5.2 Sources of Demand and Forecasting

Forecasting translates uncertain future independent demand into a usable plan. CPIM expects you to know where demand signals come from, how major techniques work, and—critically—when each method is appropriate. This section includes worked calculations you should be able to reproduce under exam time pressure.

Sources of Demand

Before choosing a model, identify the source:

• Customer orders — firm or soft bookings; strongest signal for near-term horizons in make-to-order or assemble-to-order.
• Sales forecasts / pipeline — CRM opportunity data; useful but often optimistic.
• Historical shipments or bookings — foundation for time-series models.
• Downstream POS / sell-through — richer collaborative signal than warehouse withdrawals alone.
• Marketing events — promotions, new-product launches, price changes that break historical patterns.
• External indicators — housing starts, industrial production, weather—used in causal models.

S&OP reconciles these sources into one unconstrained demand plan by family, then disaggregates to SKU as needed.

Qualitative Versus Quantitative Methods

Qualitative (judgmental) methods rely on expert opinion when history is missing or structural change invalidates the past. Examples: executive opinion, sales-force composite, Delphi method, and market research for new products. Strengths: incorporate known future events. Weaknesses: bias, politics, and inconsistency.

Quantitative methods use data. Two families matter most for CPIM:

1. Time-series — project the pattern in the item’s own history (moving average, exponential smoothing, trend and seasonal models).
2. Causal / associative — regress demand on drivers (price, advertising, leading indicators).

Simple Moving Average — Worked Example

A simple moving average (SMA) of order N is:

F(t+1) = [A(t) + A(t−1) + … + A(t−N+1)] / N

Example. Monthly demand for a spare filter: Jan 120, Feb 135, Mar 125, Apr 140. Using a 3-month SMA, the May forecast is:

F(May) = (135 + 125 + 140) / 3 = 400 / 3 = 133.3

(You drop January once April is in the window.)

Moving averages treat each period in the window equally and require storing N observations.

Exponential Smoothing — Worked Example

Simple exponential smoothing needs only the previous forecast and the latest actual:

F(t+1) = α × A(t) + (1 − α) × F(t)

where α (alpha) is between 0 and 1.

Example. Last period’s forecast F(t) = 200. Actual demand A(t) = 230. With α = 0.3:

F(t+1) = 0.3(230) + 0.7(200) = 69 + 140 = 209

With α = 0.1 the same data give:

F(t+1) = 0.1(230) + 0.9(200) = 23 + 180 = 203

Higher α puts more weight on the latest actual, so the forecast jumps more when demand spikes. Lower α trusts the old forecast and damps noise. Exam tip: if demand is volatile but the underlying level is stable, choose a lower α; if a sustained shift is underway, a higher α (or a trend model) is more appropriate.

Initialization. If no prior forecast exists, practitioners often set F(1) = A(1) or use an average of early periods. The exam usually gives F(t) explicitly so you can apply the formula directly.

Seasonality

Seasonality is a repeating pattern within the year (or week). A practical approach:

1. Compute a seasonal index for each season = average demand in that season ÷ average demand across all seasons.
2. Deseasonalize history by dividing by the index.
3. Forecast the deseasonalized series (level/trend method).
4. Reapply the seasonal index to get the seasonal forecast.

Mini example. Average Q1 demand historically is 1.2× the annual quarterly average (index = 1.20). If the deseasonalized forecast for next Q1 is 500 units, the seasonal forecast is 500 × 1.20 = 600.

Ignoring seasonality causes systematic over-forecast in troughs and under-forecast in peaks—exactly the bias pattern tracking signals later detect.

When to Use Which Method

Use this decision checklist on scenario questions:

1. Is history representative? If a redesign, new channel, or pandemic-like shock broke the series, lean qualitative or restart the model.
2. What pattern dominates? Level → SMA / simple ES; trend → trend models; seasonality → indexes or seasonal smoothing; promotions → event adjustments.
3. What is the horizon? Short-term operational forecasts favor responsive time-series; long-term capacity forecasts tolerate judgment and causal drivers.
4. What is the cost of error? High-margin, capacity-critical items justify richer models and human review; C-items may use simple defaults.
5. Can partners improve the signal? POS sharing and CPFR often beat a cleverer univariate model fed on warehouse shipments alone.

Quantitative models are tools, not oracles. The CPIM-competent planner selects a method that matches the demand pattern, documents assumptions, and overlays known events rather than blindly trusting last period’s average.