The Accuracy-First S&OP Mandate at a Specialty Foods Manufacturer
A specialty-foods manufacturer runs its monthly S&OP as a two-hour review of a single consensus volume number per product family. Over the last twelve months, family-level MAPE averaged 34% with a persistent bias of −11%, and the error is concentrated in four promotional families. Sales books quarter-end trade deals after the plan is locked. Supply builds to the plan, so the plants run a 14-week frozen schedule, finished inventory sits at 62 days against a 45-day target, OTIF is 78%, and expedite freight has reached $1.2M. The VP of supply chain wants forecast accuracy below 15% MAPE before any other change is made.
1. Evaluate the VP's sequencing - accuracy first, everything else after. What is the strongest argument against it, and what would you do over the next two S&OP cycles?
Read the worked answer
The sequencing is backwards, and partly based on a false premise.
- A large share of the 34% is a process artifact, not a forecasting skill problem. A −11% bias means the plan is systematically low, which by itself explains much of the expedite freight. Promotions booked after plan lock mean supply is being asked to build to a number that is known to be wrong at the moment it is signed. Neither is fixed by better statistical models.
- Residual error cannot be driven to 15% at family level in promotional categories in any reasonable timeframe. Waiting for it before shortening the frozen window or building mix flexibility means two more years of the current cost structure. Forecast error is a cost to be absorbed, not a precondition to be met.
- The cheap fixes are concurrent, not subsequent. De-bias the plan immediately (apply a rolling bias correction and say so), and move trade deals inside the demand review so the promo is planned as a discrete event with its own shape.
Next two cycles: (1) publish a bias-corrected plan alongside the raw consensus number, with the delta explained; (2) require promo plans to be submitted before the demand review, with finance-validated lift assumptions; (3) run a parallel track on responsiveness - shorten the frozen zone, define a capacity buffer and mix-flex plan for the promo families. Report the two tracks separately so the organization stops debating the single accuracy number.
2. Redesign the cadence and the measures. Specify what is forecast at which level and horizon, and what the executive review should actually decide.
Read the worked answer
Forecast at the level you can act on, and no lower.
| Level | Horizon | Purpose | Error measure |
|---|---|---|---|
| Family / total volume | 18–24 months | Capacity, capex, labor, supplier capacity reservation | MAPE and bias separately |
| Family × month, promo events as separate line items | 3–6 months | Rough-cut capacity, material commitments, promo build | Bias (systematic), event lift accuracy |
| SKU × location | 8 weeks (inside lead time: not a forecast, a plan) | Production scheduling, deployment | Mix error, not total error |
Measures to put on the executive page:
- Bias and MAPE separately. Bias is actionable and cheap to fix; MAPE is not. A single blended metric hides the fixable part.
- Forecast value added by family: does the demand planner beat the naive baseline? If not, stop paying for that step.
- Plan adherence and decision latency - how long between an imbalance being visible and a decision being taken. In this company the gap between OTIF 78% and the plan is usually a decision problem, not a number problem.
- Cost of error: expedite freight, overtime, write-offs, and lost sales, tagged to the family that caused them. This converts the forecast debate into a spend conversation.
The executive review should decide, not present. Agenda: the three largest demand–supply imbalances, the scenario chosen, the money authorized (expedite, buffer, capacity), and the trade-off accepted. If no decision is recorded, the meeting failed.
3. The 14-week frozen window is being defended as the only way to keep the plants stable. Assess that claim and describe how you would change the window without destabilizing production.
Read the worked answer
The 14-week freeze is longer than the error horizon justifies and locks mistakes in place. A frozen window exists to protect schedules from churn within material lead times and capacity commitments. Fourteen weeks exceeds the lead time of most packaging and many ingredients, so the company is freezing a forecast it cannot make accurately at that range - then paying expedite freight to undo the freeze.
Structure it in zones instead of one wall:
- Frozen (no change): the material lead-time horizon - roughly 4–6 weeks for constrained ingredients and packaging, longer only for genuinely long-lead items.
- Slushy (changeable within limits): ±10% by family per week, with the cost of change charged back visibly to whoever requests it.
- Open: beyond the material horizon, planned normally.
What earns the plant its stability instead of the freeze: a small finished-goods buffer on stable runners sized to mix error; a defined overtime/flex capacity envelope agreed in advance; a sequencing rule that keeps the constraint's changeovers protected; and a published promise that says how much change is legitimate. The plant's real complaint is usually unplanned change, not change itself - give it a bounded, priced channel for change and the frozen wall can come down.
Risk created: capacity whipsaw in the first two months while planners learn the new limits. Monitor schedule adherence weekly and be willing to narrow the slushy band before widening it.