Simple decay beats complex signals in customer-return timing
New certification tools show that adding behavioral and geographic signals to customer-return models provides little to no predictive gain.
Adding more data to customer-return models does not necessarily improve the prediction of when a customer will return. A new screen-and-confirm protocol demonstrates that most common conditioning signals—such as lifetime value, geography, and category—are redundant or even harmful when layered on top of a basic continuous-time decay model.
This finding challenges the prevailing industry practice of enriching temporal-point-process (TPP) models with increasingly complex feature sets. The research establishes a model-free ceiling showing that only a single-digit percentage of gap variance in return timing is predictable from any covariate. This suggests customer returns are nearly memoryless, meaning the timing of the next event depends far more on the passage of time than on the specific attributes of the user.
Companies that invest heavily in the engineering and maintenance of high-dimensional feature pipelines for churn and return prediction may be over-investing in low-yield infrastructure. If the timing of a return is largely independent of external covariates, the computational cost of processing real-time signals like recency or frequency provides diminishing returns.
The research provides a method to certify these null results by using a positive control. By planting a signal of known strength and confirming the model can recover it, the authors prove that the lack of signal in real-world data is a property of the customer behavior itself, rather than a failure of the model to detect the pattern.
Testing across benchmarks including Amazon, Taobao, and RetailRocket, as well as a live marketplace, showed that additional conditioning was statistically null or mildly harmful. The inter-event clock—the simple decay of time since the last event—remains the dominant predictor.
Operational focus should shift from feature accumulation to the refinement of the underlying decay curves. The evidence suggests that the most accurate timing models are those that prioritize the temporal gap over the demographic or behavioral profile of the customer.