Propensity to Repurchase

The time elapsed since a customer's most recent transaction is the single most predictive feature in repurchase modeling. A shorter recency window correlates strongly with higher repurchase probability, as the customer remains in an active buying state. This feature is central to RFM Analysis and Buy-Till-You-Die (BTYD) probabilistic frameworks.

• Typically measured in days or weeks from the observation point
• Often log-transformed to handle right-skewed distributions
• Combined with frequency to estimate the BG/NBD model dropout probability

The rate at which a customer accumulates transactions over a defined historical window. A decelerating purchase cadence often signals impending dormancy before recency alone can detect it. This feature captures the Poisson-Gamma mixture assumption that transaction rates vary heterogeneously across the customer base.

• Calculated as total transactions divided by the length of the observation period
• Rolling window velocities (e.g., 30-day, 90-day) capture trend shifts
• A declining velocity combined with high recency is a strong churn precursor

The standard deviation or coefficient of variation of the time gaps between consecutive purchases. Customers with highly regular purchase rhythms exhibit lower churn risk than those with erratic patterns, even if their average frequency is identical. This feature informs Survival Analysis hazard function estimation.

• Low variance indicates habitual, predictable buying behavior
• High variance suggests opportunistic or promotion-driven purchasing
• Used as a covariate in Cox Proportional Hazards Models for churn timing

The number of distinct product categories or departments a customer has purchased from historically. Broader category engagement signals deeper platform investment and higher switching costs, directly elevating repurchase propensity. This feature captures the latent class of exploratory versus habitual buyers.

• Measured as unique category count or Shannon entropy of category distribution
• A sudden narrowing of breadth may indicate competitive defection
• Cross-category buyers typically exhibit 2-3x higher Customer Lifetime Value (CLV)

Behavioral signals captured during browsing sessions that precede or occur between transactions. Metrics such as time-on-site, product detail page views, search query volume, and cart abandonment events provide leading indicators of purchase intent before a transaction materializes.

• Session count in the last 7 days is a strong short-term predictor
• Cart-to-purchase conversion lag captures decision hesitation
• These features feed Sequential User Behavior Models for next-click intent prediction

The customer's historical reaction to promotional stimuli, measured as the lift in purchase probability following a campaign exposure. This feature distinguishes organic repurchasers from those requiring incentive nudges, enabling Uplift Modeling for retention campaign targeting.

• Calculated as the difference between treated and untreated conversion rates
• Markov Chain Attribution assigns credit across email, push, and paid channels
• High organic repurchasers should be suppressed from discount campaigns to preserve margin

Marketing teams use repurchase propensity scores to segment customers into high-intent and at-risk cohorts. Instead of blasting a generic newsletter, the system triggers personalized offers only to users whose probability crosses a defined threshold.

• High propensity (>0.8): Send loyalty rewards to reinforce behavior
• Medium propensity (0.4–0.8): Deploy time-limited discount codes to nudge conversion
• Low propensity (<0.4): Suppress marketing spend to avoid churn acceleration

This approach reduces campaign costs by 30–50% while increasing conversion rates by focusing resources on persuadable customers.

E-commerce platforms integrate individual repurchase probabilities with demand forecasting models to predict stock requirements at the SKU level. A customer's propensity score for a specific consumable product (e.g., protein powder, diapers) directly informs reorder point calculations.

• Combine time-to-next-purchase estimates with propensity scores
• Trigger automated purchase reminders when both inventory is low and propensity is high
• Reduce stockouts for subscription-eligible goods by 15–25%

This application bridges the gap between behavioral prediction and supply chain execution, ensuring high-intent customers never encounter out-of-stock scenarios.

The inverse of repurchase propensity functions as a churn early warning system. Customer success teams rank accounts by their probability of not repurchasing within the expected window, enabling proactive intervention before the relationship terminates.

• Feature importance analysis reveals churn drivers (e.g., support ticket volume, delivery delays)
• High-value customers with declining propensity trigger direct outreach from account managers
• Automated service recovery workflows (refunds, expedited shipping) launch when propensity drops below a critical threshold

This shifts retention strategy from reactive firefighting to data-driven prevention, preserving customer lifetime value.

Repurchase propensity scores feed directly into contextual multi-armed bandit algorithms that determine optimal discount depth. The system balances the probability of conversion against margin preservation.

• Customers with high organic propensity receive minimal or zero discounting
• Customers with moderate propensity receive calibrated incentives to maximize incremental revenue
• The model continuously updates scores based on real-time responses to offers

This application prevents unnecessary margin erosion on customers who would have repurchased anyway, a common failure mode of blanket promotional strategies.

For subscription-based businesses, repurchase propensity models predict renewal likelihood before the contract expiration date. Finance teams use these scores to forecast monthly recurring revenue (MRR) with greater accuracy.

• Integrate propensity scores with survival analysis to estimate time-to-churn
• Trigger retention offers 30 days before expiration for low-propensity subscribers
• Feed aggregated propensity data into revenue forecasting dashboards for investor reporting

This transforms subscription management from a reactive billing event into a predictive revenue operations function.

Repurchase propensity models extend beyond repeat purchases of the same item to predict adjacent category adoption. A customer with high propensity to repurchase in Category A becomes a prime candidate for Category B introduction.

• Train separate propensity models per product category
• Identify customers with high scores in one category but zero interaction in complementary categories
• Deploy personalized cross-category recommendations during the post-purchase confirmation flow

This leverages established purchase intent to expand share of wallet without the acquisition cost of net-new customer conversion.

The inverse counterpart to repurchase propensity. A binary classification output that quantifies the likelihood a customer will discontinue their relationship within a defined future window. While repurchase propensity models the positive signal, churn scores model the hazard of attrition. These two metrics are often generated by the same underlying model architecture but interpreted through complementary lenses. Features include recency decay, support ticket frequency, and login cadence.

A foundational behavioral segmentation technique that scores customers on three dimensions:

• Recency: Time since last purchase (strongest predictor of repurchase)
• Frequency: Total number of transactions in a period
• Monetary Value: Total spend over the observation window

RFM quintiles serve as baseline features in propensity models, often outperforming demographic variables. Modern implementations use weighted RFM scores with decay functions applied to recency.

A probabilistic buy-till-you-die model that predicts future purchasing behavior in non-contractual settings. It models the transaction rate using a Poisson-Gamma mixture and the dropout probability using a Beta-Geometric distribution. The model estimates two latent parameters per customer: λ (transaction rate while alive) and p (probability of becoming permanently inactive after any purchase). These parameters directly inform repurchase propensity calculations.

A statistical framework for analyzing time-to-event data, where the event is the next purchase. The hazard function estimates the instantaneous probability of repurchase at time t given no purchase has occurred yet. Cox Proportional Hazards models incorporate time-varying covariates like browsing intensity and cart abandonment signals. Unlike binary classifiers, survival models produce a temporal curve of repurchase probability rather than a single score.

Captures temporal patterns in user actions to predict next-click intent and session outcomes. Architectures like Transformer-based sequence models and Long Short-Term Memory networks encode the order of product views, searches, and cart interactions. These embeddings reveal latent intent signals—such as comparison shopping behavior or seasonal purchase patterns—that static features miss. Sequence length and recency weighting are critical hyperparameters.

A causal inference technique that predicts the incremental impact of a specific intervention on repurchase probability. Unlike propensity models that estimate baseline likelihood, uplift models isolate the true treatment effect by modeling four segments:

• Persuadables: Respond only if treated
• Sure Things: Repurchase regardless
• Lost Causes: Won't repurchase either way
• Sleeping Dogs: Repurchase less if treated

This prevents wasteful discounting of customers who would repurchase organically.