Propensity Scoring

At its heart, a propensity model is a binary classifier that outputs a probability between 0 and 1. It answers a specific yes/no question: Will this user purchase within 7 days? Will they churn this quarter? Common algorithms include logistic regression for interpretability and gradient-boosted trees (XGBoost, LightGBM) for raw predictive power on tabular data. The output is not a segment label but a continuous score, allowing for precise rank-ordering of a customer base.

Model accuracy depends on transforming raw clickstreams into predictive features. Key feature categories include:

• Recency features: Time since last purchase, last email open, last support ticket.
• Frequency features: Purchase count in the last 30 days, page views per session.
• Velocity features: Rate of change in engagement, such as a sudden drop in login frequency.
• Derived ratios: Cart abandonment rate, wishlist-to-purchase ratio. These features are computed in real-time using windowed aggregation over event streams.

Propensity models can be deployed in two modes:

• Batch scoring: A nightly job scores the entire customer base, writing results to a CRM or CDP for next-day campaign orchestration.
• Real-time scoring: The model is served behind a low-latency REST or gRPC endpoint. When a user triggers an event (e.g., adds an item to cart), the system synchronously requests a new propensity score to power an immediate next-best-action decision. Real-time scoring requires a feature store to serve pre-computed aggregates with sub-10ms latency.

A raw model output of 0.8 does not inherently mean an 80% chance of conversion; it is a relative ranking. Probability calibration techniques like Platt scaling or isotonic regression map these raw scores to true, well-calibrated probabilities. A perfectly calibrated model means that among all users assigned a 70% propensity score, exactly 70% will perform the target action. This is critical when scores drive financial decisions like discount allocation or risk assessment.

Propensity models learn from historical examples of the action they are trying to predict.

• Positive class: Users who performed the action (e.g., completed a purchase).
• Negative class: Users who had the opportunity but did not perform the action. Defining the negative class correctly is crucial. For a purchase propensity model, negatives are users who were active and saw the product but did not buy, not users who were inactive. Observation windows must be carefully chosen: a 7-day prediction window requires labeling based on actions within the subsequent 7 days.

Accuracy is misleading for imbalanced datasets where the target action is rare (e.g., 2% purchase rate). Propensity models are evaluated using:

• AUC-ROC: Measures the model's ability to rank a random positive instance higher than a random negative one.
• Precision@K: Among the top K users scored, what percentage actually converted? This directly maps to campaign budget efficiency.
• Lift charts: Show how much better the model is at identifying converters compared to random targeting.
• Expected calibration error (ECE): Quantifies the gap between predicted probabilities and observed frequencies.

A metric quantifying the strength of a user's preference for a specific product, brand, or category based on engagement history. While propensity scoring predicts a future action (e.g., purchase), affinity scoring measures a current state of preference.

• Derived from weighted signals: views, add-to-carts, time spent, search queries
• Often serves as a feature input into propensity models
• Example: A user with high affinity for 'trail running shoes' receives a high propensity score for a new waterproof trail shoe launch

A decisioning system that consumes propensity scores to determine the single most optimal interaction to present to a customer in real-time. It combines predictive model outputs with business rules and constraints.

• Inputs: propensity scores, customer lifetime value, inventory levels, contact frequency caps
• Outputs: a ranked action (send discount, show product, suppress message)
• Example: A high churn propensity triggers a retention offer, but only if the customer hasn't received one in the last 30 days

The real-time identification of behavioral cues indicating a user's readiness to perform a high-value action. These signals are critical predictor variables in propensity models.

• Explicit signals: search queries containing 'buy,' pricing page visits, cart additions
• Implicit signals: scroll depth, hover time, return frequency, comparison tool usage
• Example: A user comparing three laptop models within 5 minutes triggers a high purchase-intent signal, elevating their conversion propensity score

The process of monitoring when the statistical properties of the target variable change over time, degrading propensity model performance. Essential for maintaining scoring accuracy in production.

• Types: sudden drift (Black Friday behavior shift) vs. gradual drift (seasonal preference changes)
• Detection methods: ADWIN, Page-Hinkley test, Kolmogorov-Smirnov test on prediction distributions
• Example: A purchase propensity model trained on pre-pandemic data drifts significantly as consumer behavior shifts to online-first

A centralized platform serving pre-computed and real-time features to propensity models at inference time with low latency. Ensures consistency between training and serving pipelines.

• Stores user aggregates: 7-day purchase count, average session duration, category affinity vectors
• Provides point-in-time correctness to prevent training-serving skew
• Example: A propensity model queries the feature store for a user's recency-frequency-monetary features in under 10ms before scoring

A predictive model estimating the total future revenue a customer will generate. Often used as a weighting factor alongside propensity scores to prioritize high-value users.

• Models: Pareto/NBD, BG/NBD for non-contractual settings; deep learning for complex patterns
• Integration: A moderate purchase propensity score for a high-CLV user may outrank a high propensity score for a low-CLV user
• Example: A luxury retailer targets a retention campaign at users with high CLV and elevated churn propensity, not just any user likely to churn