Supply Chain & Logistics Management

Forecasting Methods and Sales and Operations Planning

Module 2

Almost every supply chain decision rests on an estimate of future demand, and this lesson covers the full forecasting toolkit: error metrics, moving averages, exponential smoothing, trend and seasonality models, plus aggregate planning and S&OP.

1. Role of Forecasting

Demand Planning, Forecasting and Sales and Operations Planning: module overview infographic

Forecasting answers what customers will want, when, and where. Inventory, capacity, transportation, and sourcing decisions all depend on demand estimates; supply chain planning is fundamentally about matching supply to demand.

Process Type	Definition	Forecasting Role
Push Process	Performed in anticipation of demand, before an order is placed	Basis for manufacturing, transporting, and stocking decisions
Pull Process	Performed in response to an actual order	Needed to plan capacity and inventory availability so the system responds quickly

Paint retail case: final mixing of base paint and dyes happens after the customer selects a shade (pull), but the retailer must already stock base paint and dyes (push) because factory ordering at the moment of request would take too long. The store forecasts stocking levels, the factory forecasts base paint production, and upstream suppliers forecast raw materials.

Collaborative forecasting: independent forecasting at separate stages creates misalignment, causing stockouts or excess inventory. Partners must share planning information and align on a single common demand view. Example: a beverage company's promotion must be reflected in the joint forecast, otherwise the bottler schedules for normal weeks and a shortfall follows.

Horizon	Timeframe	Supported Decisions	Role
Long Range	Months, years	Capacity planning, warehouse locations, new product launches	Building the system
Short Range	Daily, weekly	Replenishment, staffing, production schedules, delivery planning	Running the system

Seasonal retailers integrate weather forecasts (rain raises umbrella demand; heatwaves raise cold beverage and cooler demand). Auto dealers must stock the right model and variant mix because buyers expect immediate delivery; a wrong forecast leaves sitting inventory and tied-up working capital.

Expected demand vs. forecast error: planners track both the expected level of demand and the uncertainty around it. Forecast error drives the buffers in the system. When error is high, managers choose among safety inventory, flexible capacity, faster replenishment, or demand shaping through pricing.

Organizational alignment: sales teams are optimistic, operations conservative, finance cost-focused. Planning with mismatched functional forecasts produces excess costs or poor service, so functional consensus matters.

2. Features of Forecasts and the Forecasting Process

Four Common Features of All Forecasts

Feature	Definition	Implication
Continuity Assumption	The system that created the past will continue into the future	Models cannot anticipate weather shocks, competitor moves, policy changes; manual override needed
Inherent Imperfectness	Randomness prevents perfect forecasts	Quantify error and maintain buffers
Aggregation Accuracy	Forecasts are more accurate for groups than individual items	Fluctuations cancel out over categories or regions
Horizon Sensitivity	Accuracy decreases as the horizon lengthens	Flexible chains exploit accurate short-horizon forecasts

Elements of a Good Forecast

Timely, mathematically accurate (with stated error estimates), reliable, in meaningful units, written and shared, simple enough to be trusted, and cost-effective (benefits exceed data and system costs).

Consequences of Inaccuracy

Deviation	Consequences
Under-forecasting (too low)	Shortages, stockouts, missed deliveries, production disruptions, poor service
Over-forecasting (too high)	Excess inventory, idle capacity, high holding costs, markdowns

Bad forecasts create amplification waves: a stage reacting with sudden order changes causes upstream partners to misinterpret them, amplifying variability. Mitigated by joint forecasting and information sharing.

Six-Step Forecasting Process

Determine the purpose (decision supported). 2. Establish the time horizon. 3. Obtain and clean data (errors, outliers, comparability). 4. Choose a method. 5. Generate the forecast. 6. Monitor errors and update, forming a closed feedback loop.

3. Demand Components and Forecast Accuracy

Demand Component	Definition	Examples / Treatment
Horizontal (Level)	Demand fluctuates around a stable average	Mature staples, steady B2B consumables
Trend	Long-term systematic upward or downward movement	Update the baseline to prevent chronic shortages or excess
Seasonality	Regular repeating patterns with fixed, known periodicity (daily to yearly)	Summer ice cream, monsoon umbrellas, weekend restaurant surges
Cyclical	Wave-like movements longer than a year, periodicity not fixed	Capital equipment, luxury goods, construction; managed via scenario planning
Irregular Variation	One-off spikes or drops	Severe weather, strikes, pandemics; flagged and removed from baseline
Random (Noise)	Residual after all patterns	Managed with safety stock and capacity buffers

Forecast Error Metrics

ƒForecast error

E_t = A_t - F_t

Where: A_t is actual demand and F_t is the forecast for period t. Positive error means under-forecasting; negative error means over-forecasting.

Metric	Formula	Interpretation
Mean Error (Bias)	$ME = \frac{1}{n}\sum E_t$	Systematic over- or under-forecasting
MAD	$MAD = \frac{1}{n}\sum	E_t
MSE	$MSE = \frac{1}{n}\sum E_t^2$	Penalizes large errors more heavily
MAPE	$MAPE = \frac{1}{n}\sum \frac{	E_t

ƒMean error (bias)

\mathrm{ME} = \frac{1}{n}\sum E_t

ƒMean absolute deviation (MAD)

\mathrm{MAD} = \frac{1}{n}\sum |E_t|

ƒMean squared error (MSE)

\mathrm{MSE} = \frac{1}{n}\sum E_t^2

ƒMean absolute percentage error (MAPE)

\mathrm{MAPE} = \frac{1}{n}\sum \frac{|E_t|}{A_t} \times 100

Where: n is the number of periods compared and E_t is the forecast error in period t.

Worked example. Given, over n = 8 days for a quick-commerce SKU: absolute error sum 22, squared error sum 76, absolute percentage error sum 10.26.

\mathrm{MAD} = \frac{22}{8} = 2.75 \text{ units/day}

\mathrm{MSE} = \frac{76}{8} = 9.5

\mathrm{MAPE} = \frac{10.26}{8} = 1.28\%

Answer: MAD = 2.75 units/day, MSE = 9.5, MAPE = 1.28%, a highly accurate baseline.

4. Forecasting Approaches

Category	Basis	Use Cases
Qualitative	Human judgment, expert opinion, surveys	New launches, limited history, long-term strategic shifts
Quantitative	Historical data, time series, causal regression	Day-to-day operations, stable data-rich environments

Firms often combine: a quantitative baseline plus qualitative managerial adjustments for promotions or competitor moves.

Qualitative Methods

Method	Sourcing	Strengths	Risks
Executive Opinion	Senior manager group	Cross-functional strategic insight	Strong personalities or hierarchy bias dominate
Sales Force Opinions	Sales team customer contact	Early market signals	Confusing interest with buying; quota incentives distort
Consumer Surveys	Direct customer queries	Direct demand insight for new products	Costly, slow, non-response bias, stated vs. actual gap
Delphi Method	Anonymous iterative expert questionnaires	Limits loud voices, objective consensus for long-range questions	Time-consuming, depends on expert selection

Simple and Weighted Moving Averages

A time series is a sequence of observations at regular intervals; models assume the near future behaves like the recent past.

ƒSimple moving average

F_t = \mathrm{MA}_n = \frac{A_{t-1} + A_{t-2} + \dots + A_{t-n}}{n}

Where: n is the window length and A_{t-1} to A_{t-n} are the most recent n actual demands.

Larger n smooths fluctuations but reacts slowly (systematic lag); smaller n reacts quickly but chases noise.

Worked example. Given: demand over the last three weeks was 43, 40, 41.

F_6 = \frac{43 + 40 + 41}{3} = 41.33 \text{ units}

If actual week 6 demand is 38:

F_7 = \frac{40 + 41 + 38}{3} = 39.67 \text{ units}

Answer: F6 = 41.33 units; F7 = 39.67 units.

Simple moving averages weight all observations equally, so the oldest value influences as much as the newest, creating lag. The weighted moving average assigns unequal weights (summing to 1), usually heaviest on recent data. SMA is the special case with equal weights.

Worked example. Given: weights 0.4, 0.3, 0.2, 0.1 on demands 41, 40, 43, 40.

F_6 = 0.4 \times 41 + 0.3 \times 40 + 0.2 \times 43 + 0.1 \times 40 = 16.4 + 12 + 8.6 + 4 = 41 \text{ units}

Answer: F6 = 41 units. More sensitive to recent change, but weight selection is subjective trial and error.

5. Exponential Smoothing

Each new forecast adjusts the previous forecast by a fraction of the previous forecast error.

ƒSimple exponential smoothing (error-correction form)

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

Where: α is the smoothing constant between 0 and 1, F_{t-1} is the previous forecast and A_{t-1} the previous actual demand.

ƒSimple exponential smoothing (weighted-average form)

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

The new forecast is a weighted average of the latest actual (weight α) and the previous forecast (weight 1 - α). Because the previous forecast embeds all prior demand, weights decay exponentially over history.

Alpha Value	Smoothing	Responsiveness	Best Use
Close to 0 (0.05-0.10)	High smoothing, ignores short-term errors	Slow reaction	Stable SKUs where noise dominates
Close to 1 (0.40-0.50)	Low smoothing, adjusts heavily to last error	Jumpy, chases noise	Volatile SKUs with shifting demand averages

Worked example. Given: previous forecast 42 units, actual demand 40, α = 0.10.

F = 42 + 0.1 \times (40 - 42) = 41.8 \text{ units}

If the next actual is 43:

F = 41.8 + 0.1 \times (43 - 41.8) = 41.92 \text{ units}

Answer: the new forecasts are 41.8 units, then 41.92 units. Each forecast steps in the direction of the last error. With α = 0.1 the update is "90% stick with the old forecast, 10% listen to the latest actual":

Illustration

F_t = 0.9 \times F_{t-1} + 0.1 \times A_{t-1}

Firms pick α via judgment or historical trial and error minimizing MAD, MSE, or MAPE (common values run about 0.05 to 0.5). Initialization options for the starting forecast: naive (F2 = A1), average of first few actuals, or a managerial estimate.

6. Comparing Methods

Worked example (setup). Given actuals: A1 = 42, A2 = 40, A3 = 43, A4 = 40.

Calculation (MA2)

F_3 = \frac{42 + 40}{2} = 41 \; (\text{error } +2)

F_4 = \frac{43 + 40}{2} = 41.5 \; (\text{error } -1.5)

Calculation (WMA2, weights 0.6 recent and 0.4 prior)

F_3 = 0.6 \times 40 + 0.4 \times 42 = 40.8 \; (\text{error } +2.2)

F_4 = 0.6 \times 43 + 0.4 \times 40 = 41.8 \; (\text{error } -1.8)

Calculation (single smoothing, α = 0.10, F2 = A1 = 42)

F_3 = 0.9 \times 42 + 0.1 \times 40 = 41.8

Errors from each method then feed MAD, MSE, and MAPE for the head-to-head comparison.

All models must be evaluated over the exact same historical period, excluding warm-up periods needed by moving averages (e.g., comparing MA2 and single smoothing from period 3 through 11). The best method depends on the criterion: in the comparative study, WMA2 achieved lowest MAD and lowest MAPE, while single exponential smoothing (α = 0.10) achieved lowest MSE, preferred when large misses cause severe crises.

7. Techniques for Trend

When a trend is present, flat methods lag: under an upward trend they systematically under-forecast (stockouts, expedited shipping); under a downward trend they over-forecast (excess inventory, markdowns).

ƒLinear trend forecast

F_t = a + b \times t

Where: a is the intercept at t = 0, b is the slope or average change per period, and t is the period index. Estimated by least squares regression (Excel Data Analysis - Regression, demand as Y, time index as X).

Worked example. Given: regression on 10 weeks of unit sales gives a = 699.4, b = 7.5 (sales grow 7.5 units/week).

F_{11} = 699.4 + 7.5 \times 11 = 781.9

F_{12} = 699.4 + 7.5 \times 12 = 789.4

Answer: F11 = 781.9 (782 units); F12 = 789.4 (790 units).

Practical cautions: avoid extrapolating far outside the observed range, verify linearity holds within the horizon, and confirm residuals look random (patterns indicate missing model structure).

8. Trend-Adjusted Smoothing (Holt's Method)

Holt's Method (double exponential smoothing) separately updates a level S_t (current baseline) and a trend T_t (movement per period), preventing the systematic lag of single smoothing.

ƒTrend-adjusted forecast

\mathrm{TAF}_{t+1} = S_t + T_t

Where: S_t is the smoothed level at the end of period t and T_t is the smoothed trend per period.

ƒLevel update

S_t = \mathrm{TAF}_t + \alpha \times (A_t - \mathrm{TAF}_t)

ƒTrend update

T_t = T_{t-1} + \beta \times (\mathrm{TAF}_t - \mathrm{TAF}_{t-1} - T_{t-1})

α smooths the level; β smooths the trend. High values react fast; low values stay stable.

Worked example. Given: A1..A4 = 700, 724, 720, 728, with α = 0.4 and β = 0.3.

Initialization: net change is 28 over 3 steps, so

T_4 = \frac{28}{3} = 9.33

S_4 = A_4 = 728

\mathrm{TAF}_5 = 728 + 9.33 = 737.33

With A5 = 740:

S_5 = 737.33 + 0.4 \times (740 - 737.33) = 738.40

T_5 = 9.33 + 0.3 \times (737.33 - 728 - 9.33) = 9.33

\mathrm{TAF}_6 = 738.40 + 9.33 = 747.73 \text{ units}

Answer: TAF6 = 747.73 units.

9. Seasonality

Model Type	Structure	Definition	Best Use
Additive	Demand = Trend + Seasonality	Constant absolute addition or subtraction (+20 units, -10 units)	Seasonal swings that stay constant as business scales
Multiplicative	Demand = Trend × Seasonality	Multiplier on baseline (1.2×, 0.75×)	Swings that scale proportionally with volume

ƒAdditive seasonality model

\text{Demand} = \text{Trend} + \text{Seasonality}

Where: the seasonal component is a constant absolute amount added to or subtracted from the trend, suited to swings that stay the same size as the business scales.

ƒMultiplicative seasonality model

\text{Demand} = \text{Trend} \times \text{Seasonality}

Where: the seasonal component is a multiplier on the baseline, suited to swings that scale proportionally with volume.

Multiplicative multipliers are seasonal relatives (indices): 1.2 means 20% above baseline average; 0.75 means 25% below.

Two workflows: deseasonalize (divide actual demand by the seasonal relative to reveal the clean trend) and forecast with seasonality (forecast the baseline, then multiply by the seasonal relative).

Worked example. Given seasonal relatives: Q1 = 1.2, Q2 = 1.1, Q3 = 0.75, Q4 = 0.95.

Deseasonalizing period 1 (Q1, actual 158.4) and period 3 (Q3, actual 110):

\frac{158.4}{1.2} = 132 \text{ gallons}

\frac{110}{0.75} = 146.7 \text{ gallons}

Forecasting with seasonality, using the baseline trend:

F_t = 124 + 7.5t \quad \Rightarrow \quad F_9 = 191.5, \; F_{10} = 199

Applying relatives (period 9 is Q1, period 10 is Q2):

Final_9 = 191.5 \times 1.2 = 229.8 \text{ gallons}

Final_{10} = 199 \times 1.1 = 218.9 \text{ gallons}

Answer: final forecasts are 229.8 gallons (period 9) and 218.9 gallons (period 10).

Worked example (computing relatives). Step 1: season totals Q1 = 60, Q2 = 30, Q3 = 66, Q4 = 84. Step 2: season averages 20, 10, 22, 28. Step 3: overall average:

\frac{20 + 10 + 22 + 28}{4} = 20

Step 4: relatives Q1 = 1.0, Q2 = 0.5, Q3 = 1.1, Q4 = 1.4.

Answer: Q4 is the peak at 40% above average; Q2 runs 50% below.

10. Associative Techniques

Associative forecasting models demand as a function of observable driver variables rather than time alone:

ƒSimple linear regression (associative model)

Y = a + b \times X

Where: Y is the predicted demand, X is the driver variable, and a and b are the intercept and slope estimated by regression. Multiple simultaneous predictors (price, promotional intensity, weather) require multiple linear regression, larger datasets, and care against overfitting; saturation or threshold effects require nonlinear models or transformations.

11. Aggregate Planning

Aggregate planning is an intermediate-horizon process (typically 3 to 18 months) grouping products into families or total volumes rather than SKUs, translating forecasts into capacities, production levels, and inventory strategies ahead of long-lead-time constraints.

Lever	Definition	Trade-off
Production Rate	Volume planned per period	Ramping needs workforce or overtime; flat rates need inventory or backlogs
Workforce Level	Labor force sizing	Hiring has training lead times; downsizing carries layoff and morale costs
Overtime	Capacity beyond regular hours	Short-term flexibility at an overtime wage premium
Subcontracting	External production capacity	Instant expansion at a price premium, subject to supplier constraints
Planned Inventory	Carrying stock from low to high demand periods	Stable production but ties up capital and adds holding and obsolescence costs
Backlog and Stockouts	Delayed or lost unmet demand	Minimizes inventory costs but damages service, margins, future demand

Three cost domains are balanced: capacity costs (wages, overtime premiums, hiring, training, layoffs, subcontracting), inventory costs (holding, storage, tied-up working capital, shrinkage, obsolescence), and backlog/stockout costs (wait penalties, lost margins, goodwill erosion).

12. Sales and Operations Planning (S&OP)

S&OP is the cross-functional process aligning the demand plan, supply plan, and financial plan into a single operational plan across marketing, finance, procurement, and operations.

Without S&OP, functions optimize in isolation: marketing wants holiday promotion spikes, operations wants stable rates and minimal overtime, distribution worries about warehouse throughput and trucks, procurement targets bulk buys against packaging lead times, and finance minimizes working capital. S&OP resolves these trade-offs by testing which promotions are operationally feasible and financially viable.

Strategy	Core Lever	Characteristics	Best Fit	Risks	Examples
Chase	Capacity	Production synchronized with demand via hiring/firing, shifts, subcontracting	Cheap capacity changes, expensive inventory, short lifecycles	Labor regulations, training lead times, morale	Services, call centers, gig delivery
Flexibility	Utilization	Stable workforce; capacity varied via overtime and flexible scheduling	Avoids hiring/layoff churn with output flexibility	Overtime premiums, hour limits	Manufacturing with clear overtime guidelines
Level	Inventory	Constant production and headcount; inventory or backlogs absorb swings	Stable environments with cheap inventory	Tied-up capital, storage limits, obsolescence	Commodities, capital-intensive manufacturing

In practice firms use hybrid strategies (some overtime, some inventory, some subcontracting) because the cheapest lever varies by context.

Memory hook: The three S&OP strategies by lever: "CFL bulb" - Chase changes Capacity, Flexibility flexes utilization (overtime), Level leans on inventory.

13. Exam Essentials

Expected demand vs. forecast error: expected demand is the baseline volume; forecast error is the uncertainty and directly sizes safety buffers.
SMA vs. WMA: SMA weights all window observations equally (lag); WMA emphasizes recent data (responsive but arbitrary weights).
Single smoothing vs. Holt's: single tracks one level and lags during trends; Holt's recursively updates level (S_t) and trend (T_t).
Additive vs. multiplicative seasonality: constant absolute change vs. proportional scaling factor.
SKU planning vs. aggregate planning: individual items day-to-day vs. families/volumes over 3-18 months.
Chase / Flexibility / Level: vary capacity via headcount, vary utilization via overtime, or keep production constant and buffer with inventory.
Key terms: collaborative forecasting, Delphi method, smoothing constants α (level) and β (trend), seasonal relatives, deseasonalizing, associative forecasting, S&OP.

Back to Course Take Test