Churn Probability Score

The score is a probability estimate between 0 and 1, representing the likelihood of a churn event within a defined future window (e.g., 30 days). It is the output of a binary classification model, typically using logistic regression, gradient boosting, or a neural network with a sigmoid activation. The raw score quantifies uncertainty, distinguishing a customer at 0.51 risk from one at 0.94 risk.

Every churn score is tied to a prediction horizon—the specific future period during which churn is expected. Common horizons include 7-day, 30-day, or 90-day windows. A score without a defined horizon is meaningless. The horizon is baked into the training data labeling: a customer is labeled 'churned' only if the event occurs within that window from the observation point.

In production systems, the score is computed on-demand at inference time using the latest available features. This requires a feature store to serve pre-computed aggregates (e.g., 7-day login frequency) and streaming data for real-time signals (e.g., current session duration). The entire pipeline—feature retrieval, model inference, score delivery—must execute in milliseconds to power synchronous interventions.

A raw probability alone is insufficient for action. Modern churn systems pair the score with explainability outputs that decompose the prediction into contributing factors. Techniques like SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) reveal which features—such as a sudden drop in order frequency or a spike in support ticket count—drove the score upward.

A well-calibrated churn score means that among all customers assigned a probability of 0.8, approximately 80% actually churn. Isotonic regression or Platt scaling is applied post-training to correct overconfident or underconfident raw outputs. Calibration is measured using Expected Calibration Error (ECE) and is critical for setting accurate intervention thresholds and cost-benefit analyses.

The continuous probability is converted into a binary action via a classification threshold. This threshold is not fixed at 0.5; it is optimized based on business constraints. A lower threshold increases recall (catching more churners) but raises false positives. The optimal point balances the cost of intervention against the cost of churn, often visualized using a profit curve or precision-recall curve.

A statistical framework for analyzing the expected duration until churn occurs. Unlike a static probability score, survival analysis models the hazard function over time, answering not just if a customer will churn, but when.

• Kaplan-Meier Estimator: Non-parametric method for visualizing retention curves
• Cox Proportional Hazards: Regression model that quantifies how covariates (e.g., support tickets, login frequency) shift the hazard rate
• Time-varying covariates: Allows features to change over the observation window, capturing dynamic risk profiles

The instantaneous potential per unit of time for churn to occur, given that the customer has survived up to that point. This is the mathematical foundation underlying many churn probability scores.

• Formula: h(t) = lim Δt→0 P(t ≤ T < t+Δt | T ≥ t) / Δt
• Interpretation: A hazard rate of 0.05 at week 10 means 5% of remaining customers are expected to churn in week 10
• Shapes: Can be constant, increasing (wear-out), decreasing (burn-in), or bathtub-shaped
• Directly informs expected remaining lifetime calculations

A complementary predictive score representing the probability that a customer will make a subsequent purchase. While churn probability models disengagement, repurchase propensity models re-engagement.

• Often modeled as a binary classification task using historical behavioral features
• Key features: days since last purchase, average order value trend, category exploration breadth
• Combined with churn scores, creates a two-dimensional risk-opportunity matrix for intervention targeting

A causal inference technique that predicts the incremental impact of a specific retention treatment on churn probability. This moves beyond prediction to prescriptive action.

• Four segments: Persuadables (respond only if treated), Sure Things (retain regardless), Lost Causes (churn regardless), Sleeping Dogs (churn if treated)
• Qini Curve: Measures uplift model performance by plotting incremental gains against proportion targeted
• Prevents wasteful discounting of customers who would have stayed anyway

A class of probabilistic models for non-contractual settings that jointly predicts transaction frequency and the point of permanent inactivity. These models directly estimate churn as a latent dropout event.

• BG/NBD Model: Models transaction rate (Poisson-Gamma) and dropout probability (Beta-Geometric) separately
• Pareto/NBD: Extends BG/NBD by allowing dropout to occur at any time, not just after a purchase
• Outputs include P(alive): the probability a customer is still active, directly analogous to 1 - churn probability score

Statistical techniques applied to raw cohort retention data to reduce noise and generate stable, monotonic decay curves. Critical for calibrating churn probability scores against ground truth.

• Bayesian Shrinkage: Pulls extreme cohort estimates toward the population mean, preventing overfitting from small samples
• Exponential Smoothing: Applies weighted moving averages to dampen period-over-period volatility
• Beta-Binomial Smoothing: Models retention as a binomial process with beta-distributed heterogeneity
• Essential for setting realistic churn score thresholds for intervention triggers