A €10 Million Plan Still Needs a SKU

Jan-Philipp Grabowski · Presentation takeaways from the KEYNOTION Supply Chain & Logistics Transformation Summit, Munich, 12–13 November 2026

In my talk From Broken S&OP to Zero-Shot AI, I started with a fictional winter-sports company. Its sales and operations planning (S&OP) team targets €10 million in all-mountain ski sales. The buyer has less time. If the supplier freezes capacity in February and the plan is approved in May, that revenue target arrives three months after the buying decision.

Translate revenue into something a buyer can act on

Work from the supplier commitment date backwards. For each stock-keeping unit (SKU), write down the demand quantity, required date, lead time, and supplier minimum or capacity constraint. Make the price and mix assumptions visible so Sales, Planning and Purchasing can resolve them together. Name the person who owns the SKU mix and the person who can authorise the purchase commitment.

Use a baseline to ask a better question

Say Sales plans for 1,200 units and a statistical baseline suggests 950. The difference is 250 units. Ask what could account for them. A confirmed customer order, a promotion, or new distribution might justify a higher plan. Record the reason, its timing, and when it should be reviewed.

One of my slides showed how easily a seasonal rebound can look like growth. The latest three months of synthetic demand averaged 30% more than the previous three. The estimated trend rose 0.7%. In that example, seasonality explained most of the jump. Breaking real demand into trend and seasonal components can help you decide which commercial assumptions to check.

Decide where a foundation model earns a test

A time series foundation model (TSFM) is pretrained on many time series. In a zero-shot test, it uses the item’s demand history as context without extra model training on that item. Put it alongside a baseline that fits the demand pattern.

  • Established items with seasonal demand. Start with seasonal naïve, which repeats the corresponding period from the previous season. Try exponential smoothing (ETS) or ARIMA where the history supports them.
  • New items with short histories. Test a TSFM alongside a comparable product or another launch baseline. For a launch with no sales history, start from analogues, commercial commitments, or other available launch information.
  • Intermittent demand. Give a method designed for infrequent demand, such as a Croston variant, a place in the comparison. Verify which zeros represent genuine lack of demand and where stockouts may have hidden it.
  • Large assortments. Compare the effort, compute cost, and forecast performance of one pretrained model against the existing baseline across representative item groups.

Run a pilot at the buying horizon

  1. Pick the decision. Choose one product group and a recurring purchase decision. For a buyer who orders weekly with four weeks of lead time, test the demand relevant to the next five weeks if that is the period the order must cover.
  2. Backtest the choices. Recreate several past forecast dates and give each method the same information at each cutoff. Compare error and bias at the buying horizon. For models that provide forecast intervals, check how often actual demand fell inside them. Forecasting: Principles and Practice, time series cross-validation
  3. Compare the decisions. Run both forecasts through the same ordering rules, including supplier minimums and pack sizes. Check whether anything changes in the orders and whether the changes improve service, stock exposure, or expediting under the assumptions of the test.
  4. Give the forecast an owner. Use approved data and models. Record who can override the forecast, what evidence supports the change, and who approves the demand plan and purchase commitments. Put a review date on the pilot.

The first experiment can start with one item and one question. Which decision would you make differently if you had a better estimate before the supplier’s cutoff?

Explore the examples