Demand Forecasting and Integrated Planning

Introduction

Global supply chains must commit labor, materials, transport capacity and capital before demand is fully known. Forecasting converts available evidence into an explicit estimate of future demand; planning determines how the organization should respond. Neither activity eliminates uncertainty. Their purpose is to make uncertainty visible, measurable and manageable.

International operations make this task especially difficult. Demand may vary with exchange rates, trade policy, cultural preferences, promotions, local holidays and economic conditions. Supply may be constrained by long lead times, port congestion, geopolitical disruption or limited supplier capacity. Consequently, forecasting must be integrated with aggregate planning, sales and operations planning (S&OP), integrated business planning (IBP) and collaboration with supply-chain partners.

Learning Objectives

After studying this chapter, students should be able to:

• distinguish forecasts from plans and targets;
• apply qualitative, time-series and causal forecasting methods;
• calculate and interpret forecast-error measures;
• construct an aggregate plan using inventory, backlog, capacity and workforce relationships;
• explain the roles of S&OP, IBP and collaborative planning;
• evaluate global, technological, ethical and governance issues in integrated planning.

1. Forecasting as a Decision Process

A forecast is an evidence-based estimate of a future variable. A plan specifies intended actions, while a target expresses a desired outcome. Confusing these concepts encourages political forecasting—for example, increasing a forecast merely because management wants higher sales.

Forecasts should state four dimensions:

1. Object: units, revenue, orders, shipments or another measure.
2. Horizon: operational, tactical or strategic.
3. Time bucket: day, week, month or quarter.
4. Aggregation level: stock-keeping unit (SKU), product family, region or global total.

Forecast accuracy normally declines as the horizon increases. Aggregated forecasts are often proportionally more accurate than item-level forecasts because independent variations partly offset one another. However, aggregation can conceal regional capacity constraints or divergent product trends.

1.1 Independent and dependent demand

Independent demand originates outside the organization, such as customer demand for finished products. It is usually forecast. Dependent demand is derived from the production plan through a bill of materials. If one finished unit requires four components, planned production of 500 units creates gross dependent demand of 2,000 components, subject to inventory and timing adjustments.

1.2 A disciplined forecasting cycle

A professional process includes:

1. defining the decision and forecast horizon;
2. collecting and cleansing data;
3. identifying trend, seasonality and exceptional events;
4. selecting benchmark and candidate models;
5. estimating models using training data;
6. testing them on data not used for estimation;
7. incorporating documented market intelligence;
8. measuring error and bias;
9. approving, communicating and periodically recalibrating the forecast.

A naïve benchmark is essential. A sophisticated model adds little value if it cannot outperform a simple rule such as “next period equals the latest observation.”

2. Qualitative Forecasting

Qualitative methods are useful when historical data are unavailable, structurally obsolete or insufficient—for example, when launching a product in a new country.

The Delphi method obtains anonymous forecasts from experts over several rounds. A facilitator summarizes the responses and reasoning after each round. Experts may revise their estimates, but consensus is not forced.

Judgmental adjustment should be recorded separately from the statistical baseline. Organizations can then calculate forecast value added (FVA): whether each intervention improved accuracy relative to the forecast entering that process stage.

3. Quantitative Forecasting

3.1 Time-series structure

A time series may contain a level, trend, seasonality, cycles and random variation. Decomposition models must preserve dimensional consistency.

In an additive model:

D_t is observed demand in period t, while M_t, S_t and E_t are the baseline, seasonal effect and irregular effect. All are measured in demand units. Additive decomposition is appropriate when seasonal fluctuations remain approximately constant in absolute size. Over a complete seasonal cycle, additive seasonal indices are commonly normalized to sum to zero.

In a multiplicative model:

Here M_t is the baseline demand in units, while S_t and E_t are dimensionless ratios. Multiplicative seasonality is appropriate when seasonal variation grows proportionally with the series. Seasonal factors are commonly normalized to average 1 over a cycle.

If a linear baseline is used, it may be defined as:

a is the baseline at t=0 in demand units, and b is the change in demand units per period. The level and additive trend are combined inside M_t; they must not be multiplied together as if both were dimensionless factors.

3.2 Naïve and seasonal-naïve forecasts

The naïve forecast is:

F_t+1 is the forecast for the next period and A_t is actual demand in the current period, both in the same units.

For seasonal data with cycle length m, the seasonal-naïve forecast is:

It assumes that the demand observed in the corresponding period of the previous cycle will recur.

3.3 Moving averages

A simple moving average using the most recent n periods is:

All observations have equal weight. The method smooths random variation but responds slowly to trend changes.

If demand in the latest three months was 520, 560 and 610 units, the three-period moving-average forecast is:

A weighted moving average assigns weights w_i satisfying w_i ≥ 0 and ∑w_i=1:

Greater weights on recent observations increase responsiveness but may amplify noise.

3.4 Simple exponential smoothing

For a series without pronounced trend or seasonality:

α is the smoothing constant, where 0≤α≤1. A high α responds rapidly to new observations; a low value produces greater smoothing.

Worked example. Suppose the current forecast is 500 units, actual demand is 560 units and α=0.30. Then:

Simple exponential smoothing should not be used without modification when persistent trend or seasonality is present. Holt’s method adds a trend component, while Holt–Winters methods incorporate trend and additive or multiplicative seasonality.

3.5 Causal forecasting

Causal models connect demand with explanatory variables such as price, income, advertising, weather or exchange rates. A simple linear regression is:

Y_t is demand, X_t is an explanatory variable, β₀ is the intercept, β₁ is the change in expected demand associated with a one-unit change in X_t, and ε_t is the unexplained error.

Regression forecasts rely on the stability of the estimated relationship and credible forecasts of the explanatory variables. Statistical association alone does not establish causation. Promotions are especially problematic because high expected demand may influence the decision to promote, creating endogeneity.

4. Measuring Forecast Performance

Let the forecast error be defined consistently as:

A positive error means actual demand exceeded the forecast—under-forecasting. A negative error means the forecast exceeded actual demand—over-forecasting.

4.1 Common measures

Mean error (ME):

ME measures directional bias but positive and negative errors can cancel.

Mean absolute deviation (MAD):

MAD is expressed in the original demand units.

Mean squared error (MSE) and root mean squared error (RMSE):

MSE heavily penalizes large errors but has squared units. RMSE restores the original demand units.

Mean absolute percentage error (MAPE):

MAPE is scale-free but is undefined when any A_t=0 and can be dominated by periods with very small actual demand.

Weighted absolute percentage error (WAPE):

WAPE is suitable when aggregate actual demand is nonzero. It weights errors according to actual volume.

4.2 Worked error calculation

Consider three periods:

Therefore:

The positive ME indicates slight average under-forecasting under the stated error convention.

A tracking signal is sometimes calculated as cumulative forecast error divided by MAD. Control limits are organizational decision rules rather than universal statistical constants. A signal should prompt diagnosis, not automatic model replacement.

4.3 Evaluation principles

Models should be evaluated on an out-of-sample validation period. Comparisons must use identical items, periods and horizons. Accuracy should also be segmented by region, product class and demand pattern because an attractive global average can conceal severe local errors.

5. Aggregate Planning

Aggregate planning translates expected demand into medium-term decisions about production, workforce, overtime, subcontracting, inventory and backlog. It commonly covers 3–18 months and uses product families rather than individual SKUs.

5.1 Planning strategies

• Level strategy: maintains relatively stable capacity and absorbs demand variation through inventory or backlog.
• Chase strategy: changes capacity or workforce to follow demand.
• Mixed strategy: combines stable core capacity with overtime, inventory, temporary labor or subcontracting.

A level plan supports workforce stability but may create excessive inventory. A chase plan reduces inventory but can increase hiring, training and labor-relations costs. Mixed plans are common in practice.

5.2 Inventory and backlog balance

When inventory and backlog are represented separately, the correct net-stock relationship is:

Where, for period t:

• I_t = ending inventory in units;
• B_t = ending backlog in units;
• P_t = completed regular production received in the period;
• O_t = completed overtime production;
• S_t = subcontracted or externally sourced receipts;
• D_t = demand due in the period.

All terms are measured in the same units. Beginning inventory and backlog must be included through I_t-1-B_t-1. Normally, a model imposes I_t≥0 and B_t≥0 and prevents simultaneous positive inventory and backlog for the same substitutable product and location. If orders are lost rather than backlogged, lost sales require a separate variable.

5.3 Capacity and workforce relationships

A basic workforce balance is:

W_t is the ending workforce, H_t is hiring and L_t is layoffs, all measured in workers.

If each worker can produce at most q_t units during regular time in period t, then:

This assumes comparable worker productivity and available materials and equipment. More detailed models use hours by skill, line or facility.

An aggregate plan generally minimizes relevant cost:

The coefficients represent regular production, overtime, subcontracting, inventory holding, backlog, hiring and layoff costs in compatible monetary units. The objective is subject to demand-balance, capacity, workforce and policy constraints. Service requirements, carbon limits or supplier-risk restrictions may also be included.

6. Sales and Operations Planning

S&OP is a recurring cross-functional process that balances demand and supply at an aggregate level. A typical monthly cycle includes data preparation, demand review, supply review, reconciliation and executive approval.

Figure: The monthly S&OP cycle links demand, supply, financial consequences and executive decisions.

A mature S&OP process produces:

• a consensus unconstrained demand forecast;
• a feasible supply plan;
• documented constraints and assumptions;
• alternative scenarios and financial consequences;
• assigned decisions and accountability.

A “one-number” process means that functions reconcile to a common approved plan. It does not mean uncertainty is represented by a single immutable point estimate. Ranges and scenarios remain necessary.

7. Integrated Business Planning

IBP extends S&OP by more explicitly connecting operational plans with strategy, finance, portfolio decisions and risk management. It may evaluate revenue, margin, cash flow, working capital and strategic capacity alongside physical volume.

The labels are used inconsistently across companies; therefore, process maturity matters more than terminology. A renamed S&OP meeting is not IBP unless decision scope and financial integration genuinely expand.

8. Collaborative Planning Across the Supply Chain

Collaborative planning reduces information distortion among retailers, manufacturers, logistics providers and suppliers. Collaborative Planning, Forecasting and Replenishment (CPFR) typically involves:

1. agreeing scope, responsibilities and performance measures;
2. creating a joint business plan;
3. sharing forecasts and identifying exceptions;
4. resolving material differences;
5. coordinating orders, replenishment and performance review.

Collaboration can reduce the bullwhip effect, in which order variability increases upstream. Causes include forecast updating, order batching, promotions, rationing behavior and long lead times. Effective collaboration requires timely data, compatible definitions, cybersecurity safeguards and clear rules for commercially sensitive information.

Global collaboration must also account for lead-time calendars, local holidays, customs delays, minimum order quantities and differing data-protection or competition laws.

9. Illustrative Case: Meridian Home Systems

Background

Meridian Home Systems is a fictional European producer of connected air-quality devices sold in Europe, Southeast Asia and the Middle East. Components are sourced globally, while final assembly occurs at one plant. Demand is seasonal and influenced by distributor promotions.

Problem

Sales teams had submitted optimistic forecasts without identifying promotional assumptions. Procurement used a separate forecast, and production treated the sales target as committed demand. The result was simultaneous component shortages and excess finished-goods inventory.

Analysis

A cross-functional team created a statistical baseline, separately documented promotional uplifts and consolidated demand into one product family. The approved three-month forecast was:

Initial inventory was 2,000 units and initial backlog was zero. Regular production capacity was 10,000 units per month. Overtime capacity was 3,000 units per month. Inventory holding cost was €2 per ending unit per month. Overtime carried an €8 per-unit premium, while subcontracting carried a €12 premium.

Approach

The team developed the following mixed plan:

For month 2, for example:

Total demand is 36,000 units. Total available supply is initial inventory of 2,000 plus regular production of 30,000 and overtime production of 4,000, also totaling 36,000 units.

Relevant plan cost is:

Thus, total relevant incremental cost is €42,000.

An alternative used no overtime in month 2 and subcontracted 4,000 units in month 3. Its ending inventories would be 3,000, 1,000 and 0 units. Its relevant cost would be €2(4,000)+€12(4,000)=€56,000. Under the stated assumptions, the mixed overtime plan saves €14,000.

Outcome

Meridian adopted the plan after verifying component availability and overtime labor rules. It also introduced monthly forecast-bias reporting, supplier capacity reviews and scenario ranges for promotional demand. Because this is an illustrative case, the outcome demonstrates the logic of integrated planning rather than reporting empirical company results.

Learning Points

• Forecast assumptions should be transparent and separable from targets.
• Aggregate feasibility depends on both demand and supply timing.
• Beginning inventory and backlog cannot be omitted from stock-flow calculations.
• The cheapest mathematical plan must still satisfy labor, supplier and service constraints.
• S&OP creates value through coordinated decisions, not merely forecast production.

10. Technology, AI and Contemporary Practice

Cloud planning platforms increasingly combine transactional data, point-of-sale information, weather, mobility indicators and supplier data. Machine-learning models can detect nonlinear relationships and support short-term demand sensing. Digital twins permit scenario analysis across facilities, inventory and transport networks.

These tools do not remove the need for governance. Models may fail when conditions depart from their training data, as during pandemics, wars or regulatory shocks. Excessively granular data can create false precision. Good practice includes benchmark comparison, model monitoring, explainability proportional to decision risk, controlled human overrides and retention of an audit trail.

Probabilistic forecasting is also becoming more important. Instead of reporting only expected demand, planners may estimate quantiles or prediction intervals. Capacity and inventory decisions can then reflect service objectives and the asymmetric costs of shortages and surpluses.

11. Governance, Ethics and Limitations

Forecasting systems can create ethical and legal concerns. Customer-level data must be collected and processed lawfully. Algorithmic allocation during shortages should not discriminate unfairly among regions or customers. Managers should not manipulate forecasts to achieve bonuses or justify predetermined investments.

Competition law may restrict the exchange of detailed price, customer or future-output information among market participants. Collaborative planning agreements therefore require legal review, access controls and clearly defined data purposes.

Every forecast remains conditional. Historical relationships can break, causal variables may themselves be forecast inaccurately, and aggregate plans simplify product substitution, capacity and lead times. Robust organizations combine quantitative models with documented judgment, scenarios, contingency capacity and frequent review.

Chapter Summary

Demand forecasting estimates future market requirements, while aggregate and integrated planning determine how capacity and resources should respond. Qualitative methods are valuable when reliable data are scarce; time-series and causal models are appropriate when relevant historical evidence exists. Forecasts should be evaluated out of sample using measures such as ME, MAD, RMSE, MAPE and WAPE, with error signs interpreted consistently.

Aggregate planning balances production, inventory, backlog, workforce, overtime and external sourcing over a medium-term horizon. Its stock-flow equations must include beginning inventory and backlog. S&OP reconciles demand and supply through cross-functional decisions, while IBP extends the process to finance, strategy and risk. CPFR and related collaboration can reduce information distortion across global supply chains. Technology improves scale and responsiveness, but effective planning still depends on transparent assumptions, governance, ethical data use and managerial accountability.

Key Terms

Aggregate planning: Medium-term coordination of demand, capacity, inventory, workforce and sourcing.

Backlog: Demand accepted but not yet fulfilled.

Bullwhip effect: Amplification of order variability as information moves upstream.

Causal forecasting: Forecasting based on relationships between demand and explanatory variables.

Collaborative Planning, Forecasting and Replenishment (CPFR): Joint planning and replenishment among supply-chain partners.

Demand sensing: Short-horizon forecast updating using recent, high-frequency signals.

Delphi method: Iterative anonymous expert forecasting process.

Forecast bias: Persistent directional tendency to forecast above or below actual demand.

Forecast error: Difference between actual and forecast demand.

Forecast value added (FVA): Change in forecast performance attributable to a process step.

Integrated business planning (IBP): Integration of operational, financial, strategic and risk plans.

Mean absolute deviation (MAD): Average absolute forecast error in original units.

Mean absolute percentage error (MAPE): Average absolute error as a percentage of actual demand.

Multiplicative seasonality: Seasonal variation represented by dimensionless proportional factors.

Probabilistic forecast: Forecast expressing a distribution, quantiles or intervals rather than only a point estimate.

Sales and operations planning (S&OP): Cross-functional process for reconciling aggregate demand and supply.

Seasonal index: Additive or multiplicative measure of recurring seasonal influence.

Tracking signal: Ratio used to monitor cumulative forecast error relative to an error scale.

Weighted absolute percentage error (WAPE): Total absolute error divided by total absolute actual demand.

Review Questions

Short Answer Questions

1. Distinguish a forecast, a plan and a target.
2. When is qualitative forecasting especially useful?
3. What does a negative error mean when e_t=A_t-F_t?
4. Why is MAPE unsuitable when actual demand equals zero?
5. How does IBP extend the scope of S&OP?

Long Answer Questions

1. Compare qualitative, time-series and causal forecasting methods, including their assumptions and limitations.
2. Explain additive and multiplicative decomposition with explicit attention to unit consistency.
3. Discuss level, chase and mixed aggregate-planning strategies.
4. Describe a complete monthly S&OP cycle and the responsibilities of the participating functions.
5. Evaluate the benefits, risks and governance requirements of AI-enabled global demand forecasting.

Analytical / Application Questions

1. Actual demand for four months is 200, 240, 220 and 280 units. Calculate the three-period moving-average forecast for month 5 and explain one limitation of the result.
2. A company begins a month with 500 units of inventory and 100 units of backlog. It receives 2,000 regular units and 300 overtime units against demand of 2,600 units. Calculate ending net stock and state a feasible inventory–backlog combination.
3. Design an S&OP response for a multinational company facing a 20% demand increase, a capacity-constrained supplier and uncertain port lead times. Identify assumptions, scenarios and required executive decisions.

Multiple Choice Questions with Answers

Question 1

Using simple exponential smoothing, the current forecast is 500 units, actual demand is 560 units and α=0.30. What is the next-period forecast?

A. 506 units

B. 518 units

C. 542 units

D. 560 units

Correct Answer: B. 518 units

Question 2

If forecast error is defined as e_t=A_t-F_t, what does a negative mean error indicate?

A. Average over-forecasting

B. Average under-forecasting

C. Zero forecast bias

D. Increasing seasonal demand

Correct Answer: A. Average over-forecasting

Question 3

Which equation correctly represents net inventory when unmet demand is backlogged?

A. I_t-B_t=P_t+O_t+S_t+D_t

B. I_t+B_t=I_t-1+P_t-D_t

C. I_t-B_t=I_t-1-B_t-1+P_t+O_t+S_t-D_t

D. B_t=D_t-I_t

Correct Answer: C. I_t-B_t=I_t-1-B_t-1+P_t+O_t+S_t-D_t

Question 4

In the multiplicative decomposition D_t=M_t×S_t×E_t, which statement is dimensionally correct?

A. All three factors must be measured in demand units.

B. M_t has demand units, while S_t and E_t are dimensionless.

C. Only S_t has demand units.

D. M_t must be measured in demand units squared.

Correct Answer: B. M_t has demand units, while S_t and E_t are dimensionless.

Question 5

For actual demands 100, 120 and 150 and forecasts 90, 100 and 175, what is MAPE?

A. 12.22%

B. 13.89%

C. 14.44%

D. 18.33%

Correct Answer: C. 14.44%

Question 6

A firm begins with 2,000 units of inventory, produces 10,000 regular units and faces demand of 9,000 units with no backlog or other receipts. What is ending inventory?

A. 1,000 units

B. 2,000 units

C. 3,000 units

D. 21,000 units

Correct Answer: C. 3,000 units

Question 7

Which activity is most characteristic of CPFR?

A. Keeping retailer forecasts confidential from suppliers

B. Jointly identifying forecast exceptions and coordinating replenishment

C. Replacing all forecasts with production targets

D. Eliminating performance measurement between partners

Correct Answer: B. Jointly identifying forecast exceptions and coordinating replenishment

Question 8

Why may WAPE be preferred to MAPE for an aggregate set containing periods with zero actual demand?

A. WAPE never requires forecast data.

B. WAPE uses squared errors.

C. WAPE avoids division by each individual period's actual demand, provided aggregate actual demand is nonzero.

D. WAPE always gives a lower numerical value than MAPE.

Correct Answer: C. WAPE avoids division by each individual period's actual demand, provided aggregate actual demand is nonzero.

Question 9

What most clearly distinguishes mature IBP from a narrowly operational S&OP process?

A. IBP removes the need for demand forecasts.

B. IBP integrates operational plans with finance, strategy and risk.

C. IBP is limited to daily production scheduling.

D. IBP excludes product-portfolio decisions.

Correct Answer: B. IBP integrates operational plans with finance, strategy and risk.

Question 10

Which factor can contribute to the bullwhip effect?

A. Perfectly synchronized replenishment with no order batching

B. Short, stable lead times with shared demand data

C. Order batching and repeated forecast updating

D. Complete elimination of promotions and rationing

Correct Answer: C. Order batching and repeated forecast updating