Demand Planning and Forecasting
Domain 2 tests forecast method choice, error measurement, hierarchy reconciliation, and the link from forecast error to inventory and S&OP decisions.
Forecasting rules
A forecast is not demand. Separate baseline demand from promotional lift, and track bias separately from random error. Moving averages smooth noise; exponential smoothing gives recent periods more weight through alpha; causal models use drivers such as price and weather; qualitative methods suit new products with no history. Reconcile forecasts across product, customer, and region so the plan adds up.
Error metrics
MAD is the average absolute error; MAPE is average absolute percentage error and is unstable near zero demand; bias is the average signed error. Forecast value added compares a method with a naive benchmark, so a complex model that beats nothing is rejected. Safety stock should reflect forecast error and lead-time variability, not just historical demand variability.
Worked example
Actual demand is 100, 120, 110, 130. A three-period moving average for period 4 is (100+120+110)/3 = 110. If actual period 4 is 130, the error is +20 and the absolute percentage error is 20/130 = 15.4%. A naive forecast using period 3 also gives 110 here, so the moving average adds no value in this case. That is why forecast value added compares against a naive benchmark rather than reporting accuracy in isolation.
Traps that cost marks
Using MAPE on low-volume items; treating bias as random noise; overfitting a model to history; failing to reconcile hierarchical forecasts; confusing forecast accuracy with demand variability.