Demand Planning and Forecasting
This domain tests forecast mechanics rather than opinion: which method suits which pattern, how to measure accuracy, and how to detect bias before it corrupts the plan.
Forecast characteristics
Forecasts are always wrong, should carry an error range, are more accurate for families than for single SKUs, and are more accurate over short horizons. Aggregate then disaggregate through the product/time/customer hierarchy. Distinguish independent demand (forecast it) from dependent demand (calculate it from a bill of material).
Method selection
- Qualitative: Delphi (anonymous, iterative, converges), market research, sales force composite, expert judgment. Use for new products or sparse history.
- Time series: naive, moving average, simple exponential smoothing, Holt (trend), Holt-Winters (trend plus seasonality). Exponential smoothing: F(t+1) = Ft + alpha(At - Ft). High alpha tracks change fast but amplifies noise.
- Causal: regression on price, promotion, weather, index of industrial production.
- Seasonal index = average demand in season / overall average demand; deseasonalise, forecast the base, then reseasonalise.
Accuracy and bias metrics
MAD = mean of absolute errors. MSE penalises large misses. MAPE = mean of |error|/actual, unusable when actuals are zero or tiny. RSFE = running sum of forecast errors (signed). Tracking signal = RSFE / MAD; a value drifting beyond roughly plus or minus 4 signals systematic bias, not random noise. Forecast value added compares a method against the naive benchmark.
Worked example
Actuals 100, 120, 110, 130. Simple exponential smoothing with alpha = 0.30 and F1 = 100.
- F2 = 100; error = +20. F3 = 100 + 0.3(20) = 106; error = +4. F4 = 106 + 0.3(4) = 107.2; error = +22.8.
- MAD = (20 + 4 + 22.8)/3 = 15.6. RSFE = 46.8. Tracking signal = 46.8/15.6 = 3.0.
- MAPE = (16.7% + 3.6% + 17.5%)/3 = 12.6%. Every error is positive, so the model is consistently under-forecasting and needs a bias correction, not a lower alpha.
Traps that cost marks
Using MAPE with near-zero actuals. Mixing annual and monthly demand. Re-tuning alpha on one period. Treating a low MAD as proof of an unbiased forecast. Ignoring promotion and price effects that make history a poor guide. Forgetting that a tracking signal near zero does not mean small errors, only balanced ones.