7.1 Demand Forecasting Principles & Quantitative Models

7.1 Demand Forecasting Principles & Quantitative Models

Demand forecasting is the scientific process of estimating future customer demand using historical sales data, market intelligence, statistical modeling, and expert judgment. For supply management professionals preparing for the ISM CPSM Exam 2, mastering forecasting methodology is critical. Accurate demand forecasts serve as the primary input for master production scheduling (MPS), material requirements planning (MRP), capacity planning, inventory management, and strategic procurement. Inaccurate forecasts ripple across the entire supply chain, leading to either stockouts and lost revenue or excess holding costs and obsolete inventory.

---

Qualitative vs. Quantitative Forecasting

Forecasting models fall into two broad categories: Qualitative (judgmental) and Quantitative (statistical/mathematical).

Qualitative Forecasting Methods

Qualitative methods rely on expert intuition, human judgment, and subjective assessments. They are primary choices when historical data is unavailable, non-existent, or unreliable—such as during new product introductions (NPI), major technological disruptions, or entry into unfamiliar geographic markets.

• Delphi Method: A structured, iterative forecasting technique where a panel of independent experts anonymously answers questionnaires. A coordinator summarizes the responses after each round, allowing experts to revise their estimates without peer pressure or dominant personality bias until consensus is reached.
• Executive Opinion (Management Estimate): High-level executives from sales, finance, operations, and procurement combine their experience and strategic vision to form a consensus forecast. While fast and comprehensive, it risks executive bias and groupthink.
• Sales Force Composite: A bottom-up approach where individual sales representatives project sales for their respective territories. Because reps are closest to customers, this method captures ground-level market trends, though sales reps may deliberately underestimate forecasts to lower performance quotas.
• Consumer Market Research: Direct customer feedback gathered through focus groups, surveys, and market testing to gauge buying intent for new or modified offerings.

Quantitative Forecasting Methods

Quantitative methods process objective historical data using mathematical formulas. They assume that past patterns will continue into the future.

1. Time-Series Models: Analyze historical observations over regular time intervals (e.g., daily, monthly, quarterly) to identify underlying patterns such as level, trend, seasonality, and random variation.
2. Causal (Associative) Models: Examine relationships between demand (the dependent variable) and external economic or environmental indicators (independent variables), such as housing starts, interest rates, or advertising spend.

---

Time-Series Mathematical Forecasting Models

1. Simple Moving Average (SMA)

Calculates the unweighted arithmetic mean of the $k$ most recent time periods. It filters out short-term random fluctuations but lags behind actual demand trends.

$\text{SMA}_k = \frac{\sum_{i=1}^{k} A_{t-i+1}}{k}$

Where $A_t$ is actual demand in period $t$, and $k$ is the number of periods included in the moving average.

2. Weighted Moving Average (WMA)

Assigns individual weights ($w_i$) to past periods, giving higher importance to more recent observations. The sum of all weights must equal 1.0.

$\text{WMA} = \sum_{i=1}^{k} (w_i \cdot A_{t-i+1}), \quad \text{where } \sum w_i = 1.0$

3. Single Exponential Smoothing (SES)

Applies exponentially decreasing weights to past observations. It requires only three data points: the most recent actual demand ($A_t$), the most recent forecast ($F_t$), and a smoothing constant $\alpha$ (alpha, where $0 \le \alpha \le 1$).

$F_{t+1} = F_t + \alpha (A_t - F_t) \quad \text{or} \quad F_{t+1} = \alpha A_t + (1 - \alpha) F_t$

• High $\alpha$ (e.g., 0.7 - 0.9): Gives heavy weight to recent actual demand. Responsive to rapid demand changes, but susceptible to overreacting to random noise.
• Low $\alpha$ (e.g., 0.1 - 0.3): Gives heavy weight to past forecasts. Creates a smooth, stable forecast, but responds slowly to real structural shifts.

4. Double Exponential Smoothing (Holt's Method)

Single exponential smoothing exhibits significant lag when demand displays a consistent upward or downward trend. Double exponential smoothing resolves this by maintaining two smoothing constants: $\alpha$ for the level ($L_t$) and $\beta$ (beta) for the trend ($T_t$).

$\text{Level: } L_t = \alpha A_t + (1 - \alpha)(L_{t-1} + T_{t-1})$ $\text{Trend: } T_t = \beta (L_t - L_{t-1}) + (1 - \beta) T_{t-1}$ $\text{Forecast for } m \text{ periods ahead: } F_{t+m} = L_t + m \cdot T_t$

5. Causal Modeling: Simple Linear Regression

Models linear relationships between independent variables ($X$) and dependent demand ($Y$):

$Y = a + bX$ $\text{Slope } b = \frac{n\sum (XY) - \sum X \sum Y}{n\sum X^2 - (\sum X)^2}, \quad \text{Intercept } a = \frac{\sum Y - b\sum X}{n}$

---

Forecast Error Measurement & Accuracy Metrics

Forecast error ($e_t$) is the numeric difference between actual demand ($A_t$) and forecasted demand ($F_t$) for period $t$:

$e_t = A_t - F_t$

Evaluating forecast accuracy enables supply managers to select the best statistical model, fine-tune smoothing parameters, and size safety stock appropriately.

Tracking Signal Control Limits

The Tracking Signal measures persistent bias. A positive TS indicates that actual demand consistently exceeds forecast (under-forecasting), risking stockouts. A negative TS indicates that actual demand is consistently below forecast (over-forecasting), driving excess inventory. Standard control limits are set at $\pm 3.0$ to $\pm 4.0$. If the tracking signal trips outside these limits, the forecast model is biased and must be recalibrated.

---

Worked Numerical Example: Forecasting Error & Tracking Signal

Consider a supply manager evaluating a 5-month forecast for a critical industrial component. Below is the step-by-step calculation of MAD, MAPE, MSE, RSFE, and Tracking Signal.

Step-by-Step Metric Computations:

1. Running Sum of Forecast Errors (RSFE): $\text{RSFE} = (+10) + (+5) + (-5) + (+10) + (-10) = +10$

2. Mean Absolute Deviation (MAD): $\text{MAD} = \frac{\sum \|A_t - F_t\|}{n} = \frac{40}{5} = 8.0 \text{ units}$

3. Mean Absolute Percentage Error (MAPE): $\text{MAPE} = \frac{\sum (\|A_t - F_t\| / A_t)}{n} = \frac{37.66\%}{5} = 7.53\%$

4. Mean Squared Error (MSE): $\text{MSE} = \frac{\sum e_t^2}{n} = \frac{350}{5} = 70.0 \text{ units}^2$

5. Tracking Signal (TS): $\text{TS} = \frac{\text{RSFE}}{\text{MAD}} = \frac{+10}{8.0} = +1.25$

Interpretation: The tracking signal of $+1.25$ falls comfortably within the acceptable threshold of $\pm 4.0$. This demonstrates that while individual period errors occur, the overall forecast model remains unbiased and statistically valid.