12.1 Demand Forecasting Methods and Forecast Error Tracking

Operations planning begins with an estimate of future customer demand. In industrial engineering, every downstream decision—from procurement quantities and production staffing to warehouse sizing and distribution logistics—depends fundamentally on the quality of the demand forecast. Because actual demand is inherently stochastic, industrial engineers must select appropriate mathematical forecasting techniques, track forecast errors over time, and establish statistical control limits to detect systemic bias.

---

1. Qualitative vs. Quantitative Forecasting Frameworks

Forecasting methodologies are broadly divided into two major classifications based on data availability and mathematical structure.

Forecasting Methodologies
 ├── Qualitative (Subjective / Judgmental)
 │    ├── Delphi Method (Iterative, anonymous multi-round expert panel consensus)
 │    ├── Sales Force Composite (Aggregated field sales estimates)
 │    ├── Executive Opinion (High-level managerial consensus)
 │    └── Consumer Market Research (Surveys, focus groups, concept testing)
 └── Quantitative (Objective / Mathematical)
      ├── Time Series Models (Extrapolate historical chronological patterns)
      │    ├── Simple & Weighted Moving Averages
      │    ├── Simple Exponential Smoothing (SES)
      │    ├── Trend-Adjusted Exponential Smoothing (Holt's Method)
      │    └── Classical Seasonal Decomposition
      └── Causal / Associative Models (Explanatory mathematical relationships)
           ├── Simple Linear Regression (Y = β0 + β1*X + ε)
           └── Multiple Linear Regression & Econometric Models

Qualitative Methods

Qualitative methods rely on human intuition, experience, and subjective judgment. They are critical when historical data is nonexistent (e.g., launching a revolutionary new product), when long-term strategic decisions (5–10 year horizon) render historical trends obsolete, or during severe market disruptions.

• Delphi Method: An iterative group consensus technique where geographically separated experts answer questionnaires anonymously. A neutral facilitator summarizes responses and redistributes them through multiple rounds until consensus converges. Anonymity prevents dominant individuals from unduly influencing the group.
• Sales Force Composite: Regional sales representatives estimate demand in their territories. While close to customers, estimates often suffer from optimism bias or deliberate underestimation if sales quotas are tied to forecasts.
• Executive Opinion: High-level cross-functional managers (finance, operations, marketing) develop joint estimates. Fast and strategic, but vulnerable to corporate hierarchy bias.

Quantitative Methods

Quantitative methods require historical numerical data and assume that past mathematical patterns will persist into the future. They are objective, reproducible, and easily automated within Enterprise Resource Planning (ERP) databases.

---

2. Components of a Time Series

A time series is a sequence of observations recorded at regular, uniform time intervals (hourly, daily, weekly, monthly, quarterly). Classical time series analysis assumes the observed historical demand $Y_t$ is composed of four distinct underlying components:

1. Horizontal Level ($L$): The baseline average value around which demand fluctuates when no trends or seasonal cycles exist.
2. Trend ($T$): The long-term persistent upward or downward movement in the series over an extended timeframe, driven by population growth, market adoption, or technological obsolescence.
3. Seasonality ($S$): A regular, repeating pattern of peaks and valleys that recurs over a known, fixed duration (e.g., weekly retail spikes on Saturdays, quarterly agricultural cycles, hourly power consumption).
4. Cyclical ($C$): Multi-year wave-like oscillations tied to macroeconomic expansions and recessions. Unlike seasonality, cyclical movements vary in length (typically 2–10 years) and amplitude, making them difficult to model with simple time series.
5. Random Variation / Irregular Noise ($\epsilon$): Uncorrelated, unpredictable residual noise caused by weather anomalies, equipment breakdowns, or chance occurrences ($E[\epsilon] = 0$).

Mathematical Formulation: Additive vs. Multiplicative

• Additive Model: Assumes components are independent and expressed in physical demand units: $Y_t = L + T_t + S_t + C_t + \epsilon_t$ When to use: Seasonal oscillations remain constant in magnitude regardless of overall demand level.
• Multiplicative Model: Assumes seasonal variations scale proportionally with the baseline level and trend: $Y_t = (L + T_t) \times S_t \times C_t \times \epsilon_t$ When to use: Seasonal swings amplify as baseline volume expands (standard industrial behavior).

---

3. Moving Average Forecasting Methods

Moving average models smooth out short-term random noise by calculating the arithmetic mean of the most recent historical observations.

Simple Moving Average (SMA)

For a moving average of order $k$ (spanning $k$ historical periods), the forecast for period $t+1$ is:

$F_{t+1} = \frac{1}{k} \sum_{i=0}^{k-1} A_{t-i} = \frac{A_t + A_{t-1} + A_{t-2} + \dots + A_{t-k+1}}{k}$

where $A_t$ represents the actual observed demand in period $t$.

• Parameter Selection Trade-off:
  • Large $k$ (e.g., $k = 12$): Maximum smoothing stability. Highly effective at filtering out severe white noise, but sluggish and slow to respond to genuine structural shifts or trends.
  • Small $k$ (e.g., $k = 3$): High responsiveness. Rapidly tracks recent demand shifts, but susceptible to over-reacting to transient random fluctuations.
• Fundamental Limitation: All moving averages lag behind an upward or downward trend. In an upward trend, SMA will consistently underestimate actual demand ($A_t > F_t$); in a downward trend, SMA will consistently overestimate demand ($A_t < F_t$).

Weighted Moving Average (WMA)

To improve responsiveness without discarding older data entirely, unequal weights $w_i$ are assigned to past periods, typically giving the highest weight to the most recent observation:

$F_{t+1} = \sum_{i=1}^k w_i A_{t-i+1} = w_1 A_t + w_2 A_{t-1} + \dots + w_k A_{t-k+1}$

$\text{Subject to: } \sum_{i=1}^k w_i = 1.0, \quad w_i \ge 0$

---

4. Exponential Smoothing Techniques

Exponential smoothing is the most widely implemented quantitative forecasting method in industry because it requires storing only two numbers in memory: the most recent actual demand ($A_t$) and the most recent forecast ($F_t$). Older observations are weighted with geometrically decaying importance.

Simple Exponential Smoothing (SES)

The forecast for period $t+1$ is formulated as a convex combination of actual demand in period $t$ and the forecast made for period $t$:

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

Rearranging into the intuitive error-correction form:

$F_{t+1} = F_t + \alpha (A_t - F_t) = F_t + \alpha e_t$

where $\alpha$ is the smoothing constant ($0 \le \alpha \le 1$) and $e_t = A_t - F_t$ is the forecast error in period $t$.

Exponential Decay of Weights
 F_{t+1} = α*A_t + α*(1-α)*A_{t-1} + α*(1-α)^2*A_{t-2} + α*(1-α)^3*A_{t-3} + ...
 Weights: [α]       [α(1-α)]         [α(1-α)^2]           [α(1-α)^3]
 Sum of infinite weights = α / [1 - (1 - α)] = 1.0

The Role of the Smoothing Constant $\alpha$

• High $\alpha$ (e.g., $\alpha = 0.7 - 0.9$): Puts heavy weight on the latest observation. Forecast reacts rapidly to recent changes. Equivalent to a small-period moving average.
• Low $\alpha$ (e.g., $\alpha = 0.1 - 0.2$): Puts heavy weight on past forecast history. Filters out noise, providing smooth, stable forecasts.
• Equivalent Moving Average Span ($k$): The approximate correspondence between smoothing constant $\alpha$ and moving average span $k$ is given by: $\alpha \approx \frac{2}{k + 1} \quad \Longleftrightarrow \quad k \approx \frac{2}{\alpha} - 1$ For example, $\alpha = 0.2$ behaves similarly to a $k = \frac{2}{0.2} - 1 = 9$-period moving average.

Trend-Adjusted Exponential Smoothing (Holt's Method)

Just like moving averages, simple exponential smoothing exhibits systematic lag when demand contains a persistent trend. Holt's two-parameter exponential smoothing corrects for trend by tracking both a smoothed level ($L_t$) and a smoothed trend rate ($T_t$):

1. Update Base Level ($L_t$): $L_t = \alpha A_t + (1 - \alpha)(L_{t-1} + T_{t-1})$
2. Update Trend Estimate ($T_t$): $T_t = \beta (L_t - L_{t-1}) + (1 - \beta) T_{t-1}$
3. Forecast $p$ Periods Ahead ($F_{t+p}$): $F_{t+p} = L_t + p \cdot T_t$

where $\alpha$ is the level smoothing constant and $\beta$ is the trend smoothing constant ($0 \le \alpha, \beta \le 1$).

---

5. Seasonal Indices and Classical Decomposition

When demand exhibits cyclical peaks and troughs at regular intervals within a year (or week), industrial engineers employ seasonal decomposition.

Calculating Seasonal Indices (Multiplicative Approach)

A Seasonal Index ($SI_i$) represents the ratio of average demand during season $i$ to the average demand across all seasons:

$SI_i = \frac{\bar{D}_i}{\bar{D}_{\text{overall}}}$

$\sum_{i=1}^m SI_i = m$

where $m$ is the number of distinct seasons in a full cycle (e.g., $m = 4$ for quarters, $m = 12$ for months, $m = 7$ for days of the week).

• $SI_i > 1.0$: Demand in period $i$ is above the annual average (e.g., $SI = 1.25$ indicates demand is 25% above average).
• $SI_i < 1.0$: Demand in period $i$ is below average (e.g., $SI = 0.80$ indicates demand is 20% below average).

The Forecasting Procedure with Seasonality

1. Deseasonalize Historical Data: Divide actual demand by the corresponding seasonal index to uncover the pure underlying trend and level: $\text{Deseasonalized Demand}_t = \frac{A_t}{SI_t}$
2. Estimate Trend / Baseline: Fit a linear regression or moving average to the deseasonalized data: $\hat{Y}_t = a + b t$.
3. Project Future Unadjusted Forecast: Calculate future baseline values $\hat{Y}_{t+p}$.
4. Reseasonalize: Multiply the baseline projection by the target period's seasonal index: $F_{t+p} = \hat{Y}_{t+p} \times SI_{t+p}$

---

6. Measuring Forecast Error

Forecast error in period $t$ is universally defined under NCEES standards as actual observed demand minus forecasted demand:

$e_t = A_t - F_t$

Exam Convention Alert: Always subtract Forecast from Actual ($A_t - F_t$). A positive error ($e_t > 0$) means actual demand was higher than forecasted (underforecasting $\to$ stockout risk). A negative error ($e_t < 0$) means actual demand was lower than forecasted (overforecasting $\to$ excess inventory risk).

Core Error Metrics

---

7. Tracking Signals and Forecast Control Limits

A forecasting model may have a low MAD, yet still be systematically biased if it continually under- or over-predicts demand. Industrial engineers monitor model validity using the Tracking Signal (TS).

Cumulative Forecast Error (CFE)

The algebraic sum of all historical errors, also called the Running Sum of Forecast Errors (RSFE):

$\text{CFE}_n = \sum_{t=1}^n e_t = \sum_{t=1}^n (A_t - F_t)$

Tracking Signal Formulation

The Tracking Signal divides the cumulative algebraic error by the current Mean Absolute Deviation:

$\text{TS}_t = \frac{\text{CFE}_t}{\text{MAD}_t} = \frac{\sum_{i=1}^t e_i}{\frac{1}{t}\sum_{i=1}^t |e_i|}$

Tracking Signal Control Limits
 +5 MAD ─────────────────────────────── Upper Control Limit (Severe Underforecasting)
 +4 MAD ------------------------------- Warning Threshold
  0 MAD ═══════════════════════════════ Ideal Unbiased Performance (CFE = 0)
 -4 MAD ------------------------------- Warning Threshold
 -5 MAD ─────────────────────────────── Lower Control Limit (Severe Overforecasting)

Statistical Control Limits on Tracking Signal

• In-Control Range: A tracking signal between $-4$ and $+4$ (or up to $\pm 5$ in some industries) indicates the forecasting model is in statistical control with random, zero-mean error fluctuations.
• Out-of-Control Condition: If $|\text{TS}| > 4$ (or $5$), the model is declared biased and out of control:
  • $\text{TS} > +4$: Consistent underforecasting (actuals repeatedly exceed forecasts; stockouts imminent). Model needs an upward level/trend adjustment.
  • $\text{TS} < -4$: Consistent overforecasting (forecasts repeatedly exceed actuals; inventory pile-up imminent). Model needs a downward level/trend adjustment.

Relationship Between MAD and Standard Deviation ($\sigma$)

If forecast errors are approximately normally distributed with zero mean, the standard deviation of forecast errors ($\sigma$) is mathematically related to the Mean Absolute Deviation by:

$\sigma = \sqrt{\frac{\pi}{2}} \cdot \text{MAD} \approx 1.2533 \cdot \text{MAD} \approx 1.25 \cdot \text{MAD}$

$\text{MAD} = \sqrt{\frac{2}{\pi}} \cdot \sigma \approx 0.7979 \cdot \sigma \approx 0.80 \cdot \sigma$

This fundamental identity allows engineers to translate a calculated MAD directly into standard deviations for setting safety stock and reorder points without calculating individual squared residuals.

---

8. Step-by-Step Worked Engineering Calculations

Worked Example 12.1.1: Exponential Smoothing, Error Metrics, and Tracking Signal

Problem: A manufacturing plant monitors monthly demand for an industrial electric motor. Demand over the past 4 months is recorded below. At the beginning of Month 1, the forecast was $F_1 = 110$ units. The company uses simple exponential smoothing with a smoothing constant $\alpha = 0.30$.

1. Calculate the forecasts $F_2, F_3, F_4$, and the projected forecast for Month 5 ($F_5$).
2. Compute the forecast errors $e_t$, absolute errors $|e_t|$, squared errors $e_t^2$, and absolute percentage errors $|e_t / A_t|$ for months 1 through 4.
3. Determine the MAD, MSE, MAPE, and Cumulative Forecast Error (CFE) across the 4 months.
4. Calculate the Tracking Signal at the end of Month 4 and evaluate whether the model is in statistical control.
5. Estimate the standard deviation of forecast errors ($\sigma$).

Solution:

Step 1: Compute Forecasts Using $F_{t+1} = F_t + \alpha (A_t - F_t)$

• Month 1: Given $F_1 = 110.00$. Actual $A_1 = 120$. $e_1 = A_1 - F_1 = 120 - 110 = +10.00$
• Month 2: $F_2 = 110.00 + 0.30(120 - 110) = 110.00 + 3.00 = 113.00$ $e_2 = A_2 - F_2 = 130 - 113.00 = +17.00$
• Month 3: $F_3 = 113.00 + 0.30(130 - 113.00) = 113.00 + 5.10 = 118.10$ $e_3 = A_3 - F_3 = 110 - 118.10 = -8.10$
• Month 4: $F_4 = 118.10 + 0.30(110 - 118.10) = 118.10 - 2.43 = 115.67$ $e_4 = A_4 - F_4 = 125 - 115.67 = +9.33$
• Projected Forecast for Month 5 ($F_5$): $F_5 = 115.67 + 0.30(125 - 115.67) = 115.67 + 0.30(9.33) = 115.67 + 2.80 = 118.47\text{ units}$

Step 2: Construct the Comprehensive Error Tracking Table

| $t$ | $A_t$ | $F_t$ | $e_t = A_t - F_t$ | $|e_t|$ | $e_t^2$ | $|e_t / A_t| \times 100%$ | Cumulative $e_t$ (CFE) | | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | | 1 | 120 | 110.00 | $+10.00$ | 10.00 | 100.00 | $8.33%$ | $+10.00$ | | 2 | 130 | 113.00 | $+17.00$ | 17.00 | 289.00 | $13.08%$ | $+27.00$ | | 3 | 110 | 118.10 | $-8.10$ | 8.10 | 65.61 | $7.36%$ | $+18.90$ | | 4 | 125 | 115.67 | $+9.33$ | 9.33 | 87.05 | $7.46%$ | $+28.23$ | | Sum | — | — | $+28.23$ | 44.43 | 541.66 | $36.23%$ | — |

Step 3: Evaluate Error Metrics

• MAD: $\text{MAD} = \frac{\sum |e_t|}{n} = \frac{44.43}{4} = 11.11\text{ units}$
• MSE: $\text{MSE} = \frac{\sum e_t^2}{n} = \frac{541.66}{4} = 135.42\text{ units}^2$
• RMSE: $\text{RMSE} = \sqrt{135.42} = 11.64\text{ units}$
• MAPE: $\text{MAPE} = \frac{36.23\%}{4} = 9.06\%$
• CFE (Running Sum of Forecast Errors at Month 4): $\text{CFE}_4 = +10.00 + 17.00 - 8.10 + 9.33 = +28.23\text{ units}$

Step 4: Tracking Signal Evaluation $\text{TS}_4 = \frac{\text{CFE}_4}{\text{MAD}_4} = \frac{+28.23}{11.11} = +2.54$

• Evaluation: Because $|\text{TS}_4| = 2.54 \le 4.0$, the forecasting model is in statistical control. However, the positive value ($+2.54$) reveals an underlying tendency toward underforecasting, warning the engineer that demand may be trending upward faster than $\alpha = 0.30$ can accommodate.

Step 5: Estimate Standard Deviation of Forecast Error $\sigma \approx 1.25 \times \text{MAD} = 1.25 \times 11.11 = 13.89\text{ units}$

---

Worked Example 12.1.2: Seasonal Index Decomposition and Forecasting

Problem: A manufacturer of commercial HVAC units experiences strong seasonal demand. Historical quarterly sales over the past two years are presented below. Corporate sales forecasts project total annual demand for Year 3 at 1,440 units.

1. Calculate the seasonal index for each of the four quarters using the multiplicative decomposition method.
2. Determine the quarterly demand forecasts for Year 3.

Solution:

Step 1: Determine Average Demand by Quarter and Overall Average

• Average Q1 Demand: $\bar{D}_1 = \frac{220 + 260}{2} = 240\text{ units}$
• Average Q2 Demand: $\bar{D}_2 = \frac{380 + 420}{2} = 400\text{ units}$
• Average Q3 Demand: $\bar{D}_3 = \frac{460 + 500}{2} = 480\text{ units}$
• Average Q4 Demand: $\bar{D}_4 = \frac{140 + 180}{2} = 160\text{ units}$
• Total Average Annual Cycle Demand: $240 + 400 + 480 + 160 = 1,280\text{ units}$
• Overall Average Quarterly Baseline ($\bar{D}_{\text{overall}}$): $\bar{D}_{\text{overall}} = \frac{1,280}{4} = 320\text{ units/quarter}$

Step 2: Calculate Seasonal Indices ($SI_i = \bar{D}i / \bar{D}{\text{overall}}$)

• $SI_1 = \frac{240}{320} = 0.75$
• $SI_2 = \frac{400}{320} = 1.25$
• $SI_3 = \frac{480}{320} = 1.50$
• $SI_4 = \frac{160}{320} = 0.50$
• Verification: $\sum SI_i = 0.75 + 1.25 + 1.50 + 0.50 = 4.00 = m$. (Matches cycle length).

Step 3: Forecast Year 3 Quarterly Demand

• Year 3 expected baseline quarterly demand without seasonality: $\text{Baseline Quarter} = \frac{1,440}{4} = 360\text{ units/quarter}$
• Multiply baseline by respective seasonal indices ($F_i = 360 \times SI_i$):
  • Q1 Forecast: $F_1 = 360 \times 0.75 = 270\text{ units}$
  • Q2 Forecast: $F_2 = 360 \times 1.25 = 450\text{ units}$
  • Q3 Forecast: $F_3 = 360 \times 1.50 = 540\text{ units}$
  • Q4 Forecast: $F_4 = 360 \times 0.50 = 180\text{ units}$
• Check: $270 + 450 + 540 + 180 = 1,440\text{ units}$.

---

9. NCEES Reference Handbook Tips & Realistic Exam Traps

• Error Sign Convention Trap: NCEES defines error as $e_t = A_t - F_t$ (Actual minus Forecast). If a problem gives you Actual = 100 and Forecast = 120, $e_t = -20$. A negative error indicates overforecasting. Do not write $F_t - A_t$.
• Smoothing Constant Responsiveness: Remember that higher $\alpha$ means more responsive to recent demand changes (giving weight $\alpha$ to $A_t$), while lower $\alpha$ provides smoother, more stable forecasts. On the exam, questions asking "which parameter value should be chosen if the demand process experiences rapid, permanent structural shifts?" require selecting the highest $\alpha$ (e.g., 0.7 to 0.9).
• The MAD to $\sigma$ Conversion: Memorize $\sigma \approx 1.25 \times \text{MAD}$. A common exam trap is to divide MAD by 1.25 (or multiply by 0.80) to find $\sigma$. The standard deviation of a normal distribution is always larger than its Mean Absolute Deviation ($\sigma = 1.2533 \times \text{MAD}$, while $\text{MAD} = 0.7979 \times \sigma$).
• Tracking Signal Denominator: The denominator of Tracking Signal is MAD, NOT MSE, RMSE, or $\sigma$. Tracking Signal has no units because CFE and MAD share identical units (both are in units of physical product).
• Deseasonalizing vs. Reseasonalizing: When eliminating seasonality from actual data, you divide by the seasonal index ($A_t / SI$). When creating the final operational forecast from a baseline trend, you multiply by the seasonal index ($F_t = \text{Baseline} \times SI$).