Delivery cutoff slack and delay odds statistics https://convorescue.com/guides/delivery-cutoff-slack-and-delay-odds-statistics/ ## Quick answer Delivery cutoff statistics show a tradeoff a shopper can feel: extending a promise can make its deadline easier to meet while the parcel still takes longer. In a March 2018 JD.com study, orders just after an 11 am cutoff had about 58.5% lower modeled odds of missing their promise. The comparison involved 16,710 orders and a switch from same day to next day delivery. Lower odds do not mean an equally large reduction in the percentage of orders arriving late (Balakrishnan, Choi and Epstein (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021). For a merchant, repairing a slow warehouse stage and extending a delivery promise solve different problems. The first can shorten the journey. The second changes the standard against which the journey is judged. Reading the source's local cutoff comparison helps separate those outcomes and decide what a delivery reply should explain. The historical estimates concern one product category; their main use is to clarify the measurement, rather than set a new store's cutoff. ## Key Takeaways The arrival delay odds ratio is 0.415. It is an original transformation of the cutoff coefficient, with an approximate 95% interval of 0.355 to 0.485 (Balakrishnan, Choi and Epstein (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021). The comparison covers 16,710 orders near 11 am. The regression bandwidth runs from 10 am to noon, within a same or next day delivery promise (Balakrishnan, Choi and Epstein (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021). A slow leg is not the same as a missed customer promise. The study defines a slow leg relative to the 95 th percentile of ontime orders placed at the same hour (Balakrishnan, Choi and Epstein (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021). ## What changed immediately after the delivery cutoff? The promised arrival moved from the same day to the next day. The study uses that discontinuity to investigate what extra allotted time changes, comparing orders placed close together around the threshold. Figure description: Delivery cutoff statistics for March 2018 JD.com orders: adjusted odds ratios after 11 am are 0.415 for arrival after promise, 0.657 for a slow warehouse leg, 0.093 for a slow middle leg and 0.368 for a slow last leg, with approximate 95% intervals. Figure 1: Postcutoff orders had lower modeled odds of each delay outcome. Demand bunched before the cutoff, so the local interpretation depends on the source identification assumptions. Source: Balakrishnan, Choi and Epstein, Choi dissertation Chapter 2 Table 2.2 (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021. Original analysis of the cited public source. Source periods and limits accompany the finding. ### Why is this different from comparing early and late shoppers? Orders placed throughout the day can differ in destination, product mix or operational load. The cutoff design instead examines a local jump while allowing outcomes to vary with time on either side. Its argument is that nearby orders are otherwise comparable enough to identify the effect of the promise regime. Orders visibly clustered immediately before 11 am as shoppers sought same day delivery. The source acknowledges that discontinuity and argues that fulfilment workers could not control when shoppers placed orders. Because its outcomes concern the delivery process, the authors believe this bunching does not invalidate their identification argument. That remains an assumption about the local comparison, rather than proof that shopper selection cannot matter. The source also reports small discontinuities in gifts and quantity discounts, so nearby orders must still be sufficiently comparable despite those differences. ### Why does a later promise not establish faster service? A deadline can be easier to meet simply because it is later. A merchant can raise its ontime rate while customers wait longer. Reporting only whether the promised date was met would miss that tradeoff. For a delivery conversation, keep two facts available: the elapsed journey and the promise originally shown. Reassurance should reflect the customer's remaining time, not merely a dashboard status that improved after a deadline changed. ## How should the odds ratios be read? An odds ratio compares odds, which are the probability of an event divided by the probability of its absence. It cannot be read as an equal change in the event's probability. ### What does 0.415 mean in practice? The transformed arrival coefficient implies that postcutoff odds were about 0.415 times the comparison odds at the threshold, within the fitted model. The confidence interval describes uncertainty in that estimate, not the range of possible arrival dates for one parcel. An absolute late rate needs the relevant starting probability and model context. The source coefficient alone does not supply a universal before and after percentage. Applying a 58.5% reduction directly to all late shipments would conflate odds, probability and a local historical estimate. ### Do the stage ratios compare equivalent failures? No. Warehouse, middle and last leg flags each compare a stage duration against a reference distribution for that stage and order hour. The customer arrival flag instead checks whether the order missed its promise. Their odds ratios answer different questions. That prevents an attractive but wrong inference: the middle leg's smaller odds ratio does not prove it is the most valuable stage to optimize. The baseline frequency, time saved, resources required and effect on final arrival all matter. Different outcome definitions also rule out adding the ratios into an overall improvement. ## What should a merchant measure before changing the cutoff? Measure promise compliance and journey duration together. Add operational stages when they help identify a preventable delay, while retaining the customer's actual deadline as the final service outcome. ### How can stage alerts become useful? Store the order time, warehouse exit, intermediate arrival and final delivery timestamps on the same basis. Define whether clocks use calendar time or working time. Explain what each alert means so support staff can distinguish an unusually slow stage from a confirmed missed promise. An alert should lead to a useful action, such as checking a missed collection or confirming an updated estimate. The paper's predictive analysis does not establish that every flagged order can be rescued, or that sending an extra notification prevents delay. ### What would a meaningful cutoff test preserve? Keep the original promise in the data, allow the delivery window to mature and track cancellations, contacts and elapsed time alongside late rates. A change that improves compliance but creates longer waits can still be commercially unattractive. A change that shortens waits but frequently misses an explicit date can also fail the shopper. Use the distinction alongside the carrier handoff promise analysis (https://convorescue.com/guides/delivery-promise-time-at-handoff-statistics/) and dispatch versus transit statistics (https://convorescue.com/guides/ecommerce-dispatch-and-transit-time-statistics/). Those describe different stages and populations rather than alternative estimates of this cutoff effect. ## How much does the odds result change the late-arrival probability? The reported coefficient of negative 0.88 implies an odds ratio of exp(−0.88), approximately 0.415. An odds ratio alone does not identify the probability after the cutoff. That requires a baseline probability. Holding this point odds ratio fixed, a worked baseline of 10% becomes 4.41%, a baseline of 25% becomes 12.15% and a baseline of 50% becomes 29.32%. These are mathematical scenarios, not additional JD.com observed groups. (Balakrishnan, Choi and Epstein (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021) Figure description: The same odds ratio implies different absolute and relative probability changes at different baselines. Figure 2: The same odds ratio implies different absolute and relative probability changes at different baselines. Source: Balakrishnan, Choi and Epstein (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content), Sep 2021. Original analysis of the cited public source. Source periods and limits accompany the finding. The corresponding relative probability reductions are approximately 55.9%, 51.4% and 41.4%. Each differs from the 58.5% reduction in odds, despite using the same source coefficient. The distinction becomes more consequential when the starting probability is large. Calling the odds reduction a probability reduction would overstate what the model says for those scenarios. The conversion uses p_after = OR × p_before / (1 − p_before + OR × p_before). It preserves a probability between zero and one and makes the missing baseline explicit. We do not substitute the pooled late rate from another delivery study, a different country or a different service. Those populations would not supply the missing model baseline for this cutoff comparison. There is a second interpretation limit. Orders just after the cutoff received a different promised delivery day. Lower modeled odds of missing that promise do not mean the parcel reached the customer sooner. The deadline moved as well as the order's operational circumstances. A service report needs both elapsed delivery time and performance against the commitment. A merchant investigating a cutoff should therefore retain the time the order was placed, the promise actually shown, the relevant dispatch schedule and the actual arrival. Keep orders on either side of the cutoff distinct where the operational policy changes. A generic comparison of “before” and “after” orders can contain a different promise, a different day mix and a different available service. These scenarios are useful for reading a statistical result accurately. They are not a forecast or an instruction to change a store's cutoff. The historical source has its own local comparison, regression controls and uncertainty. Applying its point odds ratio to assumed probabilities isolates one mathematical relationship while leaving those transfer questions open. ## Methodology and Sources We use Chapter 2 of MoonSoo Choi's September 2021 Harvard dissertation, which credits the delivery study to Maya Balakrishnan, MoonSoo Choi and Natalie Epstein. Table 2.2 reports logistic coefficients and standard errors for orders placed from 10 am to noon around the 11 am cutoff. We calculate exp(coefficient) and exp(coefficient plus or minus 1.96 standard errors). These approximate Wald intervals are original transformations of published estimates, not a refit of transaction data. Section 4.4(b) reports pre 11 am order bunching and explains the authors' worker control argument; this qualification remains part of the interpretation. The source uses March 2018 JD.com data from one category. Stage lateness is a relative duration definition. We do not treat conditioned stage duration comparisons as general causal effects or claim that the estimates apply to another store. ## Glossary Cutoff: The order time at which the promised delivery regime changes. Time allotment: The time available between ordering and the promised deadline. Regression discontinuity: A design comparing outcomes near a threshold that changes treatment. Bandwidth: The span of observations included around that threshold. Odds: An event probability divided by one minus that probability. Odds ratio: A comparison of odds, rather than percentage point differences. Slow leg: A stage longer than the source's reference duration threshold. Wald interval: An approximate uncertainty interval computed from a coefficient and standard error. Reproduce the odds transformations The CSV includes the reported coefficients, standard errors, transformed ratios, approximate intervals and original source locator. ## Sources Choi: Information Design in Operations Management, Chapter 2, An Empirical Study of Time Allotment and Delays in Ecommerce Delivery (https://dash.harvard.edu/server/api/core/bitstreams/a03cda89-7076-4d0b-9e0a-b04e97d434f1/content) Inspect the calculations and source distinctions Download the inputs and formulas behind this comparison. Run the calculations Use the original public dataset where required. The file includes the formulas and any worked sensitivity inputs.