Decay is Key for Customer Return Timing, New Test Confirms

In the realm of customer-return modeling, practitioners often enrich temporal-point-process (TPP) models with numerous signals like lifetime value, category, or recency. However, it's often unclear whether these additional covariates genuinely improve the prediction of event timing. This paper addresses this by providing a credible method to certify the utility of such signals. The first contribution is a "screen-and-confirm" protocol. This method involves planting a positive control (a signal of known strength) to ensure the model can detect it, thereby validating that a subsequent "null" result on real data truly means "no signal" rather than a weak detection method. This protocol was validated on various data types and a real-world dataset. The second key finding is a model-free ceiling on predictability: customer return timing is largely near-memoryless, with only a single-digit percentage of gap variance explained by covariates. Applying their protocol to several public benchmarks and a real marketplace, the researchers certified that continuous-time decay (inter-event clock) is almost entirely sufficient for predicting return timing. Most additional conditioning signals were found to be statistically null or even mildly harmful, suggesting that complexity often adds little value beyond simple decay.