Demand Data, Statistical Foundations and Data Quality

Introduction

Demand forecasting is only as reliable as the data, definitions and statistical reasoning on which it is based. In operations management, demand data are not simply a column of historical sales. They are observations generated by customers, selling systems, inventory policies, prices, promotions, competitors, weather, economic conditions and measurement processes. A recorded sale may therefore represent observed demand, but not necessarily the demand that would have occurred if the product had always been available at the desired price.

This chapter develops the foundations required to convert operational data into analytically useful evidence. It examines demand-data structures, exploratory analysis, visualization, probability, sampling, correlation and distributions. It also addresses missing values, outliers, aggregation, bias, data integration and governance. The central argument is that forecasting is partly a statistical task and partly a measurement-design task: before selecting a forecasting model, analysts must establish what has been measured, at what level, under which conditions and with what limitations.

Learning objectives

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

1. Distinguish demand, sales, orders, shipments and inventory-constrained observations.
2. Identify transactional, market and external data relevant to forecasting.
3. Select suitable levels of granularity and aggregation for managerial decisions.
4. Use descriptive statistics, visualizations, probability and distribution concepts to explore demand.
5. Explain sampling, sampling error, bias, correlation and causation.
6. Diagnose missing values, outliers, duplicates, inconsistent units and other quality problems.
7. Prepare a documented and governed dataset for forecasting and business analytics.
8. Evaluate the ethical, operational and managerial consequences of poor data quality.

1. The nature of demand data

1.1 Demand is not identical to sales

Demand is the quantity customers are willing and able to purchase under specified conditions. Sales are the transactions actually recorded. The two are equal only when the product is available, the customer can access it, prices and terms are observed, and the transaction system captures the event correctly.

Suppose a store sells 80 units on a day when 20 additional customers leave because the product is out of stock. Recorded sales are 80, but realized demand may be approximately 100. If a forecasting model treats 80 as unconstrained demand, it may systematically underforecast future requirements.

Important operational measures include:

A demand record should therefore be accompanied by contextual fields such as price, promotion, availability, lead time, channel, location and event status.

1.2 Data sources

Transactional data describe operational events. Typical fields include transaction identifier, timestamp, product, location, quantity, price, discount, customer segment, channel, order status and payment status. These data are often high volume and detailed, but they may contain duplicates, reversals, cancellations and inconsistent product codes.

Market data describe the competitive and commercial environment. Examples include competitor prices, market share, category volume, distribution coverage, customer surveys and search intensity. Market data may be sampled, delayed, proprietary or measured using different definitions from internal data.

External data capture conditions outside the firm. Examples include temperature, rainfall, holidays, school calendars, exchange rates, interest rates, unemployment, mobility and regulatory changes. External variables are useful only when their timing, geography and causal relevance are understood.

A useful data inventory records the source, owner, update frequency, grain, units, historical coverage, access restrictions, quality risks and intended use of every dataset.

1.3 The grain of a dataset

The grain is the level represented by one row. A row might represent one transaction, one product per store per day, one customer per month or one category per region per quarter. Confusing grains is a major source of analytical error.

For example, joining a transaction-level table to a daily weather table is appropriate if every transaction can be assigned to its date and location. Joining a monthly revenue table to a daily table without a clear allocation rule can create artificial repetition and inflate totals.

The grain should be stated explicitly, for example: “one row represents one SKU–store–day combination.” Every measure must then be interpreted at that level. Aggregation should follow the decision context. Replenishment may require SKU–location–day data, whereas capacity planning may require product-family–region–week data.

Aggregation reduces noise and computational burden, but it can hide important variation. Disaggregated data reveal intermittent demand, local promotions and stockouts; aggregated data may reveal stable strategic patterns. The appropriate level is the lowest level at which the decision can be made reliably and economically.

2. Exploratory data analysis

Exploratory data analysis (EDA) is the systematic examination of data before formal modelling. It combines numerical summaries, visual inspection, domain knowledge and data-quality checks. EDA is not merely a preliminary presentation exercise: it can reveal incorrect assumptions, structural breaks and measurement failures.

2.1 Descriptive statistics

For observations x₁, x₂, …, xₙ, the arithmetic mean is:

Here, n is the number of observations and x̄ has the same units as xᵢ. The mean is appropriate when values are numerical and additive, but it is sensitive to extreme observations.

The sample variance is:

The sample standard deviation is s = √s². The denominator n − 1 is used when estimating population variability from a sample under standard independent-sampling assumptions. Variance has squared units; standard deviation returns to the original units.

The median is the middle ordered observation, or the average of the two middle observations when n is even. The median is more resistant to extreme values than the mean. Quartiles divide ordered data into four parts, while the interquartile range is:

where Q₁ and Q₃ are the first and third quartiles. A common exploratory rule identifies observations below Q₁ − 1.5IQR or above Q₃ + 1.5IQR as potential outliers. This is a screening rule, not proof that a value is erroneous.

The coefficient of variation is:

provided x̄ is nonzero and meaningful. It is dimensionless and useful for comparing relative variability across series measured on different scales, but it becomes unstable when the mean is close to zero and is inappropriate when the variable can take important negative values.

Worked example: summary statistics

Weekly demand for a product is 8, 10, 10, 12 and 20 units. The mean is (8 + 10 + 10 + 12 + 20) ÷ 5 = 12 units. The ordered middle value gives a median of 10 units. Deviations from the mean are −4, −2, −2, 0 and 8; their squared sum is 16 + 4 + 4 + 0 + 64 = 88. Thus sample variance is 88 ÷ 4 = 22 units² and sample standard deviation is approximately 4.69 units. The mean exceeds the median because the unusually high value of 20 pulls the mean upward. An analyst should investigate the 20-unit observation rather than automatically delete it.

2.2 Visual analysis

A time-series plot places time on the horizontal axis and demand on the vertical axis. It can reveal trend, seasonality, level shifts, stockout periods, calendar effects and changing variance. A line chart is appropriate for ordered time observations; a bar chart is often better for comparing a small number of categories; a histogram displays the distribution of numerical values; a box plot summarizes median, quartiles and potential extreme observations; and a scatterplot examines the relationship between two numerical variables.

Good visualizations include units, time zone, aggregation level, data source and relevant annotations. A chart without these details can create false confidence. For example, a monthly chart may conceal a weekly promotion spike, while a chart of sales may conceal days with zero inventory.

A useful EDA sequence is:

1. Inspect row counts, column types, identifiers and date ranges.
2. Confirm the grain and check duplicate keys.
3. Summarize missingness and impossible values.
4. Plot demand over time and across products, locations and channels.
5. Compare sales with availability, price and promotion status.
6. Examine distributions and relationships.
7. Investigate unusual observations using source records and business knowledge.

3. Probability and distributions in demand analysis

Probability provides a language for uncertainty. A random variable represents an uncertain numerical outcome. A discrete random variable takes countable values, such as the number of orders in an hour. A continuous random variable can take values across an interval, such as delivery time, although recorded measurements are usually discretized.

For an event A, its probability P(A) lies between 0 and 1. For complementary events:

For mutually exclusive events A and B:

For general events:

Conditional probability is:

provided P(B) > 0. It expresses the probability of A given that B has occurred. In operations, the probability of high demand given a promotion may differ greatly from unconditional high-demand probability.

Two events are independent if P(A ∩ B) = P(A)P(B), equivalently P(A | B) = P(A) when the probabilities are defined. Independence should not be assumed merely because two variables are recorded in separate systems.

3.1 Common distributions

The Bernoulli distribution represents one trial with outcome 1 or 0, such as whether a customer converts. A binomial distribution counts successes in n independent trials with a constant success probability p. Its expected value is np and its variance is np(1 − p). These assumptions are often only approximate in customer and demand settings because customers may influence one another and probabilities may change over time.

The Poisson distribution is frequently used as a starting point for counts of events in a fixed interval. If X has a Poisson distribution with rate λ, then:

Its mean and variance are both λ. The model assumes a stable event rate and suitable independence conditions. If demand variance is much larger than its mean, overdispersion may indicate that a negative binomial model, a mixture model or a time-varying process is more suitable.

The normal distribution is continuous, symmetric and described by mean μ and standard deviation σ. It is useful for some aggregated errors and measurements, but raw demand is often nonnegative, skewed, intermittent or seasonal. A normal approximation should therefore be justified rather than applied automatically.

A lognormal distribution can represent positive, right-skewed quantities when the logarithm of the variable is approximately normal. It is sometimes suitable for transaction values or demand sizes, but zero demand requires special treatment.

3.2 Expected value and risk

For a discrete random variable X with possible values xᵢ and probabilities pᵢ:

Expected value is a probability-weighted average, not a guarantee of the outcome. If demand is 100 units with probability 0.6 and 200 units with probability 0.4, expected demand is 0.6×100 + 0.4×200 = 140 units. A manager should not interpret 140 as necessarily occurring; it supports planning under uncertainty.

4. Sampling, estimation and uncertainty

A population is the complete set of units relevant to a question; a sample is the observed subset. A census observes all population units, while sampling observes some. Transaction systems may appear to provide a census but still omit offline purchases, unrecorded cancellations or customers who never reach the transaction stage.

A probability sample gives units a known, nonzero selection probability. Simple random sampling gives each unit an equal selection probability. Stratified sampling divides the population into groups, such as regions or customer segments, and samples within each group. Cluster sampling selects groups, such as stores, and observes units within selected groups. Convenience samples are easy to obtain but can be strongly biased.

The standard error of a sample mean under independent simple random sampling is approximately:

where s is sample standard deviation and n is sample size. The formula assumes observations are reasonably independent and the sample design is compatible with the calculation. Time-series observations often violate independence because nearby periods can be correlated.

A large sample can reduce random sampling error but cannot remove systematic bias. For example, a very large sample of online customers may still misrepresent customers who purchase only in physical stores. Analysts should report sampling frame, selection method, response rate, weighting and limitations.

5. Correlation, dependence and causation

Covariance measures joint variation. The Pearson correlation coefficient is:

For variables with nonzero variation, r lies between −1 and 1. A positive value indicates that high values of one variable tend to occur with high values of the other; a negative value indicates the opposite linear tendency. Correlation is unit-free, but it measures association, not causation.

Correlation can be distorted by outliers, nonlinear relationships, aggregation and common trends. Ice-cream sales and temperature may correlate because temperature affects both, whereas two unrelated trending variables may correlate merely because time increases. A promotion may be associated with higher sales because managers target promotions at products that are already expected to sell well, creating confounding by promotional allocation.

Scatterplots, lagged relationships, stratification and controlled experiments can improve interpretation. Causal claims require a credible design, such as randomization, a natural experiment, a difference-in-differences design or a carefully justified causal model. Forecasting may benefit from a predictive relationship even when causal interpretation is weak, but managerial decisions require clarity about that distinction.

6. Data quality and preparation

Data quality is multidimensional. Common dimensions include accuracy, completeness, consistency, timeliness, validity, uniqueness, integrity and fitness for purpose. A dataset may be accurate at the transaction level but not fit for a monthly capacity decision if the required location or product hierarchy is absent.

6.1 Missing values

Missingness should be classified rather than treated uniformly. A value may be missing because it was not applicable, not recorded, unavailable at extraction, suppressed for privacy or genuinely zero. A blank sales quantity is not necessarily zero demand.

Three conceptual mechanisms are commonly distinguished:

• Missing completely at random: missingness is unrelated to observed and unobserved values.
• Missing at random: missingness depends on observed information after conditioning.
• Missing not at random: missingness depends on the unobserved value or an unmeasured process.

Methods include deletion, constant or rule-based replacement, mean or median imputation, regression imputation, multiple imputation and model-based treatment of missingness. Each method introduces assumptions. For example, replacing missing demand with zero may be appropriate when a validated system explicitly records “no demand,” but dangerous when blank means “data unavailable.” Missingness indicators and sensitivity analyses can show whether conclusions change under alternative treatments.

6.2 Outliers and unusual observations

An outlier is an observation that differs substantially from the rest of a dataset. It may be an error, a rare but valid event, a structural change or evidence of a different population. The correct response is investigation: compare with source records, check units and timestamps, examine related variables and consult process owners. Deleting an observation merely because it harms model fit can remove precisely the events that operations must plan for.

Robust alternatives include medians, trimmed summaries, winsorization, robust regression and explicit event flags. Any alteration should be recorded in a data lineage log, with original values preserved.

6.3 Duplicates, invalid values and reconciliation

Duplicate identifiers can arise from repeated extraction, partial updates or legitimate line items. A transaction ID may not be unique if an order contains several products; the true key may be transaction ID plus line number. Validation rules should identify negative quantities, impossible dates, invalid product codes, prices below zero, inconsistent currencies and quantities that violate packaging units.

Reconciliation compares totals across systems. For example, the sum of invoice lines should be compared with the general ledger or order-management total after accounting for tax, returns and timing. Differences should be explained, not silently adjusted.

6.4 Data integration

Integration requires common identifiers, definitions and time conventions. Product hierarchies may change; stores may open, close or relocate; currencies require exchange-rate rules; and timestamps may use different time zones. A master-data table should provide stable keys and effective dates for changes. Historical restatement may be necessary when a product is reclassified, but both original and restated versions should be retained where auditability matters.

7. Bias and measurement risk

Bias is systematic deviation from the quantity of interest. Selection bias occurs when observed records differ systematically from the target population. Survivorship bias occurs when failed products, closed stores or lost customers are excluded. Confirmation bias occurs when analysts search selectively for evidence supporting an expected result. Aggregation bias occurs when group-level relationships are incorrectly applied to individuals or local units.

Measurement bias arises when the recording process changes the observed value. A promotion can increase recorded sales while reducing margin; a stockout can reduce sales while latent demand remains high; a change in barcode scanning can create an apparent structural break. Analysts should distinguish the business phenomenon from the measurement mechanism.

Fairness is also relevant. Forecasts used for staffing, credit, service access or inventory allocation may systematically disadvantage certain locations or customer groups if historical data reflect unequal access or prior discriminatory policies. Governance should require purpose specification, access control, privacy protection, explainability appropriate to the decision and monitoring for harmful outcomes.

8. An analytical preparation workflow

A defensible workflow links business purpose to data design:

1. Define the decision. Specify what will be forecast, for whom, at what horizon and at what cost of error.
2. Define the target. State whether the target is sales, unconstrained demand, orders, shipments or consumption.
3. Specify grain and keys. Document one-row meaning, identifiers, time zone, units and aggregation rules.
4. Inventory sources. Record ownership, provenance, refresh rate, coverage and restrictions.
5. Profile the data. Measure missingness, duplicates, ranges, cardinality, distributions and date gaps.
6. Reconcile and validate. Compare with trusted totals and enforce business rules.
7. Investigate anomalies. Use source records and domain experts; preserve original values.
8. Construct analytical variables. Create lagged demand, prices, promotions, calendar indicators and external variables only using information available at the forecast time.
9. Document assumptions. Record imputation, exclusions, transformations and version history.
10. Monitor after deployment. Track data drift, missingness, schema changes, forecast performance and operational impacts.

Data leakage must be prevented. A feature is leaked when it contains information unavailable at the time the forecast would have been made. For example, using end-of-week inventory status to predict demand at the beginning of that week can produce deceptively strong validation results.

9. Illustrative case: FreshMart chilled-food forecasting

Background

FreshMart, a fictional regional supermarket, experienced repeated stockouts for chilled salads in summer and high waste in winter. Its analytics team possessed point-of-sale transactions, store inventory, promotions, local temperature and competitor-price data.

Problem

The first forecasting model used daily sales by product and store. It performed well statistically in ordinary weeks but underforecast high-temperature periods and treated stockout days as low demand. The business needed a more credible demand dataset for replenishment.

Analysis

The team first defined the grain as SKU–store–day. It found that 7% of records had missing inventory status, 2% contained duplicate transaction lines, and several stores reported timestamps in local time while the warehouse system used UTC. A distribution review showed strong right skew and many zero-demand observations for slow-moving products. A time plot revealed that sales dropped sharply on days with zero ending inventory.

The team compared sales with availability and classified observations into available, constrained and unknown states. It retained raw sales, but created an analytical demand-status field. Missing inventory status was not automatically imputed as available; records were flagged for sensitivity analysis. Promotional prices were aligned to the date and store, and temperature was joined using store geography and local calendar date.

Correlation between temperature and sales was positive, but the team avoided claiming that temperature alone caused all increases. Promotions were more frequent during holidays, so holiday and promotion indicators were examined jointly. Validation used rolling time periods rather than random row splitting, thereby respecting temporal ordering.

Approach and solution

FreshMart produced two related targets: observed sales for financial reporting and adjusted demand for replenishment analysis. Constrained observations were excluded from a baseline unconstrained-demand model only when the evidence of stockout was reliable; otherwise, they were retained with an uncertainty flag. Missing and anomalous records were documented in a data-quality register.

The company introduced automated checks for duplicate keys, impossible prices, missing store identifiers, unexplained total differences and abrupt changes in the proportion of constrained observations. Forecast users received both a point estimate and an explanation of data limitations.

Outcome

The revised process reduced systematic underforecasting during hot periods and made the source of uncertainty visible to planners. It did not eliminate forecast error, because weather forecasts, customer substitution and managerial reactions remained uncertain. Its principal improvement was methodological: the organization stopped confusing recorded sales with unconstrained demand and established a repeatable data-governance process.

Learning points

1. A forecast target must be defined before a model is selected.
2. Inventory information is essential for interpreting sales.
3. Time zones, keys and grain can materially alter results.
4. Extreme observations require investigation, not automatic deletion.
5. Data-quality controls are operational controls because they affect replenishment, waste and service.

10. Technology and contemporary practice

Modern analytics platforms support automated profiling, schema monitoring, data catalogs, lineage tracking and feature stores. Machine-learning systems can detect unusual patterns, but automated detection does not establish whether an observation is erroneous or meaningful. Human review remains important for structural breaks, policy changes and ethical consequences.

Artificial intelligence can assist with anomaly triage, entity matching and natural-language documentation. However, AI-generated transformations may reproduce historical bias, infer sensitive attributes or conceal assumptions. Organizations should apply access controls, privacy-by-design principles, retention limits and audit trails. Synthetic data can support development where privacy is important, but synthetic records do not automatically preserve the joint distributions or rare events needed for reliable forecasting.

Contemporary data governance increasingly emphasizes data contracts: explicit agreements between producers and consumers concerning schema, definitions, quality thresholds, refresh times and change notification. For forecasting teams, this is especially valuable because a silent change in product hierarchy or inventory coding can appear as a sudden change in customer demand.

Challenges and limitations

Demand data are inherently observational and affected by managerial intervention. Prices, promotions, inventory and assortment are not randomly assigned. Historical patterns may fail after a competitor enters, a regulation changes, a supply disruption occurs or customer behavior shifts. Distributional assumptions are approximations, and correlations can be unstable across locations and periods.

No cleaning rule is universally correct. Removing outliers may improve statistical stability but remove genuine peaks. Imputation may preserve sample size but create artificial certainty. Aggregation can improve signal-to-noise ratio but conceal local constraints. Consequently, preparation decisions should be evaluated against the decision objective, operational consequences and sensitivity to assumptions.

Best-practice checklist

• Define demand and the forecast target operationally.
• State the dataset grain and primary key.
• Preserve raw data and maintain transformation lineage.
• Separate zero, missing, not applicable and unknown values.
• Check availability before interpreting low sales.
• Use visualizations alongside numerical summaries.
• Investigate outliers with domain evidence.
• Align timestamps, units, currencies and product hierarchies.
• Avoid random validation splits for temporal forecasting problems.
• Test for leakage and document information availability at forecast time.
• Report uncertainty, sampling limitations and potential bias.
• Monitor quality after deployment rather than treating preparation as a one-time task.

Chapter Summary

Demand forecasting begins with a defensible representation of demand. Sales, orders, shipments and consumption are related but distinct measures, and stockouts can make sales an underestimate of underlying demand. Transactional, market and external data must be evaluated according to their grain, timing, definitions, provenance and decision relevance.

Exploratory analysis uses means, medians, variation measures, distributions and visualizations to reveal patterns and problems. Probability and expected value provide a framework for uncertainty, while sampling concepts explain estimation error and bias. Correlation describes association but does not establish causation. Distributional assumptions must be matched to the nature of the data, especially when demand is skewed, intermittent or overdispersed.

Data-quality work includes handling missing values, investigating outliers, identifying duplicates, reconciling totals, integrating identifiers and documenting transformations. Bias can arise from selection, measurement, aggregation and historical decision processes. A rigorous workflow defines the decision and target, confirms grain, profiles and validates data, prevents leakage, documents assumptions and monitors quality over time. Effective analytics therefore combines statistical competence with operational knowledge, governance and ethical judgment.

Key Terms

Demand: Quantity customers are willing and able to purchase under specified conditions.

Observed sales: Transactions recorded as completed sales.

Grain: The level of detail represented by one row in a dataset.

Transactional data: Records of operational events such as orders, sales and shipments.

Market data: Information about competitors, customers, categories and market conditions.

External data: Information originating outside the organization, such as weather or economic indicators.

Exploratory data analysis: Systematic examination of data using summaries, visualizations and domain reasoning.

Mean: Arithmetic average of numerical observations.

Median: Middle value of ordered observations.

Variance: Average squared deviation used to describe dispersion.

Standard deviation: Square root of variance, expressed in the original units.

Interquartile range: Difference between the third and first quartiles.

Probability: Numerical representation of uncertainty for an event.

Conditional probability: Probability of one event given that another event has occurred.

Sampling bias: Systematic difference between the sample and the target population.

Correlation: Standardized measure of linear association between two variables.

Outlier: Observation unusually distant from the main pattern of data.

Data leakage: Use of information unavailable at the time a forecast would have been produced.

Data lineage: Record of data origins, transformations and destinations.

Data governance: Policies and controls governing data quality, access, use and accountability.

Review Questions

Short Answer Questions

1. Why might observed sales be lower than underlying customer demand?
2. What is meant by the grain of a dataset?
3. Distinguish a median from a mean in terms of sensitivity to extreme values.
4. Why does correlation not establish causation?
5. Give two possible meanings of a blank demand value.

Long Answer Questions

1. Explain how transactional, market and external data can be combined for demand analysis. Discuss at least four integration risks.
2. Describe an exploratory data-analysis process for a SKU–store–day demand dataset.
3. Explain the principal forms of sampling and the distinction between random error and systematic bias.
4. Compare the use of normal, Poisson and negative-binomial reasoning for demand data.
5. Develop a data-quality governance framework for a forecasting function.

Analytical / Application Questions

1. A product records daily sales of 40, 42, 41, 0 and 43 units. Inventory was unavailable on the zero-sales day. Explain how you would represent this observation for financial reporting and for unconstrained-demand forecasting.
2. A manager observes a strong positive correlation between promotions and sales. Propose an analysis that examines whether the relationship is causal, addressing confounding and timing.
3. A forecasting dataset combines transaction timestamps in UTC with weather observations recorded in local time. Design a preparation and validation procedure that prevents incorrect matching and documents the assumptions.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes demand from observed sales?

A. Demand is always equal to shipments.

B. Observed sales may be lower than demand when products are unavailable.

C. Demand includes only transactions that have been paid for.

D. Observed sales are independent of price and inventory.

Correct Answer: B. Observed sales may be lower than demand when products are unavailable.

Question 2

A dataset has one row for each product–store–day combination. What does this describe?

A. Its distribution

B. Its sampling frame

C. Its grain

D. Its covariance

Correct Answer: C. Its grain

Question 3

For demand values 8, 10, 10, 12 and 20, what is the arithmetic mean?

A. 10

B. 11

C. 12

D. 14

Correct Answer: C. 12

Question 4

Which measure is generally least sensitive to a single extreme observation?

A. Mean

B. Variance

C. Standard deviation

D. Median

Correct Answer: D. Median

Question 5

If events A and B are mutually exclusive, which relationship is correct?

A. P(A ∪ B) = P(A)P(B)

B. P(A ∪ B) = P(A) + P(B)

C. P(A ∩ B) = P(A) + P(B)

D. P(A | B) = 1

Correct Answer: B. P(A ∪ B) = P(A) + P(B)

Question 6

A Poisson model is commonly used as a starting point for which type of observation?

A. Counts of events in a fixed interval

B. Unbounded negative prices

C. Categorical product names

D. A deterministic time schedule

Correct Answer: A. Counts of events in a fixed interval

Question 7

What is the principal limitation of a correlation coefficient?

A. It can be calculated only for categorical variables.

B. It proves that one variable causes another.

C. It measures only linear association and does not establish causation.

D. It always changes when units are converted.

Correct Answer: C. It measures only linear association and does not establish causation.

Question 8

Which action is most appropriate when an unusually high demand value is detected?

A. Delete it immediately.

B. Replace it with the mean without investigation.

C. Investigate its source and operational context.

D. Treat it as missing.

Correct Answer: C. Investigate its source and operational context.

Question 9

Which situation is an example of data leakage?

A. Using last year's demand to forecast this year's demand

B. Using a promotion indicator known before the forecast date

C. Using end-of-week inventory information to predict demand at the beginning of that week

D. Using a product identifier in a transaction table

Correct Answer: C. Using end-of-week inventory information to predict demand at the beginning of that week

Question 10

If demand is 100 units with probability 0.6 and 200 units with probability 0.4, what is expected demand?

A. 120 units

B. 140 units

C. 160 units

D. 300 units

Correct Answer: B. 140 units