Delivery promises and carrier selection in cross-region commerce: a design for logistics intelligence

1Context and motivation

An order is not finished when it is paid for; it is finished when it arrives. The delivery date shown at checkout is a forecast, but the customer reads it as a commitment, and what happens when it is missed is documented. In archival data from an online retailer, previously loyal customers who experienced a fulfilment delay placed orders less often and in smaller amounts afterwards, and their measured anxiety rose¹. Order fulfilment cycle time is also linked to whether customers refer the retailer to others². Customers who shop online rate fulfilment, meaning whether the order arrives as and when promised, as one of the four core dimensions of electronic service quality³.

Speed and reliability both carry value. When a retailer opened a new distribution centre and deliveries to part of its market became faster without announcement, online sales there rose with each business day removed from delivery time, with a positive spillover to its physical stores⁴. When a large retail platform temporarily removed a high-quality delivery provider as a shipping option, sales of the affected products fell during the removal, and the loss was larger for expensive products, where customers need reliable delivery to protect valuable items⁵.

Cross-region commerce makes the problem harder. A merchant selling across countries ships on lanes with very different transit times, scan coverage, customs steps and failure rates, and a single shipping table cannot describe them. We propose a logistics layer that treats each promise as a decision made from a forecast distribution, and each carrier and service-level choice as a lane-specific decision with explicit costs. This paper describes that design and the reasons for each part. We have not deployed or measured it.

2Design principles

• Promises come from distributions, not point estimates. A mean delivery time says nothing about how often a promise built on it will be missed. The design forecasts the distribution of delivery time and sets the promise at a chosen quantile of it⁶.
• The cost of being late and the cost of promising long are stated. Both are parameters a merchant or operator sets, and the promise follows from them by a rule that can be inspected.
• Lanes are the unit of analysis. A lane is an origin node, a destination area, a carrier and a service level. Forecasts, reliability scores and service menus are kept per lane, pooled across similar lanes where data are thin.
• Forecasts are scored as forecasts. Distributional forecasts are evaluated with proper scoring rules, so that a model cannot score well by being confidently wrong⁷.
• Service by geography and merchant size is monitored. Optimising only for cost or sales can quietly withdraw fast options from remote areas and from small merchants. The design reports service-level availability by those groups rather than leaving it implicit.

3Architecture

3.1Event record

The layer consumes order and fulfilment events (order creation, pick, pack, handover to carrier, delivery confirmation, return, cancellation), carrier scan events with delay codes, cost records (transport, surcharges, handling), inventory by node with item size and weight, and context such as destination type, customs requirements and calendar holidays. Every promise shown to a customer is logged with the forecast that produced it, so that promised and actual delivery can be compared per lane afterwards.

3.2Delivery-time forecasting

The forecasting component predicts, for a new order, the distribution of time from order to delivery on each candidate lane. Two families of method support this.

• Quantile forests. Random forests can estimate the full conditional distribution of a response, not only its conditional mean, so conditional quantiles of delivery time can be read from the same fitted model⁸. Delivery-time promising has been studied directly in this frame: tree-based distributional forecasts built from operational predictors, with a split rule that uses an asymmetric loss and a cost-sensitive rule that picks the promised day from the predicted distribution⁶.
• Refinement of route-based estimates. Where a carrier or routing engine already supplies an estimate, a learned model can correct it. Estimates conditioned on a planned route can be unreliable when the actual route is not known in advance, and a residual network that post-processes them has been reported to reduce mean and median absolute error at large scale⁹.

Lanes with little history borrow strength from similar lanes through shared features (distance band, carrier, customs regime, destination type). A lane whose recent forecasts score poorly is flagged and falls back to a wider, more conservative promise until enough new data arrive.

3.3Promise rule

The promise rule turns a forecast distribution into a date. Let a be the cost of each day the promise is longer than the actual delivery (customers deterred by a long window) and b the cost of each day the delivery is later than promised (contacts, refunds, lost repeat orders). Under these linear costs, the expected cost is minimised by promising the quantile of the delivery-time distribution at level b / (a + b). This is the asymmetric piecewise linear loss for which a quantile is the optimal point forecast⁷, and it is the structure behind the cost-sensitive promise rule studied for online retail⁶. Table 2 works this rule through for three hypothetical lanes.

The rule is deliberately simple. It makes one thing explicit that fixed shipping tables hide: a shorter promise is a choice to accept more late deliveries, and the ratio of the two costs decides how many.

3.4Service-level and carrier selection

For each order, the selection step decides which service levels to offer, which to make the default, and which carrier fulfils each. It scores candidate options on forecast delivery time, lane reliability (the probability of delay, loss or damage), cost to serve including dimensional weight, and tracking coverage. Three rules apply before any learned score:

• an option is not offered if its forecast quantile at the promise level is no faster than a cheaper option on the same lane;
• an express label is shown only where the lane's distribution supports it at the chosen quantile;
• on lanes with high loss or dispute probability, a tracked service is the default.

Route-level decisions inside a carrier's network are out of scope; that is the vehicle routing problem and its variants, for which a mature body of exact and heuristic methods exists¹⁰. The layer treats carriers as services with forecast performance, not as routes to optimise.

3.5Inventory placement

Delivery time depends on where stock sits before an order is placed. For merchants with stock in more than one node, the design recommends placement using the forecast lane distributions as inputs. Stock decisions at linked locations interact, and the structure of optimal policies for a serial multi-echelon system, where each installation is supplied by the next, was characterised early in the inventory literature¹¹. The design uses that framing for recommendations only; it does not move stock automatically.

3.6Governance

The governance component reports, per destination area and per merchant size band, which service levels are available, the promise length shown and the share of promises met. A change to the promise rule or the selection rules that narrows availability for a group is flagged for review before it is rolled out. Logistics telemetry includes addresses and delivery patterns, so the design keeps records in their region, restricts access by role, and retains only the fields forecasting needs.

Table 1 maps each component to the problem it addresses and the evidence behind it.

Component	Problem it addresses	Evidence
Logged promise per order	Missed promises cannot be measured or attributed later	1
Distributional delivery-time forecast	A point estimate hides how often a promise will be missed	8, 6
Refinement of route-based estimates	Route-conditioned estimates are unreliable when the route is unknown	9
Quantile promise rule from stated costs	Promise length is set by habit, not by the cost of lateness	7, 6
Proper scoring of forecasts	Overconfident forecasts score well on point metrics	7
Reliability in carrier scoring	Fast but unreliable options lose sales on valuable items	5
Delivery speed in service design	Slow delivery suppresses online and offline sales	4
Fulfilment as a service quality dimension	Logistics treated as cost only, not as part of service quality	3, 2
Multi-echelon framing for placement	Stock decisions at linked nodes made one node at a time	11
Carriers treated as services, routing out of scope	Scope creep into network routing	10

Source: our design, with the works cited in the table.

4A worked promise model

Table 2 applies the promise rule to three hypothetical lanes. The quantiles of delivery time and the two costs are inputs we chose to show the arithmetic; they are not measurements of any lane or carrier.

Lane (hypothetical)	Median (input)	80th pct (input)	90th pct (input)	95th pct (input)	Cost per day promised too long, a (input)	Cost per day late, b (input)	Promise level b / (a + b) (computed)	Promise (computed)
Domestic express	1	1	2	2	1	4	0.80	1 day
Domestic standard	3	4	5	6	1	9	0.90	5 days
Cross-border tracked	6	9	11	14	1	19	0.95	14 days

Source: Worked model; inputs in the table.

Three points follow from the arithmetic. First, the promise depends on the spread of the distribution, not on its median: the cross-border lane's median is 6 days, but at a promise level of 0.95 its promise is 14 days. Second, the promise level is set by the cost ratio alone: raising b from 4 to 9 with a fixed moves the level from 0.80 to 0.90. Third, the rule promises a date that is missed, by construction, on about 1 minus the promise level of orders if the forecast is calibrated; on the cross-border lane that is about 5 in 100. Calibration therefore matters as much as accuracy, which is why the design scores forecasts with proper scoring rules⁷.

The linear cost model is an assumption. If lateness beyond some number of days triggers a refund, the cost of lateness is not linear and the rule needs the full loss function instead of a single ratio.

5Illustrative scenarios

The scenarios below are hypothetical. They show how the parts would interact; they are not results.

Next-day labels where they hold. A merchant offering next-day delivery in two countries would see the express label kept on lanes whose forecast 80th percentile is one day, and replaced by a two-to-three-day standard promise on lanes where it is not. Whether customers then buy more or less is exactly what an experiment would have to show.

Metro and remote destinations. On a lane set where metropolitan deliveries have a tight distribution and remote ones a wide distribution, the rule would give a short promise in the city and a longer one elsewhere. Governance reporting would show the longer promise and the share met for remote areas side by side, so that the difference is a visible decision rather than a silent one.

High-risk cross-border lanes. On a lane with frequent loss in transit, the selection step would default to a tracked service. The cost of tracking would appear in the cost-to-serve figure, and an experiment would compare disputes and repeat orders against an untracked default.

6Evaluation protocol

Before any claim about outcomes, a deployment would need the following.

• Forecast evaluation. Distributional forecasts are scored on held-out orders with a proper scoring rule, such as the quantile loss at each promise level, and checked for calibration: the observed share of deliveries within each quantile should match its level, per lane⁷.
• Promise experiments. The quantile promise rule is compared with the existing shipping table in randomised experiments with a pre-registered target metric (conversion), guardrail metrics (late-delivery rate, contact rate, cancellations) and a minimum detectable effect¹².
• Repeat behaviour. Because the documented effects of a missed promise appear in later orders, outcomes are tracked over a follow-up window long enough to observe repeat purchase, not only at checkout¹.
• Distribution of service. Every experiment reports results by destination area and merchant size band, not only on average.

7Limitations

• No measurements. We have not measured the effect of this design on conversion, late deliveries, contacts or repeat orders. Every benefit it implies is a hypothesis for the protocol above.
• Evidence from other settings. The cited effects of delays, speed and delivery quality come from particular retailers, platforms and countries. Whether their size carries over to other lanes and categories is unknown.
• Linear costs. The promise rule assumes linear costs of early and late delivery; refund thresholds, perishable goods and scheduled deliveries break that assumption.
• Thin lanes. New and low-volume lanes have little history, forecasts there are uncertain, and pooling across similar lanes can hide lane-specific problems such as a single customs point.
• Carrier data quality. Scan coverage and delay codes differ by carrier and country. Missing scans bias forecasts toward the lanes that report well.
• Special categories. Cold chain, oversized items and regulated goods need constraints this design does not model.
• Behavioural response. Customers may adjust to promises over time, and carriers may adjust to being scored. The design does not model either feedback.

8Conclusion

We have described a logistics layer that sets delivery promises from forecast distributions with a quantile rule driven by stated costs, chooses service levels and carriers per lane, recommends inventory placement, and reports service by geography and merchant size. Each part rests on published work, summarised in Table 1, and Table 2 shows the promise arithmetic with its inputs. What the design lacks is evidence of its own effect. The next step is to score delivery-time forecasts on logged orders for one lane set and run a promise experiment against the existing shipping table.