Demand Analysis and Inventory Forecasting

Introduction

Inventory decisions are fundamentally decisions under uncertainty. An organisation must decide how much stock to hold, when to replenish it, and how to allocate scarce materials among competing users. These decisions depend on expectations about future demand. A forecast is therefore not merely a prediction produced by a statistical technique; it is an input to purchasing, production, capacity planning, warehouse design, working-capital management and customer-service policy.

Demand analysis begins by identifying what creates the requirement for an item. Some demand originates outside the organisation, such as customer purchases. Other demand is generated internally by the requirement for components, subassemblies, packaging or maintenance materials. The distinction between independent and dependent demand is central because the two categories require different planning systems. Independent demand is usually forecast from historical and market information, whereas dependent demand is calculated from production schedules, bills of material and usage relationships.

Forecasts are always imperfect. The professional task is not to eliminate error, which is impossible, but to measure, interpret and manage it. A sound forecasting system combines appropriate data, a suitable method, managerial judgement, collaboration with supply-chain partners and continuous evaluation.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Distinguish independent demand from dependent demand and select appropriate planning approaches.
2. Identify level, trend, seasonality, cycles, irregular variation and structural changes in demand data.
3. Explain forecast error, bias, accuracy and uncertainty.
4. Apply moving-average and exponential-smoothing methods.
5. Analyse seasonal demand and understand causal forecasting.
6. Explain collaborative forecasting and its governance requirements.
7. Interpret forecast results for inventory and managerial decisions.
8. Evaluate the limitations, ethical implications and technological developments associated with forecasting.

1. Demand as the Basis of Inventory Control

Demand is the quantity of an item that customers or internal users require during a specified period. It should be distinguished from sales. Sales may be lower than true demand when stockouts, credit restrictions, poor availability or lost orders prevent customers from purchasing the quantity they wanted. If observed sales are used without adjustment, a forecasting system may systematically underestimate future demand.

Demand data should be described with four attributes:

• Item definition: the stock-keeping unit, material code or product family being analysed.
• Time interval: day, week, month or another consistent period.
• Location and channel: plant, warehouse, region, customer segment or sales channel.
• Demand event: shipment, order, consumption, withdrawal or estimated lost demand.

The choice of aggregation is important. Forecasting at a product-family level may be more stable than forecasting every individual stock-keeping unit, but it may conceal different replenishment requirements. Conversely, highly disaggregated forecasts can produce excessive noise and administrative effort. A useful principle is to forecast at the level at which a decision is made, while maintaining reconciliation with higher-level totals.

Independent and Dependent Demand

Independent demand is demand for an item that is determined primarily by external customers or external market conditions. Examples include finished appliances, replacement parts sold directly to customers and standard office supplies. It is uncertain and commonly estimated using time-series, causal or judgemental methods.

Dependent demand is demand for an item that is derived from the demand for another item. A bicycle’s requirement for tyres, tubes and brake assemblies depends on the number and configuration of bicycles planned for production. Dependent demand is often calculated rather than forecast.

If a production plan requires 500 finished units and each unit uses 2 bearings, gross bearing demand is:

where D_b is gross bearing demand in units, Q_p is the planned production quantity in finished units, and q_b is bearing usage per finished unit. If 80 usable bearings are already available and 20 are expected to be received before production, net additional demand is:

where N_b is net requirement, I_b is usable inventory on hand and S_b is scheduled receipt available in the relevant period. The calculation assumes that the bill of material is accurate, production yield is known, inventory is usable and timing is aligned.

The distinction is not absolute. A spare part may have independent demand when sold for after-sales service, but dependent demand when installed during planned maintenance. The same material can therefore require different planning rules in different contexts.

Demand, Usage and Inventory Transactions

A forecast must be based on a clearly defined demand signal. Shipment history may be appropriate for distribution planning, while production consumption may be more appropriate for a raw-material forecast. Returns, cancellations, transfers, promotional orders and emergency purchases should be identified rather than blindly treated as ordinary demand.

Data cleansing is therefore a managerial activity, not merely an information-technology task. Relevant adjustments include correcting duplicate transactions, separating stockout periods, identifying one-time project demand and recording product substitutions. A demand history that is mathematically clean but operationally misinterpreted can produce an inaccurate forecast with an appearance of precision.

2. Patterns and Variability in Demand

Demand observations can be represented as D_t, the actual demand in period t. Common components are:

• Level: the underlying average magnitude of demand.
• Trend: a persistent upward or downward movement.
• Seasonality: a recurring pattern associated with calendar or operational cycles.
• Cycle: longer-term fluctuation, often related to economic or industry conditions.
• Irregular variation: random or exceptional movement not explained by the regular pattern.
• Structural change: a lasting shift caused by a new technology, competitor, regulation, contract or business model.

A simple additive representation is:

where L_t is the level, T_t is the trend component, S_t is the seasonal component and ε_t is irregular error. This form is appropriate when seasonal fluctuations are approximately constant in absolute size. A multiplicative representation is:

which is more appropriate when seasonal variation grows in proportion to the level. In practice, decomposition is an approximation; the components may interact and may not be uniquely identifiable.

Variability and Intermittency

Variability is the extent to which demand differs across periods. High variability increases the difficulty of forecasting and generally increases the inventory required to achieve a given service level. Intermittent demand contains many zero-demand periods and occasional non-zero requirements. It is common in maintenance spares and low-volume components.

The coefficient of variation is a dimensionless measure of relative dispersion:

where σ_D is the standard deviation of demand and μ_D is mean demand. It is useful for comparing items with different scales, although it becomes unstable when the mean is close to zero. For intermittent demand, the average interval between non-zero demands and the size of those demands should also be examined.

Demand variability should not automatically be interpreted as random noise. It may reflect changing prices, promotions, capacity constraints, weather, competitor actions or an unstable ordering policy. Diagnosis should precede method selection.

3. Forecast Error and Accuracy

Let A_t denote actual demand and F_t the forecast made for period t. Forecast error is conventionally defined as:

A positive error means actual demand exceeded the forecast; a negative error means the forecast exceeded actual demand. Some organisations use the reverse sign, so the sign convention must always be stated.

Measures of Error

For n periods, mean error or mean forecast error is:

ME indicates bias. A positive value under the convention above indicates systematic underforecasting.

Mean absolute deviation is:

MAD is expressed in demand units and is easy to interpret. Mean squared error is:

MSE gives greater weight to large errors and is expressed in squared demand units. Root mean squared error is:

RMSE returns the measure to the original demand units.

Mean absolute percentage error is:

MAPE is interpretable as a percentage but is undefined when actual demand is zero and can become excessively large when actual demand is close to zero. Weighted absolute percentage error or scale-dependent measures are often preferable for intermittent demand. No single accuracy measure is universally best.

A tracking signal can be used to identify persistent bias:

where cumulative running sum of forecast errors is:

The tracking signal is a diagnostic, not a universal pass–fail rule. Control limits should reflect the process and the consequences of false alarms.

Worked Example: Comparing Two Forecasts

Suppose actual demand for four months is 100, 120, 80 and 100 units. Forecast A is 90, 110, 90 and 100; Forecast B is 105, 115, 85 and 95.

For Forecast A, errors are 10, 10, −10 and 0. Therefore:

• ME = (10 + 10 − 10 + 0) ÷ 4 = 2.5 units.
• MAD = (10 + 10 + 10 + 0) ÷ 4 = 7.5 units.
• MSE = (100 + 100 + 100 + 0) ÷ 4 = 75 units².
• RMSE = √75 ≈ 8.66 units.

For Forecast B, errors are −5, 5, −5 and 5. Therefore:

• ME = 0 units.
• MAD = 5 units.
• MSE = 25 units².
• RMSE = 5 units.

Forecast B is better on all these measures and has no net bias in the four observations. Nevertheless, zero bias does not imply that every period was forecast accurately, nor does historical superiority guarantee future superiority.

Forecast accuracy should be evaluated using a holdout period or rolling-origin testing where possible. Evaluating a method on the same observations used to fit it can produce an overly optimistic assessment.

4. Time-Series Forecasting Methods

Naïve Forecast

The naïve method uses the most recent actual demand as the next forecast:

It is appropriate as a benchmark and can perform well when demand follows a random-walk-like process. More complex methods should demonstrate improvement over this baseline.

Simple Moving Average

A moving average of the most recent k observations is:

The method smooths random fluctuations. A larger k produces greater smoothing but responds more slowly to changes; a smaller k responds faster but is more sensitive to noise. The method assumes that recent observations are relevant and that no strong trend or seasonal pattern requires explicit treatment.

A weighted moving average assigns weights w_i to observations:

with the condition:

Weights are normally non-negative, and larger weights are commonly assigned to more recent observations. The units remain those of demand because the weights are dimensionless.

Example. If demand in the last three months was 100, 120 and 110 units, the three-month moving-average forecast for the next month is:

F_t+1 = (100 + 120 + 110) ÷ 3 = 110 units.

Simple Exponential Smoothing

Simple exponential smoothing updates the forecast as:

where 0 < α ≤ 1 is the smoothing constant. The actual demand and prior forecast must have the same units. A high α places more emphasis on recent demand and responds rapidly; a low α produces more smoothing.

An equivalent error-correction form is:

The method is appropriate for level demand without a systematic trend or seasonality. It is not a complete trend model merely because it reacts to changes.

Example. If F_t = 110, A_t = 120 and α = 0.30, then:

F_t+1 = 110 + 0.30(120 − 110) = 113 units.

The initial forecast affects early results. It should be selected transparently, for example by using the first observation, an initial average or an optimisation procedure evaluated through time-based validation.

Trend-Adjusted Exponential Smoothing

When demand has a persistent trend, Holt’s method can be used. One common formulation is:

Here ℓ_t is the estimated level, b_t is the estimated trend, α and β are smoothing constants between 0 and 1, and m is the number of periods ahead. The method assumes that the recent trend is sufficiently stable to extrapolate. Dampened-trend variants reduce the risk of projecting an unrealistic trend indefinitely.

5. Seasonality and Causal Forecasting

Seasonality is a repeating pattern with a known or discoverable period, such as monthly demand for air conditioners or weekly demand for food-service packaging. Seasonal analysis requires enough historical cycles to distinguish a recurring pattern from an isolated event.

A seasonal index expresses demand in a period relative to the average level. Under a multiplicative model, an index greater than 1 indicates above-average demand. If four quarterly indices are used, their average should normally equal 1, so their sum is 4. Forecasts are generated by combining a deseasonalised level or trend with the relevant seasonal index:

The multiplicative form assumes that seasonal amplitude increases with the level. An additive form is:

which assumes a roughly constant seasonal amount.

Illustrative calculation. Suppose the deseasonalised monthly forecast is 2,000 units and the index for December is 1.25. The multiplicative forecast is 2,500 units. If the seasonal effect were additive and equal to 500 units, the forecast would also be 2,500 units, but the two models would diverge at other base levels. The model should be selected from data and operational understanding, not convenience.

Causal forecasting explains demand using predictor variables such as price, advertising expenditure, temperature, economic activity, installed equipment base or customer orders. A simple linear model is:

where D_t is demand, X_t is a causal variable, a is the intercept, b is the expected change in demand for a one-unit change in X_t, and ε_t is unexplained error. Multiple regression may include several predictors, but causal interpretation requires more than a high statistical association. Timing, omitted variables, reverse causality and multicollinearity must be considered.

A causal model is useful only if future values of the predictors are themselves available or forecastable. For example, a demand model depending on future price requires a credible pricing plan. Managers should distinguish correlation from causation and test models outside the estimation sample.

6. Collaborative Forecasting

Forecast quality often improves when relevant supply-chain participants share information. Collaborative planning, forecasting and replenishment (CPFR) typically involves agreement on data, joint identification of events, preparation of forecasts, reconciliation of differences and review of performance.

Important information may include point-of-sale data, inventory at customer locations, promotion calendars, new-product introductions, planned plant shutdowns, customer project schedules and supplier capacity constraints. Sharing only historical orders can perpetuate the bullwhip effect, in which small changes in final demand produce increasingly large fluctuations upstream.

A robust collaborative process includes:

1. Common definitions: agreement on item codes, units, time buckets, demand ownership and forecast horizons.
2. Data governance: validation, access controls, version history and rules for correcting data.
3. Baseline forecast: an independently generated statistical forecast that provides an objective reference.
4. Event review: documented adjustments for promotions, contracts, shutdowns or exceptional projects.
5. Consensus process: reconciliation of differences using evidence rather than hierarchy alone.
6. Accountability: identification of who owns the forecast and who approves overrides.
7. Performance review: measurement of bias, accuracy, service and inventory consequences.

Collaboration does not mean averaging every participant’s opinion. It means integrating information unavailable to a single organisation while controlling incentives and judgemental bias. A sales team may overstate demand to protect availability or achieve targets; a purchasing department may understate it to limit commitments. Transparent override logs and post-event review help distinguish useful knowledge from motivated bias.

7. Forecasting and Inventory Decisions

The forecast is not itself the inventory policy. Inventory decisions also depend on lead time, service objectives, order constraints, supplier reliability, review frequency and the cost of shortage and surplus.

If demand during lead time is approximately normally distributed and lead-time demand has mean μ_L and standard deviation σ_L, a simplified safety-stock expression is:

where SS is safety stock and z is the standard-normal factor associated with the desired cycle-service probability. The reorder point is:

These formulas assume that the selected probability model is reasonable, demand and lead-time uncertainty have been represented correctly, and the service measure is defined consistently.

If period demand has standard deviation σ_D, average lead time is L periods, and lead time is constant, then:

If demand rate d and lead time L are both variable and independent, a commonly used approximation is:

where σ_Lt is the standard deviation of lead time. If demand and lead time are correlated, the covariance term should not be ignored. These relationships are approximations and should not be applied mechanically to intermittent or highly non-normal demand.

Forecast uncertainty should also be separated from process variation where possible. A model that appears accurate on average may still produce inadequate protection for a critical item if its errors are biased, clustered or concentrated during peak seasons.

8. Illustrative Case: Forecasting a Seasonal Packaging Material

Background

A fictional beverage producer purchases printed cartons for a flagship product. Carton demand is independent at the finished-product level but dependent on the production plan once the monthly production schedule is approved. Sales increase substantially during the summer, and the supplier has a six-week lead time.

Problem

The procurement team used a six-month moving average. This reduced random fluctuations but consistently underforecast May–August demand. The result was expedited purchases, higher freight cost and occasional production rescheduling. The team proposed increasing safety stock without examining the forecasting method.

Analysis

The demand history showed a recurring seasonal pattern rather than merely random variability. May–August demand was approximately 30–45% above the annual monthly average. The six-month moving average lagged the seasonal rise because it included several low-demand winter months. Forecast errors were positive during the peak and negative after the peak, indicating timing bias rather than simple random error.

The team compared three approaches: a three-month moving average, simple exponential smoothing and a seasonal index applied to a level forecast. A rolling-origin evaluation over two previous years used MAD and bias. The seasonal method produced the lowest average absolute error and substantially reduced peak-season bias. The team also corrected historical shipments for one exceptional customer promotion and separately recorded the promotion calendar for future planning.

Approach and Solution

The agreed process was:

• The statistical baseline used the seasonal model.
• Sales provided documented promotion and customer-event information.
• Production converted the finished-goods forecast into carton requirements using the approved bill of material.
• Procurement reviewed supplier capacity and the six-week lead-time assumption.
• Safety stock was recalculated using lead-time demand variability rather than an arbitrary percentage of annual usage.
• Forecast overrides required a reason code and were reviewed after the season.

Outcome

In the following season, forecast bias during the peak months declined and emergency freight was reduced. Inventory increased modestly before the season but excess cartons after the season were lower because the forecast also captured the seasonal decline. The case demonstrates that improving the forecast can be more effective than simply increasing inventory.

Learning Points

1. A method that smooths noise may still be unsuitable for seasonal demand.
2. Forecast and dependent-demand calculations should be connected but not confused.
3. Safety stock should reflect uncertainty and service objectives, not compensate blindly for systematic bias.
4. Collaboration is most valuable when participants contribute information that is both timely and verifiable.

9. Limitations, Governance and Contemporary Developments

Forecasting models can fail because of regime changes, new products, discontinued items, price shocks, pandemics, regulatory changes or supplier disruption. Historical data may encode discrimination or unequal service patterns if allocation decisions have historically favoured particular customers. Governance should therefore address not only accuracy but also fairness, traceability and business consequences.

Artificial intelligence and machine-learning methods can process high-dimensional data, detect nonlinear relationships and automate forecasts across thousands of items. However, they require reliable data, appropriate validation and operational explainability. A more complex model is not automatically better. For low-volume intermittent items, a simple method with expert review may be more robust than a data-intensive algorithm. Forecasts generated by AI should be monitored for drift, leakage, unstable features and unexplained overrides.

Best practice includes:

• define the demand signal and forecast purpose before choosing a method;
• maintain a naïve benchmark;
• validate using time-ordered data rather than random splits when chronology matters;
• measure both accuracy and bias;
• evaluate consequences for service, working capital, waste and obsolescence;
• segment items by demand behaviour and business criticality;
• document assumptions, overrides and data changes;
• review forecasts at a frequency consistent with demand and lead-time dynamics;
• integrate forecasts with material planning, capacity and supplier collaboration;
• use human judgement to add information, not to conceal incentives or accountability.

Chapter Summary

Demand analysis establishes what is being forecast, for whom, at what location and over what time horizon. Independent demand arises from external market requirements and is commonly forecast, whereas dependent demand is derived from parent-item plans and is generally calculated through bills of material and material requirements planning. The same item may have either type of demand in different contexts.

Demand patterns include level, trend, seasonality, cycles, irregular variation and structural change. Forecast error is the difference between actual and forecast demand. ME identifies bias, MAD provides an interpretable absolute measure, MSE and RMSE emphasise large errors, and percentage measures require caution when actual demand is zero or small.

Naïve forecasts provide essential benchmarks. Moving averages smooth observations but can lag turning points. Exponential smoothing updates forecasts according to a smoothing constant, while trend-adjusted and seasonal methods address systematic movement. Causal models use explanatory variables but require careful interpretation and credible future predictor values.

Collaborative forecasting integrates information from supply-chain partners through common definitions, governed data, baseline forecasts, documented events and performance review. Forecasts must be translated into inventory policies that consider lead time, service objectives, variability and cost. Contemporary analytics and AI offer useful capabilities but do not remove the need for data quality, validation, transparency and managerial accountability.

Key Terms

Independent demand: Demand determined primarily by external customers or markets.

Dependent demand: Demand derived from the requirement for another item.

Demand signal: The operational measure used to represent demand, such as orders, shipments or consumption.

Level: The underlying average magnitude of demand.

Trend: A persistent upward or downward movement in demand.

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

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

Forecast: An estimate of future demand made using information available at a specified time.

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

Bias: Persistent tendency to overforecast or underforecast.

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

RMSE: Root mean squared error, the square root of mean squared error.

Moving average: A forecast formed from the average of recent observations.

Exponential smoothing: A recursive method that weights recent observations and prior forecasts.

Seasonal index: A factor or effect representing demand associated with a seasonal period.

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

Collaborative forecasting: Forecasting that integrates information and decisions from supply-chain participants.

Safety stock: Inventory held to protect against uncertainty in demand or replenishment.

Reorder point: The inventory position at which replenishment is triggered.

Forecast governance: The rules for data, assumptions, overrides, accountability and review.

Review Questions

Short Answer Questions

1. Define independent demand and dependent demand, giving one example of each.
2. What is the difference between forecast error and forecast bias?
3. State one advantage and one limitation of a moving-average forecast.
4. Why can MAPE be problematic when actual demand is zero or very small?
5. What is the purpose of a baseline forecast in collaborative forecasting?

Long Answer Questions

1. Explain the major components of a demand pattern and discuss how each affects inventory planning.
2. Compare moving averages, simple exponential smoothing and trend-adjusted exponential smoothing, including their assumptions and limitations.
3. Explain how seasonal indices are developed and distinguish additive from multiplicative seasonality.
4. Discuss the role of causal forecasting in materials management and the dangers of interpreting correlation as causation.
5. Describe a governed collaborative forecasting process and explain how it can reduce supply-chain distortion.

Analytical / Application Questions

1. Actual demand for three periods is 240, 270 and 210 units. Calculate the three-period moving-average forecast for the next period. Explain one reason why this forecast might be unsuitable if the demand is strongly seasonal.
2. A forecast of 500 units is made for a period in which actual demand is 560 units. Calculate the forecast error using e_t = A_t − F_t and interpret its sign. If the item is critical and lead time is long, explain two inventory implications.
3. A manufacturer plans 1,200 finished units, each requiring 4 components. It has 600 usable components on hand and 500 scheduled to arrive before production. Calculate the net component requirement and state two assumptions that must be checked before releasing the purchase order.

Multiple Choice Questions with Answers

Question 1

Which statement best describes dependent demand?

A. Demand generated primarily by external consumer preferences

B. Demand derived from the planned requirement for another item

C. Demand that occurs only during seasonal periods

D. Demand that is always uncertain and intermittent

Correct Answer: B. Demand derived from the planned requirement for another item

Question 2

Using the convention e_t = A_t − F_t, what does a positive forecast error indicate?

A. The forecast exceeded actual demand

B. Actual demand exceeded the forecast

C. Actual demand was zero

D. The forecast was unbiased

Correct Answer: B. Actual demand exceeded the forecast

Question 3

Actual demand in the last three periods was 90, 120 and 150 units. What is the three-period moving-average forecast for the next period?

A. 110 units

B. 120 units

C. 130 units

D. 150 units

Correct Answer: B. 120 units

Question 4

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

A. Respond more quickly to recent observations

B. Ignore the most recent observation

C. Include an explicit seasonal cycle

D. Independent of the previous forecast

Correct Answer: A. Respond more quickly to recent observations

Question 5

Which measure is especially useful for identifying systematic overforecasting or underforecasting?

A. Mean error

B. Mean absolute deviation only

C. Inventory turnover only

D. Order quantity

Correct Answer: A. Mean error

Question 6

A multiplicative seasonal index of 1.20 applied to a base forecast of 2,500 units produces a forecast of:

A. 2,080 units

B. 2,500 units

C. 2,700 units

D. 3,000 units

Correct Answer: D. 3,000 units

Question 7

Which condition most directly supports using a causal forecasting model?

A. A relevant explanatory variable is available or can be forecast

B. All demand observations are zero

C. The item has no identifiable demand drivers

D. Historical data are deliberately excluded

Correct Answer: A. A relevant explanatory variable is available or can be forecast

Question 8

If a production plan requires 800 units and each finished unit uses 3 components, gross component demand is:

A. 267 units

B. 800 units

C. 2,400 units

D. 3,200 units

Correct Answer: C. 2,400 units

Question 9

What is a principal purpose of collaborative forecasting?

A. To replace all statistical methods with opinions

B. To integrate relevant information held by different supply-chain participants

C. To guarantee that forecast error is zero

D. To eliminate the need for data governance

Correct Answer: B. To integrate relevant information held by different supply-chain participants

Question 10

A reorder point under a simplified uncertainty model is commonly expressed as:

A. Average lead-time demand minus safety stock

B. Average lead-time demand plus safety stock

C. Annual demand divided by unit price

D. Forecast error multiplied by order quantity

Correct Answer: B. Average lead-time demand plus safety stock