From Customer Profiles to Dynamic Monitoring: Purchase-Sequence Signals and Fresh-Item Repurchase in Online Grocery Shopping

Abstract

Customer summaries record how much a household has bought, not how its orders unfolded. We reconstruct ten-order histories for 1,477,553 target orders from 83,755 Instacart users and derive two sequence measures from each history: the coefficient of variation of interpurchase gaps, and the mean Jaccard overlap of consecutive fresh-item baskets. The outcome is the share of fresh items in a target basket that the household had bought before. Fractional-logit models split each measure into a customer mean and a window-level deviation. Timing irregularity is negative in both comparisons. Basket overlap is not: a one-standard-deviation increase in a customer's average overlap corresponds to 1.77 percentage points more repurchase, whereas the same increase relative to that customer's own average corresponds to 0.84 points less. The sign reversal survives eight-, ten- and twelve-order windows, strictly non-overlapping windows, a fixed complete-user sample, two opposite treatments of censored gaps, a narrower fresh-product definition, and an outcome whose reference period is aligned to the same ten orders. It also survives a customer fixed-effects linear model and a control for the preceding order's repurchase intensity. A customer's usual sequence and a recent departure from it are therefore distinct signals, and storing them as one similarity score discards the second.