Demand Management and Forecasting

Introduction

Demand management is the coordinated process of understanding, shaping, prioritising and responding to customer demand so that organisational resources are used effectively. Forecasting is a central analytical activity within this process, but it is not the same as demand management. A forecast estimates what may happen; demand management determines how the organisation will interpret, influence and fulfil that demand.

In supply chains, demand is often uncertain, intermittent, seasonal, price-sensitive and influenced by competitors, promotions, weather, regulation and macroeconomic conditions. Forecast errors can create excess inventory, stockouts, unstable production schedules, emergency transportation and poor customer service. Conversely, a well-designed demand-management process links market intelligence with operations, procurement, finance, sales and executive decision-making.

Forecasting should therefore be treated as a managerial system rather than merely a statistical exercise. The most sophisticated model cannot compensate for poor data, unrecorded promotions, structural market change or dysfunctional incentives. The objective is not to eliminate uncertainty—an impossible task—but to measure it, communicate it and make decisions that are robust to it.

Learning Objectives

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

1. Explain the role of demand management in supply-chain strategy.
2. Distinguish independent, dependent, derived, latent and observed demand.
3. Select and apply qualitative, time-series and causal forecasting methods.
4. Calculate and interpret forecast accuracy, bias and uncertainty measures.
5. Explain demand sensing, collaborative forecasting and Sales and Operations Planning (S&OP).
6. Evaluate the assumptions, limitations, governance risks and managerial implications of forecasting systems.

1. Demand Management in the Supply Chain

1.1 Demand, sales and consumption

Sales history is not always equivalent to true customer demand. If a product was unavailable, recorded sales may understate demand. If a retailer purchased inventory in anticipation of a promotion, shipment data may overstate immediate consumer consumption. Important distinctions include:

• Observed sales: transactions recorded by the organisation.
• True demand: quantity customers would have purchased under the prevailing offer and availability conditions.
• Consumption: the rate at which the end customer uses a product.
• Shipments: quantities dispatched between supply-chain members.
• Orders: quantities requested, which may be cancelled, delayed or inflated.

A demand-management team should identify which demand signal is appropriate for each decision. Production planning may use unconstrained demand, whereas replenishment may require point-of-sale consumption and current inventory. Forecasting a distorted signal can institutionalise past stockouts or channel-order volatility.

1.2 Types of demand

1.3 The demand-management cycle

A mature process normally includes demand sensing, data preparation, statistical forecasting, review of causal assumptions, collaborative reconciliation, scenario analysis, consensus approval and performance learning. The process should be continuous rather than a once-per-month spreadsheet exercise.

Figure: An integrated demand-management cycle.

2. Forecasting Foundations

A forecast is a statement about a future value or distribution of values, conditional on information available at the time it is produced. A point forecast gives one expected value; a probabilistic forecast gives a range or probability distribution. Forecasts should always be associated with a horizon, aggregation level, data version and forecast origin.

The forecast horizon is the period into the future covered by the forecast. The forecast origin is the date on which it was generated. A forecast made twelve months ahead usually has greater uncertainty than one made one week ahead. Forecasts at an aggregate product-family level are often more accurate proportionally than forecasts for individual stock-keeping units (SKUs), although decisions may require both levels.

A useful decomposition of a demand series is:

where Y_t is observed demand at time t, T_t is trend, S_t is seasonal variation, C_t is cyclical or business-cycle variation, and I_t is irregular variation. This additive form assumes that component effects are measured in the same units and combine additively. If seasonal variation grows with the level of demand, a multiplicative representation is more appropriate:

Multiplicative models require non-negative observations and are unsuitable without modification when observations are zero or negative. Logarithmic transformations can sometimes convert multiplicative relationships into additive ones, but transformation changes the interpretation of errors and requires careful back-transformation.

3. Qualitative Forecasting Methods

Qualitative methods are valuable when historical data are limited, markets are changing, products are new or the forecast depends on knowledge not yet encoded in databases. They are not inherently inferior to quantitative methods; however, they are vulnerable to anchoring, overconfidence, groupthink, political influence and selective use of evidence.

3.1 Executive opinion and expert judgement

A panel of experienced managers can assess market changes, competitor behaviour and strategic events. A structured process should record assumptions, separate independent estimates before discussion and document reasons for adjustments. Unstructured consensus can allow seniority to substitute for evidence.

3.2 Sales-force composite

Sales representatives estimate demand in their territories, often using customer conversations and pipeline information. The method can identify local intelligence but may be biased by incentives, optimism, sandbagging or insufficient visibility beyond the immediate customer relationship. Calibration against historical accuracy is essential.

3.3 Market research and customer surveys

Surveys, interviews, conjoint analysis, test marketing and social-listening data may reveal preferences and latent demand. Stated intention is not identical to purchase behaviour, so research should be triangulated with revealed behaviour and controlled experiments.

3.4 Delphi method

The Delphi method uses several confidential rounds of expert estimates. A facilitator aggregates responses and provides feedback without identifying individuals. Repeated rounds can reduce dominance effects, but convergence does not guarantee correctness; experts may share the same unexamined assumptions.

3.5 Scenario planning

Scenario planning is particularly useful under structural uncertainty. Instead of claiming to predict one future, managers construct coherent alternatives—for example, high inflation, supply disruption or rapid technology adoption—and test capacity, inventory and sourcing decisions against them.

4. Time-Series Forecasting

Time-series methods use patterns in the historical series itself. They are most defensible when the future is expected to resemble the past after accounting for level, trend and seasonality.

4.1 Naïve and seasonal-naïve methods

The naïve forecast uses the most recent observation:

where F_t+1 is the forecast for the next period and Y_t is the latest actual demand. It is a strong benchmark for series with random-walk behaviour.

A seasonal-naïve method uses demand from the corresponding prior season:

where m is the seasonal period and h is the forecast horizon. For monthly data with annual seasonality, m = 12. These methods are simple, transparent and difficult to beat consistently without genuine information.

4.2 Moving averages

A simple moving average over k periods is:

All observations receive equal weight. A short window responds quickly but is more volatile; a long window is smoother but slower to recognise change. The method assumes local level stability and does not explicitly model trend or seasonality.

A weighted moving average is:

Weights must be non-negative and sum to one if the forecast is intended to be a convex weighted average. Greater weight on recent observations increases responsiveness.

4.3 Exponential smoothing

Simple exponential smoothing updates the forecast using the latest error:

where 0 < α ≤ 1 is the smoothing parameter. A high α reacts strongly to recent changes; a low value produces greater smoothing. The method is appropriate for a series with a changing level but no systematic trend or seasonality.

For data with trend, Holt’s method maintains level and trend components. One common form is:

Here ℓ_t is the estimated level, b_t is the estimated trend, β is the trend-smoothing parameter with 0 < β ≤ 1, and h is the number of periods ahead. Forecasts can become unrealistic if a linear trend is extrapolated indefinitely.

Seasonal exponential-smoothing models add a seasonal component. Additive seasonality is appropriate when seasonal amplitude is approximately constant; multiplicative seasonality is more suitable when amplitude changes with the series level. Model selection should be based on out-of-sample performance and operational plausibility, not merely in-sample fit.

4.4 Intermittent demand

For spare parts and slow-moving items, zero-demand periods make ordinary averages and percentage errors problematic. Croston-type methods separately estimate demand size and the interval between non-zero demands. Variants such as SBA and TSB address some bias and obsolescence issues, but no method is universally best. Inventory policy should also consider lead time, criticality, repairability and the cost of stockout.

5. Causal and Predictive Forecasting

Causal forecasting explains demand using drivers such as price, promotions, advertising, income, weather, calendar effects, competitor actions and installed equipment. A simple linear regression is:

Y_t is demand, X_jt are explanatory variables, β₀ is the intercept, β_j are coefficients, and ε_t is the unexplained error. The model assumes, among other things, an adequately specified relationship, meaningful measurement of predictors, stable parameters over the relevant horizon and errors whose dependence structure is appropriately handled.

A coefficient is not automatically causal. If price falls during a promotion targeted at weak-demand periods, a naïve regression may infer the wrong price effect. Endogeneity, omitted variables, multicollinearity, reverse causality and structural breaks must be considered. Experiments, natural experiments, instrumental-variable methods or difference-in-differences designs may be needed when causal claims matter.

For a price model, own-price elasticity is often defined as:

where Q is quantity and P is price. It is dimensionless. A negative value commonly indicates that quantity decreases as price rises, but the magnitude and interpretation depend on market conditions and the estimation design.

Machine-learning models can capture nonlinearities and interactions using trees, boosting, neural networks or other algorithms. They may improve predictive accuracy when large, high-quality datasets exist, but they can be difficult to explain, unstable under distribution shift and blind to unrecorded commercial events. A forecasting system should compare complex models with transparent benchmarks and use time-based validation rather than random splitting of temporally ordered data.

6. Forecast Accuracy, Bias and Uncertainty

Forecast evaluation must distinguish the forecast error from the forecast itself. Define the error for period t as:

Positive error means actual demand exceeded the forecast; negative error means the forecast was higher than actual demand.

The formulas assume consistent units and a clearly defined evaluation window. MAPE is undefined when Y_t = 0 and can become disproportionately large when actual demand is close to zero. WAPE is often more stable for aggregated demand but may hide poor performance on low-volume items. No single metric should determine model choice.

Bias is different from dispersion. A forecast can have low average signed error while making large positive and negative errors that cancel. A useful tracking signal is:

where cumulative running sum of forecast errors is RSFE = ∑e_t, and MAD is mean absolute deviation, often calculated as ∑\|e_t\| ÷ n. The tracking signal is a diagnostic rather than a universal pass/fail rule; thresholds must reflect the business context and sample size.

Forecast uncertainty should be communicated through prediction intervals, scenarios or quantiles. A point forecast of 10,000 units does not mean that 10,000 is certain. For a normally distributed forecast error with standard deviation σ, an approximate two-sided 95% interval is:

This approximation assumes an appropriate model, sufficiently stable error behaviour and an approximately normal distribution. In practice, empirical, bootstrap or model-specific intervals may be preferable, especially with skewed, intermittent or promotion-driven demand. Prediction intervals concern future observations; confidence intervals concern uncertainty about an estimated parameter and should not be confused.

Evaluation should use rolling-origin or walk-forward validation. At each origin, the model is fitted only with information available then, forecasts are generated for the required horizon, and errors are collected. This prevents look-ahead bias and reflects actual deployment.

7. Demand Sensing and Collaborative Forecasting

7.1 Demand sensing

Demand sensing uses near-real-time signals—point-of-sale data, web searches, orders, inventory positions, weather, traffic, promotions and logistics events—to update a near-term forecast. It is particularly relevant where demand changes faster than the normal planning cycle. Demand sensing does not create information from nothing. It can amplify noisy signals, double-count information already embedded in orders or confuse temporary variation with structural change.

Good practice includes signal-quality tests, explicit event calendars, causal feature governance, latency measurement, human review for exceptions and monitoring of forecast performance by horizon. A demand-sensing model should be judged on incremental value over an established baseline, not on technical novelty.

7.2 Collaborative forecasting

Collaborative forecasting shares information across manufacturers, distributors, retailers and customers. Collaborative Planning, Forecasting and Replenishment (CPFR) typically involves agreement on a process for strategy, demand planning, supply planning, execution and performance review. Collaboration can reduce the bullwhip effect when partners share consumption data and coordinate replenishment rather than independently reacting to distorted orders.

Collaboration requires clear ownership. Partners should define the forecast hierarchy, data standards, exception rules, event assumptions, confidentiality arrangements and responsibility for overrides. A consensus forecast is not necessarily an average of every participant’s number. It should be a documented decision based on evidence and explicit trade-offs.

7.3 S&OP and integrated business planning

S&OP reconciles demand and supply at an aggregate level, generally over a medium-term horizon. Typical stages include data review, demand review, supply review, reconciliation of alternatives and executive decision. The process identifies capacity constraints, inventory risks, financial consequences and policy choices. Its value lies not only in the forecast but in making cross-functional trade-offs visible and actionable.

8. Worked Example: Forecasting and Error Interpretation

A distributor records monthly demand of 100, 120 and 110 units. A three-period moving-average forecast for month 4 is:

F₄ = (100 + 120 + 110) ÷ 3 = 110 units.

Suppose actual demand in month 4 is 130 units. The error is:

e₄ = 130 − 110 = 20 units.

Assume forecasts and actual demand for four months are as follows:

The mean error is ME = (−5 + 5 − 2 + 20) ÷ 4 = 4.5 units, indicating under-forecasting on average. The MAE is MAE = (5 + 5 + 2 + 20) ÷ 4 = 8 units. The RMSE is approximately √((25 + 25 + 4 + 400) ÷ 4) = √113.5 ≈ 10.65 units. The large month-4 error makes RMSE substantially larger than MAE, signalling an unusually large miss. Management should investigate whether month 4 involved a promotion, stock recovery or data problem before changing the entire model.

9. Illustrative Case: Meridian Home Appliances

Background

Meridian Home Appliances sells refrigerators through national retailers and an online channel. Retailer orders are volatile because each retailer uses different replenishment rules. The company’s monthly forecast was produced by sales managers and was frequently adjusted upward before major promotions.

Problem

Meridian experienced simultaneous excess inventory of slow-moving models and stockouts of two popular models. The operations director initially proposed replacing the spreadsheet with an artificial-intelligence forecasting platform. An analysis showed that retailer shipment data included forward buying, while point-of-sale data were available only for 60% of volume. Promotions were recorded inconsistently, and forecast accuracy was measured using MAPE, which produced extreme values for low-volume models.

Analysis

The project team first established a data hierarchy: point-of-sale data where available, otherwise validated retailer shipments adjusted for inventory movements. It created a promotion calendar, separated new-product launches from mature products and classified SKUs by demand pattern. Seasonal-naïve and exponential-smoothing models were used as benchmarks. A causal model was then tested for price, promotion, housing activity and weather. Forecasts were evaluated with rolling-origin MAE, WAPE and signed error by channel and horizon.

The analysis found that the sales-force override improved forecasts during documented retailer promotions but worsened them in ordinary months. Rather than banning judgement, Meridian introduced an override reason code and required evidence for material changes. The company also shared a baseline forecast and promotion assumptions with key retailers in a monthly CPFR process.

Approach and outcome

For the next six months, Meridian used a segmented approach: seasonal-naïve forecasts for stable seasonal models, exponential smoothing for changing-level models, intermittent-demand methods for service parts and causal adjustments for planned promotions. It adopted a consensus meeting within S&OP and tracked both statistical error and business outcomes such as stockouts, obsolete inventory and expedite cost.

The improvement was not attributed to the algorithm alone. Data correction, common definitions, disciplined overrides and retailer collaboration were equally important. The case demonstrates that forecasting maturity is organisational as well as technical.

Learning points

1. Shipment history may be a distorted proxy for consumption.
2. Baselines are necessary for judging whether complexity adds value.
3. Forecast overrides should be transparent, reason-coded and evaluated.
4. Accuracy metrics must suit the demand pattern and decision purpose.
5. Collaboration works when data ownership and assumptions are explicit.

10. Challenges, Governance and Best Practice

Forecasts influence procurement, labour, supplier commitments and customer allocation, so governance is essential. Organisations should maintain version control, document model assumptions, retain forecast snapshots and separate actuals from revised historical data. Access to customer and point-of-sale data must comply with contracts, privacy law and competition requirements. Sharing aggregated demand information can support efficiency, but sharing competitively sensitive information may create legal risk.

AI-based forecasting raises additional issues: explainability, data provenance, model drift, discriminatory allocation outcomes, cybersecurity and human accountability. A model should not silently override commercial policy or allocate scarce products without an approved governance framework. Automated systems need monitoring for drift, missing signals, changed product definitions and anomalous events.

Best practice includes:

• define the decision, horizon, unit and aggregation level before selecting a method;
• establish a transparent benchmark, often naïve or seasonal-naïve;
• cleanse data for stockouts, returns, forward buying, substitutions and promotions;
• use causal knowledge where events are known rather than forcing all variation into a time series;
• validate with time-respecting, out-of-sample procedures;
• report accuracy, bias and uncertainty by segment and horizon;
• distinguish statistical forecasts from managerial overrides;
• measure forecast value through inventory, service and financial outcomes;
• use scenarios when structural uncertainty is high;
• review models after product, process, market or policy changes.

Forecasting is ultimately a decision-support capability. An accurate forecast that arrives too late, is not trusted or cannot be converted into supply action has limited value. Conversely, a transparent forecast with acknowledged uncertainty can support better decisions even when the market remains unpredictable.

Chapter Summary

Demand management connects market demand with supply-chain decisions. It requires clear distinctions among consumption, orders, shipments, sales and constrained demand. Qualitative methods contribute expert and market knowledge, especially for new products and structural change. Time-series methods use historical level, trend and seasonality, while causal methods incorporate drivers such as price, promotion and external conditions.

Forecast evaluation should use signed error for bias and absolute or squared measures for magnitude. MAPE has important limitations for zero and low demand. Prediction intervals and scenarios communicate uncertainty more responsibly than point forecasts alone. Rolling-origin validation protects against look-ahead bias.

Demand sensing can improve near-term responsiveness when high-frequency signals are reliable, but it can also amplify noise. Collaborative forecasting, CPFR and S&OP align partners and functions around shared assumptions and trade-offs. Effective forecasting depends as much on data governance, process discipline and managerial accountability as on statistical technique or artificial intelligence.

Key Terms

Demand management: The coordinated process of sensing, shaping, planning and responding to demand.

Forecast horizon: The future period covered by a forecast.

Forecast origin: The date or information set available when a forecast is generated.

Independent demand: Demand determined primarily by external customers.

Dependent demand: Demand calculated from demand for another item or product.

Intermittent demand: Demand containing many zero-demand periods.

Naïve forecast: A forecast based on the most recent observation.

Seasonal-naïve forecast: A forecast using demand from the corresponding prior season.

Moving average: An average of observations in a rolling historical window.

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

Trend: Persistent upward or downward movement in a demand series.

Seasonality: A systematic pattern repeating at a known calendar frequency.

Causal forecasting: Forecasting based on explanatory demand drivers.

Forecast error: The difference between actual demand and forecast demand.

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

MAE: Mean absolute error, measuring average error magnitude in original units.

WAPE: Weighted absolute percentage error using total absolute error relative to total demand.

Prediction interval: A range intended to contain a future observation with a stated probability.

Demand sensing: Use of recent, high-frequency signals to update near-term demand estimates.

CPFR: Collaborative Planning, Forecasting and Replenishment among supply-chain partners.

Review Questions

Short Answer Questions

1. Why are shipments not always an accurate measure of true customer demand?
2. What is the difference between a forecast origin and a forecast horizon?
3. State one strength and one limitation of qualitative forecasting.
4. Why can MAPE be unsuitable for intermittent demand?
5. What is the purpose of a forecast baseline?

Long Answer Questions

1. Compare naïve, moving-average, exponential-smoothing and causal forecasting methods. Explain when each is appropriate.
2. Explain how forecast accuracy, bias and uncertainty differ, and recommend measures for a multi-SKU supply chain.
3. Discuss the assumptions and limitations of regression-based demand forecasting.
4. Explain how S&OP and CPFR can improve demand-supply alignment.
5. Evaluate the benefits and risks of using AI-based demand sensing in a consumer-goods supply chain.

Analytical / Application Questions

1. A product has actual demand of 80, 100 and 120 units in three consecutive periods. Calculate the three-period moving-average forecast for period 4 and explain one limitation of this forecast.
2. A forecast series has errors of −10, 5, 15 and −10 units. Calculate ME and MAE, and interpret whether the results show bias.
3. Design a forecasting and collaboration process for a retailer whose order history is volatile but whose point-of-sale data are available weekly. Specify data, models, metrics, governance and S&OP actions.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes true demand from observed sales?

A. True demand is always lower than observed sales

B. Observed sales may understate true demand when stockouts occur

C. Observed sales automatically include lost sales

D. True demand refers only to shipped quantities

Correct Answer: B. Observed sales may understate true demand when stockouts occur

Question 2

For a stable series without trend or seasonality, which method provides a strong transparent benchmark?

A. Naïve forecasting

B. Delphi forecasting

C. Multiple regression only

D. Scenario planning only

Correct Answer: A. Naïve forecasting

Question 3

In simple exponential smoothing, increasing α generally makes the forecast:

A. Less responsive to recent observations

B. More responsive to recent observations

C. Automatically seasonal

D. Independent of the latest actual demand

Correct Answer: B. More responsive to recent observations

Question 4

If actual demand is 150 units and the forecast is 130 units, using e_t = Y_t − F_t, the forecast error is:

A. −20 units

B. 20 units

C. 280 units

D. 0.87 units

Correct Answer: B. 20 units

Question 5

Which metric is especially problematic when actual demand can be zero?

A. MAE

B. RMSE

C. MAPE

D. ME

Correct Answer: C. MAPE

Question 6

A positive mean forecast error over many periods usually indicates:

A. Systematic over-forecasting

B. Systematic under-forecasting

C. Absence of uncertainty

D. Perfect calibration

Correct Answer: B. Systematic under-forecasting

Question 7

What is the main purpose of rolling-origin validation?

A. To use future data when fitting historical forecasts

B. To mimic deployment while avoiding look-ahead bias

C. To guarantee zero forecast error

D. To remove all seasonal patterns

Correct Answer: B. To mimic deployment while avoiding look-ahead bias

Question 8

Which situation most strongly supports using a causal forecasting model?

A. Demand is driven by documented price and promotion changes

B. There is no historical or explanatory data

C. All observations are random and unrelated to external factors

D. The forecast horizon is undefined

Correct Answer: A. Demand is driven by documented price and promotion changes

Question 9

Demand sensing primarily refers to:

A. Replacing all forecasts with executive opinion

B. Using recent high-frequency signals to update near-term demand estimates

C. Ignoring point-of-sale information

D. Forecasting only annual demand

Correct Answer: B. Using recent high-frequency signals to update near-term demand estimates

Question 10

Which practice best supports responsible managerial overrides?

A. Allowing undocumented changes by senior managers

B. Prohibiting every adjustment regardless of context

C. Recording the reason and evidence for each material override

D. Replacing statistical forecasts with sales targets

Correct Answer: C. Recording the reason and evidence for each material override