Retail promotion forecasting: run a bounded planner pilot

A four-store homeware promotion starts in three weeks. The supplier needs an order now, but the team has only two comparable events and a proposed model that predicts a 25% uplift. The relevant question is whether that estimate improves the buying decision under the available evidence.

Start with a bounded planner pilot: one category, a fixed decision date, a defined promotion window and a baseline the buyer understands. Keep the original forecast and purchase recommendation so later results can be compared with what was actually known.

Turn the forecast into a buying scenario

Assume baseline demand of 100 units per store during the promotion: 4 × 100 = 400 units. A proposed 25% uplift gives 400 × 1.25 = 500 units. This is a scenario, not a measured model advantage.

Across the stores, 160 units are eligible stock and 80 confirmed inbound units will arrive before the event. The planner wants 40 units left after the promotion. Under those assumptions:

Required new receipts = 500 demand + 40 ending stock − 160 current stock − 80 inbound = 300 units.

The supplier ships packs of 24 with a minimum order of 240. Rounding 300 upward gives 13 × 24 = 312 units, which meets the minimum. Total available supply becomes 552, leaving 52 units if demand is exactly 500.

A forecast difference becomes a commitment difference. The planner should review the cost of excess and shortage before choosing a scenario, including whether the supplier can deliver in time. A lower forecast error does not automatically justify a larger irrevocable order.

Separate demand from sales constrained by stock

Retain daily item-store sales, selling price, promotion details, availability, returns, receipt dates and any known closures. A store that sold 60 units and stocked out may have had more than 60 units of demand. Do not label all missed sales as known demand, either: lost-demand estimates need their own method and uncertainty.

Shopify's inventory definitions help identify why on-hand and sellable quantities differ. The pilot needs historical availability at the decision-relevant grain, not only today's stock total.

For every feature, ask whether it existed before the order deadline. Final promotion sales, later stock receipts and revised event labels cannot be used to score a forecast as if the buyer knew them earlier. Forecasting: Principles and Practice explains forecast accuracy and rolling-origin evaluation using earlier observations to predict later ones. Match the evaluation horizon to the lead time and buying decision.

1. Freeze the decision inputsFour stores; baseline 400; stock 160; inbound 80.
2. Compare scenarios10%, 25% and 40% uplift imply demand of 440, 500 and 560.
3. Apply buying constraintsOrders become 240, 312 and 360 after packs and minimums.
4. Evaluate the outcomeCompare error, available stock, leftovers and planner overrides.

Exception: Sparse history, unavailable stock or a changed promotion mechanic keeps the result as a labelled scenario for human review.

Check the error calculation without inventing a winner

An illustrative evaluation exercise has actual demand observations of 100, 120, 80 and 100 units. A baseline predicts 90, 110, 90 and 110. Absolute errors are 10 in each case, giving mean absolute error 40 ÷ 4 = 10 units.

A candidate predicts 105, 115, 85 and 105. Its absolute errors are five in each case, giving 20 ÷ 4 = 5 units. The candidate performs better on this invented four-observation exercise. It is not evidence that the proposed production model will outperform a baseline.

Also inspect signed error. The baseline total is 400, matching the actual total of 400 despite location-level misses. The candidate total is 410, over by ten even while its absolute error is lower. A network total can hide store allocation problems, and one metric can hide bias.

Choose the smallest useful pilot

Begin with comparable-event or seasonal baselines. Consider a more complex model only when there is enough relevant history and a repeatable evaluation. Keep the same data cutoffs, assortment and horizons for both methods. Hold out complete future events where possible rather than randomly mixing rows from the same promotion into training and evaluation.

The demand planning manager owns forecast use; the buyer owns the commitment; finance approves the cost basis. Record overrides with a reason such as a changed promotion, supplier risk or store closure. Do not edit the original forecast to make the final decision look model-generated.

Acceptance requires reproducible inputs and calculations, visible low-evidence cases and no automatic PO release during the pilot. Report availability and unsold units beside forecast errors. Compare the recommendation with the replenishment policy and the category budget before approving purchases.

Preserve the decision, then learn from it

The pilot succeeds when the team can see whether a forecast helped a specific purchase under known constraints. A polished uplift chart is insufficient without the order calculation and the later stock outcome.

1. Choose one promotion family, forecast horizon and simple baseline.
2. Freeze historical inputs at each simulated decision date.
3. Convert forecasts into constrained order scenarios and retain planner decisions.
4. Review error, shortages, leftovers and override reasons before expanding the model's authority.