Demand Planning and Forecasting
Demand planning and forecasting tests whether you can choose a forecast method, measure its error honestly, and turn the result into a plan. Questions often cover moving averages, exponential smoothing, seasonality, forecast error metrics, bias, and S&OP.
Core rules
- A forecast is always wrong; the goal is a bounded, unbiased error that supports decisions. Forecast horizon and aggregation must match the decision.
- Time-series methods decompose demand into level, trend, seasonality, and noise. Exponential smoothing weights recent demand more heavily; the smoothing constant alpha controls responsiveness.
- Causal and regression models use drivers such as price, promotions, or weather. They are only as good as the driver data and the assumed relationship.
- MAD is the average absolute error. MAPE expresses absolute error as a percentage of actual demand. Bias is the average signed error; a tracking signal reveals persistent bias.
- S&OP reconciles demand, supply, and financial plans in a monthly cycle; it is not a one-time forecast approval meeting.
Worked example
Actual demand is 100 in period 1 and 120 in period 2. Forecasts were 90 and 110. Errors are 10 and 10, so MAD = (10 + 10) / 2 = 10 units. MAPE = [(10 / 100) + (10 / 120)] / 2 = (0.10 + 0.0833) / 2 = 9.17%. Bias = (10 + 10) / 2 = +10, meaning the forecast consistently understated demand. A low MAD with a large bias is dangerous because the planner may miss a systematic shortfall.
Traps that cost marks
- Computing MAPE when actual demand is zero or near zero; the percentage explodes or is undefined.
- Judging a model by in-sample fit while ignoring out-of-sample accuracy.
- Comparing MAPE across products with very different volumes without weighting.
- Treating forecast error as planner failure when the demand itself is inherently volatile.
- Ignoring bias because absolute error looks acceptable.