Demand Forecasting for Inventory Decisions

Introduction

Inventory decisions are decisions about the future. A purchasing officer must decide how much to order before actual customer demand is known; a warehouse manager must set replenishment levels before the next sales period; and a supplier must plan production before receiving complete information from downstream customers. Demand forecasting provides a systematic estimate of that future demand.

A forecast is not a promise and it is not a prediction of exactly what will happen. It is an estimate produced from available information, assumptions and a chosen method. Good forecasting reduces avoidable stockouts, excess inventory, emergency purchases, obsolescence and inefficient use of working capital. It does not eliminate uncertainty. The purpose is to make decisions that are better informed and more transparent.

Forecasting is especially important in procurement and inventory management because replenishment decisions are made before demand is observed. The appropriate method depends on the nature of demand, the time horizon, data quality, product characteristics and the cost of forecast errors.

Learning objectives

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

1. Explain the role and limitations of demand forecasting in inventory decisions.
2. Distinguish qualitative and quantitative forecasting methods.
3. Calculate and interpret moving-average and exponential-smoothing forecasts.
4. Recognise trend and seasonality and use seasonal information appropriately.
5. Measure forecast error using common accuracy indicators.
6. Explain how collaborative demand information improves supply-chain decisions.
7. Select and govern a forecasting process for practical procurement use.

1. Demand, forecasts and the inventory decision

Demand is the quantity customers are willing and able to purchase during a specified period. Observed sales are not always equal to true demand. If an item was out of stock, recorded sales may understate what customers wanted. Promotions, rationing, substitutions, price changes and cancelled orders can also distort the data.

A forecast should therefore identify its:

• Item or product family;
• location or market;
• time bucket, such as day, week or month;
• horizon, such as the next week or next twelve months; and
• purpose, such as purchasing, workforce planning or capacity planning.

Forecasts support several inventory decisions. The forecast during supplier lead time affects the reorder point. The forecast over a review cycle affects the order quantity. Longer-term forecasts influence supplier contracts, capacity and safety-stock policy.

A basic reorder-point relationship is:

Here, ROP is the reorder point in units, μ_L is expected demand during lead time in units, and SS is safety stock in units. This relationship assumes that the item is replenished when its inventory position reaches the reorder point and that lead-time demand and safety stock have been defined consistently. A forecast is used to estimate μ_L; it does not by itself determine safety stock.

Forecasting and inventory control must be connected. A highly accurate forecast can still produce poor service if lead time, order constraints or safety stock are ignored. Conversely, a sound inventory policy may tolerate a moderately inaccurate forecast if uncertainty is explicitly covered.

2. Sources and patterns of demand

Demand patterns commonly contain four elements:

• Level: the typical magnitude of demand.
• Trend: a persistent upward or downward movement.
• Seasonality: a systematic pattern that repeats at a known interval, such as higher beverage demand in summer.
• Random variation: irregular movement that cannot be reliably anticipated.

A fifth influence, often called a cycle, is a broader fluctuation associated with economic or business conditions. In practice, it is difficult to distinguish cycles from structural changes using short data series.

Demand may be independent or dependent. Independent demand comes from external customers and must generally be forecast. Dependent demand is derived from a known production plan or bill of materials; for example, the demand for screws depends on the planned production of tables. Dependent demand is often calculated from the master production schedule rather than forecast independently.

Data preparation is essential. The analyst should investigate missing periods, duplicate transactions, stockout periods, returns, one-time projects, promotions, price changes and product substitutions. A forecast calculated from corrupted data may be mathematically correct but operationally misleading.

3. Qualitative forecasting

Qualitative methods rely primarily on informed judgement rather than a numerical time series. They are useful when historical data are unavailable, unreliable or no longer representative—for example, when launching a new product, entering a new market or responding to a major regulatory change.

Important approaches include:

• Expert opinion: estimates from experienced buyers, sales staff, planners or technical specialists.
• Sales-force composite: sales representatives estimate demand in their territories; the estimates are reviewed and aggregated.
• Customer or market research: interviews, surveys, pilot sales and market tests provide information about likely adoption.
• Delphi technique: several experts provide independent estimates, receive controlled feedback and revise their views over successive rounds. Anonymity reduces pressure to conform.
• Analogy: demand for a new item is estimated by comparing it with a similar product or market.

Qualitative judgement is not automatically superior to a statistical method. It can incorporate information that is absent from historical data, but it is vulnerable to optimism, anchoring, availability bias, sales incentives and group pressure. A professional process records the rationale, assumptions and confidence level of each adjustment. Judgement should adjust a baseline forecast for identifiable information, not replace analysis without explanation.

4. Quantitative time-series methods

Quantitative time-series methods use historical observations. They are most useful when the future is expected to resemble the past in a reasonably stable way. A method should be selected according to demand behaviour, not convenience.

4.1 Naïve forecast

For a stable series, a simple benchmark is:

F_t+1 is the forecast for period t+1, and A_t is actual demand in period t. The method assumes the latest observation is the best available estimate of the next period. It is easy to understand and valuable as a benchmark, although it reacts strongly to unusual observations.

4.2 Simple moving average

A moving average uses the most recent n observations:

The observations must use the same unit and time interval. A longer window smooths random variation but responds slowly to a genuine change. A shorter window responds more quickly but is more variable. A moving average is most appropriate when demand has no strong trend or seasonality, unless the data are first adjusted for these effects.

Worked example. Weekly demand for a component was 100, 120 and 110 units in the last three weeks. A three-week moving-average forecast for week 4 is:

If week 4 demand is 130 units, the forecast error defined as actual minus forecast is 20 units. The error indicates that actual demand exceeded the forecast; it is not itself a percentage accuracy measure.

4.3 Weighted moving average

A weighted moving average gives greater importance to selected observations:

The weights are non-negative and must sum to 1. For example, weights of 0.5, 0.3 and 0.2 give the greatest importance to the most recent period. The method is still a moving average, not a guarantee of improvement. Weights should be tested using historical holdout data or chosen through a documented planning process.

4.4 Simple exponential smoothing

Simple exponential smoothing updates the previous forecast using the latest error:

α is the smoothing constant, where 0 < α ≤ 1. A_t is actual demand and F_t is the forecast for period t. All quantities use the same units. A high α responds quickly to new observations; a low α produces more smoothing. This method assumes a reasonably stable level and no explicitly modelled trend or seasonality.

Worked example. Suppose the forecast for period 1 is 200 units, actual demand in period 1 is 240 units, and α = 0.3. Then:

The forecast increases because actual demand was above the previous forecast. The initial forecast affects early results, so analysts should document how it was established and allow sufficient history before judging performance.

4.5 Trend and seasonality

Simple moving averages and simple exponential smoothing are not designed to represent a persistent trend or repeating seasonal pattern. A method that ignores seasonality may systematically overforecast some periods and underforecast others.

A practical seasonal approach is the seasonal index. For season s, an index can be estimated as:

An index above 1 indicates demand above the overall average; an index below 1 indicates demand below it. For a complete seasonal cycle, the average of the seasonal indices should be approximately 1, subject to rounding.

A deseasonalised observation can be calculated as:

A baseline forecast is then multiplied by the relevant future seasonal index:

D_t is deseasonalised demand, B_t+1 is the baseline forecast without seasonality, and SI_s,t+1 is the index for the future season. This multiplicative approach assumes that seasonal variation grows approximately in proportion to the level. If seasonal fluctuations are roughly constant in absolute units, an additive approach may be more suitable.

A trend can be incorporated through methods such as Holt’s trend-adjusted exponential smoothing. Such methods require careful initialization and monitoring because they can overreact to temporary changes. At diploma level, the key principle is to use a method that represents the observed pattern and to test it against simpler benchmarks.

5. Forecast error and accuracy

Forecast error must be defined consistently. A common signed error is:

Positive error means actual demand exceeded the forecast. Signed errors reveal bias, but positive and negative errors may cancel. Common measures include:

Here, n is the number of evaluated periods. MAD and RMSE are expressed in demand units; MSE is expressed in squared units; MAPE is a percentage. MSE and RMSE penalise large errors more heavily than MAD. MAPE is not suitable when actual demand is zero and can be unstable when actual demand is very small. In those cases, use MAD, RMSE or a measure designed for intermittent demand.

A useful bias measure is mean error:

A positive ME indicates average underforecasting under the stated error convention; a negative ME indicates average overforecasting. Accuracy should be measured on periods not used to tune the method. Comparing methods on the same holdout periods avoids misleading conclusions.

Forecast performance should be evaluated in business terms. A small unit error for a low-value item may matter less than a small error for a critical spare part. Service level, stockout cost, holding cost, obsolescence risk and supplier flexibility should influence method selection.

6. Collaborative demand information

Forecasting improves when relevant partners share information. Collaborative planning, forecasting and replenishment (CPFR) is a structured approach in which supply-chain partners agree on objectives, exchange data, develop a common forecast and manage exceptions.

Useful information includes point-of-sale data, promotions, planned price changes, inventory availability, open orders, marketing campaigns, product launches, supplier constraints and customer forecasts. Point-of-sale data are often more informative than shipments because shipments may reflect a retailer’s replenishment policy rather than final consumption.

A practical collaborative process is:

1. Define the product, location, time bucket and planning horizon.
2. Establish data ownership, definitions and update frequency.
3. Produce an independent statistical baseline.
4. Add documented information about promotions, events, constraints and known customer commitments.
5. Reconcile differences through an exception-based meeting.
6. Publish one agreed forecast and record assumptions.
7. Compare the forecast with actual demand and improve the process.

Collaboration does not mean accepting every partner estimate. It means making information visible, resolving differences and assigning accountability. Forecast ownership, version control and a clear override policy prevent multiple conflicting numbers from circulating in the organisation.

7. Illustrative case: MetroOffice Supplies

Background. MetroOffice Supplies distributes printer cartridges to business customers. Monthly demand has increased gradually, and demand is higher in September and January because of budget cycles. The procurement team previously used the prior month’s sales as the next-month forecast.

Problem. The method produced stockouts during seasonal peaks and excess stock after the peaks. Sales staff also submitted optimistic estimates before large promotions, but these estimates were not documented or compared with actual customer consumption.

Analysis. The planner first removed two months in which the product was unavailable for ten days and replaced recorded sales with an estimate based on confirmed backorders and comparable weeks. The team plotted thirty-six months of demand and identified trend and seasonality. It compared a naïve method, a three-month moving average and an exponential-smoothing baseline on the last six months of historical data. The seasonal method produced lower MAD and fewer systematic errors during peak months.

Approach. MetroOffice adopted a seasonal baseline, reviewed monthly by the planner. Sales provided promotion dates and expected customer quantities; procurement provided supplier lead times and minimum order quantities. The agreed forecast was adjusted only when a documented event justified the change. A dashboard displayed forecast, actual demand, MAD, bias, stockouts and inventory value.

Outcome. The case is illustrative rather than a claim of measured industry performance. Its expected operational benefit is that seasonal demand is represented explicitly, while collaboration captures information not visible in historical sales. The company still needs safety stock because forecast error and supplier lead-time uncertainty remain.

Learning points. Forecasting is a process, not merely a formula. Clean data, a suitable baseline, documented judgement, shared information and post-period evaluation are all necessary. Forecast accuracy should be considered alongside service and cost outcomes.

8. Implementation, governance and technology

A professional forecasting process should include:

• a defined forecast calendar and cut-off date;
• consistent item, location and unit-of-measure codes;
• separation of baseline forecasts from human overrides;
• a documented reason for each material override;
• performance reporting by product family and location;
• review of bias as well as absolute accuracy; and
• escalation rules for exceptional demand or supply disruption.

Forecasting software can automate calculations, compare models, detect outliers and generate alerts. Machine-learning tools may be useful where there are many products and explanatory variables, but complexity does not guarantee accuracy. A transparent simple method may be more appropriate when data are limited or decisions must be explained to managers and suppliers.

Artificial intelligence should be governed carefully. Users should check whether training data contain stockout bias, whether the model changes without approval, and whether recommendations can be traced to relevant data. Personal or commercially sensitive information must be protected. Forecasts should not be used to manipulate suppliers, conceal poor service or transfer unreasonable risk to smaller partners.

Challenges and limitations

Forecasts become less reliable when the horizon increases, demand is intermittent, products are new, markets are disrupted or historical relationships have changed. Common failure modes include forecasting shipments instead of consumption, treating stockout sales as true demand, ignoring product substitution, selecting a method because it is fashionable, and judging accuracy without considering the forecast horizon.

Forecast error is also asymmetric in its consequences. Underforecasting a critical medical component may be far more costly than overforecasting it, while overforecasting a perishable product may create severe waste. Consequently, the “best” statistical accuracy measure is not necessarily the best inventory decision. Managers should connect forecast performance to the cost and service objectives of the supply chain.

Chapter Summary

Demand forecasting estimates future customer requirements so that procurement and inventory decisions can be made before demand is observed. Forecasts must specify the item, location, time period and horizon, and they should be distinguished from actual demand, which may be distorted by stockouts or promotions. Qualitative methods are valuable for new products and structural change but require controlled, documented judgement. Quantitative methods include naïve forecasts, moving averages and exponential smoothing; their assumptions determine when they are appropriate. Trend and seasonality should be represented rather than ignored. Forecast errors can be assessed through ME, MAD, MSE, RMSE and MAPE, with attention to bias, zero-demand limitations and business consequences. Collaborative information such as point-of-sale data, promotions and supplier constraints improves the forecast when definitions, ownership and overrides are governed. Forecasting supports, but does not replace, safety-stock and inventory-policy decisions.

Key Terms

Forecast: An estimate of future demand for a defined item, location and period.

Demand: Quantity customers are willing and able to purchase.

Independent demand: Demand originating from external customers.

Dependent demand: Demand derived from another item or production plan.

Level: The typical magnitude of a demand series.

Trend: Persistent upward or downward movement.

Seasonality: A repeating demand pattern at a known interval.

Naïve forecast: A forecast using the latest actual observation.

Moving average: The average of a specified number of recent observations.

Weighted moving average: A moving average assigning different weights to observations.

Exponential smoothing: A method that updates a forecast using a smoothing constant.

Smoothing constant: The value α controlling responsiveness in exponential smoothing.

Seasonal index: A factor representing demand relative to an overall average.

Forecast error: The difference between actual and forecast demand.

Bias: Persistent tendency to overforecast or underforecast.

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

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

CPFR: Collaborative planning, forecasting and replenishment.

Safety stock: Inventory held to protect against uncertainty.

Reorder point: Inventory position at which replenishment is initiated.

Review Questions

Short Answer Questions

1. Why is a forecast not the same as a promise of future demand?
2. State two situations in which qualitative forecasting is appropriate.
3. What is the effect of increasing the number of periods in a moving average?
4. What does a positive mean error indicate when error is defined as actual minus forecast?
5. Why can recorded sales understate true demand?

Long Answer Questions

1. Explain how demand level, trend, seasonality and random variation affect method selection.
2. Compare simple moving averages and exponential smoothing, including their assumptions and responsiveness.
3. Explain the purpose and limitations of seasonal indices.
4. Discuss how forecast accuracy measures should be connected to inventory cost and service objectives.
5. Describe a controlled CPFR process for a retailer and its supplier.

Analytical / Application Questions

1. A product had demands of 80, 100 and 120 units in the last three weeks. Calculate the three-week moving-average forecast for week 4. If actual week 4 demand is 130 units, calculate the signed error using actual minus forecast.
2. The previous forecast was 500 units, actual demand was 560 units and α = 0.25. Calculate the next exponential-smoothing forecast and interpret the change.
3. A product has frequent zero-demand periods and occasional large orders. Explain why MAPE may be unsuitable and recommend two alternative evaluation or planning approaches.

Multiple Choice Questions with Answers

Question 1

Which situation most strongly supports using a qualitative forecast?

A. A stable product with five years of clean weekly data

B. A new product with no sales history

C. A mature product with measured seasonality

D. A component whose demand is fully derived from a production schedule

Correct Answer: B. A new product with no sales history

Question 2

Weekly demand in the last three weeks was 90, 110 and 100 units. What is the three-week moving-average forecast for week 4?

A. 90 units

B. 95 units

C. 100 units

D. 110 units

Correct Answer: C. 100 units

Question 3

Using α = 0.2, a previous forecast of 300 units and actual demand of 350 units, what is the next exponential-smoothing forecast?

A. 300 units

B. 310 units

C. 320 units

D. 350 units

Correct Answer: B. 310 units

Question 4

If error is defined as actual demand minus forecast, what does a positive mean error generally indicate?

A. Persistent overforecasting

B. Persistent underforecasting

C. No forecast variation

D. Zero demand

Correct Answer: B. Persistent underforecasting

Question 5

Which measure is expressed in the original demand units and averages absolute errors?

A. MAPE

B. MSE

C. MAD

D. Seasonal index

Correct Answer: C. MAD

Question 6

Why can MAPE be problematic for intermittent demand?

A. It cannot compare percentages

B. It is undefined when actual demand is zero

C. It always penalises large errors twice

D. It requires seasonal indices

Correct Answer: B. It is undefined when actual demand is zero

Question 7

A seasonal index of 1.25 indicates that the relevant season’s demand is approximately:

A. 25% below the overall average

B. Equal to the overall average

C. 25% above the overall average

D. 125 units regardless of the average

Correct Answer: C. 25% above the overall average

Question 8

Which information is generally most useful for distinguishing customer consumption from retailer replenishment behaviour?

A. Point-of-sale data

B. Supplier bank details

C. Warehouse lighting records

D. Employee attendance data

Correct Answer: A. Point-of-sale data

Question 9

What is a key limitation of a simple moving average when demand has a sudden upward trend?

A. It cannot use numerical data

B. It tends to lag behind the changing level

C. It always produces zero forecasts

D. It automatically removes seasonality

Correct Answer: B. It tends to lag behind the changing level

Question 10

In the relationship ROP = μ_L + SS, what does SS represent?

A. Supplier selling price

B. Seasonal sales

C. Safety stock

D. Statistical sample

Correct Answer: C. Safety stock