Propensity Scoring

Estimates the probability a customer will defect within a defined window. By identifying high-risk users, businesses can trigger proactive retention offers.

• Input Signals: Declining login frequency, reduced cart sizes, negative sentiment in support tickets, and price comparison browsing.
• Intervention: Trigger a Next-Best-Action model to offer a loyalty discount or a direct call from a high-touch agent.
• Example: A telecom provider reduces involuntary churn by 15% by offering a tailored data plan upgrade to users with a churn propensity score above 0.7.

Calculates the likelihood of a user converting at a specific price point or with a specific incentive, enabling margin optimization.

• Mechanism: A Contextual Bandit uses the propensity score as a context feature to decide whether to show a 10% or 20% coupon.
• Goal: Maximize revenue by offering the minimum discount required to convert a price-sensitive user while maintaining full margin for price-insensitive users.
• Example: An e-commerce platform boosts profit margins by 5% by suppressing discounts for users with a high organic purchase propensity.

Predicts the probability that a user who has left items in their cart will complete the purchase without intervention.

• Temporal Decay: The model weighs the recency of the abandonment event heavily; a session abandoned 10 minutes ago has a higher rescue probability than one abandoned 3 days ago.
• Action: If the organic conversion propensity is low, trigger a personalized push notification or email containing the exact abandoned items within 30 minutes.
• Example: A fashion retailer recovers 12% of abandoned carts by targeting only users whose predicted recovery propensity dropped below 0.4.

Estimates the likelihood a customer will purchase a specific complementary product given their current ownership and recent browsing behavior.

• Sequential Modeling: Uses Sequential User Behavior Modeling to understand the typical product journey (e.g., phone → case → screen protector).
• Application: Rank the "Frequently Bought Together" widget not just by global popularity, but by the user's specific propensity to buy the accessory.
• Example: A consumer electronics store increases attachment rate by 8% by promoting a specific lens kit to camera buyers with a high propensity score for portrait photography.

Predicts the probability a user will engage with a specific piece of content or communication channel.

• Channel Optimization: A user may have a high purchase propensity but a low email engagement propensity. The model routes the offer via push notification or SMS instead.
• Content: Predicts propensity to click on a blog post vs. a video vs. a case study, enabling true 1:1 content personalization.
• Example: A B2B SaaS company increases lead conversion by 20% by routing high-intent whitepaper downloads to sales, while nurturing low-intent blog readers with email drips.

Estimates the probability that a transaction, account login, or return request is fraudulent or abusive.

• Real-Time Scoring: The propensity score is generated at the point of transaction using Streaming Data Pipelines to analyze device fingerprint, velocity checks, and geolocation anomalies.
• Action: If the fraud propensity exceeds a threshold, the system triggers step-up authentication (MFA) or blocks the transaction entirely.
• Example: A fintech platform reduces false positives by 40% by using a propensity model that distinguishes between a genuine user traveling and a fraudster using a VPN.

The workhorse algorithm for propensity scoring due to its direct output of well-calibrated probabilities between 0 and 1. Unlike tree-based models that require Platt scaling for probability estimates, logistic regression natively models the log-odds of a binary outcome as a linear combination of features.

• Produces inherently interpretable coefficients that quantify each feature's contribution
• Supports L1/L2 regularization to handle multicollinearity in high-dimensional customer data
• Serves as the baseline model against which more complex architectures are benchmarked

A typical e-commerce churn model might use features like days since last purchase, support ticket frequency, and session depth to output a churn probability score.

Ensemble methods like XGBoost, LightGBM, and CatBoost often outperform logistic regression on propensity tasks with complex, non-linear feature interactions. These models sequentially build decision trees where each new tree corrects the residual errors of the ensemble.

• Automatically capture feature interactions without manual engineering
• Handle missing values natively through sparsity-aware split finding
• Require careful probability calibration using isotonic regression or Platt scaling

In practice, a gradient boosted propensity model for subscription renewal might discover that the interaction between contract length and recent usage decline is a powerful churn signal that a linear model would miss.

A causal inference technique that uses estimated propensity scores to create balanced treatment and control groups in observational studies. PSM pairs each treated unit with one or more untreated units that have similar propensity scores, simulating the covariate balance achieved by random assignment.

• Reduces selection bias when A/B testing is infeasible or unethical
• Relies on the conditional independence assumption: all confounders are observed
• Common matching algorithms include nearest neighbor, caliper, and kernel matching

A retail bank might use PSM to evaluate a new mobile app feature by matching users who adopted it with non-adopters who had identical propensity scores based on demographics and prior behavior.

A re-weighting method that corrects for selection bias by assigning each observation a weight equal to the inverse of its probability of receiving the treatment it actually received. This creates a pseudo-population where treatment assignment is independent of observed covariates.

• Weights are calculated as 1 / P(Treatment | Covariates) for treated units
• Extremely high weights from near-zero propensities can cause variance inflation
• Stabilized weights multiply by the marginal treatment probability to reduce variance

IPW is foundational to off-policy evaluation in recommendation systems, where historical click data collected under a production policy must be re-weighted to estimate the performance of a new candidate policy.

The process of aligning a model's predicted probabilities with the true empirical frequencies of outcomes. A perfectly calibrated propensity model will have 10% of customers with a 0.1 churn score actually churn.

• Reliability diagrams plot predicted probability against observed frequency to visualize miscalibration
• Expected Calibration Error (ECE) quantifies the weighted average gap across probability bins
• Modern neural networks often suffer from overconfidence and require temperature scaling

Calibration is critical when propensity scores feed directly into business rules, such as triggering a retention offer only when churn probability exceeds a specific threshold tied to intervention cost.

The predictive power of propensity models depends heavily on temporal feature engineering that captures the recency, frequency, and trajectory of customer actions.

• Recency features: days since last purchase, last login, or last support contact
• Velocity features: rate of change in spending, session frequency, or engagement depth
• Cohort-normalized features: behavior expressed relative to peers in the same acquisition cohort
• Decay-weighted aggregates: exponential smoothing of historical metrics to emphasize recent behavior

A subscription propensity model might engineer a feature representing the 7-day moving average of daily active minutes, divided by the user's 90-day baseline, to detect engagement decline before it manifests in churn.