virtual Drift: The distribution of input features shifts without changing the underlying relationship. Example: A new mobile app feature attracts a different user demographic.
• Detection: Monitoring the Population Stability Index (PSI) and model performance metrics (AUC, log-loss) on live data streams triggers retraining pipelines.

A statistical framework for modeling time-to-event data, directly addressing not just if a customer will churn, but when. This provides a more granular risk profile than binary classification.

• Kaplan-Meier Estimator: A non-parametric method to estimate the survival function (probability of not churning) over time, handling censored data (customers who haven't churned yet).
• Cox Proportional Hazards Model: A semi-parametric regression model that quantifies the impact of features on the instantaneous risk (hazard) of churning.
• Output: A hazard curve over time, enabling time-sensitive interventions (e.g., 'This customer has a 60% chance of churning within the next 7 days').

A set of techniques to estimate the performance of a new churn intervention policy using historical data collected by a different, often suboptimal, policy. This is essential for safely evaluating retention strategies without costly A/B tests.

• Inverse Propensity Scoring (IPS): Corrects for selection bias by re-weighting observed outcomes by the inverse probability of the action taken by the logging policy.
• Doubly Robust Estimation: Combines IPS with a direct outcome model, providing an unbiased estimate if either the propensity model or the outcome model is correctly specified.
• Application: Estimating the true churn reduction of a new discount strategy using data from the old, non-personalized strategy.