Demand Forecasting and Inventory Data

Introduction

Inventory decisions are made under uncertainty. An entrepreneur must decide how much to purchase, when to reorder, which supplier commitments to make, and how much working capital to allocate before the future demand is known. Demand forecasting does not remove uncertainty; it makes assumptions about uncertainty explicit and provides a disciplined basis for action.

In inventory control, a forecast is not merely a predicted sales number. It is an input to decisions about replenishment, capacity, service levels, supplier contracts, safety stock, promotions, and cash flow. A statistically accurate forecast can still produce poor business decisions if it is at the wrong product-location level, ignores lead-time variation, or is not translated into an appropriate inventory policy.

This chapter examines qualitative, time-series, causal, and judgmental forecasting methods; evaluates forecast error, bias, variability, and data quality; and shows how forecasts become operational inventory decisions.

Learning Objectives

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

1. Distinguish demand, sales, orders, and inventory consumption data.
2. Select and apply qualitative, time-series, causal, and judgmental forecasting methods.
3. Calculate and interpret forecast error, MAD, MSE, RMSE, MAPE, bias, and tracking signals.
4. Diagnose data-quality problems and demand variability.
5. Translate forecasts and uncertainty into reorder points, safety stock, and replenishment decisions.
6. Critically evaluate forecasting technology, managerial judgment, and ethical governance.

1. The Forecasting Problem in Inventory Systems

1.1 What is being forecast?

The relevant variable is usually future demand during a specified time bucket, such as units demanded per day or weekly customer orders. Several measures must be distinguished:

A forecast must specify the item, location, time horizon, time bucket, unit of measure, and demand event. For example, “forecast of 1,000 units” is incomplete; “forecast of unconstrained weekly demand for SKU A at the Manchester distribution centre for the next 12 weeks” is operationally meaningful.

1.2 Forecast horizon and aggregation

Short-horizon forecasts support purchase orders and workforce scheduling. Medium-horizon forecasts support supplier capacity, budgets, and promotions. Long-horizon forecasts support product portfolio and facility decisions. Aggregating demand across products or locations often reduces relative variability, but it can conceal local stockout risks. Forecasts should therefore be produced at the lowest level at which decisions are made, while reconciliation ensures that item-level forecasts remain consistent with product-family and enterprise totals.

A useful distinction is between a forecast and a plan. A forecast describes what is likely under stated assumptions. A plan describes what the organisation intends to do, including pricing, promotions, assortment, and supply constraints. Treating a constrained sales plan as an unbiased demand forecast can systematically understate true market demand.

2. Forecasting Approaches

2.1 Qualitative forecasting

Qualitative methods are appropriate when historical data are limited, the product is new, structural change is expected, or expert knowledge is important. Common methods include:

• Executive opinion: senior managers form a combined estimate. It is fast but vulnerable to hierarchy and optimism bias.
• Sales-force composite: sales representatives estimate customer or territory demand. It captures market intelligence but may be influenced by targets and incentives.
• Market research: surveys, interviews, test markets, and customer panels provide direct evidence. Stated intentions may not equal actual purchases.
• Analogy: demand is inferred from a comparable product, market, or launch. The analogy must be adjusted for price, distribution, technology, and customer differences.
• Delphi method: independent, anonymous expert estimates are iteratively aggregated. Anonymity reduces status effects, although consensus can still become artificial.

Qualitative forecasts should document the evidence, assumptions, participants, and confidence range. They should not be presented as objective merely because a committee approved them.

2.2 Time-series forecasting

Time-series methods infer future demand from patterns in the demand sequence. They are useful when the future resembles the process that generated the past. Important components are:

• Level: the central magnitude of demand.
• Trend: persistent upward or downward movement.
• Seasonality: a repeating pattern with a known period, such as weekly or annual cycles.
• Cycle: longer, less regular movements associated with economic or industry conditions.
• Irregular variation: random or unexplained movement.

#### Naïve forecast

The simplest benchmark sets the next forecast equal to the most recent observation:

Here, F_t+1 is the forecast for period t+1, and A_t is actual demand in period t. It is particularly useful as a benchmark and can perform well for a random walk, but it does not explicitly model seasonality or trend.

#### Moving average

A k-period moving average is:

The method assumes that recent observations are representative and that each included observation deserves equal weight. A larger k smooths random variation but responds more slowly to genuine change. The forecast has the same units as demand.

#### Weighted moving average

A weighted moving average is:

The weights w_i are non-negative in the usual application, and recent periods commonly receive larger weights. The weights must sum to one so that units are preserved and the result remains a weighted average.

#### Simple exponential smoothing

For a series without strong trend or seasonality, simple exponential smoothing is:

α is the smoothing constant, A_t is actual demand, and F_t is the previous forecast. A high α responds rapidly to new observations; a low α produces greater smoothing. The method assumes a relatively stable level and is unsuitable without modification for strong trend or seasonality.

The equivalent error-correction form is F_t+1 = F_t + α(A_t − F_t). The term in parentheses is the latest forecast error.

#### Trend and seasonality

When a series has trend, Holt’s double exponential smoothing estimates level and trend. When it has repeating seasonal effects, Holt–Winters methods add a seasonal component. A multiplicative seasonal model is appropriate when seasonal amplitude increases with the level; an additive model is more appropriate when seasonal deviations remain approximately constant. These methods require sufficient history—usually several seasonal cycles—and should be tested against simpler benchmarks.

Seasonal indices can be estimated by comparing observations with a centred moving average, then normalising indices so that their average is 1 for a multiplicative model. Seasonal adjustment must be based on the correct calendar: retail weeks, public holidays, fiscal periods, and movable holidays may differ.

2.3 Causal and explanatory forecasting

Causal methods explain demand using variables such as price, advertising, income, weather, competitor actions, availability, and economic indicators. A simple linear model is:

Here, y_t is demand in period t, x_jt are explanatory variables, β₀ is the intercept, β_j are estimated effects, and ε_t is unexplained error. The model assumes that the relationship is adequately specified and that the data support estimation. Correlation alone does not establish causation.

For example, historical sales may rise during advertising campaigns, but advertising may have been increased precisely when demand was expected to rise. Omitted variables, reverse causality, multicollinearity, changing prices, and stockouts can make coefficients misleading. Regression forecasts should therefore be combined with business knowledge and validated out of sample.

For count demand, intermittent demand, or nonlinear response, alternatives may include generalized linear models, Croston-type methods, dynamic regression, gradient boosting, or probabilistic models. Method complexity is justified only when it improves decisions or forecast performance after implementation costs are considered.

2.4 Judgmental adjustment and consensus forecasting

Judgment is unavoidable when a new promotion, competitor entry, supply disruption, or regulatory change is not represented in historical data. A sound process begins with a statistical baseline and records adjustments separately. This preserves accountability and permits post-event learning.

A useful governance sequence is:

1. Generate an independent statistical baseline.
2. Present relevant external information and assumptions.
3. Request documented, evidence-based adjustments.
4. Challenge extreme adjustments and identify incentives.
5. Approve a forecast range as well as a point estimate.
6. Compare the final forecast with the baseline and later measure outcomes.

Consensus is not automatically superior to an independent forecast. Groupthink, anchoring, and power differences can reduce accuracy. Independent estimates followed by structured discussion are generally more defensible than an unstructured meeting.

3. Measuring Forecast Performance

Let actual demand in period t be A_t and the forecast made for that period be F_t. Define the signed forecast error as:

Under this convention, positive error means under-forecasting and negative error means over-forecasting. Organisations must state their convention because some systems define error as forecast minus actual.

3.1 Common accuracy measures

Here, n is the number of evaluated periods. MAPE is undefined when A_t = 0 and becomes unstable when actual demand is close to zero. Weighted MAPE can be dominated by high-volume items, while unweighted averages can give tiny items disproportionate influence. The metric should reflect the decision cost of errors.

A worked example illustrates the calculations. Suppose actual demand for three weeks is 100, 120, and 80 units; forecasts are 90, 130, and 100 units. Errors are 10, −10, and −20 units. Therefore:

• MAD = (10 + 10 + 20) ÷ 3 = 13.33 units.
• MSE = (100 + 100 + 400) ÷ 3 = 200 units².
• RMSE = √200 ≈ 14.14 units.
• MAPE = (10% + 8.33% + 25%) ÷ 3 ≈ 14.44%.
• Mean error = (10 − 10 − 20) ÷ 3 = −6.67 units.

The negative mean error indicates net over-forecasting under the stated convention, even though the first period was under-forecast.

3.2 Bias and tracking signal

Bias is systematic error rather than random fluctuation. A simple measure is cumulative forecast error:

A tracking signal commonly used in practice is:

CFE is in demand units and MAD is also in demand units, so the tracking signal is dimensionless. Large positive values suggest persistent under-forecasting; large negative values suggest persistent over-forecasting. Control limits such as ±4 or ±5 are managerial heuristics, not universal statistical laws. They should trigger investigation, not automatic model replacement.

3.3 Variability and uncertainty

Forecast accuracy is not the same as forecast certainty. For a sample of demand observations, sample standard deviation is:

where Ā is the sample mean. The denominator n − 1 is used when estimating population variability from a sample. Demand variability, forecast residual variability, and lead-time variability are distinct and should not be casually substituted for one another.

Evaluation should use a time-ordered holdout period. Randomly mixing future observations into a training set creates leakage. Rolling-origin evaluation—repeatedly training on earlier data and testing on the next period—better represents operational forecasting. Accuracy should be examined by item, location, horizon, season, demand class, and lifecycle stage rather than only as one corporate average.

4. Data Quality for Forecasting

Forecasting quality is bounded by data quality. A data audit should examine:

• Completeness: missing periods, missing locations, and unrecorded transactions.
• Accuracy: incorrect quantities, prices, dates, units, or product identifiers.
• Consistency: conflicting item codes, calendars, currencies, and units of measure.
• Timeliness: late postings, delayed returns, and backdated adjustments.
• Validity: impossible negative quantities, duplicate invoices, or invalid dates.
• Uniqueness: duplicate orders and repeated shipment records.
• Lineage: ability to trace a forecast input to its source and transformation.

Promotions, stockouts, recalls, assortment changes, and one-time contracts are not automatically “outliers” to delete. They may represent genuine causal events. A stockout produces censored sales: recorded sales are less than unconstrained demand. Analysts should flag such periods and, where credible, estimate lost demand using substitute products, waiting lists, backorders, or comparable locations.

Master-data governance is especially important. A change in pack size can create an apparent demand jump; a product relaunch can make historical observations non-comparable. Version-controlled calendars, documented transformations, and reconciliation to financial and operational totals are basic controls.

5. Translating Forecasts into Inventory Decisions

A demand forecast becomes useful only when linked to a replenishment policy. Suppose average demand per period is d, replenishment lead time is L periods, and demand and lead time are sufficiently stable. The expected demand during lead time is:

where DLT is lead-time demand in units. If the target service policy includes safety stock SS, the reorder point is:

The reorder point is the inventory position at which replenishment should be initiated. Inventory position equals on-hand inventory plus on-order inventory minus backorders or other committed quantities, according to the organisation’s policy.

5.1 Safety stock under a basic uncertainty model

If period demand is independent with standard deviation σ_d, and lead time is constant, the standard deviation of demand during lead time is:

For a cycle-service target associated with z-score z, safety stock is:

This formula assumes independent, identically distributed demand, constant lead time, a stable distribution, and that the chosen z-score appropriately represents the desired service probability. It is not appropriate without adjustment when demand is autocorrelated, lead time is variable, demand is intermittent, or the cost of shortage is asymmetric.

If demand and lead time are independent but both variable, with mean demand μ_d, lead-time mean μ_L, demand standard deviation σ_d, and lead-time standard deviation σ_L, a commonly used approximation is:

The first term represents demand variation over the average lead time; the second represents variation caused by lead-time uncertainty. The approximation neglects covariance and higher-order effects. If demand and lead time are correlated, a covariance term is needed and empirical simulation may be preferable.

5.2 Forecast uncertainty and inventory risk

Safety stock based only on historical demand variation may understate risk when the forecast itself is uncertain. Forecast residuals should be assessed at the replenishment horizon. A promotional forecast, for example, may have a wider prediction interval than a stable base-demand forecast. Managers should distinguish:

• expected demand, used for the central replenishment estimate;
• uncertainty around expected demand, used for safety stock or risk analysis;
• business tolerance for shortage, used to select service targets;
• financial and obsolescence costs, used to limit excess inventory.

Service level is multidimensional. Cycle service level is the probability of no stockout in a replenishment cycle; fill rate is the proportion of demand fulfilled immediately. The same safety stock can produce different measures depending on order sizes and shortage patterns.

6. Worked Application: UrbanHarvest Meal Kits

The following is an illustrative case. UrbanHarvest is a small meal-kit entrepreneur selling refrigerated weekly boxes through an online platform. One popular kit has demand histories of 420, 460, 500, 440, 480, and 520 boxes over six recent weeks. A simple operational baseline uses the three-week moving average.

Problem

The company must forecast week 7 and decide whether to reorder packaging. Supplier lead time is two weeks. Packaging demand is one set per kit. Management is concerned that a recent influencer campaign may have changed the level of demand.

Analysis

The week-7 forecast is:

The full history shows an upward tendency, while the three-week average smooths that tendency and may lag. A naïve forecast would be 520 boxes. The appropriate decision is not to select the higher number simply because it is more convenient. The team should compare models using rolling-origin validation and examine campaign reach, website traffic, conversion, and pre-orders.

Assume that, after reviewing campaign evidence, the approved week-7 demand forecast is 500 boxes per week. Historical forecast residuals at the two-week horizon have a standard deviation of 45 boxes per week. If weekly demand is treated as independent and lead time is constant, lead-time uncertainty is:

For an approximate 95% cycle service target, use z = 1.645 for a one-sided normal quantile:

Expected demand during lead time is:

Thus:

If inventory position falls to approximately 1,105 packaging sets, an order should be initiated, subject to order multiples, minimum order quantities, shelf life, and supplier reliability. The calculation is a decision aid, not a guarantee. If the influencer campaign creates a structural break, historical residuals may understate future uncertainty. A scenario analysis might therefore evaluate base, moderate-growth, and high-growth cases separately.

Outcome and learning points

UrbanHarvest adopts the statistical baseline as a transparent starting point, documents the campaign adjustment, and monitors pre-orders daily. It also asks the supplier for a shorter lead time and a flexible order amendment. The case demonstrates four principles: forecast level and forecast uncertainty are different; judgment should be recorded rather than hidden; service targets have inventory consequences; and supplier flexibility can be more valuable than a marginal improvement in point-forecast accuracy.

7. Technology, AI, and Governance

Enterprise resource planning, point-of-sale, warehouse, customer relationship, and supplier systems can provide automated data pipelines. Forecasting platforms may use hierarchical reconciliation, probabilistic forecasts, machine learning, and external signals such as weather or web searches. AI can detect nonlinear patterns and interactions, but it cannot repair biased transaction data by itself.

A responsible forecasting system should provide:

• traceable source data and transformations;
• versioned models and reproducible forecasts;
• prediction intervals, not only point estimates;
• monitoring for drift, bias, and performance deterioration;
• human review for exceptional events;
• access controls and protection of customer and supplier data;
• clear ownership for overrides and decisions.

Ethical issues include discriminatory allocation when scarce stock is prioritised using problematic customer proxies, unfair supplier treatment based on opaque scores, privacy risks from individual-level purchasing data, and automation bias. Forecasts should support accountable decisions rather than create an excuse to avoid responsibility. Contractual and regulatory requirements concerning personal data, auditability, product safety, and financial reporting must be incorporated into system design.

8. Best Practices and Critical Limitations

A mature forecasting process generally:

1. Defines the demand event, unit, hierarchy, horizon, and decision owner.
2. Establishes naïve and seasonal benchmarks before adopting complex models.
3. Separates unconstrained demand from observed sales where stockouts occur.
4. Uses time-based backtesting and evaluates both accuracy and bias.
5. Measures performance with metrics aligned to business costs.
6. Records promotions, price changes, launches, and exceptional events.
7. Converts forecast distributions and lead-time risk into inventory policies.
8. Reviews forecast overrides and learns from them.
9. Monitors supplier reliability, not only demand variability.
10. Retires models that no longer improve decisions.

No method is universally best. MAPE can mislead for low-volume items; regression can confuse correlation with causality; machine learning can overfit; expert judgment can be politically distorted; and historical data can fail during structural change. The correct criterion is not mathematical sophistication but robust decisions under the organisation’s actual uncertainty, costs, and capabilities.

Chapter Summary

Demand forecasting is a structured process for estimating future demand at a specified item, location, horizon, and time bucket. Inventory analysis must distinguish demand from sales, shipments, orders, and consumption, particularly when stockouts censor observed sales. Qualitative methods are valuable for new products and structural change; time-series methods exploit level, trend, and seasonality; causal models incorporate explanatory variables; and judgmental adjustments address information absent from historical data.

Forecast errors should be defined consistently. MAD and RMSE express error in units, MSE penalises large errors, percentage measures support scale comparison but fail near zero, and cumulative error and tracking signals help reveal bias. Forecast evaluation should use time-ordered validation and should examine variability by item, location, horizon, and demand class.

Data quality involves completeness, accuracy, consistency, timeliness, validity, uniqueness, and lineage. Promotions and stockouts require investigation rather than automatic deletion. Forecasts become inventory decisions through lead-time demand, safety stock, reorder points, service-level policies, and supplier constraints. Basic safety-stock formulas depend on explicit assumptions and must be adapted when demand, lead time, or forecast uncertainty is complex. Technology and AI can improve scale and pattern recognition, but governance, transparency, privacy, and human accountability remain essential.

Key Terms

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

Sales: Transactions actually completed or recorded; not necessarily unconstrained demand.

Time series: Ordered observations indexed by time.

Naïve forecast: A forecast using the latest actual observation as the next prediction.

Moving average: An average of recent observations used to smooth random variation.

Exponential smoothing: A method that gives declining weights to older observations.

Trend: Persistent directional movement in a series.

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

Causal forecast: A forecast based on explanatory variables believed to influence demand.

Judgmental adjustment: A documented expert modification to a statistical baseline.

Forecast error: The difference between actual demand and forecast demand under a stated convention.

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

RMSE: Root mean squared error, which emphasises large errors while retaining demand units.

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

Tracking signal: Cumulative forecast error divided by MAD.

Lead-time demand: Expected demand during replenishment lead time.

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

Reorder point: Inventory position at which replenishment is initiated.

Service level: A measure of the probability or proportion of demand supplied without shortage.

Forecast leakage: Use of information unavailable at the time the forecast would have been made.

Review Questions

Short Answer Questions

1. Why is observed sales history not always an accurate measure of market demand?
2. What is the principal effect of increasing the window length in a moving average?
3. Under the convention e_t = A_t − F_t, what does a negative error indicate?
4. Why can MAPE be problematic for intermittent or zero demand?
5. Name four dimensions of data quality relevant to forecasting.

Long Answer Questions

1. Compare qualitative, time-series, causal, and judgmental forecasting methods, including appropriate use cases and limitations.
2. Explain how trend and seasonality affect the selection of a forecasting method.
3. Discuss why forecast evaluation should use time-ordered holdout or rolling-origin testing rather than random data splitting.
4. Explain the assumptions behind the basic safety-stock formula and describe situations in which those assumptions fail.
5. Discuss how governance can reduce the risks associated with forecast overrides and AI-based forecasting.

Analytical / Application Questions

1. Actual weekly demand is 200, 240, and 220 units, while forecasts are 190, 250, and 210 units. Calculate MAD, MSE, RMSE, MAPE, and mean signed error using e_t = A_t − F_t. Interpret the result.
2. A retailer has mean daily demand of 80 units, daily standard deviation of 12 units, constant lead time of 4 days, and a one-sided service quantile z = 1.645. Using the basic independent-demand model, calculate safety stock and reorder point. State the assumptions.
3. A product’s sales fall sharply during two weeks because of a stockout, but the forecasting system treats those observations as normal demand. Diagnose the likely consequences and propose a data and inventory-policy response.

Multiple Choice Questions with Answers

Question 1

Which measure is most directly affected by stockouts because it records fulfilled transactions rather than unconstrained customer requirements?

A. Demand

B. Sales

C. Forecast error

D. Safety stock

Correct Answer: B. Sales

Question 2

A three-period moving-average forecast uses demands of 90, 110, and 100 units. What is the next forecast?

A. 90 units

B. 95 units

C. 100 units

D. 110 units

Correct Answer: C. 100 units

Question 3

Under the convention e_t = A_t − F_t, an error of −25 units means that the forecast was:

A. 25 units below actual demand

B. 25 units above actual demand

C. Exactly accurate

D. 25 percent below actual demand

Correct Answer: B. 25 units above actual demand

Question 4

Which metric penalises large forecast errors more strongly because errors are squared?

A. MAD

B. MAPE

C. MSE

D. Mean signed error

Correct Answer: C. MSE

Question 5

Why is MAPE unsuitable when actual demand is zero?

A. It has squared units

B. It requires a lead-time estimate

C. Its denominator becomes zero

D. It always reports negative values

Correct Answer: C. Its denominator becomes zero

Question 6

A tracking signal with a large positive value most commonly indicates:

A. Persistent under-forecasting under the chapter’s error convention

B. Persistent over-forecasting under the chapter’s error convention

C. No forecast bias

D. A reduction in lead-time variability

Correct Answer: A. Persistent under-forecasting under the chapter’s error convention

Question 7

If mean daily demand is 50 units and constant lead time is 3 days, expected lead-time demand is:

A. 53 units

B. 100 units

C. 150 units

D. 1,500 units

Correct Answer: C. 150 units

Question 8

Under the basic constant-lead-time model, if daily demand standard deviation is 10 units and lead time is 4 days, the standard deviation of lead-time demand is:

A. 5 units

B. 20 units

C. 40 units

D. 160 units

Correct Answer: B. 20 units

Question 9

Which practice best prevents future information from contaminating forecast evaluation?

A. Randomly mixing all observations before splitting the data

B. Using time-ordered holdout or rolling-origin testing

C. Deleting all unusual observations

D. Evaluating only the final forecast approved by management

Correct Answer: B. Using time-ordered holdout or rolling-origin testing

Question 10

What is the primary purpose of a reorder point?

A. To determine product selling price

B. To identify when replenishment should be initiated

C. To calculate supplier profit

D. To eliminate all demand variability

Correct Answer: B. To identify when replenishment should be initiated