Demand Analysis and Forecasting for Inventory Decisions

Introduction

Inventory decisions are decisions under uncertainty. A firm must decide how much stock to hold before it knows exactly how much customers will buy, when replenishment will arrive, or whether demand will shift because of price, competition, seasonality, or disruption. Forecasting does not eliminate uncertainty; it provides a structured estimate of future demand and a basis for measuring the uncertainty that inventory policy must absorb.

The central managerial question is not simply, “What will demand be?” It is: “What demand should be protected during the replenishment lead time, and how much uncertainty protection is economically justified?” This chapter develops that logic from demand diagnosis through forecasting and translation into reorder points, safety stock, and service policies.

Learning Objectives

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

1. Distinguish trend, seasonality, cycles, irregular variation, and intermittent demand.
2. Estimate lead-time demand and explain the effects of demand and lead-time variability.
3. Apply moving-average, exponential-smoothing, trend, causal, and intermittent-demand methods.
4. Measure forecast error and select appropriate accuracy metrics.
5. Translate forecast uncertainty into safety stock, reorder-point, and service-level parameters.
6. Evaluate forecasting methods critically, including their organizational, technological, and ethical implications.

1. Demand as an Inventory Variable

Demand is the quantity requested or consumed during a specified time interval. It must be defined with a consistent unit, time bucket, product level, and customer population. For example, “monthly demand for finished product A at the national distribution centre” is more useful than “sales of A.” Sales may be constrained by stockouts, while true demand may have been higher.

Three distinctions are particularly important:

• Dependent demand is derived from the demand for another item, as in components required for a production schedule. Material requirements planning commonly forecasts the parent item and calculates component requirements.
• Independent demand arises from external customers and must generally be forecast.
• Demand and sales are not always identical. Lost sales, substitutions, rationing, and backorders can conceal demand.

A demand series can be represented as D_t, the observed demand in period t. A useful conceptual decomposition is:

where T_t is trend, S_t is seasonal variation, C_t is longer-term cyclical variation, and I_t is irregular or random variation. This additive form is suitable when seasonal fluctuations have approximately constant absolute size. A multiplicative form,

is more suitable when seasonal variation grows in proportion to the level of demand. A logarithmic transformation can often convert a multiplicative structure into an additive one, but the transformation must be reversed carefully when forecasts are returned to the original units.

Demand patterns

Before selecting a method, analysts should plot the series, examine stockout periods, identify promotions and price changes, and aggregate data at a decision-relevant frequency. Excessive aggregation hides intermittency; excessive disaggregation creates noise.

2. Lead-Time Demand and Uncertainty

Inventory is consumed while replenishment is being obtained. If lead time is L periods, lead-time demand is:

For constant lead time and independent demand with mean μ_D per period and standard deviation σ_D,

and

The second relationship assumes independent demand across periods and constant lead time. It follows because variances add: Var(D_L) = L × σ_D². If demand is autocorrelated, the variance must include covariance terms; the simple square-root rule may then understate or overstate risk.

When both demand and lead time vary, let lead time have mean μ_L and standard deviation σ_L. A widely used approximation is:

This assumes demand periods are approximately independent, lead time is independent of demand, and the variability is adequately represented by means and variances. The first term represents demand uncertainty during average lead time; the second represents the effect of uncertain duration. A more exact treatment requires a distribution for the random sum of demands and may be obtained by simulation or empirical convolution.

If demand and lead time are correlated, an additional covariance effect is relevant. In practice, long lead times may occur precisely when demand is high because of congestion or capacity shortages. Ignoring this dependence can create systematic understocking.

3. Forecasting Methods

3.1 Naïve and seasonal-naïve forecasts

A naïve forecast uses the latest observation:

It is a valuable benchmark, especially for non-stationary series. A seasonal-naïve forecast uses the observation from the same seasonal position in the previous cycle:

where m is the seasonal period and h is the forecast horizon. A sophisticated method should not be adopted unless it improves on a credible benchmark.

3.2 Moving averages

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

It smooths random variation but reacts slowly to a trend. A weighted moving average assigns weights w_i:

Higher weights on recent observations increase responsiveness but may increase sensitivity to noise.

3.3 Simple exponential smoothing

For a series with a changing level but no systematic trend or seasonality:

where 0 < α < 1. A high α responds quickly to recent change; a low value produces stronger smoothing. The method can also be written as an update to the estimated level l_t:

with the next forecast equal to l_t. Initialisation matters, particularly in short series.

3.4 Trend and seasonal exponential smoothing

When trend exists, Holt’s method estimates level and trend:

where β is the trend-smoothing parameter. For seasonal data, Holt–Winters methods add a seasonal component. Additive seasonality is appropriate when the seasonal amplitude is approximately constant; multiplicative seasonality is appropriate when it scales with the level. Parameter selection should use rolling-origin validation rather than visual preference alone.

3.5 Regression and causal forecasting

Causal methods explain demand using predictors such as price, advertising, weather, economic activity, installed base, or planned promotions. A simple model is:

where β₀ is the intercept, β_i are effects of predictors X_i, and ε_t is unexplained error. Coefficients should be interpreted cautiously: association is not automatically causation. Promotion variables may be endogenous because managers promote products when demand is already expected to fall or rise. Forecasting also requires future values of the predictors; a model dependent on an unavailable future price or weather value is operationally incomplete.

Causal models are useful for scenario analysis, but they require diagnostics for residual autocorrelation, changing variance, multicollinearity, omitted variables, and structural breaks. Out-of-sample performance is more relevant than in-sample fit.

3.6 Intermittent demand

Intermittent demand contains many zero observations and occasional positive demands. It is common for spare parts, maintenance items, and slow-moving stock. Two distinct processes should be separated:

1. Occurrence process: whether demand occurs.
2. Size process: the quantity demanded when it occurs.

Croston-type methods estimate the average non-zero demand size and the average interval between non-zero demands. The original method is useful conceptually but has known bias properties and does not naturally provide a complete probabilistic forecast. Variants such as SBA and TSB adjust bias or allow the probability of demand occurrence to decay. Method choice should reflect the service objective: forecasting average demand is not the same as estimating the probability of a stockout in a long lead time.

For intermittent items, report demand frequency, average non-zero size, variance, lead-time demand distribution, and forecast accuracy by item class. A zero forecast may have a low mean absolute error while being operationally unacceptable if a single failure causes a costly shutdown.

4. Measuring Forecast Error

Let e_t = D_t − F_t be the forecast error. A positive error means actual demand exceeded the forecast. Common measures are:

Mean error measures bias; MAE is in demand units; MSE and RMSE penalise large errors more heavily. Percentage measures can be unstable when actual demand is zero. MAPE is therefore unsuitable for many intermittent series. A scaled measure such as MASE compares the model’s MAE with the MAE of a naïve benchmark and is more useful across products with different units, although its denominator must be defined carefully.

Forecast accuracy must be evaluated at the level at which decisions are made. A forecast can be inaccurate for individual stores but accurate after regional aggregation. This is the aggregation effect. Conversely, aggregate accuracy can conceal serious errors for critical stock-keeping units.

Bias monitoring is essential. A simple tracking signal is:

where RSFE = ∑e_t is running sum of forecast errors and MAD is mean absolute deviation, commonly calculated over the same monitoring window. Large persistent values indicate systematic under- or over-forecasting. Control limits should be established using the organisation’s statistical governance rather than treated as universal constants.

5. From Forecasts to Inventory Policy

Safety stock and reorder point

For continuous review, the reorder point is commonly written:

where μ_DL is expected lead-time demand and SS is safety stock. Under an approximately normal lead-time-demand distribution and a cycle-service target, safety stock is:

where z is the standard-normal quantile corresponding to the desired probability of no stockout during the protection period. Thus:

This formula is an approximation. It assumes that the lead-time-demand distribution is adequately represented by a normal distribution, that the forecast mean is unbiased, and that the selected service measure is a cycle service level. It is inappropriate to use it mechanically for highly skewed, intermittent, lumpy, or bounded demand.

If the forecast itself has error variance σ_F², the protection-period variance should include forecast uncertainty rather than only observed demand variation. A practical approach is to estimate the empirical distribution of lead-time forecast errors from historical rolling forecasts. This avoids the common mistake of treating forecast error as if it were already included in ordinary demand standard deviation.

Service measures

• Cycle service level: probability of no stockout in a replenishment cycle.
• Fill rate: proportion of demand units supplied immediately from stock.
• Ready rate: proportion of demand events satisfied immediately.

These measures are not interchangeable. A policy can achieve a high cycle service level while having a lower fill rate if occasional stockouts are very large. The inventory objective should therefore specify the service metric, review period, and consequences of shortage.

Worked example

A component has mean weekly demand of 240 units and weekly standard deviation of 60 units. Lead time is constant at 3 weeks. The organisation targets a 95% cycle service level, for which z ≈ 1.645. Assuming independent weekly demand and an approximately normal lead-time distribution:

The result means that a replenishment order is triggered when inventory position reaches approximately 891 units. It does not mean that 891 units are always held physically: inventory position includes on-hand stock, on-order stock, and relevant backorders. The calculation should be revised if demand is seasonal, lead time varies, or the item has a non-normal demand distribution.

6. An Illustrative Case: Service Parts at Meridian Pumps

Background and problem

Meridian Pumps, a fictional industrial-equipment manufacturer, experienced repeated stockouts of a low-volume seal kit despite carrying substantial total inventory. The item’s monthly history contained many zero-demand months and occasional requirements of 1–6 kits. The planner used a 12-month moving average and a fixed two-month safety factor.

Analysis

The moving average estimated average demand adequately but failed to represent occurrence timing. A review found that demand was related to the installed base, equipment age, and preventive-maintenance campaigns. In addition, emergency orders had been recorded as zero-demand months when the requested quantity could not be fulfilled, causing censored data.

The team separated normal service demand from emergency and lost demand, estimated the probability of at least one requirement during the two-month lead time, and modelled positive demand sizes separately. It also classified the item as critical because a shortage could stop a customer’s production line.

Approach and solution

The firm used an intermittent-demand method as a baseline and compared it with an empirical bootstrap distribution of historical lead-time demand. The bootstrap preserved the observed frequency and size of non-zero demand. A criticality-adjusted policy targeted a higher service level than the monetary ABC class alone would imply. Suppliers were also asked to provide lead-time percentiles rather than a single nominal lead time.

Forecasts were evaluated by rolling-origin tests, while inventory performance was simulated using fill rate, emergency shipments, average inventory, and shortage cost. Master-data controls were introduced so that backorders and substitutions were not mistaken for zero demand.

Outcome and learning points

The revised policy increased average stock modestly but reduced emergency shipments substantially. The principal improvement did not come from selecting a more complicated algorithm; it came from correcting the demand history, separating occurrence from size, recognising item criticality, and measuring the service outcome that mattered.

The case illustrates five principles: data quality precedes model sophistication; forecast error must be evaluated at the decision horizon; intermittent demand requires specialised treatment; service targets should reflect consequence as well as value; and policy parameters must be monitored after implementation.

7. Implementation, Governance, and Technology

A sound forecasting process includes demand-data ownership, documented calendars, promotion flags, exception handling, forecast version control, and a clear approval process for managerial overrides. Overrides should be recorded with reasons and later evaluated; otherwise judgmental intervention becomes invisible and unlearnable.

Machine-learning methods can model nonlinear relationships, many predictors, and interactions. They are useful when there is sufficient clean history and when future predictor values are available. However, they can overfit, drift when customer behaviour changes, and create opaque decisions. A machine-learning forecast should be benchmarked against simple methods, validated through time-based splits, monitored for drift, and governed for explainability and fairness. Algorithms must not use protected or ethically inappropriate attributes, and customer data should be collected and retained lawfully.

Contemporary practice increasingly combines statistical forecasts with probabilistic prediction intervals, scenario analysis, hierarchical reconciliation, and simulation. Probabilistic forecasts are especially valuable because inventory decisions require a distribution of possible demand, not merely a point estimate. Forecasts at store, regional, and national levels may be reconciled so that their totals are coherent, while planners retain the ability to represent local events.

Best practice is therefore to:

1. Define the demand signal and correct for stockouts and substitutions.
2. Segment items by volume, variability, intermittency, criticality, and lifecycle.
3. Use simple benchmarks and rolling-origin validation.
4. Choose error measures consistent with the inventory decision.
5. Estimate lead-time demand directly where demand or lead time is non-normal.
6. Separate forecast uncertainty from process variation where possible.
7. Link safety stock to an explicit service measure and shortage cost.
8. Monitor bias, drift, overrides, and realised service continuously.
9. Re-estimate parameters after structural changes, launches, and end-of-life events.

Challenges and Limitations

Forecasts are weakened by short histories, changing product definitions, promotions, cannibalisation, new-product uncertainty, stockout censoring, and unstable lead times. Historical accuracy does not guarantee future validity when the demand-generating process changes. Statistical confidence intervals may also understate risk when extreme events, dependence, or model uncertainty are ignored.

There is a fundamental trade-off between responsiveness and stability. A highly responsive model may chase random noise; a highly stable model may react too slowly to a genuine shift. The appropriate balance depends on lead time, replenishment flexibility, shortage cost, product lifecycle, and the reversibility of the inventory decision.

Chapter Summary

Demand forecasting for inventory is a decision-support discipline rather than a promise of certainty. Demand should be diagnosed for level, trend, seasonality, cycles, intermittency, and structural change. Lead-time demand is the relevant protection-period quantity, and its variance depends on demand variability, lead-time variability, dependence, and forecast error. Naïve, moving-average, exponential-smoothing, causal, and intermittent-demand methods each have appropriate domains and limitations. Forecast error should be measured with bias, absolute, squared, and scaled measures selected for the decision context. Reorder points combine expected lead-time demand with safety stock, but normal-distribution formulas are approximations whose assumptions must be checked. Effective implementation integrates data governance, service definitions, validation, simulation, managerial judgment, and continuous monitoring.

Key Terms

Demand: Quantity requested or consumed in a specified period.

Independent demand: Demand arising from external customers.

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

Lead-time demand: Total demand during the replenishment lead time.

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

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

Intermittent demand: Demand containing many zero periods and occasional positive observations.

Structural break: A persistent change in the process generating demand.

Naïve forecast: Forecast based on the latest observation or corresponding seasonal observation.

Exponential smoothing: Forecasting method that assigns geometrically declining weights to older observations.

Causal forecast: Forecast based on explanatory variables related to demand.

Forecast error: Difference between actual demand and forecast demand.

Bias: Persistent tendency to overforecast or underforecast.

MAE: Mean absolute error, expressed in demand units.

Lead-time-demand standard deviation: Measure of uncertainty in demand during replenishment lead time.

Safety stock: Inventory held to protect against uncertainty.

Reorder point: Inventory position at which replenishment is initiated.

Cycle service level: Probability of avoiding a stockout during a replenishment cycle.

Fill rate: Proportion of demand units supplied immediately from stock.

Forecast drift: Deterioration in forecast performance as the operating environment changes.

Review Questions

Short Answer Questions

1. Distinguish independent demand from dependent demand.
2. What is the difference between trend and seasonality?
3. Why can MAPE be unsuitable for intermittent demand?
4. State the assumptions behind σ_DL = σ_D√L.
5. Define cycle service level and fill rate.

Long Answer Questions

1. Explain how demand decomposition supports inventory policy design.
2. Compare moving averages, simple exponential smoothing, and causal regression.
3. Discuss why lead-time variability should be included in safety-stock analysis.
4. Explain the limitations of normal-distribution reorder-point formulas.
5. Design a governance process for forecasting in a multi-echelon supply chain.

Analytical / Application Questions

1. Weekly demand has mean 500 units and standard deviation 80 units; constant lead time is 4 weeks. Calculate expected lead-time demand and its standard deviation under independent-demand assumptions. Explain one reason the result may be unreliable.
2. An item has frequent zero demand, high criticality, and a long supplier lead time. Recommend a forecasting and inventory-policy approach, including the performance measures you would monitor.
3. A forecast has persistent negative errors during promotions. Diagnose likely causes and propose changes to data, modelling, validation, and managerial override procedures.

Multiple Choice Questions with Answers

Question 1

Which demand pattern contains many zero observations and occasional positive demands?

A. Trend demand

B. Intermittent demand

C. Seasonal demand

D. Cyclical demand

Correct Answer: B. Intermittent demand

Question 2

If weekly demand is independent with standard deviation 50 units and constant lead time is 4 weeks, what is the lead-time-demand standard deviation?

A. 50 units

B. 100 units

C. 200 units

D. 250 units

Correct Answer: B. 100 units

Question 3

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

A. More responsive to recent observations

B. Independent of recent observations

C. Unable to use an initial forecast

D. Automatically seasonal

Correct Answer: A. More responsive to recent observations

Question 4

Which measure is expressed in the original demand units and does not square errors?

A. MSE

B. RMSE

C. MAE

D. MAPE

Correct Answer: C. MAE

Question 5

The usual reorder-point expression under a normal lead-time-demand approximation is:

A. ROP = safety stock − mean lead-time demand

B. ROP = mean lead-time demand + safety stock

C. ROP = mean demand ÷ lead time

D. ROP = forecast error × lead time

Correct Answer: B. ROP = mean lead-time demand + safety stock

Question 6

A causal forecasting model primarily uses:

A. Only the latest demand observation

B. Explanatory variables such as price or promotion

C. Only non-zero demand sizes

D. A fixed safety factor without demand data

Correct Answer: B. Explanatory variables such as price or promotion

Question 7

What does a positive forecast error defined as actual demand minus forecast indicate?

A. Actual demand exceeded the forecast

B. Forecast exceeded actual demand

C. The forecast was exactly correct

D. Lead time was zero

Correct Answer: A. Actual demand exceeded the forecast

Question 8

Which service measure is the proportion of demand units supplied immediately from stock?

A. Cycle service level

B. Fill rate

C. Forecast bias

D. Tracking signal

Correct Answer: B. Fill rate

Question 9

The formula σ_DL = √(μ_Lσ_D² + μ_D²σ_L²) is intended to approximate uncertainty when:

A. Both demand and lead time vary

B. Demand is always zero

C. Lead time is always exactly one period

D. There is no forecast error or demand variation

Correct Answer: A. Both demand and lead time vary

Question 10

Why should a forecasting method be compared with a naïve benchmark?

A. To guarantee zero forecast error

B. To determine whether added complexity produces useful improvement

C. To remove all seasonality automatically

D. To eliminate the need for inventory records

Correct Answer: B. To determine whether added complexity produces useful improvement