Why forecasting matters
Every operating decision — how much to buy, how many to schedule, when to add capacity — rests on an estimate of future demand. Forecasts are always wrong to some degree; the job is to keep the error small, measured, and understood.
Qualitative methods
- Market research
- Surveys and interviews with current and potential customers.
- Sales force estimates
- Aggregated forecasts from the people closest to the customer, adjusted for optimism or quota bias.
- Executive opinion
- A judgment reached by a group of senior managers pooling experience.
- Expert judgment
- Outside specialists with knowledge of the industry or technology.
- Delphi method
- Repeated anonymous rounds of expert estimates with feedback between rounds until estimates converge.
Qualitative methods suit new products, new markets, and long horizons where there is no relevant history.
Time series components
- Trend: gradual long-term movement up or down.
- Seasonality: a repeating pattern tied to the calendar.
- Cycles: longer swings tied to economic conditions.
- Random variation: unexplained residual movement — never forecast it, measure it.
Averaging and smoothing
Formula
Naive forecast: F(t+1) = A(t)n-period moving average: F(t+1) = ( sum of the last n actuals ) ÷ nWeighted moving average: F(t+1) = Σ (weight × actual), weights summing to 1Exponential smoothing: F(t+1) = F(t) + α [ A(t) − F(t) ]where α is between 0 and 1: a high α reacts fast, a low α smooths hardWorked example
Worked example: exponential smoothing, α = 0.2
| Period | Actual A(t) | Forecast F(t) | Error A − F |
|---|---|---|---|
| 1 | 100 | 100 | 0 |
| 2 | 110 | 100 | +10 |
| 3 | 105 | 102 | +3 |
| 4 | — | 102.6 | — |
Formula
F(3) = 100 + 0.2(110 − 100) = 102F(4) = 102 + 0.2(105 − 102) = 102.6Trend projection and least squares
Formula
ŷ = a + bx where x is the time periodb = [ Σxy − n x̄ ȳ ] ÷ [ Σx² − n x̄² ]a = ȳ − b x̄Worked example
Worked example: least-squares trend
| x (period) | y (demand) | xy | x² |
|---|---|---|---|
| 1 | 100 | 100 | 1 |
| 2 | 110 | 220 | 4 |
| 3 | 120 | 360 | 9 |
| 4 | 130 | 520 | 16 |
| 5 | 140 | 700 | 25 |
| Σ = 15 | Σ = 600 | Σ = 1,900 | Σ = 55 |
Formula
x̄ = 3, ȳ = 120b = (1,900 − 5(3)(120)) ÷ (55 − 5(9)) = 100 ÷ 10 = 10a = 120 − 10(3) = 90ŷ = 90 + 10x, so the period-6 forecast is 90 + 60 = 150Measuring forecast error
Formula
Error = Actual − ForecastMAD = Σ | Actual − Forecast | ÷ nMSE = Σ ( Actual − Forecast )² ÷ nMAPE = [ Σ ( | error | ÷ Actual ) ÷ n ] × 100Tracking signal = Σ errors ÷ MADWorked example
Worked example: MAD and MSE
Three periods produce errors of +5, −3, and +4 units.
Formula
MAD = (5 + 3 + 4) ÷ 3 = 4.0 unitsMSE = (25 + 9 + 16) ÷ 3 = 16.67MSE penalizes large misses more heavily than MAD, so use it when one big error hurts more than several small ones. A tracking signal drifting persistently positive or negative means the forecast is biased, not merely noisy.
