Retail Demand Forecasting

Introduction

Retail demand forecasting is the disciplined estimation of future customer demand for a product, category, channel, store, region, or enterprise over a specified time horizon. It is a central capability in retail analytics because almost every operational decision depends on an expectation of future sales. Buying quantities, replenishment, warehouse capacity, labour scheduling, pricing, promotion planning, transport, working capital, and financial budgeting are all influenced by forecasts.

A forecast is not a promise about the future. It is an estimate conditional on information, assumptions, and a chosen method. The purpose of forecasting is therefore not to eliminate uncertainty but to quantify it sufficiently to support better decisions. A technically accurate model can still create poor business outcomes if it forecasts the wrong measure, ignores stock-outs, fails to account for promotions, or is not translated into an executable replenishment decision.

Retail demand is particularly difficult to forecast. It is affected by seasonality, holidays, promotions, competitor actions, product substitution, weather, economic conditions, assortment changes, store openings, stock availability, and changing consumer preferences. Demand is also observed at multiple levels: item–store–day, item–region–week, category–channel–month, and total-company level. These levels must often be forecast consistently.

This chapter develops baseline, seasonal, hierarchical, causal, time-series, regression, and machine-learning approaches. It also addresses intermittent demand, new products, promotions, forecast bias, accuracy measurement, governance, and managerial use.

Learning Objectives

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

1. Explain the role, scope, and limitations of retail demand forecasting.
2. Construct naïve, seasonal-naïve, moving-average, exponential-smoothing, and trend-based forecasts.
3. Distinguish time-series, causal, hierarchical, and machine-learning forecasting approaches.
4. Select suitable methods for intermittent demand, new products, promotions, and multiple aggregation levels.
5. Calculate and interpret common forecast-accuracy and bias measures.
6. Design a forecasting workflow that incorporates data quality, validation, governance, and human judgement.
7. Evaluate forecasting results in relation to inventory, service, profitability, and ethical considerations.

1. The Nature of Retail Demand

1.1 Demand, sales, and lost sales

Observed sales are not always equal to true customer demand. If an item is unavailable, customers may postpone purchase, substitute another item, or leave the store. A simple representation is:

This relationship assumes that every unit demanded can be fulfilled until inventory is exhausted and that sales are recorded accurately. When inventory is zero, observed sales may understate demand substantially. Forecasting sales without identifying stock-outs can therefore produce a systematic underforecast.

Retail analysts should distinguish at least four quantities:

• Demand: the quantity customers would purchase under the prevailing price, promotion, assortment, and availability conditions.
• Sales: the quantity actually transacted.
• Orders: the quantity requested by stores, customers, or downstream partners.
• Shipments: the quantity physically dispatched.

The appropriate target depends on the decision. A replenishment model may require unconstrained demand, whereas a cash-flow forecast may require expected sales revenue.

1.2 Components of a demand series

A demand series can be represented conceptually as the combination of several components:

where y_t is demand at time t, T_t is trend, S_t is seasonal variation, C_t is cyclical or business-cycle variation, and I_t is irregular variation. This additive form is useful when seasonal fluctuations are approximately constant in absolute size.

A multiplicative representation is:

It is more appropriate when seasonal variation grows in proportion to the level of demand. For example, a seasonal uplift of 30% is naturally multiplicative. Multiplicative models require positive observations; additive models can accommodate zero or negative transformed values more easily.

Retail demand is also characterized by several forms of dependence:

• Autocorrelation: current demand is related to previous demand.
• Seasonality: a pattern repeats at a known calendar frequency.
• Trend: the underlying level changes over time.
• Cross-sectional dependence: related products, stores, or channels influence one another.
• Exogenous dependence: demand responds to prices, promotions, weather, holidays, or macroeconomic variables.

1.3 Forecast horizons

The horizon should be linked to the decision lead time. A daily store-replenishment forecast may cover 7–28 days, a buying plan may cover several months, and an annual budget may cover multiple years. Short horizons usually benefit from recent sales and operational variables; longer horizons require assumptions about assortment, price, competition, and market growth.

Forecasts should also be probabilistic where uncertainty matters. A point forecast gives one expected value, while a prediction interval gives a range of plausible outcomes. Inventory and capacity decisions often require quantiles or full demand distributions rather than only the mean.

2. Forecasting Architecture and Data Preparation

A professional forecasting system begins before model selection. The analyst must define the unit of observation, forecast target, horizon, hierarchy, and decision context.

2.1 Granularity and the retail hierarchy

A typical hierarchy may be:

Enterprise
 ├── Channel
 │    ├── Region
 │    │    ├── Store
 │    │    │    └── SKU
 │    └── E-commerce
 └── Category
      └── Subcategory

In practice, product and location hierarchies may intersect. Forecasts can be generated at the lowest level and aggregated upward, or produced at several levels and reconciled so that totals agree.

2.2 Essential data fields

A forecasting dataset commonly contains date, product, store or channel, units sold, revenue, price, promotion indicators, inventory, stock-out status, product lifecycle status, holidays, weather, and competitor information. Data should be aligned to a consistent calendar and time zone.

Important preparation activities include:

• correcting duplicate transactions and coding errors;
• distinguishing zero demand from missing data;
• identifying stock-outs and store closures;
• separating returns from sales;
• recording price changes and promotional depth;
• accounting for assortment additions and discontinuations;
• treating extreme observations using business investigation rather than automatic deletion;
• preventing future information from entering the training data.

The last issue is data leakage. For example, using a final monthly sales total to predict an earlier week, or using a promotion flag that was decided after the forecast date, creates an unrealistically optimistic evaluation.

2.3 Forecast origin and rolling validation

Randomly splitting retail time series into training and test sets is generally inappropriate because it allows future patterns to influence the past. A stronger design uses rolling-origin evaluation. A model is trained through time t, forecasts the next horizon, then the origin advances and the process repeats.

This approach estimates performance under conditions closer to actual deployment. It also reveals whether a method works consistently or only during a particular period.

3. Baseline and Seasonal Forecasts

Baselines are not trivial exercises. They provide transparent benchmarks against which sophisticated models must demonstrate value.

3.1 Naïve forecast

The one-step naïve forecast is:

where y_t is the most recently observed demand and ŷ_t+1 is the forecast for the next period. It is suitable for stable series with little systematic change and is often difficult to beat at very short horizons.

A drift version extrapolates the average change between the first and latest observations:

where h is the forecast horizon, y₁ is the first observation, and t is the number of observations available. The method assumes that the average historical change continues into the future.

3.2 Seasonal-naïve forecast

For a series with seasonal period m, the seasonal-naïve forecast is:

For weekly data with annual seasonality, m may be approximately 52; for monthly data with annual seasonality, it is 12. The method assumes that the corresponding period from the previous season is the best available estimate.

3.3 Moving averages

A simple moving-average forecast with window length k is:

The forecast is the arithmetic mean of the most recent k observations. A larger k smooths random variation but responds slowly to changes; a smaller k responds faster but is noisier. Moving averages are most appropriate when the series has a relatively stable level and no strong trend or seasonality.

3.4 Exponential smoothing

Simple exponential smoothing updates the level recursively:

and forecasts future periods as:

Here ℓ_t is the estimated level and α is a smoothing parameter satisfying 0 < α < 1. A high α gives greater weight to recent demand. The initial level must be specified or estimated.

For trend, Holt’s method uses level and trend components:

where b_t is the estimated trend and β satisfies 0 < β < 1. Trend forecasts can become unrealistic over long horizons, so damped trends are often preferable in retail planning.

4. Time-Series Models

4.1 Autoregressive models

An autoregressive model of order p is:

where c is an intercept, φ_i are coefficients, and ε_t is an error term with mean approximately zero. The model assumes that past demand contains information about current demand after accounting for the specified structure. Stationarity conditions are required for stable long-run behaviour; an unrestricted trend or structural break can invalidate a simple autoregression.

4.2 ARIMA and seasonal ARIMA

ARIMA models combine autoregression, differencing, and moving-average terms. Differencing transforms a non-stationary series, for example:

A seasonal difference with period m is y_t − y_t−m. Seasonal ARIMA extends the model with seasonal autoregressive, differencing, and moving-average terms. These models are useful when temporal dependence is strong and explanatory variables are limited, but they require careful diagnostics and can struggle with abrupt promotion effects or assortment changes.

4.3 Intermittent demand

Intermittent demand contains many zero observations and occasional positive demands. It occurs with slow-moving spare parts, niche products, or low-volume store–SKU combinations. Standard methods may produce biased or unstable results because the timing and size of demand are separate processes.

Croston-type methods estimate the average demand size and average interval between non-zero demands. Let z_j denote the size of the jth non-zero demand and p_j the interval since the previous non-zero demand. When demand occurs, estimates may be updated as:

The forecast is approximately z_j/p_j. Different Croston variants correct known bias properties. These methods should be compared with simple alternatives such as a zero forecast, empirical mean, or a two-stage model that predicts demand occurrence and positive quantity separately.

Intermittent-demand accuracy should not rely solely on percentage errors, because division by zero is undefined and a single large error can dominate interpretation.

5. Causal and Regression Forecasting

Time-series models explain demand using its own history. Causal models add variables believed to influence demand:

Here the coefficients estimate associations conditional on the included variables. A coefficient is not automatically a causal effect. Causal interpretation requires attention to confounding, reverse causality, treatment assignment, and model specification.

5.1 Price and promotion variables

Price may be represented as a continuous variable, a log price, or a relative price index. Promotions can include a binary indicator, discount depth, display, feature placement, campaign type, or lagged and lead effects. A log-log model provides an elasticity interpretation:

where P_t is price and β₁ is the approximate percentage change in demand associated with a 1% change in price, under the model assumptions. Log models require positive values and require careful treatment of zero sales.

Promotion effects are rarely isolated. A promoted product may cannibalize another product, attract incremental category demand, or cause customers to stockpile. Therefore, evaluating only promoted-unit uplift can overstate profitability.

5.2 Lagged and distributed effects

Advertising or promotions may influence demand over multiple periods. A distributed-lag model can include:

The coefficients represent contemporaneous and lagged associations. The analyst must ensure that all lagged variables are available at forecast creation time.

5.3 Machine learning

Tree-based ensembles, gradient boosting, regularized regression, and neural networks can model nonlinearities and interactions. Common features include lagged demand, rolling means, price, promotion, store attributes, calendar variables, weather, product age, and category-level demand.

Machine learning is particularly useful when many related series and predictors are available. However, complexity does not guarantee accuracy. Models can overfit, extrapolate poorly outside the training range, and become difficult to explain. Time-based cross-validation, feature-availability controls, model monitoring, and benchmark comparison are essential.

A useful design is a global model trained across many products and stores, with product and location identifiers represented through carefully managed features. Global models can share information across sparse series, but they must not erase genuine differences among stores or categories.

6. Hierarchical Forecasting and Reconciliation

Retail decisions are made at multiple levels, but independently generated forecasts may not add up. If store forecasts sum to 9,800 units while the category forecast is 10,000, planners need a coherent set of numbers.

6.1 Bottom-up and top-down methods

• Bottom-up: forecast the most detailed item–store series and aggregate upward. This preserves local detail but may be noisy for sparse series.
• Top-down: forecast an aggregate and allocate it to lower levels using historical proportions. This is stable at the top level but may miss changing local patterns.
• Middle-out: forecast an intermediate level and distribute upward and downward.
• Optimal reconciliation: adjust base forecasts using an estimate of forecast-error covariance so that forecasts satisfy aggregation constraints while retaining useful information from several levels.

Let b be the vector of base forecasts and S the summing matrix that maps bottom-level series to all hierarchy levels. A coherent forecast can be represented as:

where g is a vector of reconciled bottom-level forecasts. The exact reconciliation estimator depends on the selected method and covariance assumptions. The essential requirement is that the resulting forecasts obey the hierarchy.

6.2 Grouped structures

Some retail structures are not strictly hierarchical. A product may be classified simultaneously by brand, category, price tier, and channel. These are grouped time series. Reconciliation must respect multiple aggregation constraints, which increases computational and governance complexity.

7. New Products and Product Lifecycle Changes

A new product has little or no own-history data. Forecasting should therefore use analogues, attributes, market research, planned distribution, and category-level information.

A practical approach is:

1. Define the launch scope, stores, channel, price, and planned promotion.
2. Identify comparable products based on category, brand, pack size, price tier, and customer segment.
3. Estimate an initial adoption curve or analogous sales profile.
4. Adjust for distribution, cannibalization, seasonality, and marketing support.
5. Update rapidly using early sales while distinguishing availability from true demand.

Analogue selection should be systematic rather than based only on managerial intuition. Hierarchical or global machine-learning models can partially pool information across products, reducing extreme forecasts for sparse items. Bayesian approaches are also useful because they formally combine prior information with emerging data, although their implementation requires appropriate modelling expertise.

Discontinued products create the opposite problem. Their historical sales should not be blindly used to forecast continuing assortment. Product substitutions and successor items must be represented explicitly.

8. Measuring Accuracy, Bias, and Uncertainty

No single metric is universally best. The metric must reflect the business decision, volume distribution, and cost of errors.

8.1 Error definitions

Let actual demand be y_t and forecast be ŷ_t. Forecast error is:

A positive error means demand exceeded the forecast, indicating underforecasting.

Mean absolute error is:

It is expressed in the original demand units and is easy to interpret.

Root mean squared error is:

RMSE penalizes large errors more strongly than MAE and is useful when extreme misses are particularly costly.

Mean absolute percentage error is:

MAPE is undefined when y_t = 0 and can be distorted when actual demand is small. It should not be the sole metric for intermittent demand.

Weighted absolute percentage error, often called WAPE, is:

It is defined when total actual demand is positive and gives greater influence to high-volume observations. It can hide poor performance on low-volume items.

Mean error is:

It measures average signed error and is a direct indicator of bias. A positive ME means systematic underforecasting; a negative ME means systematic overforecasting.

A scaled measure such as mean absolute scaled error compares model error with a naïve benchmark. For a non-seasonal benchmark, one form is:

The denominator is calculated on the training data. A value below 1 indicates improvement over the corresponding naïve benchmark under the selected scaling scheme.

8.2 Worked example

Suppose actual weekly demand is 100, 120, 80, and 100 units, while forecasts are 90, 130, 100, and 95 units. Errors are 10, −10, −20, and 5 units.

The MAE is:

The ME is:

Thus, the forecasts overstate demand by 3.75 units on average. The absolute error may be acceptable operationally, but the negative bias could create excess inventory if it persists.

8.3 Accuracy is decision-specific

A forecast with lower RMSE may be preferable for capacity planning, while a forecast with lower weighted error may be preferable for high-volume replenishment. Stock-out costs, markdown costs, holding costs, and service-level targets should guide evaluation. Accuracy should be reported by horizon, category, store cluster, lifecycle stage, and demand class—not only as one enterprise average.

9. Forecast Bias and the Human–Model Interface

Bias is systematic, not merely random error. Common causes include optimistic sales plans, unrecorded stock-outs, deliberate safety buffers embedded in the forecast, promotions that were cancelled, and incentives that reward avoiding stock-outs but ignore excess inventory.

Bias monitoring should include:

• mean error and cumulative error;
• bias by product, store, category, and horizon;
• forecast versus replenishment overrides;
• performance before and after promotions;
• changes in bias following model releases.

Human judgement can improve forecasts when it adds information unavailable to the model, such as a confirmed competitor closure or a supply disruption. It can harm forecasts when overrides are unaudited, systematically optimistic, or applied inconsistently. A robust process records the reason, magnitude, author, timing, and subsequent outcome of every material override.

Forecasts should not be adjusted merely to meet a budget or target. Planning and forecasting are related but distinct: a forecast estimates likely demand, whereas a plan may express an aspirational or constrained objective.

10. Applied Case: FreshBite Grocery’s Promotional Forecasting

Background

FreshBite Grocery operates 85 supermarkets and an online channel. Its ready-meal category experiences strong weekday patterns, payday effects, and promotional peaks. The company historically used the previous week’s sales as the forecast for each store–SKU combination.

Problem

The baseline produced frequent stock-outs during two-for-one promotions and excessive waste after promotions ended. Analysts also discovered that several stores had zero recorded sales because products were unavailable. The category manager requested a more accurate and coherent forecast without creating an opaque system that store managers could not understand.

Analysis

The analytics team first separated sales from availability. It added stock-on-hand, delivery, and stock-out indicators, then flagged observations where demand was censored. Exploratory analysis showed a weekly seasonal pattern and different store-level responses to promotions. A simple comparison was conducted among:

1. a naïve forecast;
2. a seasonal-naïve forecast using the corresponding day of the previous week;
3. exponential smoothing with weekday effects;
4. a regression model containing price discount, promotion type, payday, holiday, store, and lagged demand;
5. a gradient-boosting model using the same variables plus rolling demand and category features.

Rolling-origin validation was used, with evaluation separated into ordinary weeks and promotional weeks. WAPE, MAE, ME, and stock-out rates were reported. The team avoided using post-period promotion information and excluded periods in which the item was unavailable from naïve demand comparisons.

Approach and solution

The selected architecture combined a transparent regression model for the baseline with a machine-learning residual model. The regression estimated systematic effects of calendar and promotion variables, while the residual model captured nonlinear store and product interactions. Forecasts were reconciled so that store-level forecasts summed to category and enterprise totals.

For new ready-meals, analogue products supplied an initial forecast. For low-volume items, a two-stage intermittent-demand method estimated the probability of a sale and the quantity conditional on a sale. The system generated both a point forecast and a prediction interval for replenishment planning.

Store managers could override forecasts only through a governed interface. Each override required a reason code such as local event, competitor closure, supply constraint, or data correction.

Outcome

In the pilot, the combined model reduced promotional-week absolute error relative to the previous-week baseline and reduced negative forecast bias. However, the largest operational improvement came from correcting stock-out data and using promotion calendars reliably. The company did not deploy the machine-learning model as an autonomous decision-maker; it used it as one component of a monitored forecasting process.

Learning points

The case illustrates five principles. First, demand data must be distinguished from constrained sales. Second, evaluation must reflect the conditions in which the forecast will be used. Third, promotions require explicit variables and separate validation. Fourth, hierarchy reconciliation prevents conflicting numbers. Fifth, governance and override records are as important as algorithm selection.

11. Implementation, Technology, and Governance

A production forecasting platform normally includes data ingestion, feature engineering, model training, forecast generation, reconciliation, accuracy monitoring, workflow approval, and integration with ordering or planning systems. Cloud computing and automated pipelines can support thousands of item–location series, but automation should not obscure responsibility.

Artificial intelligence is useful for feature discovery, global forecasting, scenario simulation, and anomaly detection. Generative AI may assist with explanations or natural-language summaries, but it should not be trusted to invent demand drivers, alter forecasts without traceability, or expose confidential customer and commercial data.

Governance should address:

• data ownership and quality controls;
• versioning of models and features;
• reproducibility of forecasts;
• access controls and audit logs;
• privacy and lawful use of customer-level data;
• fairness when forecasts influence product availability across communities;
• explainability for material decisions;
• monitoring for drift and performance deterioration;
• human accountability for overrides and deployment.

Forecasts can produce social consequences. A model trained on historical purchasing may allocate less inventory to communities whose prior sales were constrained by poor availability. Dynamic pricing and personalized promotions may also create fairness concerns. Analysts should test for systematic disparities and distinguish low observed sales from low underlying demand.

12. Best-Practice Forecasting Process

A high-quality retail forecasting process follows a disciplined sequence:

1. Define the decision, target variable, horizon, and unit of analysis.
2. Map the product, location, channel, and time hierarchy.
3. Audit sales, inventory, promotions, prices, returns, and missingness.
4. Establish naïve and seasonal-naïve benchmarks.
5. Explore trend, seasonality, intermittency, structural breaks, and outliers.
6. Select candidate models appropriate to the demand pattern and data availability.
7. Use rolling-origin validation and prevent leakage.
8. Evaluate both statistical accuracy and operational outcomes.
9. Reconcile forecasts across required hierarchy levels.
10. Produce uncertainty information for inventory and capacity decisions.
11. Govern judgemental overrides and document assumptions.
12. Monitor bias, drift, data quality, and business impact after deployment.

The best model is not necessarily the most complex one. It is the model or ensemble that produces sufficiently accurate, unbiased, stable, explainable, and actionable forecasts at an acceptable total cost.

Chapter Summary

Retail demand forecasting estimates future demand to support buying, replenishment, pricing, promotion, labour, capacity, and financial decisions. Observed sales may be censored by stock-outs and therefore cannot automatically be treated as true demand. Demand series may contain level, trend, seasonality, cycles, autocorrelation, intermittent zeros, and external effects.

Naïve and seasonal-naïve forecasts provide essential benchmarks. Moving averages and exponential smoothing are useful for stable series, while autoregressive and ARIMA models capture temporal dependence. Regression and causal models incorporate price, promotion, calendar, weather, and other drivers, although association must not be confused with causation. Machine-learning models can capture nonlinear relationships but require time-aware validation and strong governance.

Hierarchical forecasting ensures that forecasts at store, region, channel, category, and enterprise levels are coherent. Intermittent demand requires methods that distinguish demand occurrence from demand size. New products require analogues, attributes, market knowledge, and rapid updating. Accuracy should be assessed using multiple measures, including MAE, RMSE, WAPE, scaled error, and signed bias. No metric is universally appropriate.

Effective forecasting is a socio-technical process. Data quality, promotion execution, inventory availability, human overrides, governance, ethical review, and operational integration are as important as model choice. Forecasts should be treated as evidence for decisions, accompanied by uncertainty and continuously monitored after deployment.

Key Terms

Demand forecasting: Estimating future customer demand over a stated horizon.

Observed sales: Transactions recorded during a period, which may be constrained by inventory availability.

Stock-out: A condition in which inventory is unavailable to satisfy demand.

Time series: A sequence of observations ordered in time.

Seasonality: A recurring pattern associated with a known calendar frequency.

Trend: A persistent upward or downward movement in the underlying demand level.

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

Seasonal-naïve forecast: A forecast based on the observation from the corresponding previous season.

Exponential smoothing: A forecasting family that weights recent observations more heavily through recursive updating.

Autoregression: A model relating current demand to previous demand values.

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

Causal forecast: A forecast using explanatory variables believed to influence demand.

Feature leakage: The use of information unavailable at the time a forecast would actually be made.

Hierarchical forecasting: Forecasting across related aggregation levels such as SKU, store, region, and enterprise.

Forecast reconciliation: Adjusting forecasts so that they satisfy aggregation constraints.

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

Forecast bias: Systematic overforecasting or underforecasting.

WAPE: Weighted absolute percentage error based on total absolute error divided by total actual demand.

Rolling-origin validation: Repeated time-ordered training and testing using successive forecast origins.

Prediction interval: A range intended to contain a future observation with a stated level of uncertainty.

Review Questions

Short Answer Questions

1. Why can observed sales understate true demand in a retail store?
2. State the difference between a naïve forecast and a seasonal-naïve forecast.
3. What does a positive mean forecast error indicate when error is defined as actual demand minus forecast demand?
4. Why is MAPE unsuitable as the sole metric for series containing many zero observations?
5. What is the purpose of forecast reconciliation?

Long Answer Questions

1. Explain the major components of retail demand and discuss how each affects model selection.
2. Compare exponential smoothing, ARIMA, regression, and machine-learning approaches for retail forecasting.
3. Explain how stock-outs, promotions, and product substitutions should be represented in a forecasting system.
4. Discuss the advantages and limitations of bottom-up, top-down, and reconciled hierarchical forecasting.
5. Design a governance framework for human overrides and post-deployment forecast monitoring.

Analytical / Application Questions

1. A retailer has daily demand with a strong weekly pattern, frequent promotions, and several stock-out days. Design a validation and modelling strategy, including the target variable, benchmarks, features, and evaluation metrics.
2. A new private-label product has no sales history but will launch in 40 stores at a price below the leading brand. Explain how you would construct its first eight-week forecast and update it after launch.
3. A model has MAE lower than the incumbent model but consistently negative mean error. Analyse the managerial risks and recommend whether the model should be deployed unchanged.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand from observed sales?

A. Demand is always lower than observed sales.

B. Observed sales may be lower than demand when inventory is unavailable.

C. Demand includes only online transactions.

D. Observed sales automatically include lost sales.

Correct Answer: B. Observed sales may be lower than demand when inventory is unavailable.

Question 2

A seasonal-naïve forecast for monthly data with annual seasonality normally uses which observation for the forecast of next January?

A. The previous January

B. The previous December

C. The average of all Januaries and Februaries

D. The most recent month only

Correct Answer: A. The previous January

Question 3

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

A. Systematic overforecasting

B. Systematic underforecasting

C. No forecast error

D. Seasonality has been removed

Correct Answer: B. Systematic underforecasting

Question 4

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

A. MAE

B. RMSE

C. MAPE

D. WAPE

Correct Answer: A. MAE

Question 5

Why should random train-test splitting generally be avoided for retail time series?

A. It always reduces sample size to zero.

B. It can allow future information to influence model training.

C. It prevents calculation of revenue.

D. It eliminates seasonality from the data.

Correct Answer: B. It can allow future information to influence model training.

Question 6

Which situation is most characteristic of intermittent demand?

A. High sales every day with no zeros

B. A constant upward trend with no variation

C. Many zero-demand periods interrupted by positive demand

D. Demand observed only during promotions

Correct Answer: C. Many zero-demand periods interrupted by positive demand

Question 7

What is the principal purpose of hierarchical forecast reconciliation?

A. To remove all uncertainty from forecasts

B. To ensure forecasts at related levels add up consistently

C. To replace all store forecasts with one enterprise forecast

D. To guarantee that every SKU has identical demand

Correct Answer: B. To ensure forecasts at related levels add up consistently

Question 8

In the worked example, what was the mean error for errors of 10, −10, −20, and 5 units?

A. −3.75 units

B. 3.75 units

C. 11.25 units

D. 45 units

Correct Answer: A. −3.75 units

Question 9

Which practice best prevents promotion-related data leakage?

A. Using the final realised promotion outcome before the forecast date

B. Using only promotion information available when the forecast is created

C. Removing all promotion variables from every model

D. Replacing promotion effects with random numbers

Correct Answer: B. Using only promotion information available when the forecast is created

Question 10

Why can a lower MAE model still be unsuitable for deployment?

A. MAE is never a valid metric.

B. The model may have systematic bias or poor performance on critical segments.

C. Lower MAE proves the model is overfitted.

D. MAE can only be calculated for annual data.

Correct Answer: B. The model may have systematic bias or poor performance on critical segments.