Forecast Bias and Override Culture at a Beverage Distributor
A regional beverage distributor forecasts 1,400 SKUs weekly. Overall MAPE is 31%, but it is 22% for non-promoted weeks and 48% for promoted weeks. Promotions drive 15% of volume. The forecast has a -8% bias, meaning actual demand exceeds forecast. Planners override the statistical forecast on 62% of SKU-weeks; an audit shows overridden forecasts are less accurate than the system forecast in 57% of cases. Sales incentives reward volume, and sales submits its own 'commit' number that planning often adopts. Forecast accuracy is measured only at total company level in the monthly S&OP pack, so SKU-level error is invisible to leadership. A new 12-week energy-drink launch is coming, with no history.
1. Diagnose why forecast performance is poor despite a statistical system.
Read the worked answer
Three mechanisms explain the poor result. First, the metric is aggregated, so a 31% SKU-week error can look acceptable at company level; planners cannot see where error concentrates. Second, the -8% bias is structural: sales' volume-based incentive and its 'commit' number encourage optimistic forecasts, and planning adopts them. The 62% override rate with worse accuracy in 57% of cases shows overrides are not adding value; they are often relationship management. Third, promotions, 15% of volume, are forecast with the same process as base demand, producing 48% error.
The statistical system is being blamed for a process and incentive problem. Fix measurement first: report MAPE and bias by SKU family and by planner, and compute forecast value added (FVA) for overrides against the naive and statistical benchmarks. Do not reward a low MAPE on low-volume SKUs without weighting by volume or cost. Also check whether the system is receiving clean promotion and inventory data; a model cannot separate lift if promotional weeks are not flagged.
2. Propose a revised forecasting and S&OP process for the next two quarters, including specific metrics and how to forecast the new launch.
Read the worked answer
Separate the forecast into baseline, promotional lift, and launch. For baseline, use statistical models by demand pattern (smooth, seasonal, intermittent) and publish a prediction interval. For promotions, require a pre-promo plan with baseline, lift, cannibalization, and post-promo dip; measure promo forecast error separately. Establish an override log with reason codes and require FVA evidence; overrides that do not beat the system over a rolling 13 weeks are removed.
Replace the sales 'commit' with a consensus demand plan owned by planning, with sales accountable for assumptions, not a single number. Measure at SKU-week: MAPE, bias, and weighted error; add FVA and bias by family to the S&OP pack. For the 12-week launch, use analog SKUs, assume a range not a point, set initial allocation by store cluster, and plan a week-4 and week-8 reread with sell-through triggers. Freeze the near-term window only for execution, not for learning. Report forecast accuracy publicly by family so the override debate becomes evidence-based.