Demand Forecasting and Sales and Operations Planning

Introduction

Operations decisions are made before the future is known. A manufacturer must decide how much to produce, a hospital how many staff to schedule, and a retailer how much inventory to position. Demand forecasting provides an informed estimate of future requirements; Sales and Operations Planning (S&OP) converts that estimate into an integrated plan for demand, supply, inventory, workforce, finance and strategic priorities.

Forecasting is not prediction in the sense of certainty. It is a disciplined process for reducing uncertainty, making assumptions explicit and coordinating decisions. A useful forecast may be statistically accurate but operationally inappropriate if it ignores capacity, lead times or commercial commitments. Conversely, a forecast with modest statistical accuracy may support excellent decisions when it correctly identifies risks, ranges and scenarios.

Learning Objectives

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

1. Distinguish demand, sales, orders and forecasts.
2. Select qualitative, quantitative, time-series and causal forecasting approaches.
3. Calculate and interpret forecast-error measures.
4. Recognise bias, uncertainty, intermittency and aggregation effects.
5. Design a collaborative forecasting and S&OP process.
6. Use scenarios to balance demand, capacity, inventory and strategic objectives.
7. Evaluate technology, governance and ethical issues in forecasting.

1. The Nature of Demand Forecasting

Demand is the quantity customers would purchase under specified conditions. Sales are the transactions actually completed. Orders are commitments received by the firm. These measures can differ because of stockouts, promotions, price changes, lost sales, cancellations, allocation rules and delivery constraints. For example, observed sales during a stockout may understate true demand.

A forecast should therefore specify:

• the item or product family;
• the location and market;
• the time bucket and horizon;
• the demand measure being forecast;
• the information available when the forecast is prepared; and
• the decision the forecast will support.

Forecast horizons commonly include short term (days to weeks), medium term (months to approximately two years) and long term (multiple years). Short-term forecasts support replenishment and scheduling. Medium-term forecasts support aggregate capacity, workforce and inventory planning. Long-term forecasts inform facilities, technology, product portfolios and strategy.

Forecasts can be hierarchical: total market demand may be disaggregated into regions, channels, products and locations. Forecasts at an aggregate level are often more accurate because random variation cancels out. However, operational execution requires forecasts at the level where decisions are made. A sound system reconciles forecasts across the hierarchy rather than allowing contradictory numbers.

Forecasting principles

Several principles are especially important:

1. Forecasts are normally wrong; the objective is to quantify and manage error.
2. Aggregated forecasts are generally more accurate than highly disaggregated forecasts.
3. Shorter horizons are usually more accurate than longer horizons.
4. Forecasts should be separated from targets. A target expresses ambition; a forecast expresses expected outcomes.
5. Forecasting should be judged by decision relevance, not only statistical accuracy.
6. Human judgement is valuable when structural change, missing information or one-off events make historical patterns unreliable.

2. Qualitative and Quantitative Methods

Qualitative methods

Qualitative methods are appropriate when historical data are limited, the product is new, the market is changing structurally, or expert knowledge contains information not yet present in databases.

Executive opinion obtains a forecast from senior managers. It is quick but vulnerable to hierarchy, optimism and political pressure. Sales-force composite forecasting uses local customer knowledge, but salespeople may understate forecasts to obtain easier targets or overstate them to secure inventory. Customer surveys and market research investigate purchase intentions, although stated intention may not become actual demand. Delphi methods obtain several anonymous expert rounds, summarising responses between rounds to reduce dominance effects. Analogy forecasting uses the adoption pattern of a comparable product or market.

Qualitative methods should not be dismissed as unscientific. Their quality depends on structured evidence, explicit assumptions, independent challenge and retrospective evaluation. A documented judgemental adjustment can be audited; an undocumented adjustment cannot.

Quantitative methods

Quantitative forecasting uses numerical observations and formal models. Two broad families are time-series and causal models.

Time-series models explain demand using its own historical pattern: level, trend, seasonality, cycles and irregular variation. Causal models explain demand using drivers such as price, income, advertising, weather, population, competitor activity or economic indicators.

A practical forecasting system often combines methods. A baseline statistical forecast may be reviewed by commercial experts, while causal variables and scenario assumptions are used for major changes. Combination can reduce model-specific risk, but combining forecasts does not automatically improve accuracy; it must be evaluated out of sample.

3. Time-Series Forecasting

Let y_t denote observed demand in period t, and F_t the forecast made for that period. Forecasting must avoid using information that was unavailable at the time the decision was made.

Naïve forecast

The simplest model uses the last observation:

This is appropriate as a benchmark, especially for stable or highly persistent series. For seasonal demand, a seasonal naïve model uses the observation from the corresponding prior season:

where m is the seasonal cycle length and h is the forecast horizon. For monthly data with annual seasonality, m = 12.

Moving average

A simple moving average of the most recent k observations is:

The method smooths random variation but responds slowly to trend. A larger k produces greater smoothing and slower responsiveness. The units of the forecast are the same as the units of demand.

Weighted moving average

A weighted moving average assigns greater importance to selected observations:

with w_i ≥ 0 and ∑w_i = 1. The weights must be dimensionless and sum to one so that the result retains demand units.

Exponential smoothing

Simple exponential smoothing is suitable for a series with level but no systematic trend or seasonality:

where 0 < α ≤ 1 is the smoothing constant. A high α reacts quickly to recent changes; a low value produces a smoother forecast. The model assumes that the underlying level changes gradually and that residual errors are not systematically biased.

The equivalent updating form is:

The term y_t − F_t is the current forecast error. This form makes the adjustment logic explicit.

For trend, Holt’s method estimates a level and trend. One common form is:

Here ℓ_t is the estimated level, b_t the trend, β the trend smoothing constant, and h the horizon. These formulas are applicable when trend is reasonably stable and should be treated cautiously when structural breaks occur.

For trend and seasonality, Holt–Winters methods add a seasonal component. Additive seasonality is appropriate when seasonal fluctuations have roughly constant absolute size; multiplicative seasonality is more suitable when seasonal variation increases with the level. Model selection should be based on validation performance and operational interpretation rather than automatic fitting alone.

Worked example: exponential smoothing

Suppose demand in period 1 is 100 units and the initial forecast for period 2 is 100. Demand in period 2 is 120, and α = 0.30. The forecast for period 3 is:

If demand in period 3 is 110, then:

The forecast adjusts toward actual demand but does not copy every fluctuation.

4. Causal Models and Regression

A causal model relates demand to explanatory variables. A linear regression with one driver can be written as:

where y_t is demand, x_t is the explanatory variable, β₀ is the intercept, β₁ is the expected change in demand for a one-unit change in x_t, and ε_t is unexplained error. A multiple regression is:

Regression is useful for what-if analysis, but association is not proof of causation. A price coefficient may be distorted if price reductions occur during weak demand. Explanatory variables may also be correlated with one another, creating multicollinearity. Residuals should be examined for changing variance, autocorrelation, outliers and nonlinearity. Forecast evaluation must use a holdout period or rolling-origin validation rather than relying only on in-sample fit.

For a log-linear model, coefficients can be interpreted approximately as elasticities. However, transformations change interpretation and require careful back-transformation when forecasts are converted to original units. The analyst must also ensure that future values of causal variables are available or forecast separately.

5. Forecast Accuracy, Bias and Uncertainty

Define the error for period t as:

Positive error means actual demand exceeded the forecast. A negative error means the forecast was too high.

Common measures

Here n is the number of evaluated periods and y_t ≠ 0 is required for MAPE. MAPE becomes unstable or undefined when actual demand is zero or close to zero. MAE and RMSE are preferable when zeros are common. WAPE can obscure poor performance on low-volume items because high-volume items dominate the denominator.

Error measures are not interchangeable. RMSE is useful when large misses are particularly costly. MAE is easier to interpret operationally. ME detects systematic bias but can conceal large offsetting errors. Accuracy should therefore be reported with at least one scale-dependent measure and one bias measure, alongside service and financial outcomes.

A forecast interval expresses uncertainty. Under assumptions of approximately independent, normally distributed errors with constant variance, a simple interval can be represented as forecast ± a multiple of estimated standard error. These assumptions often fail in supply chains because errors can be seasonal, correlated and affected by promotions. Prediction intervals should therefore be validated empirically and communicated as ranges, not certainties.

Bias and tracking signals

Persistent positive or negative errors indicate bias. A commonly used tracking signal is:

where RSFE = ∑e_t is the running sum of forecast errors and MAD = (∑|e_t|)/n is mean absolute deviation. A large absolute value may prompt investigation, but thresholds are context-dependent and should not be used mechanically. Causes include trend omission, discontinued products, systematic promotional effects, data defects and deliberate managerial overrides.

6. Demand Sensing, Intermittency and Special Events

Intermittent demand contains many zero observations and occasional nonzero demand, common in spare parts. Conventional percentage metrics and ordinary moving averages can perform poorly. Croston-type methods separately estimate the size of demand and the interval between demands; variants such as SBA and TSB address some bias and obsolescence issues. These methods require careful validation because no single intermittent-demand method is universally best.

Promotions, holidays, weather events, launches and disruptions should be represented explicitly. Historical sales during a promotion should not be treated as ordinary baseline demand unless the promotion will recur. A useful decomposition is:

baseline demand + event effect + distribution effect + random variation.

Demand sensing uses near-real-time signals such as point-of-sale data, web searches, orders, traffic, weather and supply-chain status. Faster data can improve responsiveness, but it can also amplify noise and create frequent replanning. Governance should define which signals are trusted, how anomalies are filtered and when a forecast is allowed to change.

7. Collaborative Planning and Forecast Governance

Forecasting is organisational as well as statistical. Sales, marketing, finance, operations, procurement and customers possess different information and incentives. Collaboration is valuable when it adds genuine knowledge, not when it simply averages opinions.

A disciplined process includes:

1. Data preparation: establish definitions, cleanse outliers, identify stockouts and document product and customer changes.
2. Statistical baseline: generate a reproducible forecast with a specified horizon and version.
3. Demand review: incorporate promotions, customer intelligence, launches and market events using documented assumptions.
4. Challenge and reconciliation: test adjustments against evidence and reconcile product, regional and corporate levels.
5. Supply review: compare demand with capacity, materials, labour, logistics and financial constraints.
6. Executive review: resolve trade-offs, approve scenarios and assign owners.
7. Measurement and learning: monitor accuracy, bias, service, inventory, overrides and forecast value added.

Forecast value added asks whether each process step improves the forecast relative to the prior step. A judgemental override that consistently worsens accuracy should be redesigned or removed, even if it reflects senior opinion.

Data governance requires common calendars, product hierarchies, ownership, access controls, version management and audit trails. Forecast overrides should record who changed the number, why, expected effect and expiry date.

8. Sales and Operations Planning

S&OP is a recurring cross-functional process that aligns market demand and organisational supply with business strategy. It is normally conducted at an aggregate or product-family level over a rolling medium-term horizon, often monthly. The objective is not merely to create one number; it is to expose gaps and make decisions about trade-offs.

A typical S&OP cycle contains:

• Portfolio and data review: products, customers, events, actuals and assumptions.
• Demand planning: unconstrained demand forecast and alternative demand scenarios.
• Supply planning: capacity, inventory, labour, suppliers, production, distribution and constraints.
• Pre-S&OP reconciliation: evaluate alternatives using service, cost, cash, revenue and risk.
• Executive S&OP: approve a feasible plan, policies, decisions and escalation actions.
• Implementation and feedback: translate aggregate decisions into master scheduling, procurement and deployment.

The distinction between unconstrained demand and constrained supply is essential. If a forecast is reduced merely because capacity is insufficient, the organisation may lose visibility of unmet market opportunity. The demand plan should show what customers are expected to require; the supply plan should show what can be delivered; the gap should be explicit.

Capacity and balancing alternatives

When demand exceeds available capacity, managers may use overtime, subcontracting, inventory, alternative sites, postponement, price changes, allocation, hiring or capital investment. Each alternative has cost, lead-time, quality, resilience and strategic implications.

A simplified aggregate-planning objective can be expressed as:

where P is regular production, O overtime, H hiring, L layoffs, I inventory and B backorders; the c terms are their respective costs. This is a conceptual objective, not a universal complete model. Real models require flow-balance constraints, workforce limits, capacity, inventory bounds, service requirements and non-linear costs.

For a period, a basic inventory balance is:

where I_t is ending inventory, S_t supply received during the period and D_t demand fulfilled or allocated. If backorders are allowed, the balance must distinguish positive inventory from unmet demand; otherwise the equation can conceal service failures.

9. Scenario Analysis and Strategic Trade-offs

A single-point plan hides uncertainty. Scenario analysis constructs internally consistent alternatives, such as base, high-demand, low-demand, supply-disruption and strategic-investment scenarios. Each scenario should specify drivers, assumptions, trigger indicators, decisions, financial consequences and owners.

Sensitivity analysis changes one assumption while holding others constant. Scenario analysis changes a coherent set of assumptions. Both are useful, but neither assigns probabilities automatically. If probabilities are credible, expected value may be calculated; if not, robust or minimax-regret approaches may be more defensible.

A scenario plan should answer:

• What could change?
• How early would the organisation detect it?
• What capacity, inventory or supplier decision must be made now?
• What is the trigger for switching response?
• Which strategic priorities are protected?

The purpose is not to predict every future but to reduce response time and prevent avoidable surprises.

Figure: The iterative S&OP cycle integrates demand, supply, reconciliation and executive decisions.

10. Applied Case: Northstar Medical Devices

Background. Northstar Medical Devices sells a disposable diagnostic cartridge through hospitals and distributors. Monthly demand is seasonal, promotions are irregular, and a critical imported component has a twelve-week lead time. Sales frequently submit optimistic targets, while operations produce a conservative plan because capacity is limited.

Problem. Stockouts occurred during winter, followed by excess inventory in spring. The organisation measured forecast accuracy against sales, even when stockouts prevented sales. Product-family forecasts did not reconcile with regional plans, and overrides were not recorded.

Analysis. A cross-functional team separated observed sales from known lost demand, created a seasonal naïve benchmark and compared it with exponential smoothing and a regression containing procedure volumes and hospital tender events. Rolling-origin validation showed that the regression improved performance only when tender information was genuinely available before the forecast date. A bias review found systematic underforecasting before winter.

Approach. Northstar introduced a monthly S&OP cycle. Demand planning produced a baseline and an unconstrained event-adjusted plan. Sales documented tender probabilities and expiry dates for overrides. Supply planning modelled component lead time, labour capacity and alternate suppliers. The executive meeting approved a base plan and a high-demand scenario with pre-authorised overtime and supplier escalation triggers.

Outcome. The organisation did not eliminate uncertainty, but it made the demand–capacity gap visible earlier. Inventory was positioned before the winter peak, while the high-demand response avoided committing to permanent capacity before evidence justified it. Performance was evaluated using MAE, bias, service level and inventory, rather than MAPE alone.

Learning points. Forecast quality depends on demand definitions, information timing, cross-functional governance and response design. A better forecast is useful only when the organisation can act on it.

11. Technology, AI and Responsible Use

Enterprise planning systems can automate data integration, hierarchical reconciliation, event calendars, exception alerts and scenario modelling. Machine-learning models may identify nonlinear relationships and interactions across large datasets. They are most valuable where data are abundant, stable enough for learning and linked to a decision process.

AI does not remove the need for causal reasoning or governance. Risks include leakage of future information into training data, unstable performance after structural change, biased customer or regional allocations, opaque recommendations, privacy violations and automation bias. Models should be benchmarked against simple methods, validated on future-like data, monitored for drift and accompanied by explanations appropriate to the decision. Human review should focus on exceptions and assumptions rather than mechanically overriding every statistical forecast.

12. Challenges, Limitations and Best Practices

Forecasting is limited by structural breaks, new products, discontinued items, missing data, behavioural changes, supply-induced sales distortion and strategic manipulation. No accuracy metric captures every operational consequence. A small error on a critical spare part may matter more than a large error on a low-value item.

Best practices include:

• define the decision, horizon, granularity and demand measure first;
• establish a transparent statistical benchmark;
• correct for stockouts, promotions and calendar effects;
• use rolling validation and compare against naïve methods;
• report both error and bias;
• distinguish forecasts, targets and scenarios;
• reconcile across product and geographic hierarchies;
• document judgemental overrides and their expiry;
• connect forecast uncertainty to inventory, capacity and service decisions;
• use S&OP to resolve trade-offs, not merely to circulate reports; and
• monitor whether forecasting activity creates measurable decision value.

Chapter Summary

Demand forecasting estimates future requirements under uncertainty, while S&OP integrates that estimate with capacity, inventory, supply, finance and strategy. Qualitative methods contribute market and expert knowledge, whereas quantitative methods provide reproducible baselines. Naïve, moving-average, exponential-smoothing, trend-seasonal and causal models each have assumptions and limitations. Forecast errors should be measured with complementary metrics because MAE, RMSE, bias, MAPE and WAPE answer different questions. Intermittent demand, stockouts, promotions and structural change require specialised treatment. Collaborative governance makes assumptions visible and prevents targets from being confused with forecasts. S&OP uses demand and supply reviews, reconciliation and executive decisions to manage gaps. Scenario analysis converts uncertainty into preparedness through triggers and response options. Technology and AI can improve scale and responsiveness, but require validation, accountability, privacy protection and human judgement.

Key Terms

Demand forecast: An estimate of future customer requirements for a defined item, place and period.

Sales: Transactions completed and recorded by the organisation.

Forecast error: Actual demand minus forecast demand.

Bias: Persistent tendency to overforecast or underforecast.

Time series: Observations ordered by time.

Seasonality: A recurring pattern linked to a calendar or fixed cycle.

Trend: Persistent directional movement in a series.

Exponential smoothing: A method that updates forecasts using weighted recent observations.

Causal model: A model relating demand to explanatory drivers.

MAE: Mean absolute error, expressed in demand units.

RMSE: Root mean squared error, which gives greater weight to large errors.

MAPE: Mean absolute percentage error, unsuitable when actual demand is zero or near zero.

WAPE: Volume-weighted absolute percentage error.

Intermittent demand: Demand containing many zero periods and occasional requirements.

Unconstrained demand: Expected demand before considering supply limitations.

S&OP: A cross-functional process aligning demand, supply and strategy.

Scenario analysis: Evaluation of coherent alternative futures and responses.

Forecast value added: Assessment of whether a process step improves the forecast.

Demand sensing: Use of recent operational and market signals to update expectations.

Forecast hierarchy: Nested levels such as total, region, product family and item.

Review Questions

Short Answer Questions

1. Distinguish demand, sales and orders.
2. State two assumptions of simple exponential smoothing.
3. Why can MAPE be problematic for intermittent demand?
4. What is the purpose of an unconstrained demand plan?
5. Name four stages in a typical S&OP cycle.

Long Answer Questions

1. Compare qualitative forecasting with time-series and causal forecasting, including appropriate applications and limitations.
2. Explain how bias, MAE and RMSE provide different information for operations decisions.
3. Describe how a firm should design governance for judgemental forecast overrides.
4. Explain how S&OP balances capacity, inventory, service and financial objectives.
5. Discuss the benefits and risks of using AI in demand forecasting.

Analytical / Application Questions

1. A product has monthly demand of 90, 110 and 100 units. Using a three-period moving average, calculate the forecast for the next month and explain one limitation of the method.
2. A forecast system has persistent positive errors before a seasonal peak. Diagnose two possible causes and recommend a validation approach.
3. A supplier disruption threatens a high-demand scenario. Design a base, disruption and recovery scenario, including trigger indicators and at least three response options.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a forecast from a target?

A. A forecast expresses expected outcomes, while a target expresses an intended outcome.

B. A forecast is always lower than a target.

C. A target is calculated only from historical data.

D. A forecast is a financial budget and a target is a sales transaction.

Correct Answer: A. A forecast expresses expected outcomes, while a target expresses an intended outcome.

Question 2

If demand in the last three periods was 90, 110 and 100 units, what is the three-period moving-average forecast for the next period?

A. 90 units

B. 100 units

C. 105 units

D. 110 units

Correct Answer: B. 100 units

Question 3

In simple exponential smoothing, what is the effect of increasing α while holding other information constant?

A. The forecast gives more weight to the most recent actual observation.

B. The forecast becomes independent of recent demand.

C. Seasonality is automatically removed.

D. The forecast must equal the long-run average.

Correct Answer: A. The forecast gives more weight to the most recent actual observation.

Question 4

Which measure is specifically designed to reveal systematic overforecasting or underforecasting?

A. Mean absolute error

B. Mean error

C. Root mean squared error

D. Mean absolute percentage error

Correct Answer: B. Mean error

Question 5

Why is MAPE unsuitable when actual demand is zero?

A. It squares the error.

B. It requires a causal variable.

C. It divides by actual demand.

D. It can only be calculated for annual data.

Correct Answer: C. It divides by actual demand.

Question 6

What is the principal purpose of an S&OP process?

A. To replace all operational schedules with a single annual budget.

B. To align demand, supply, resources and strategy through cross-functional decisions.

C. To maximise production regardless of demand.

D. To allow sales targets to override capacity constraints.

Correct Answer: B. To align demand, supply, resources and strategy through cross-functional decisions.

Question 7

Which situation most strongly supports using a causal forecasting model?

A. Demand is completely random and no explanatory information exists.

B. A reliable future driver, such as planned price or advertising, materially affects demand.

C. All observations are zero.

D. The only available datum is the latest demand observation.

Correct Answer: B. A reliable future driver, such as planned price or advertising, materially affects demand.

Question 8

What is the key difference between unconstrained demand and constrained supply planning?

A. Unconstrained demand shows expected market requirements; constrained supply shows what can be delivered.

B. Unconstrained demand is always a sales target; constrained supply is always a forecast.

C. Unconstrained demand excludes customers; constrained supply excludes capacity.

D. There is no practical difference.

Correct Answer: A. Unconstrained demand shows expected market requirements; constrained supply shows what can be delivered.

Question 9

Which practice best supports accountable judgemental overrides?

A. Allowing undocumented changes by senior managers.

B. Recording the reason, owner, expected effect and expiry date of each override.

C. Replacing the statistical baseline permanently after every override.

D. Measuring only whether the override increased the sales target.

Correct Answer: B. Recording the reason, owner, expected effect and expiry date of each override.

Question 10

A high RMSE relative to MAE most plausibly indicates what?

A. Errors are all exactly equal.

B. The forecast has no errors.

C. Some errors are relatively large and are being penalised strongly by RMSE.

D. Actual demand contains no variation.

Correct Answer: C. Some errors are relatively large and are being penalised strongly by RMSE.