6.1 Demand Forecasting Methods: Moving Averages, Exponential Smoothing & Error Metrics

Demand Forecasting Methods: Moving Averages, Exponential Smoothing & Error Metrics

Core Principle: Operational capacity planning, master scheduling, and inventory systems depend fundamentally upon demand forecasting. Industrial engineers must select forecasting architectures that minimize both random variance and systematic bias, establishing statistical tracking mechanisms to trigger model recalibration when process dynamics shift.

Every downstream operation in production and logistics—from aggregate workforce planning and master production scheduling to safety stock sizing and material procurement—originates from a demand forecast. Because forecasts are virtually never 100% accurate, industrial and systems engineers focus on three core objectives: selecting appropriate mathematical forecasting models, quantifying uncertainty, and implementing statistical control mechanisms to detect structural shifts in customer demand.

---

1. Qualitative vs. Quantitative Forecasting Architectures

Forecasting methodologies are broadly divided into qualitative (judgmental) and quantitative (mathematical) paradigms based on the availability and relevance of historical data.

The Delphi Method

Among qualitative methods, the Delphi method is prominent in systems engineering. Developed by the RAND Corporation, it achieves objective expert consensus while eliminating the interpersonal biases of group discussions:

1. Facilitator Panel Selection: A steering coordinator recruits an independent panel of domain experts.
2. Iterative Questionnaire Rounds: Experts answer anonymous questionnaires regarding technological timelines or market projections.
3. Statistical Feedback Summary: The facilitator compiles responses into statistical summaries (median, interquartile range) and qualitative rationales, returning them to the panel.
4. Convergence: Panelists review the anonymized rationales and revise their estimates over multiple rounds (typically 3 to 4 iterations) until opinions converge to a stable consensus.

---

2. Classical Time Series Decomposition

A chronological sequence of historical demand observations $A_t$ (where $t = 1, 2, \dots, n$) can be decomposed into five structural mathematical components:

1. Level ($L_t$): The baseline or central tendency of the series if all variations were eliminated.
2. Trend ($T_t$): The persistent, long-term upward or downward secular slope over multiple periods.
3. Seasonality ($S_t$): Repeating, predictable cyclical patterns linked to calendar intervals (e.g., weekly shifts in parcel processing, quarterly surges in retail manufacturing).
4. Cyclicality ($C_t$): Wavelike macroeconomic expansions and contractions that span multi-year durations with varying amplitudes and non-fixed periodicity.
5. Random Noise ($\epsilon_t$): Uncorrelated, non-deterministic random variation assumed to follow an independent and identically distributed normal distribution $\epsilon_t \sim \mathcal{N}(0, \sigma^2)$.

Time Series Composition
Observed Demand A_t = Level (L) + Trend (T) + Seasonality (S) + Cyclical (C) + Noise (ε)

Demand ^
       |                                            / Combined Time Series
       |                                    /\     /
       |                            /\     /  \   /
       |                    /\     /  \   /    \ /
       |            /\     /  \   /    \ /      v
       |    /\     /  \   /    \ /
       |   /  \   /    \ /
       |  /    \ /
       | /                                  ---------------- Secular Linear Trend (T)
       +--------------------------------------------------------> Time (t)

Additive vs. Multiplicative Models

• Additive Decomposition: Assumes seasonal fluctuations possess constant absolute amplitude regardless of the baseline level:

$A_t = L_t + T_t + S_t + C_t + \epsilon_t$

• Multiplicative Decomposition: Assumes seasonal variation scales proportionally with the baseline demand level:

$A_t = (L_t + T_t) \times S_t \times C_t \times \epsilon_t$

In industrial supply chains, the multiplicative model is widely applied because seasonal swings typically scale with business growth.

---

3. Moving Averages & Weighted Moving Averages

When demand exhibits no persistent trend or seasonality, moving averages filter out high-frequency random noise.

Simple Moving Average (SMA)

The $n$-period Simple Moving Average forecasts demand for period $t$ by taking the unweighted arithmetic mean of the preceding $n$ actual observations:

$SMA_n(t) = \frac{1}{n} \sum_{i=1}^n A_{t-i} = \frac{A_{t-1} + A_{t-2} + \dots + A_{t-n}}{n}$

• The Stability vs. Responsiveness Trade-off:
  • Small $n$ (e.g., $n = 3$): Rapidly tracks real shifts in demand level, but reacts excessively to transient random noise.
  • Large $n$ (e.g., $n = 12$): Highly stable and smooths out noise effectively, but severely lags behind genuine shifts in baseline level.

Weighted Moving Average (WMA)

The Weighted Moving Average assigns unequal weights $w_i$ to historical observations, typically weighting recent periods more heavily:

$WMA(t) = \sum_{i=1}^n w_i A_{t-i} = w_1 A_{t-1} + w_2 A_{t-2} + \dots + w_n A_{t-n}$

Subject to the normalization condition:

$\sum_{i=1}^n w_i = 1.0, \quad w_i \ge 0$

---

4. Simple Exponential Smoothing (SES)

Simple Exponential Smoothing (SES) is an autoregressive moving average technique that applies geometrically declining weights to past demand. It requires storing only two values: the prior forecast $F_{t-1}$ and the prior actual demand $A_{t-1}$.

Mathematical Formulation

The fundamental recurrence relation is:

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

Where:

• $F_t$ is the forecast for period $t$.
• $A_{t-1}$ is the actual demand observed in period $t-1$.
• $F_{t-1}$ is the forecast generated for period $t-1$.
• $(A_{t-1} - F_{t-1}) = e_{t-1}$ is the forecast error of the preceding period.
• $\alpha$ is the smoothing constant, bounded by $0 \le \alpha \le 1$.

Expanding the Geometric Series

Substituting $F_{t-1} = \alpha A_{t-2} + (1 - \alpha)F_{t-2}$ recursively demonstrates how historical weights decay:

$F_t = \alpha A_{t-1} + \alpha(1 - \alpha) A_{t-2} + \alpha(1 - \alpha)^2 A_{t-3} + \dots + (1 - \alpha)^k F_{t-k}$

Because $(1 - \alpha) < 1$, the influence of past observations decreases exponentially with age.

Weight Assigned to Historical Observations (Exponential Decay)
Weight ^
       |  * [α]
       |   \
       |    * [α(1-α)]
       |     \
       |      * [α(1-α)²]
       |       \
       |        * [α(1-α)³]
       |         \----
       +---------------------------------------------> Lag Periods into Past
          t-1   t-2   t-3   t-4   t-5

Selection of the Smoothing Constant $\alpha$

• The equivalent moving average span $n$ that matches the average age of observations under SES is:

$n \approx \frac{2 - \alpha}{\alpha} \iff \alpha = \frac{2}{n + 1}$

• Low $\alpha$ ($0.05 \le \alpha \le 0.15$): Corresponds to a large moving average window ($n \approx 12$ to $39$). High filtering stability; ideal for stable products with high random noise.
• High $\alpha$ ($0.30 \le \alpha \le 0.50$): Corresponds to a short window ($n \approx 3$ to $6$). Rapid response; ideal for dynamic, low-noise demand environments.
• Response to Step Changes: When demand undergoes an instantaneous step change of magnitude $\Delta$, the SES forecast tracks toward the new level with asymptotic convergence:

$F_{t+k} = F_t + \Delta \left( 1 - (1 - \alpha)^k \right)$

---

5. Trend-Adjusted Exponential Smoothing (Holt's Model)

When a time series exhibits a systematic trend, Simple Exponential Smoothing exhibits a perpetual lag: it consistently underestimates demand during an upward trend and overestimates demand during a downward trend. The steady-state lag under a linear trend with slope $T$ is:

$\text{Lag}_{SES} = \frac{1 - \alpha}{\alpha} T$

To eliminate this lag, Charles C. Holt introduced two-parameter linear exponential smoothing, which separates the baseline level from the trend component.

Holt's Mathematical Formulation

At the conclusion of period $t$, upon observing actual demand $A_t$, two smoothing equations update the state variables:

1. Level Equation ($L_t$):

$L_t = \alpha A_t + (1 - \alpha)(L_{t-1} + T_{t-1})$

1. Trend Equation ($T_t$):

$T_t = \beta (L_t - L_{t-1}) + (1 - \beta) T_{t-1}$

1. Forecast Generation for $k$ Periods Ahead ($F_{t+k}$):

$F_{t+k} = L_t + k T_t$

Where:

• $L_t$ is the smoothed level at period $t$.
• $T_t$ is the smoothed trend (slope per period) at period $t$.
• $\alpha$ is the level smoothing parameter ($0 < \alpha < 1$).
• $\beta$ is the trend smoothing parameter ($0 < \beta < 1$).
• $k$ is the forecast lead time horizon ($k \ge 1$).

Seasonal Decomposition and Winter's Model Overview

For series exhibiting both trend and seasonality, Winter's Three-Parameter Model incorporates a seasonal multiplicative factor $S_t$:

$L_t = \alpha \left(\frac{A_t}{S_{t-p}}\right) + (1 - \alpha)(L_{t-1} + T_{t-1})$ $T_t = \beta (L_t - L_{t-1}) + (1 - \beta) T_{t-1}$ $S_t = \gamma \left(\frac{A_t}{L_t}\right) + (1 - \gamma) S_{t-p}$ $F_{t+k} = (L_t + k T_t) \times S_{t-p+k}$

Where $p$ is the seasonal period (e.g., $p = 4$ for quarters, $p = 12$ for months), and $\gamma$ is the seasonal smoothing parameter.

---

6. Forecast Error Metrics and Bias Tracking

The forecast error in period $t$ is defined as:

$e_t = A_t - F_t$

Evaluating forecast models requires measuring both dispersion (magnitude of error) and bias (directional skew).

Error Metric Hierarchy
Forecast Error e_t = Actual (A_t) - Forecast (F_t)
├── Dispersion Metrics (Error Magnitude)
│   ├── MAD = (1/n) Σ |e_t|              [Linear penalty; robust to outliers; σ ≈ 1.25 MAD]
│   ├── MSE = (1/n) Σ (e_t)²             [Quadratic penalty; heavily penalizes large errors; σ ≈ √MSE]
│   └── MAPE = (1/n) Σ (|e_t|/A_t) × 100% [Scale-independent percentage; enables cross-SKU comparisons]
└── Bias Metrics (Directional Skew)
    ├── Mean Error (ME) = (1/n) Σ e_t    [Detects persistent over/under-forecasting]
    ├── RSFE = Σ e_t                     [Cumulative sum of errors]
    └── Tracking Signal (TS) = RSFE / MAD [Statistical control limit: ±4 to ±5 MAD]

Detailed Mathematical Definitions

1. Mean Absolute Deviation (MAD): Measures the average absolute error magnitude:

$MAD = \frac{1}{n} \sum_{t=1}^n |A_t - F_t| = \frac{1}{n} \sum_{t=1}^n |e_t|$

For normally distributed errors with zero mean, the standard deviation of demand $\sigma$ is related to MAD by:

$\sigma \approx \sqrt{\frac{\pi}{2}} \times MAD \approx 1.2533 \times MAD \iff MAD \approx 0.7979 \times \sigma$

1. Mean Squared Error (MSE): Measures variance by penalizing large outlier errors quadratically:

$MSE = \frac{1}{n} \sum_{t=1}^n (A_t - F_t)^2 = \frac{1}{n} \sum_{t=1}^n e_t^2$ $\sigma \approx \sqrt{MSE}$

1. Mean Absolute Percentage Error (MAPE): Evaluates relative error, normalizing by actual demand volume to allow comparison across high-volume and low-volume products:

$MAPE = \frac{1}{n} \sum_{t=1}^n \frac{|A_t - F_t|}{A_t} \times 100\%$

Tracking Signal ($TS$) and Statistical Control Limits

A model may have a low MAD but still suffer from severe systematic bias (e.g., consistently under-forecasting demand). The Running Sum of Forecast Errors ($RSFE$), also called cumulative forecast error, tracks bias accumulation:

$RSFE_t = \sum_{i=1}^t e_i = \sum_{i=1}^t (A_i - F_i)$

The Tracking Signal ($TS_t$) normalizes $RSFE$ by the contemporary Mean Absolute Deviation:

$TS_t = \frac{RSFE_t}{MAD_t} = \frac{\sum_{i=1}^t (A_i - F_i)}{\frac{1}{t} \sum_{i=1}^t |A_i - F_i|}$

• Interpretation:
  • $TS_t = 0$: Perfect unbiased calibration.
  • $TS_t > 0$: Actual demand consistently exceeds forecast (under-forecasting; risks stockouts).
  • $TS_t < 0$: Actual demand consistently falls below forecast (over-forecasting; causes excess inventory).
  • Control Limits: Operational control bounds are typically set at $\pm 4.0$ to $\pm 5.0$ MAD. If $|TS_t| > 4.0$, the probability that the error is due solely to random variation is under 3% (equivalent to approximately $\pm 3\sigma$). Exceeding these bounds signals structural change in demand or model obsolescence, requiring immediate retuning of parameters or switching models.

---

7. Comprehensive Worked Numerical Example

A precision manufacturing plant tracks monthly orders for an industrial gearbox. The actual demands observed over a 6-month period are:

• Period 1: $A_1 = 120$ units
• Period 2: $A_2 = 135$ units
• Period 3: $A_3 = 130$ units
• Period 4: $A_4 = 150$ units
• Period 5: $A_5 = 165$ units
• Period 6: $A_6 = 160$ units

We evaluate and compare three forecasting models:

1. Model A: 3-period Simple Moving Average ($SMA_3$).
2. Model B: 3-period Weighted Moving Average ($WMA_3$) with weights $w_1 = 0.50$ (lag 1), $w_2 = 0.30$ (lag 2), and $w_3 = 0.20$ (lag 3).
3. Model C: Simple Exponential Smoothing ($SES$) with $\alpha = 0.30$, initialized with $F_1 = 120.0$.

Step-by-Step Forecast Generation

1. Calculating $SMA_3$:

• $F_4 = \frac{A_3 + A_2 + A_1}{3} = \frac{130 + 135 + 120}{3} = \frac{385}{3} = 128.33$
• $F_5 = \frac{A_4 + A_3 + A_2}{3} = \frac{150 + 130 + 135}{3} = \frac{415}{3} = 138.33$
• $F_6 = \frac{A_5 + A_4 + A_3}{3} = \frac{165 + 150 + 130}{3} = \frac{445}{3} = 148.33$
• Forecast for Period 7: $F_7 = \frac{A_6 + A_5 + A_4}{3} = \frac{160 + 165 + 150}{3} = 158.33$

2. Calculating $WMA_3$:

• $F_4 = 0.50(130) + 0.30(135) + 0.20(120) = 65.0 + 40.5 + 24.0 = 129.50$
• $F_5 = 0.50(150) + 0.30(130) + 0.20(135) = 75.0 + 39.0 + 27.0 = 141.00$
• $F_6 = 0.50(165) + 0.30(150) + 0.20(130) = 82.5 + 45.0 + 26.0 = 153.50$
• Forecast for Period 7: $F_7 = 0.50(160) + 0.30(165) + 0.20(150) = 80.0 + 49.5 + 30.0 = 159.50$

3. Calculating $SES$ ($\alpha = 0.30$):

• $F_1 = 120.00$
• $F_2 = 120.00 + 0.30(120 - 120.00) = 120.00$
• $F_3 = 120.00 + 0.30(135 - 120.00) = 120.00 + 4.50 = 124.50$
• $F_4 = 124.50 + 0.30(130 - 124.50) = 124.50 + 1.65 = 126.15$
• $F_5 = 126.15 + 0.30(150 - 126.15) = 126.15 + 7.155 = 133.31$
• $F_6 = 133.31 + 0.30(165 - 133.31) = 133.31 + 9.507 = 142.82$
• Forecast for Period 7: $F_7 = 142.82 + 0.30(160 - 142.82) = 142.82 + 5.154 = 147.97$

Comparative Forecast and Error Ledger Table

Analysis and Engineering Findings

1. Evaluation of SES Error Metrics Across All 6 Periods: $MAD = \frac{0 + 15 + 5.5 + 23.85 + 31.69 + 17.18}{6} = \frac{93.22}{6} = 15.54 \text{ units}$ $MSE = \frac{0 + 225 + 30.25 + 568.82 + 1004.26 + 295.15}{6} = \frac{2123.48}{6} = 353.91 \text{ units}^2$ $\text{RMSE} = \sqrt{353.91} = 18.81 \text{ units}$ $MAPE = \frac{0.00\% + 11.11\% + 4.23\% + 15.90\% + 19.21\% + 10.74\%}{6} = \frac{61.19\%}{6} = 10.20\%$

2. Tracking Signal Analysis: In Period 6, $RSFE_6 = +93.22$ and $MAD_6 = 15.54$, yielding a Tracking Signal of: $TS_6 = \frac{+93.22}{15.54} = +6.00$ Because $TS_6 = +6.00 > +4.0$, the SES model has breached the upper control limit. Every single forecast error from Period 2 onward was strictly positive ($e_t > 0$), proving that actual demand grew systematically while the SES model lagged behind. The plant must transition from Simple Exponential Smoothing to Holt's trend-adjusted model to account for the secular growth rate of approximately $+8$ to $+10$ units/month.

---

8. Causal Forecasting with Simple Linear Regression

Time series methods use only past demand. A causal model forecasts demand from a leading indicator, such as housing starts for a window manufacturer or flight schedules for an aircraft parts supplier. The simplest causal model is a least-squares line $\hat{y} = a + bx$:

$b = \frac{n\sum xy - \sum x \sum y}{n\sum x^2 - \left(\sum x\right)^2}, \qquad a = \bar{y} - b\bar{x}$

$r = \frac{n\sum xy - \sum x \sum y}{\sqrt{\left[n\sum x^2 - \left(\sum x\right)^2\right]\left[n\sum y^2 - \left(\sum y\right)^2\right]}}$

The coefficient of determination $r^2$ is the fraction of the variation in $y$ explained by $x$.

Worked example: A window plant records regional housing starts $x$ (thousands) and its window orders $y$ (hundreds of units) for five quarters:

$b = \frac{5(22{,}650) - 75(1{,}450)}{5(1{,}193) - 75^2} = \frac{4{,}500}{340} = 13.24, \qquad a = 290 - 13.24(15) = 91.5$

With $\sum y^2 = 432{,}950$, the correlation is $r = 0.978$ and $r^2 = 0.957$, so housing starts explain about 96% of the variation in orders. If next quarter's housing starts are forecast at 22 thousand, the order forecast is $91.5 + 13.24(22) \approx 382.6$ hundred units.