Forecasting demand

Almost every supply chain decision starts with a forecast: how much will customers want, and when? Forecasts are always wrong by some amount; the aim is to be less wrong, and to know how wrong you usually are.

The simplest method is a moving average: the average of the last few periods. If the last three months' sales were 120, 90 and 150, the three-month moving average, and the forecast for next month, is (120 + 90 + 150) ÷ 3 = 120. A longer average is smoother but slower to notice a real change.

Exponential smoothing updates last period's forecast by a fraction of how wrong it was: new forecast = old forecast + α × (actual − old forecast). α (alpha) is between 0 and 1. If you forecast 100, sold 120, and α is 0.3, the next forecast is 100 + 0.3 × 20 = 106. A bigger α reacts faster; a smaller one ignores more of the noise. Real forecasts also add trend and seasonality, but the idea is the same.

To measure accuracy, look at the errors (actual − forecast). MAD, the mean absolute deviation, averages the errors ignoring their sign: errors of 10, −5, 15 and −10 give a MAD of (10 + 5 + 15 + 10) ÷ 4 = 10. MAPE puts the error as a percentage of actual demand, so you can compare a fast seller with a slow one: forecast 90, actual 100 is a 10% error.

Bias is the plain average of the errors, signs included. If it stays well above zero you keep forecasting too low (and running short); well below zero, too high (and piling up stock). A forecast with a small bias and a known MAD is something you can plan safety stock around.