Capstone — The Evening Delivery Promise

An integrated HarborMart case from decision contract to monitored rollout

Capstone — The Evening Delivery Promise

The decision

HarborMart's regional manager asks: “Can the risk model reduce evening delivery failures?” Your task is not to praise or reject the model. You must design a defensible decision system.

At checkout, the system may:

  1. accept the requested window;
  2. offer a later promise;
  3. reroute to a nearby fulfilment centre when feasible; or
  4. call additional picker capacity for a future slot.

The customer must see the promise before payment. Information observed after that moment is unavailable to the decision.

Evidence pack A: eight-week operating summary

There were 24,000 completed evening orders and 312 cancellations after acceptance.

WindowCompleted ordersLateLate rate
16:00–17:006,0003305.5%
17:00–18:007,0005608.0%
18:00–19:006,50071511.0%
19:00–20:004,50077017.1%
total24,0002,3759.9%

This table does not prove that the time window causes lateness. Store, distance, basket and capacity mix also change through the evening.

Evidence pack B: data dictionary

FieldGrain or timingAudit note
orders.order_idone accepted orderprimary key; cancelled orders remain
order_itemsmany rows per orderjoining before aggregation duplicates orders
score_at_checkoutstored before promiselegitimate historical prediction
scheduled_picker_minutesknown at checkoutlegitimate capacity feature
actual_picker_minutesfinal shift recordleakage for checkout prediction
weather_forecast_at_checkoutforecast version stored at checkoutlegitimate if timestamp is preserved
actual_weatherobserved lateroutcome-time information
arrival_timeobserved for completed deliverymissing for cancellations
customer_service_notefree text after deliverypossible leakage and sensitive content
service_areaoperational zoneuse for error and impact audit

The late label is arrival_time > promise_end. Decide explicitly how cancellation, re-promise and missing arrival should enter the business outcome; silently dropping them rewards avoidance.

Evidence pack C: model audit

The candidate model was evaluated on the latest four weeks, after all training periods.

Risk bandOrdersMean predicted riskObserved late rate
below 5%8,0003.0%3.0%
5%–10%7,0007.0%7.0%
10%–20%5,50014.0%14.0%
above 20%3,50025.0%25.0%
MeasureConstant-risk baselineCandidate
Brier score0.0890.078
AUROC0.5000.720

Aggregate calibration hides one issue:

Service areaOrdersMean predictionObserved rate
central and inner19,2009.7%8.5%
outer4,80010.6%15.4%

The outer-area result may reflect model error, changed road conditions, weak labels or a different data-generating process. Diagnose before changing a threshold.

Evidence pack D: candidate actions

The finance team values a prevented late delivery at £9 after refunds, support time and estimated retention effects. Treat this figure as an assumption requiring sensitivity analysis.

Action over eight weeksDirect costEstimated consequenceEvidence quality
add one picker, 17:00–20:00£2,160420 fewer late orders, interval 250–580randomised shift pilot
reroute up to 600 eligible orders£1,440210 fewer late ordersadjusted observational comparison
offer later windows to top 300 risks£1,800 lost contribution90 fewer recorded late ordersmodel simulation only
keep current rule£0baselineobserved history

Capacity allows only one additional picker. Rerouting is unavailable for chilled baskets and adds a median 11 minutes to travel. A re-promised order is not “on time” under the original promise; retain both original- and revised-promise measures.

Your six decisions

A. Frame

Write an analytics contract naming owner, decision time, unit, eligible population, actions, objective, constraints, baseline, horizon and review rule. State one descriptive, predictive, causal and prescriptive claim the evidence can or cannot support.

B. Audit

Specify the order-level analysis table. Show how you will prevent item-join duplication, temporal leakage and silent cancellation exclusion. Add five executable data checks.

C. Diagnose

Produce one decision display that separates volume from late rate and exposes the outer-area problem. Include counts, denominators and uncertainty. Give two rival process explanations.

D. Predict

Decide whether the candidate improves on the baseline for this decision. Evaluate temporal validity, calibration, ranking, subgroup error and the cost/capacity range. Do not select a threshold from AUROC alone.

E. Prescribe

Recommend an action or staged combination. Calculate expected net value, test the £9 assumption and use the intervention evidence quality. State why “offer later windows to top risks” may optimise the reported metric while harming service.

F. Deploy and learn

Design a shadow stage and a controlled rollout. Specify five monitors, two guardrails, one pause condition, one rollback rule, owners and a causal value evaluation. Explain how selective labels will arise.

Deliverable template

This is a teaching template, not an institutional rule:

  1. executive decision memo: recommendation, expected value range, material uncertainty and next decision;
  2. technical appendix: contract, data audit, validation, policy calculation and sensitivity;
  3. control sheet: monitoring, fairness/privacy review, incident and rollback rules;
  4. reproducibility record: data snapshot, code/environment, assumptions and generated outputs.

Evaluation rubric

CriterionStrong work demonstrates
decision logicobjective, alternatives, constraints and counterfactual align
evidence disciplinedescription, prediction, causation and prescription are not conflated
technical validitygrain, time, denominator, baseline, calibration and uncertainty are audited
operational realismaction is feasible; ownership, latency, feedback and rollback are explicit
responsibilityaffected groups, privacy, security, review and contestability shape the design
communicationconclusion is concise, traceable and honest about what is unknown

Postgraduate work should add an identification diagram, a formal estimand, robustness to model misspecification and decision-regret comparison.

Instructor checkpoint — not a unique solution

The additional picker has point-estimate gross value 420(£9)=£3,780420(£9)=£3,780 and net value £1,620. At the pilot interval endpoints, net value ranges from £90 to £3,060; break-even value per prevented late order is about £5.14. That makes a staged picker rollout defensible, not certain.

Rerouting has point-estimate net value 210(£9)£1,440=£450210(£9)-£1,440=£450, but weaker causal evidence and eligibility constraints. Re-promising may reduce the recorded late rate by redefining service rather than improving it. Preserve original-promise outcomes.

A strong answer investigates outer-area calibration, keeps cancellations in a guardrail, logs proposed and executed actions, and randomises rollout where feasible. It does not feed actual_picker_minutes, actual_weather or service notes into a checkout-time model.

Use the formula map, then compare your reasoning with the worked pitfalls.

Copyright © 2026