Governance, Ethics, Strategy and the Future of Forecasting

Introduction

Demand forecasting is no longer only a statistical activity performed by analysts. In contemporary organizations, forecasts influence procurement, production, inventory, staffing, pricing, capital allocation, customer service, and financial reporting. As forecasting systems become more automated and incorporate machine learning, external data, generative artificial intelligence, and real-time decision rules, their risks also become more significant. A forecast can be statistically accurate yet operationally harmful if it embeds discrimination, exposes personal data, creates unjustified confidence, or triggers decisions that the organization cannot explain or reverse.

Forecast governance is therefore the system of accountability, controls, documentation, standards, and decision rights used to ensure that forecasting is reliable, lawful, ethical, secure, and strategically aligned. It connects analytical quality with organizational responsibility. This chapter examines model risk management, documentation, monitoring, privacy, cybersecurity, fairness, regulation, analytics capability development, and emerging technologies that may reshape forecasting.

The central managerial principle is that a forecast is not merely a number. It is an evidence-based representation of uncertainty used to support a decision. Governance must therefore address both the model and the decision system surrounding it.

Learning Objectives

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

1. Explain the purpose and components of forecasting governance.
2. Apply a model risk management lifecycle to demand-forecasting systems.
3. Design documentation, validation, monitoring, privacy, security, and fairness controls.
4. Evaluate regulatory and ethical issues in the use of forecasting analytics.
5. Develop a strategic roadmap for analytics capability and organizational adoption.
6. Critically assess generative AI, probabilistic forecasting, digital twins, autonomous planning, and climate-related uncertainty.
7. Distinguish forecast accuracy from decision quality and responsible use.

1. Forecast Governance as an Organizational Capability

Governance establishes who may develop, approve, modify, use, and retire a forecasting model. It also establishes what evidence is required before a model is trusted. Effective governance does not seek to eliminate all uncertainty; it makes uncertainty visible and manageable.

A useful governance architecture has five mutually reinforcing layers:

Governance should be proportional to risk. A model used to suggest reorder quantities for low-value office supplies does not require the same controls as one used to allocate scarce medical products, determine employee staffing, or forecast credit-sensitive customer behavior. Risk generally increases with the scale of consequences, the degree of automation, the sensitivity of data, the difficulty of reversal, and the number of affected stakeholders.

Forecasts, decisions, and accountability

Forecast error is not always model failure. Demand may be affected by an unforeseen strike, regulatory change, competitor action, or extreme weather. Conversely, a numerically accurate model may be badly governed if users apply it outside its validated population or interpret a probabilistic interval as a guarantee.

Accountability should therefore be assigned across the complete decision chain:

• Data owners are responsible for source definitions, permissions, and quality.
• Model developers are responsible for design choices, reproducibility, and technical documentation.
• Independent validators challenge assumptions, test performance, and assess limitations.
• Business owners determine whether the forecast is appropriate for the decision.
• Operators and planners provide contextual information and document overrides.
• Senior management and the board establish risk appetite and resource priorities.

A governance committee should not replace professional judgment. Its role is to make judgment explicit, reviewable, and consistent.

2. Model Risk Management for Forecasting

Model risk is the possibility of adverse consequences resulting from incorrect model design, erroneous implementation, inappropriate use, or misunderstanding of model outputs. In forecasting, model risk includes omitted variables, data leakage, structural breaks, unstable relationships, inappropriate aggregation, excessive automation, and misuse of confidence or prediction intervals.

A practical model risk lifecycle contains the following stages:

1. Inventory and classification: register every material forecasting model, its owner, purpose, data sources, users, and risk tier.
2. Specification: define the target variable, forecast horizon, decision use, population, frequency, constraints, and success criteria.
3. Development: construct the model using controlled data and reproducible procedures.
4. Validation: independently assess conceptual soundness, data, implementation, performance, and limitations.
5. Approval: authorize deployment subject to documented conditions.
6. Deployment and change control: release a version-controlled model and manage changes through approval procedures.
7. Monitoring: track accuracy, calibration, data quality, drift, overrides, incidents, and business outcomes.
8. Retirement: remove obsolete models, archive evidence, and communicate replacement arrangements.

Validation beyond accuracy

Validation should include at least four dimensions:

• Conceptual validation: Does the model represent a plausible demand-generating process?
• Empirical validation: Does it perform adequately on out-of-sample or rolling-origin data?
• Implementation validation: Does the deployed code produce the intended result?
• Use validation: Are users applying it within the documented scope?

For point forecasts, common error measures include mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE). If actual demand is y_t and the forecast is \hat{y}_t for periods t = 1, 2, ..., n, then:

MAE is expressed in the same units as demand and weights all absolute errors linearly. RMSE is:

RMSE is also expressed in demand units, but it penalizes large errors more heavily. MAPE is:

MAPE is undefined when y_t = 0 and can be unstable when actual demand is close to zero. Consequently, it should not be used automatically for intermittent or low-volume demand. Metrics must be selected according to the decision, cost structure, and data characteristics rather than convenience.

Forecast validation should also examine bias. Mean forecast error can be written as:

The sign convention must be stated. Under this definition, a positive MFE means that actual demand tends to exceed the forecast, indicating underforecasting. A model with low MAE but persistent bias may create systematic stockouts or excess capacity.

For probabilistic forecasts, accuracy alone is insufficient. A prediction interval should have appropriate coverage, meaning that the observed outcome falls within the stated interval at approximately the advertised rate over a suitable evaluation sample. Calibration asks whether predicted probabilities correspond to observed frequencies. A 90% interval that contains actual demand only 55% of the time is not reliable, even if its average width appears attractive.

3. Documentation and Reproducibility

Documentation is a control, not an administrative afterthought. A forecasting model should be reproducible by a competent reviewer who was not involved in its original development.

A model card or equivalent record should include:

• business purpose and decision supported;
• model owner, technical owner, approver, and users;
• target definition and forecast horizon;
• data sources, extraction dates, transformations, and lineage;
• missing-value, outlier, and anomaly treatment;
• features, parameters, algorithms, and benchmark models;
• training, validation, and test design;
• accuracy, bias, calibration, and stability results;
• assumptions, exclusions, known failure modes, and limitations;
• human-override rules and escalation thresholds;
• privacy, security, fairness, and regulatory assessment;
• version history, approval status, and retirement criteria.

Documentation should distinguish facts from judgments. For example, “the model has a MAE of 420 units in the 2024 rolling-origin evaluation” is an empirical statement. “The model is suitable for replenishment decisions” is a judgment that requires a business and risk rationale.

Data lineage is particularly important. Analysts should be able to trace a forecast from its published output back through code, model version, feature values, source systems, and extraction time. Without lineage, an organization cannot reliably investigate an anomalous forecast or reproduce a previous decision.

4. Monitoring in Production

Performance observed during development is not a guarantee of future performance. Demand processes evolve because of changing prices, products, channels, customer behavior, competitors, macroeconomic conditions, and climate patterns. Production monitoring should therefore cover four categories.

Data monitoring

Checks should identify missing data, delayed feeds, unexpected ranges, duplicate records, unit changes, schema changes, and abnormal category frequencies. A sudden fall in observed demand may represent genuine market decline, a supply stockout, or a broken transaction interface. These causes have different implications and must not be treated as equivalent.

Model monitoring

Model monitoring examines forecast error, bias, prediction-interval coverage, feature drift, residual patterns, and performance by product, geography, customer segment, and horizon. Aggregate accuracy can conceal severe failure in a strategically important segment.

Process monitoring

Organizations should monitor the frequency and direction of human overrides, exception volumes, approval delays, and the percentage of forecasts generated on schedule. A high override rate may indicate valuable local knowledge, poor model design, weak trust, or incentives that conflict with the model.

Outcome monitoring

The ultimate question is whether forecasts improve decisions. Relevant outcomes may include service level, inventory cost, waste, expedite spending, production stability, revenue, and customer experience. Forecast accuracy and business performance should be analyzed together because operational constraints can prevent a good forecast from producing a good outcome.

A monitoring policy should define thresholds and responses. For example, a persistent increase in bias may trigger investigation; a broken data feed may trigger automatic fallback to a validated baseline; a severe security event may require suspension of automated decisions. Thresholds should be established before incidents occur, while allowing expert review of unusual circumstances.

5. Privacy, Security, and Regulatory Responsibility

Forecasting systems increasingly use customer transactions, location data, online behavior, loyalty records, workforce information, and supplier data. The fact that information can improve a forecast does not establish that it should be collected or used.

Privacy principles

Responsible data use commonly incorporates:

• Purpose limitation: collect and use data for specified legitimate purposes.
• Data minimization: use no more personal data than necessary.
• Transparency: communicate relevant uses in understandable language.
• Accuracy: correct materially inaccurate information.
• Retention limitation: retain information only as long as justified.
• Security: protect data against unauthorized access, alteration, and disclosure.
• Rights and accountability: provide appropriate access, correction, objection, or review mechanisms where required by law.

Legal requirements differ by jurisdiction and sector. Depending on the organization and use case, relevant regimes may include comprehensive data-protection laws, sector-specific health or financial rules, employment law, consumer-protection law, cybersecurity obligations, records requirements, and emerging AI regulation. Managers must obtain jurisdiction-specific legal advice rather than assuming that a single global checklist is sufficient.

Security controls

Forecasting infrastructure should use role-based access, strong authentication, encryption in transit and at rest, secrets management, network segmentation, logging, vulnerability management, backup and recovery procedures, and tested incident response. Model integrity matters as much as data confidentiality: an attacker who changes a feature pipeline or model file may manipulate inventory, pricing, or procurement decisions without stealing information.

Third-party and generative-AI services create additional exposure. Contracts should address data use, retention, subcontractors, breach notification, model training, intellectual property, service continuity, and audit rights. Sensitive data should not be entered into an external system merely because the interface is convenient.

6. Fairness, Explainability, and Human Oversight

Forecasting can affect groups unevenly even when it does not explicitly use protected characteristics. Historical sales may reflect unequal access, discriminatory service, different marketing exposure, or supply constraints. A model trained on such data can reproduce those patterns.

Fairness assessment begins by identifying affected groups and the allocation mechanism. Relevant questions include:

• Does the model systematically underforecast demand for a community, region, or customer group?
• Does a service-allocation rule give some groups longer waiting times or lower availability?
• Are errors concentrated where historical data are sparse?
• Are proxy variables encoding sensitive characteristics?
• Does the business decision amplify a small forecasting difference into a large unequal outcome?

There is no single fairness metric that is appropriate for every forecasting problem. Measures may compare error, bias, service level, access, or resource allocation across groups. Trade-offs can arise: improving parity in one metric may worsen another, and statistical parity may not be suitable when legitimate differences in demand exist. Fairness is therefore a socio-technical assessment, not merely a model-score adjustment.

Explainability should be matched to the decision. Users may need a global explanation of major drivers, a local explanation of an individual forecast, an uncertainty statement, and an account of what would change the result. Explanations must not falsely imply causal relationships when the model is predictive rather than causal.

Human oversight is meaningful only when humans have authority, competence, time, and information to intervene. A nominal “human in the loop” who routinely accepts automated recommendations without review does not provide effective control. Override processes should record the reason, evidence, result, and whether the intervention should inform future model improvement.

7. Strategic Development of Analytics Capability

Organizations often attempt to purchase advanced technology before establishing data quality, governance, process ownership, or analytical skills. Sustainable capability develops across people, processes, platforms, and culture.

A maturity pathway may be described as follows:

Capability development requires a portfolio rather than one large technology project. Useful investments include data engineering, statistical and machine-learning expertise, domain knowledge, change management, model-risk skills, visualization, and executive education. Training should include not only how to use a tool but also when not to trust it.

A strategic analytics roadmap should identify high-value use cases, risk tiers, required data, dependencies, benefits, controls, and stopping criteria. Pilot projects should be evaluated against a credible baseline and operational outcome, not only against a technical benchmark. Benefits should account for implementation cost, behavioral change, forecast uncertainty, and unintended consequences.

8. Emerging Directions in Forecasting

Generative artificial intelligence

Generative AI can assist with extracting information from unstructured documents, generating code, explaining forecast results, creating scenario narratives, and providing conversational access to analytical systems. It may reduce barriers between managers and data, but it introduces risks of hallucination, unverifiable reasoning, prompt injection, data leakage, unstable outputs, and inappropriate automation.

Generative AI should not be treated as an authoritative forecasting engine merely because its language is fluent. Numerical forecasts should normally be generated and validated by controlled analytical methods, with generative AI used as a bounded interface or support tool. Outputs require grounding in approved data, citation or traceability where feasible, testing, access controls, and human review.

Probabilistic forecasting

A point forecast hides uncertainty. Probabilistic forecasting represents a distribution or a set of quantiles, allowing decision-makers to evaluate multiple risk levels. If F_t(y) denotes the forecast cumulative distribution for future demand y at time t, then a predictive interval can be derived from selected quantiles of that distribution. The interval is useful only if its coverage and interpretation are validated.

Probabilistic forecasts support inventory and capacity decisions because different decisions have different costs of underprediction and overprediction. They also encourage honest communication about uncertainty. However, users may find distributions more difficult to interpret, and poor calibration can create a false appearance of sophistication.

Digital twins

A digital twin is a continuously updated digital representation of a physical system, process, or network that supports monitoring, simulation, and decision analysis. In operations, a twin may represent facilities, inventory, transportation, capacity, and demand scenarios. Its value depends on data integration, valid process logic, timely updates, and clearly defined decision use. A visually impressive simulation is not a reliable twin if its assumptions and synchronization are weak.

Autonomous planning

Autonomous planning systems can sense conditions, generate forecasts, recommend actions, execute approved rules, and learn from outcomes. Appropriate autonomy is bounded by guardrails such as inventory limits, spending authority, service constraints, segregation of duties, approval thresholds, rollback capability, and emergency shutdown. Autonomy should be graduated: low-risk repetitive decisions may be automated, while high-impact or irreversible decisions require human authorization.

Climate-related uncertainty

Climate change creates both acute shocks and structural shifts in demand. Heatwaves, floods, storms, water scarcity, changing disease patterns, migration, insurance costs, and policy transitions can alter demand and supply relationships. Historical data may be a poor guide when the underlying environment is changing.

Organizations should combine historical models with scenario analysis, stress testing, external climate indicators, adaptation plans, and expert judgment. Climate scenarios are not single-point predictions; they are structured representations of plausible futures under specified assumptions. Decision-makers should examine robustness: whether a strategy performs acceptably across several futures rather than optimizing only for the most likely one.

9. Integrated Applied Case: Northstar Medical Supplies

The following is an illustrative case.

Background

Northstar Medical Supplies distributes diagnostic consumables to hospitals and clinics. It forecasts weekly demand for approximately 8,000 products. The company historically used regional spreadsheets, but it introduced a machine-learning system using order history, hospital characteristics, promotions, lead times, and local disease indicators.

Problem

After deployment, aggregate MAE improved by 12%. However, several rural clinics experienced repeated stockouts, while central hospitals received excess inventory. The model also used detailed patient-linked demand signals supplied by a third-party data broker. Planners frequently overrode the model but did not record why.

Analysis

The governance review identified five issues. First, the evaluation was dominated by high-volume hospitals, masking poor performance in low-volume clinics. Second, stockouts were treated as low demand, creating censored observations and reinforcing underforecasting. Third, the data broker’s permission for the proposed use was unclear. Fourth, access controls allowed more users than necessary to view sensitive features. Fifth, no formal threshold existed for suspending automated replenishment.

The team segmented evaluation by clinic type and calculated MAE, bias, and service-level outcomes. Under the stated MFE convention, rural clinics showed a strongly negative MFE, meaning the forecast exceeded actual recorded sales. Investigation found that recorded sales were suppressed by previous stockouts; the apparent overforecast was therefore misleading. The relevant business question was unconstrained demand, not merely observed transactions.

Approach and solution

Northstar created a cross-functional governance group involving supply planning, compliance, information security, clinical representatives, and data science. It removed unnecessary patient-linked variables, documented lawful purpose and retention, and restricted access. The model was retrained using stockout flags and censored-demand methods, with separate validation for clinic types.

The company introduced a champion–challenger process: the production model was compared with a transparent statistical baseline and a planner-adjusted alternative. Prediction intervals were reported for high-risk products. Automated replenishment was limited by spending and service guardrails, and orders outside thresholds required approval. Every override required a short reason code and optional narrative.

Monitoring dashboards tracked data freshness, forecast bias by clinic type, interval coverage, stockout rates, override patterns, and security incidents. A fallback baseline was activated when data feeds failed.

Outcome

The revised system produced a smaller aggregate accuracy improvement than the original model, but rural stockouts declined, excess inventory was reduced, and planners reported greater trust because the system explained uncertainty and recorded exceptions. The organization also avoided continuing a questionable data practice. The case demonstrates that responsible forecasting may prioritize equitable service, lawful data use, and resilience over a single aggregate accuracy statistic.

Learning points

1. Observed sales are not always unconstrained demand.
2. Aggregate metrics can conceal segment-level harm.
3. Data permission and minimization are part of model quality.
4. Human overrides are evidence about the decision system, not merely noise.
5. A less accurate but better-governed system may create superior organizational outcomes.

10. Best-Practice Control Framework

A graduate-level practitioner should be able to convert principles into operating controls. The following sequence is suitable for a material forecasting application:

1. Define the decision, affected stakeholders, acceptable risk, and success measures.
2. Register the model and classify its impact, sensitivity, and automation level.
3. Establish data ownership, lineage, quality rules, lawful purpose, and retention.
4. Develop transparent benchmarks before selecting a more complex method.
5. Validate with time-appropriate designs, segment analysis, stress tests, and uncertainty assessment.
6. Document assumptions, limitations, versions, approvals, and user responsibilities.
7. Implement security, access control, change management, and incident response.
8. Test fairness and distributional effects in the actual decision context.
9. Deploy gradually with fallback procedures and bounded automation.
10. Monitor data, model, process, outcome, and ethical indicators.
11. Review material changes and emerging risks through independent challenge.
12. Retire or redesign models that no longer meet their purpose.

The strongest governance systems are enabling rather than merely restrictive. They allow an organization to experiment safely, compare alternatives, learn from errors, and scale valuable analytics without losing accountability.

Challenges and Limitations

Governance itself has limitations. Excessive approval layers can delay useful innovation and encourage shadow systems. Documentation may become ceremonial if nobody challenges its claims. Fairness metrics cannot resolve political or ethical disagreement about what a fair allocation means. Monitoring can detect drift without identifying its cause. Privacy controls may restrict data that would produce social value, while weak controls may create unacceptable intrusion. Finally, no framework can remove uncertainty from a changing world.

Managers must therefore treat governance as a continuing process of inquiry. The appropriate question is not whether a model is “safe” in the abstract, but whether its purpose, evidence, controls, and consequences remain acceptable under current conditions.

Chapter Summary

Forecasting governance integrates strategy, data management, model validation, operational controls, ethics, security, and accountability. Model risk management requires an end-to-end lifecycle from registration and specification through validation, deployment, monitoring, change control, and retirement. Accuracy metrics such as MAE and RMSE must be interpreted with bias, calibration, segment performance, and business outcomes.

Documentation and lineage make forecasts reproducible and reviewable. Production monitoring must examine data quality, model performance, process behavior, and operational outcomes. Privacy requires purpose limitation, minimization, transparency, retention control, and lawful handling; security must protect both information and model integrity. Fairness and explainability must be assessed in the context of the decisions that forecasts influence, with meaningful human oversight.

Strategic capability develops through aligned people, processes, platforms, and culture. Generative AI offers useful assistance but requires grounding and controls. Probabilistic forecasting makes uncertainty explicit, digital twins support simulation, autonomous planning requires bounded authority, and climate change demands scenario-based and robust decision-making. The future of forecasting is therefore not defined by automation alone, but by trustworthy integration of analytics with responsible organizational judgment.

Key Terms

Forecast governance: The policies, roles, controls, and decision rights governing forecasting systems.

Model risk: The possibility of harmful consequences arising from model error, misuse, or misunderstanding.

Model inventory: A controlled register of forecasting models and their ownership, purpose, and risk classification.

Model validation: Independent assessment of conceptual soundness, data, implementation, performance, and use.

Data lineage: The traceable history of data from source through transformation to forecast output.

Concept drift: Change in the relationship between inputs and the outcome being forecast.

Data drift: Change in the distribution, quality, or structure of model inputs.

Calibration: Agreement between predicted probabilities or intervals and observed frequencies.

Forecast bias: Systematic tendency to overforecast or underforecast.

MAE: Mean absolute error, an average absolute forecast error in demand units.

RMSE: Root mean squared error, which weights large errors more heavily.

Probabilistic forecast: A forecast representing a distribution, probabilities, or quantiles rather than only one value.

Prediction interval: A range intended to contain a future observation at a stated coverage level.

Data minimization: Limiting data collection and use to what is necessary for a legitimate purpose.

Fairness: The ethically and legally relevant assessment of how model errors and decisions affect groups.

Explainability: The ability to communicate how a model produces or supports an output.

Human oversight: Competent, empowered human review and intervention in an automated system.

Digital twin: A continuously updated digital representation used to monitor or simulate a physical system.

Autonomous planning: Planning in which software senses, predicts, recommends, or executes actions with limited intervention.

Climate-related uncertainty: Uncertainty arising from changing climate conditions, impacts, policies, and adaptation responses.

Review Questions

Short Answer Questions

1. What is model risk, and how does it differ from forecast error?
2. Why is data lineage important in a governed forecasting system?
3. State two limitations of MAPE.
4. What is meant by meaningful human oversight?
5. Identify three controls relevant to secure forecasting infrastructure.

Long Answer Questions

1. Explain the model risk management lifecycle for a demand-forecasting application.
2. Compare data monitoring, model monitoring, process monitoring, and outcome monitoring.
3. Discuss how privacy and fairness considerations should influence the design of forecasting systems.
4. Evaluate the strategic capabilities required to progress from fragmented analytics to responsible autonomous planning.
5. Critically assess the potential benefits and risks of generative AI in demand forecasting.

Analytical / Application Questions

1. A retailer reports improved aggregate RMSE after deploying a machine-learning forecast, but stockouts increase in low-volume regions. Design an investigation and governance response.
2. A manufacturer wants to use customer location, browsing behavior, and climate projections to forecast demand. Propose a framework for assessing legality, privacy, fairness, security, and business value before deployment.
3. Design a bounded-autonomy policy for an inventory-planning system. Specify which decisions may be automated, which require approval, the monitoring indicators required, and the conditions for fallback or shutdown.

Multiple Choice Questions with Answers

Question 1

Which statement best describes model risk in demand forecasting?

A. It is only the numerical difference between actual and forecast demand.

B. It is the possibility of harmful consequences from incorrect design, implementation, use, or interpretation of a model.

C. It is the cost of storing historical demand data.

D. It is the unavoidable variability in customer demand.

Correct Answer: B. It is the possibility of harmful consequences from incorrect design, implementation, use, or interpretation of a model.

Question 2

Under the chapter’s definition of MFE, a positive mean forecast error indicates that:

A. Forecasts tend to exceed actual demand.

B. Actual demand tends to exceed forecasts.

C. The model has no systematic bias.

D. The forecast interval is too wide.

Correct Answer: B. Actual demand tends to exceed forecasts.

Question 3

Which error metric is expressed in the same units as demand and penalizes all absolute errors linearly?

A. MAPE

B. RMSE

C. MAE

D. Prediction-interval coverage

Correct Answer: C. MAE

Question 4

Why can MAPE be unsuitable for intermittent demand?

A. It requires normally distributed errors.

B. It is undefined when actual demand is zero and unstable near zero.

C. It always gives a result in physical units.

D. It ignores the forecast horizon.

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

Question 5

Which control most directly supports reproducibility of a forecast?

A. Increasing the number of dashboard colors

B. Maintaining data lineage, code versions, model versions, and transformation records

C. Removing all human overrides

D. Using only generative AI for documentation

Correct Answer: B. Maintaining data lineage, code versions, model versions, and transformation records

Question 6

What is the principal purpose of calibration assessment for a probabilistic forecast?

A. To determine whether predicted probabilities or intervals correspond to observed frequencies.

B. To convert all forecasts into point estimates.

C. To eliminate the need for data-quality checks.

D. To guarantee that every forecast is correct.

Correct Answer: A. To determine whether predicted probabilities or intervals correspond to observed frequencies.

Question 7

Which practice best reflects data minimization?

A. Collecting every potentially predictive variable indefinitely

B. Using only the personal data necessary for a specified legitimate purpose

C. Sharing all source data with every planner

D. Retaining data permanently in case it becomes useful

Correct Answer: B. Using only the personal data necessary for a specified legitimate purpose

Question 8

Why can aggregate forecast accuracy conceal an ethical or operational problem?

A. Aggregate metrics always exaggerate errors.

B. High-volume segments can dominate the metric while smaller groups experience systematic harm.

C. Accuracy metrics cannot be calculated for time series.

D. Ethical issues occur only when accuracy is low overall.

Correct Answer: B. High-volume segments can dominate the metric while smaller groups experience systematic harm.

Question 9

Which is an appropriate guardrail for autonomous inventory planning?

A. Unlimited automated purchasing authority

B. Removal of rollback capability

C. Spending limits, approval thresholds, and fallback procedures

D. Disabling all monitoring after deployment

Correct Answer: C. Spending limits, approval thresholds, and fallback procedures

Question 10

Why are climate scenarios useful in forecasting strategy?

A. They provide one certain prediction of future climate conditions.

B. They replace all historical data and managerial judgment.

C. They represent plausible futures and help test whether decisions remain robust under uncertainty.

D. They eliminate structural change from demand processes.

Correct Answer: C. They represent plausible futures and help test whether decisions remain robust under uncertainty.