Customer Churn Prediction in Non-Subscription Retail: Survival Analysis Applied to PoS Transaction Histories
Apply Cox proportional hazards and accelerated failure-time models to inter-purchase intervals, adapting subscription-churn methods to discretionary retail.
Key Takeaways
- Non-subscription retail churn is fundamentally different from subscription churn because there is no explicit cancellation event, requiring probabilistic definitions of customer attrition.
- Survival analysis models such as Cox proportional hazards can estimate the probability that a customer has permanently defected based on their elapsed time since last purchase.
- Recency-frequency-monetary (RFM) features derived from PoS data serve as strong predictors in churn models, but transaction-level behavioral features improve discrimination further.
Defining Churn Without Cancellation Events
In subscription-based businesses, churn is unambiguous: a customer cancels their subscription, generating a clear event that marks the end of the relationship. Non-subscription retail enjoys no such clarity. A customer who has not visited a store in three months may have defected to a competitor, moved out of the area, or simply had no need for the products during that period. This definitional ambiguity has profound methodological consequences. Binary classification approaches that label customers as churned or retained based on an arbitrary inactivity threshold — for example, no purchase in 90 days — impose a sharp boundary on what is inherently a continuous phenomenon. The threshold choice directly affects model performance metrics and business decisions: too short a threshold generates excessive false positives among customers with naturally low purchase frequency, while too long a threshold delays intervention until reactivation becomes unlikely. Survival analysis offers a more principled framework by modeling the time-to-event distribution directly, treating the question not as "has this customer churned?" but as "what is the probability that this customer will make another purchase within the next k days, given their observed history?" askbiz.co employs survival-based churn models that continuously estimate each customer retention probability without requiring arbitrary threshold definitions.
Cox Proportional Hazards for Retail Applications
The Cox proportional hazards (PH) model is a semi-parametric survival model that estimates the hazard function — the instantaneous rate of the event occurring at time t, conditional on survival to t — as a function of covariates without specifying the baseline hazard distribution. In the retail churn context, the "event" is the next purchase, and the hazard represents the instantaneous probability of a customer making a purchase at time t given that they have not purchased since their last transaction. Covariates include RFM metrics (recency, purchase frequency, average transaction value), behavioral features (category diversity, brand loyalty indices, discount sensitivity), and temporal features (day-of-week and seasonal purchase patterns). The proportional hazards assumption — that covariate effects multiply the baseline hazard by a constant factor — may be violated for certain retail covariates. For example, the effect of promotional sensitivity on purchase hazard likely varies over time as promotional calendars shift. Time-varying coefficients or stratified Cox models address this limitation. The model produces individual-level survival curves that estimate the probability of repurchase within any given time horizon, enabling targeted retention interventions. askbiz.co fits Cox PH models on PoS transaction histories, generating customer-level churn risk scores that update with each new transaction or elapsed day of inactivity.
Feature Engineering From Transaction Histories
The predictive power of churn models depends critically on the features extracted from raw transaction data. Beyond the classical RFM triumvirate, several feature categories enhance discrimination. Inter-purchase interval statistics — mean, standard deviation, coefficient of variation, and trend — capture the regularity and trajectory of customer engagement. A customer whose inter-purchase intervals are steadily lengthening exhibits a different risk profile than one whose intervals are stable but long. Basket composition features measure category breadth, brand concentration, and the ratio of staple to discretionary items, under the hypothesis that customers who purchase across more categories are more deeply engaged and less likely to churn. Transaction timing features, including preferred day of week, time of day, and seasonal purchase patterns, enable the model to distinguish true inactivity from expected gaps in purchase cycles. Payment method consistency and discount utilization rates provide additional behavioral signals. Critically, all features must be computed relative to the customer segment rather than in absolute terms: a monthly purchase frequency that signals high engagement for a furniture store indicates potential churn for a grocery store. askbiz.co automatically engineers these features from PoS transaction logs, normalizing them against store-specific and category-specific baselines.
Accelerated Failure Time Models as Alternatives
While the Cox PH model is the most widely used survival analysis framework, accelerated failure time (AFT) models offer complementary advantages for retail churn prediction. AFT models directly model the logarithm of survival time as a linear function of covariates, providing an intuitive interpretation: a positive coefficient means the covariate accelerates the time to next purchase (desirable), while a negative coefficient means it decelerates purchase timing (indicating elevated churn risk). Common distributional assumptions include Weibull, log-normal, and log-logistic. The Weibull AFT model is particularly useful because its hazard function can be monotonically increasing, decreasing, or constant depending on the shape parameter, accommodating different customer engagement dynamics. Log-normal AFT models capture the common empirical observation that inter-purchase intervals are right-skewed with a heavy tail. Model selection between AFT specifications and the Cox PH model can be guided by AIC/BIC criteria or cross-validated concordance indices. In practice, ensemble approaches that average predictions from multiple survival models often outperform any single specification. askbiz.co evaluates multiple survival model specifications for each store and selects the best-performing model based on out-of-sample concordance, retraining as new transaction data accumulates.
Operationalizing Churn Predictions for Retention
Translating churn probability estimates into effective retention actions requires bridging the gap between statistical output and business decision-making. The survival model produces a time-varying churn probability for each customer, but the business must decide at what probability threshold to trigger an intervention, what form the intervention takes, and how to allocate limited retention budget across at-risk customers. Expected customer lifetime value (CLV) provides the economic framework for these decisions: the optimal retention investment for a customer is bounded by the product of their churn probability, their expected future CLV, and the estimated probability that the intervention successfully prevents churn. This last quantity — intervention effectiveness — is notoriously difficult to estimate and typically requires controlled experimentation. A/B testing of retention offers (personalized discounts, loyalty rewards, re-engagement communications) among at-risk customers enables causal estimation of intervention effects. Without such testing, the business risks spending retention budget on customers who would have returned anyway (wasted spend) or on customers who are irrecoverably lost (futile spend). askbiz.co integrates churn scores with estimated CLV to prioritize retention interventions, and supports A/B testing frameworks that measure the incremental impact of retention campaigns on customer reactivation rates.