Forecasting and Demand Planning

Introduction

Operations decisions are made before the future is known. Capacity, inventory, workforce, procurement, transportation, budgeting, and customer-service commitments all depend on expectations about future demand. Forecasting is the disciplined process of developing those expectations; demand planning is the broader managerial process of translating them into an agreed, actionable view of demand and coordinating the organization’s response.

A forecast is not a promise and not a prediction of a single inevitable outcome. It is an estimate conditional on information, assumptions, and a defined horizon. Good forecasting therefore combines statistical evidence with market knowledge, makes uncertainty visible, and establishes processes for learning from error. At graduate level, the central issue is not selecting the most sophisticated algorithm. It is designing a forecasting system that is accurate enough, explainable, timely, resilient to change, and connected to decisions.

Learning Objectives

By the end of this chapter, students should be able to:

1. Distinguish forecasting, demand planning, and decision-making.
2. Select and apply judgmental, time-series, causal, and scenario-based methods.
3. Calculate and interpret forecast-error measures, bias, and uncertainty.
4. Diagnose seasonality, trend, intermittency, structural change, and outliers.
5. Design collaborative demand planning and sales and operations planning (S&OP) processes.
6. Evaluate the managerial, ethical, technological, and governance implications of forecasting.

1. Foundations of Demand Forecasting

Demand is the quantity customers would purchase under stated conditions. Sales are the quantity actually recorded. The distinction matters: stockouts, allocation rules, lost customers, price changes, and delivery constraints can make observed sales a poor measure of unconstrained demand. A forecast should therefore identify whether it estimates observed sales, unconstrained market demand, orders, shipments, or consumption.

Forecasts also require a horizon and a level of aggregation. Strategic forecasts may cover years and support facility or technology decisions. Tactical forecasts commonly cover months and support aggregate capacity, budgets, and supplier agreements. Operational forecasts may cover days or weeks and support replenishment, scheduling, and labor deployment. Forecasts are often more reliable at aggregate levels because random variation partially cancels across products, locations, or customers.

The forecasting system can be represented as a sequence:

Figure: Forecasting is a closed learning cycle linking evidence, judgment, decisions, and error evaluation.

Forecast quality has at least four dimensions:

• Accuracy: closeness of estimates to outcomes.
• Bias: systematic over- or under-forecasting.
• Timeliness: availability early enough to influence decisions.
• Usability: clarity, stability, explainability, and fit with planning processes.

There is usually a trade-off between granularity and reliability. A forecast for a single slow-moving item at one location may be highly variable, whereas a product-family forecast for the entire region may be stable. The appropriate level is the level at which decisions are made, with reconciliation across levels where necessary.

2. Main Forecasting Approaches

2.1 Judgmental forecasting

Judgmental methods use expert knowledge when historical data are limited, unreliable, or made irrelevant by structural change. Methods include executive opinion, sales-force composites, customer surveys, market research, Delphi panels, analogy, and structured consensus meetings.

Judgment is particularly useful for new products, regulatory changes, major promotions, geopolitical disruption, and technology substitution. It is not automatically superior to statistical methods. Experts may anchor on recent events, exhibit optimism, protect targets, or double-count the same information. Good practice separates an unbiased statistical baseline from documented adjustments. Each adjustment should state its magnitude, rationale, evidence, owner, and expiry date.

The Delphi method obtains several rounds of anonymous estimates and feedback, reducing dominance by senior participants. It is useful for long-range or highly uncertain questions, but it can be slow and sensitive to panel composition.

2.2 Time-series forecasting

Time-series methods infer future values from historical patterns. A simple naïve forecast uses the latest observation:

Here, A_t is actual demand in period t, and F_t+1 is the forecast for the next period. The method is appropriate as a benchmark and can be effective for random-walk demand. It does not explicitly model trend or seasonality.

A moving average of order k is:

It smooths random variation, but it lags when demand trends and gives equal weight to all included observations. The units of the forecast are the same as demand units.

A weighted moving average is:

Weights reflect recency or perceived relevance. The formula assumes the weights are fixed for the forecast calculation and sum to one.

Exponential smoothing updates the forecast recursively:

The smoothing constant α controls responsiveness. Higher α reacts quickly to new observations but may amplify noise; lower α produces a smoother forecast but responds slowly to change. Simple exponential smoothing is designed for level data without systematic trend or seasonality.

For trend, Holt’s method uses level and trend components:

Here, L_t is the estimated level, T_t is the trend, α and β are smoothing parameters between zero and one, and m is the number of periods ahead. Holt’s method assumes a reasonably persistent trend and no explicit seasonal component.

Seasonal demand may be modeled with seasonal indices. Under a multiplicative model:

Seasonal indices are commonly normalized so their average over a complete seasonal cycle equals one. Multiplicative seasonality is appropriate when seasonal amplitude grows with the level; additive seasonality is more suitable when seasonal deviations are approximately constant in absolute units.

More advanced methods include ARIMA, state-space models, exponential-smoothing state-space models, and machine-learning models. Complexity should be justified by improved out-of-sample performance or decision value, not by technical novelty.

2.3 Causal and explanatory forecasting

Causal methods relate demand to explanatory variables such as price, advertising, income, weather, competitor actions, distribution, and macroeconomic indicators. A basic linear model is:

Y_t is demand, X_jt are explanatory variables, β_j are coefficients, and ε_t is unexplained error. Coefficients should not automatically be interpreted as causal effects: omitted variables, reverse causality, confounding, multicollinearity, and simultaneous business decisions can produce misleading relationships.

For price elasticity, a common interpretation is:

Elasticity is dimensionless. A negative price elasticity is typical for ordinary demand, but the magnitude varies by segment, time horizon, substitutes, and stock availability. Causal analysis requires sound research design, appropriate controls, and validation outside the estimation sample.

2.4 Scenario-based forecasting

A scenario is a coherent, internally consistent description of a possible future, not simply a high, medium, and low number. Scenarios are valuable when uncertainty is structural or when historical relationships may fail. A robust scenario process identifies driving forces, critical uncertainties, plausible combinations, operational implications, early-warning indicators, and contingent actions.

Probabilistic forecasts can represent uncertainty directly. If scenario outcomes are D_i with probabilities p_i, expected demand is:

The expected value is not necessarily the most likely outcome and should not replace risk analysis. Capacity and inventory decisions may require a service percentile, downside estimate, or robust plan rather than the mean.

3. Forecast Accuracy, Bias, and Uncertainty

Let actual demand be A_t and forecast demand be F_t. Forecast error is conventionally:

With this convention, positive error means under-forecasting and negative error means over-forecasting. Organizations must state their convention because the opposite sign is also used.

Common measures include:

MAD and RMSE are in demand units. RMSE penalizes large errors more heavily. Percentage measures can be useful across products but are unstable when actual demand is zero or near zero:

Weighted absolute percentage error is often preferable for portfolios:

WAPE remains interpretable at aggregate level but can conceal poor performance on low-volume items. No single metric is universally best.

Bias can be measured by mean error:

A tracking signal compares cumulative error with MAD:

It can flag persistent bias, but control limits should be calibrated to the process and not treated as universal laws. Bias may arise from optimistic sales inputs, systematic promotion omissions, structural breaks, or incentives tied to targets.

Forecast accuracy should be assessed using rolling-origin or time-series cross-validation. Randomly splitting observations can leak future information into the training set. Comparisons should use the same forecast horizon, information availability, aggregation level, and treatment of promotions. Statistical significance is useful, but managerial significance also matters: a small accuracy gain may be irrelevant if it adds cost or reduces explainability.

Uncertainty grows with horizon and may be represented by prediction intervals. A prediction interval concerns a future observation, whereas a confidence interval concerns an estimated parameter or mean. Intervals depend on model assumptions, including error independence or correctly modeled autocorrelation, stable variance, and an appropriate distributional approximation. They should be widened when structural uncertainty, data revisions, or scenario uncertainty is material.

4. Collaborative Demand Planning and S&OP

Collaborative demand planning integrates information held by sales, marketing, finance, operations, procurement, logistics, and customers. Its purpose is not to average opinions mechanically; it is to expose assumptions, resolve conflicts, and create accountability.

A mature monthly S&OP cycle commonly includes:

1. Data and portfolio review: validate history, events, exclusions, and product status.
2. Statistical baseline: produce an unconstrained demand forecast.
3. Demand review: incorporate promotions, customer intelligence, launches, losses, and risks.
4. Supply review: test capacity, materials, labor, inventory, and logistics constraints.
5. Reconciliation: evaluate alternative responses, such as overtime, outsourcing, inventory build, allocation, or price changes.
6. Executive review: decide trade-offs and approve one integrated plan.
7. Execution and learning: monitor assumptions, outcomes, service, inventory, and forecast bias.

A useful distinction is between the unconstrained demand plan—what customers are expected to require—and the constrained operating plan—what the organization can profitably and feasibly supply. Conflating the two hides shortages and encourages sandbagging. S&OP should make the gap explicit.

Consensus does not mean equal authority for every estimate. Roles should specify who owns the baseline, who may adjust it, what evidence is required, who approves overrides, and how conflicts are escalated. Forecast overrides should be logged so their value can be evaluated retrospectively.

5. Worked Example: Comparing Simple Forecasts

Suppose monthly demand for five months is 100, 120, 110, 130, and 140 units. A two-period moving-average forecast is available from month 3 onward.

For month 3:

F₃ = (100 + 120)/2 = 110

For month 4:

F₄ = (120 + 110)/2 = 115

For month 5:

F₅ = (110 + 130)/2 = 120

Errors using e_t = A_t − F_t are 0, 15, and 20 units. Therefore:

MAD = (|0| + |15| + |20|)/3 = 11.67 units.

ME = (0 + 15 + 20)/3 = 11.67 units, indicating under-forecasting on average. The next forecast is:

F₆ = (130 + 140)/2 = 135 units.

The positive mean error does not prove that the model is wrong; the series may be trending upward, making a moving average systematically lag. A trend method or causal information should be tested against the same historical holdout periods.

6. Illustrative Case: Northstar Appliances

Background. Northstar Appliances, a fictional mid-sized manufacturer, sells countertop air purifiers through retailers and its online channel. Sales rise in winter and during pollution alerts. Retail promotions are recorded inconsistently, and stockouts occurred during the previous winter.

Problem. The sales team’s annual forecast was 22% above the statistical baseline. Operations could not obtain enough filters and proposed a large inventory build. Finance questioned whether the additional demand reflected genuine market growth or retailer forward-buying.

Analysis. A cross-functional team first reconstructed available-to-promise demand by identifying stockout weeks and separating shipments from consumer sell-through. It created a seasonal baseline and evaluated it using rolling-origin tests. The team found that the sales forecast systematically overestimated non-promotional demand, while the baseline underestimated alert-driven spikes. A causal model using air-quality alerts improved short-horizon forecasts, but its performance deteriorated when alert data were unavailable in advance.

Approach. Northstar adopted a layered process. The statistical seasonal forecast became the baseline. Sales could submit customer-specific promotion and listing information through a standard event calendar. Marketing supplied expected media intensity; operations supplied filter constraints; and an executive review approved scenarios: normal winter, severe pollution season, and supplier disruption. The company maintained separate unconstrained demand and constrained supply plans. It also introduced forecast-value-added reviews to determine whether each adjustment improved accuracy.

Outcome. In the next season, the organization reduced unexplained overrides, improved visibility of stockout risk, and used pre-approved contingent actions, including earlier filter orders and selective allocation to high-priority channels. The main improvement was not a single superior algorithm. It was the governance system that made assumptions visible and connected forecasts to decisions.

Learning points. First, observed sales must be corrected for lost demand where possible. Second, causal variables are useful only when their future availability is realistic. Third, scenarios are appropriate when rare events dominate risk. Finally, collaboration creates value when it adds distinct information rather than political pressure.

7. Technology, AI, and Governance

Cloud planning platforms can automate data pipelines, hierarchical reconciliation, event calendars, exception management, and forecast monitoring. Machine learning can detect nonlinear relationships and high-dimensional interactions, especially where many explanatory variables and sufficient history exist. However, predictive performance may deteriorate after market change, and black-box outputs can be difficult to challenge.

Responsible use of AI requires data lineage, version control, access controls, explainability appropriate to the decision, drift monitoring, human review, and documented override rights. Models should not use protected or sensitive customer attributes without legitimate purpose and governance. Forecasts that affect worker scheduling, customer allocation, or credit-related decisions may create fairness concerns. Organizations should also guard against automation bias: accepting a model because it appears objective.

A sound model-risk process includes champion–challenger comparisons, back-testing, stress testing, monitoring of residuals and bias by segment, and retirement criteria. Cybersecurity and supplier data confidentiality are operational concerns, not merely IT concerns.

8. Challenges and Best Practices

Forecasting is difficult because demand is affected by promotions, substitution, cannibalization, new products, intermittent orders, supply shortages, changing customer behavior, and data quality problems. Structural breaks cannot always be detected promptly. Aggregation can hide local shortages, while excessive disaggregation creates noise and administrative cost.

Best practices include:

• define the forecast object, horizon, unit, and decision use;
• establish a transparent statistical baseline;
• treat judgment as evidence-based adjustment, not replacement by hierarchy;
• evaluate models out of sample and against simple benchmarks;
• measure both accuracy and bias;
• report uncertainty and scenario assumptions;
• distinguish demand from shipments constrained by supply;
• reconcile forecasts across product, location, and organizational levels;
• document promotions, launches, stockouts, and extraordinary events;
• connect forecast metrics to service, inventory, revenue, and profit outcomes;
• review overrides and forecast value added;
• design S&OP as a decision process, not a meeting calendar.

Chapter Summary

Forecasting estimates future demand under explicit information and assumptions; demand planning integrates that estimate with market intelligence, organizational judgment, constraints, and decisions. Judgmental methods are valuable under novelty and structural change but require protection against bias. Time-series methods exploit level, trend, and seasonality; causal methods use explanatory variables but require careful interpretation; scenario planning addresses uncertainty that historical patterns cannot adequately represent.

Forecast errors should be measured with clearly defined conventions. MAD, MSE, RMSE, MAPE, WAPE, mean error, and tracking signals answer different questions and have different limitations. Accuracy evaluation should be out of sample and aligned with the decision horizon. Collaborative demand planning and S&OP reconcile unconstrained demand with feasible supply and create an accountable cycle of review and learning. Technology and AI can improve scale and responsiveness, but governance, explainability, data quality, and human challenge remain essential.

Key Terms

Forecasting: Estimating future values or events for decision-making.

Demand planning: Coordinating forecasts, market intelligence, assumptions, and decisions about future demand.

Forecast horizon: The future period covered by a forecast.

Naïve forecast: A forecast that uses the latest observed value as the next estimate.

Moving average: A forecast based on the average of recent observations.

Exponential smoothing: A recursive method that weights recent observations more heavily.

Seasonality: A repeating pattern associated with a calendar or operating cycle.

Trend: Persistent systematic movement in the level of a series.

Causal forecasting: Forecasting based on explanatory variables believed to influence demand.

Scenario: A coherent description of a plausible future and its implications.

Forecast error: The difference between actual and forecast demand.

Bias: Persistent tendency to over- or under-forecast.

MAD: Mean absolute deviation, the average absolute forecast error.

RMSE: Root mean squared error, which penalizes large errors more strongly.

MAPE: Mean absolute percentage error.

WAPE: Weighted absolute percentage error based on aggregate absolute error.

Unconstrained demand: Expected customer requirements without supply limitations.

S&OP: Sales and operations planning, an integrated cross-functional planning process.

Forecast override: A deliberate adjustment to a statistical baseline.

Forecast value added: The measurable improvement produced by a forecasting step or intervention.

Review Questions

Short Answer Questions

1. Why should sales and unconstrained demand be distinguished?
2. What is the effect of increasing α in simple exponential smoothing?
3. Why can MAPE be problematic for intermittent demand?
4. What is the difference between forecast accuracy and forecast bias?
5. What is the purpose of an executive S&OP review?

Long Answer Questions

1. Compare judgmental, time-series, causal, and scenario-based forecasting, including their assumptions and appropriate applications.
2. Explain how an organization should evaluate forecast accuracy using rolling-origin testing and multiple metrics.
3. Discuss how seasonality, trend, promotions, stockouts, and new products complicate demand estimation.
4. Design a monthly collaborative demand-planning and S&OP process for a multi-channel manufacturer.
5. Critically evaluate the benefits and governance risks of using AI in demand forecasting.

Analytical / Application Questions

1. A product has actual demand of 80, 100, and 120 units, while forecasts were 90, 95, and 110. Using e_t=A_t−F_t, calculate ME and MAD and interpret the signs.
2. A retailer faces three demand scenarios: 800 units with probability 0.50, 1,100 units with probability 0.30, and 1,500 units with probability 0.20. Calculate expected demand and explain why the retailer may still plan above the expected value.
3. A sales team increases a statistical baseline by 25% for a promotion but provides no evidence or event dates. Propose a governance process to evaluate the override and decide whether it should enter the approved demand plan.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand planning from forecasting?

A. Demand planning is limited to calculating moving averages.

B. Demand planning integrates forecasts, market information, constraints, and decisions.

C. Demand planning applies only to long-term capital investment.

D. Demand planning eliminates uncertainty from future demand.

Correct Answer: B. Demand planning integrates forecasts, market information, constraints, and decisions.

Question 2

In the convention e_t=A_t−F_t, a positive error indicates that the forecast has:

A. Overestimated demand

B. Underestimated demand

C. Exactly matched demand

D. Used a seasonal index

Correct Answer: B. Underestimated demand

Question 3

What is a principal limitation of a two-period moving average when demand is rising steadily?

A. It cannot produce a numerical forecast.

B. It always exaggerates the trend.

C. It tends to lag the rising demand.

D. It requires causal variables.

Correct Answer: C. It tends to lag the rising demand.

Question 4

If α is increased in simple exponential smoothing, the forecast generally becomes:

A. More responsive to recent observations

B. Unaffected by recent observations

C. Less responsive to recent observations

D. Identical to a seasonal forecast

Correct Answer: A. More responsive to recent observations

Question 5

Which measure penalizes large forecast errors relatively more heavily?

A. ME

B. MAD

C. RMSE

D. Median error

Correct Answer: C. RMSE

Question 6

Why should random train-test splitting be avoided for many time-series forecasting evaluations?

A. It always increases sample size.

B. It can allow future information to leak into model training.

C. It prevents calculation of MAD.

D. It removes all seasonality.

Correct Answer: B. It can allow future information to leak into model training.

Question 7

A causal forecast differs from a purely time-series forecast because it:

A. Uses explanatory variables such as price or advertising

B. Never requires historical demand data

C. Cannot be evaluated out of sample

D. Assumes all errors are zero

Correct Answer: A. Uses explanatory variables such as price or advertising

Question 8

What is the main purpose of separating unconstrained demand from constrained supply in S&OP?

A. To hide capacity shortages

B. To make the gap between expected requirements and feasible supply visible

C. To ensure sales always determines capacity

D. To eliminate the need for scenarios

Correct Answer: B. To make the gap between expected requirements and feasible supply visible

Question 9

For scenario outcomes D_i with probabilities p_i, expected demand is calculated as:

A. The largest scenario only

B. The simple average regardless of probabilities

C. The probability-weighted sum of outcomes

D. The smallest scenario only

Correct Answer: C. The probability-weighted sum of outcomes

Question 10

Which is a sound governance practice for AI-based forecasts?

A. Accept every model output without review

B. Remove version control to accelerate changes

C. Monitor drift, document data lineage, and permit informed human challenge

D. Use protected attributes whenever they improve fit

Correct Answer: C. Monitor drift, document data lineage, and permit informed human challenge