Forecasting Process and Organizational Frameworks

Introduction

Forecasting is not simply the production of a number. It is an organizational process through which an enterprise converts information about demand, supply, markets, customers, and managerial intentions into a time-phased view of the future. That view supports decisions about procurement, production, inventory, capacity, workforce, transportation, finance, and customer service.

A technically accurate statistical forecast can still fail if the forecasting problem is poorly defined, data ownership is unclear, commercial information is excluded, approval responsibilities are ambiguous, or published forecasts are not monitored. Conversely, a modest model embedded in a disciplined process may produce substantial value because it creates a common planning language and makes assumptions visible.

This chapter presents an end-to-end forecasting process. It integrates statistical baseline generation, collaborative review, governance, rolling forecasts, forecast-value-added analysis, publication, performance monitoring, and continuous improvement. The emphasis is managerial as well as analytical: forecasting is treated as a cross-functional capability rather than an isolated activity of the supply-chain analyst.

Learning Objectives

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

1. Define a forecasting problem in terms of decisions, horizons, aggregation, and uncertainty.
2. Design an end-to-end process with clear data ownership and decision rights.
3. Distinguish a statistical baseline from collaborative adjustments and an approved plan.
4. Explain rolling forecasts and select appropriate forecast-performance measures.
5. Apply forecast-value-added principles to evaluate interventions in the process.
6. Diagnose governance, data, behavioral, and technological risks.
7. Recommend a practical framework for continuous forecasting improvement.

1. Forecasting as a Managed Business Process

A forecast is a conditional statement about future events. It depends on the information available, the time at which the forecast is made, the level of aggregation, and the definition of demand. For example, a forecast of weekly customer orders for a product family is different from a forecast of daily production requirements for a specific stock-keeping unit (SKU). Neither is universally superior; each serves a different decision.

The forecasting process should therefore begin with the decision rather than with a favored algorithm. A useful design question is: What decision will change because of this forecast? If the answer is unclear, the organization may be generating numbers that consume effort without improving action.

Forecasts commonly support three planning horizons:

The same organization may require all three. Short-term forecasts often use detailed time buckets and item-location combinations, whereas long-term forecasts may use product families, regions, and monetary values. Aggregation reduces noise but can conceal local constraints; disaggregation increases detail but also increases data and process complexity.

Forecast object and forecast horizon

The forecast object identifies what is being predicted: for example, unconstrained customer demand, shipments, orders, consumption, or revenue. These are not interchangeable. Shipments may be limited by inventory availability, while orders may be distorted by order batching. A demand forecast intended for replenishment should generally represent the demand that would occur under adequate supply, not merely the sales that were fulfilled.

The forecast horizon is the future interval covered by the forecast. It should reflect the longest relevant decision lead time plus a planning margin. If a supplier lead time is eight weeks and capacity decisions require twelve weeks of notice, a forecast horizon shorter than twelve weeks cannot support those decisions adequately.

2. Defining the Forecasting Problem

A formal problem definition should be documented before model selection. At minimum, it should specify:

• the decision or decisions being supported;
• the forecast object and unit of measure;
• the product, customer, geographic, and channel hierarchy;
• the time bucket, such as day, week, or month;
• the forecast horizon and update frequency;
• the information cutoff date;
• the treatment of stockouts, returns, cancellations, and one-time events;
• the accountable owner and principal users; and
• the performance measures and acceptable error tolerance.

A useful distinction is between demand sensing, demand planning, and supply planning. Demand sensing uses very recent signals to improve near-term visibility. Demand planning estimates expected market or customer demand. Supply planning converts that demand view into a feasible response given capacity, inventory, sourcing, and logistics constraints. Confusing these roles can cause a constrained shipment plan to be mislabeled as a demand forecast.

Forecasting assumptions

Forecasts are conditional on assumptions. Examples include the continuation of a promotion, a planned product launch, an exchange-rate scenario, or the absence of supply restrictions. Assumptions should be recorded with an owner, effective period, source, and confidence level. Unrecorded assumptions become difficult to challenge and can create false precision.

Forecasting does not eliminate uncertainty. It organizes uncertainty for decision-making. A point forecast should therefore be complemented, where decisions justify it, by a range, scenario, or probability distribution. A safety-stock decision, for example, requires information about forecast error variability and replenishment lead time, not only the expected demand value.

3. Data Ownership and Information Governance

Data ownership is a governance responsibility, not merely a systems designation. The owner is accountable for the definition, quality, access, retention, and appropriate use of a data element. In a forecasting process, ownership is often distributed:

A data dictionary should define fields, units, granularity, refresh timing, permissible values, and responsible parties. For example, “sales” may mean invoiced quantity, shipped quantity, or booked order quantity. Without a precise definition, different functions may believe they are analyzing the same measure when they are not.

Data quality should be assessed along several dimensions: completeness, accuracy, timeliness, consistency, uniqueness, and validity. A simple quality score can help prioritize remediation, but a score should not replace investigation. Missing data may be more serious for a high-value item than for a low-impact item, even if the percentage missing is the same.

Bias caused by censored demand

When an item is out of stock, observed sales may be lower than true demand. Treating every low-sales period as weak demand can bias the baseline downward. Organizations may use stockout flags, lost-sales estimates, customer orders, substitution records, or inventory availability indicators to identify censored observations. Any imputation method should be documented and tested because an aggressive correction can create an upward bias.

Access, privacy, and security

Customer-level forecasts may contain commercially sensitive or personal information. Access should follow least-privilege principles, and data should be retained only as long as justified. Forecast collaboration must not become an uncontrolled mechanism for sharing confidential customer prices, competitive intelligence, or personal data. Governance should also define how externally sourced signals and generative-AI tools may be used, especially where confidential data could be exposed.

4. The End-to-End Forecasting Cycle

A robust process separates analytical production from managerial judgment while allowing structured interaction between them. The following sequence is adaptable to different industries.

Stage 1: Initiate and define

The process owner confirms the scope, calendar, forecast object, horizon, required hierarchy, and decision users. Exceptions such as new products, discontinued items, and major events are identified at this stage.

Stage 2: Prepare and validate data

The analyst extracts data at the required grain, reconciles totals to authoritative systems, detects missing periods and outliers, and records transformations. Changes to historical data should be traceable. A model should not silently revise history without an explanation.

Stage 3: Generate the statistical baseline

The baseline is the forecast produced by a documented method before commercial overrides. Appropriate methods depend on the demand pattern. Intermittent demand, strong seasonality, short product histories, and structural breaks require different treatment. A baseline should be reproducible: another qualified analyst should be able to obtain the same result from the same inputs and model version.

A general point forecast for period t may be denoted by F_t, while actual demand is A_t. A simple seasonal-naïve baseline, for example, uses the observation from the corresponding prior seasonal cycle:

Here, s is the seasonal cycle measured in time buckets. This method is appropriate when seasonal behavior is reasonably stable and a recent seasonal observation is informative. It is not a universal model and should be compared with alternatives.

For a moving-average baseline with window length k:

The units of F_t and A_t are the same, such as cases per week. The method assumes that the recent average is a useful estimate of the next period and does not explicitly model trend or seasonality.

Stage 4: Conduct diagnostic review

Before collaborative review, the analyst examines forecast errors, bias, exceptional observations, model failures, and reconciliation across the hierarchy. The goal is not to manipulate the forecast to fit expectations but to identify information that the baseline cannot reasonably know.

Stage 5: Perform collaborative review

Sales, marketing, operations, finance, procurement, and relevant country or category teams review the baseline using structured inputs. Valid reasons for adjustment may include a confirmed promotion, a tender award, a product launch, a customer closure, or a verified market change. “The number feels too low” is a concern, not evidence; it should trigger a search for an explicit driver.

Adjustments should be recorded as quantity, percentage, or both, with rationale, owner, start and end dates, confidence, and expected business consequence. A useful rule is that adjustments should be made at the level where the information exists and then reconciled to the level used for decisions.

Stage 6: Reconcile and approve

The consensus forecast is checked for internal consistency, financial alignment, capacity implications, and conflicts between functions. Approval rights should be explicit. Consensus does not mean that every participant receives the outcome they prefer; it means that the organization has a documented, authoritative view and understands material disagreements.

Stage 7: Publish and communicate

The approved forecast is published to defined recipients through a controlled system. Publication should include version, timestamp, horizon, units, assumptions, confidence or scenario information, and any known exclusions. Downstream systems should not consume an unpublished working version by accident.

Stage 8: Monitor outcomes

As actuals become available, the organization measures error, bias, process adherence, and business consequences. Monitoring should distinguish model performance from execution performance. A forecast may be accurate while service is poor because supply failed; conversely, excellent service may hide an over-forecast supported by excess inventory.

Stage 9: Learn and improve

The cycle closes with a review of recurring errors, ineffective overrides, data defects, model drift, and process delays. Improvement actions should be assigned owners and due dates. Continuous improvement is weakened when the organization discusses performance but does not change a process, data source, model, or decision rule.

5. Rolling Forecasts and the Forecast Calendar

A rolling forecast always extends into the future as time passes. When the current period closes, it is removed from the horizon and a new future period is added. The forecast therefore maintains a constant forward-looking window.

Suppose a twelve-month forecast is issued in January for February through January of the following year. In February, the February period is closed using actual results, the remaining periods are updated, and a new February period of the following year is added. The rolling structure prevents the organization from waiting for an annual planning cycle to recognize changes.

Rolling forecasts do not imply that every forecast should be completely revised every day. Excessive updating can create instability, confuse users, and encourage reaction to noise. Update frequency should match demand volatility, decision lead time, data latency, and the cost of change. A hierarchical cadence may be appropriate: weekly short-term updates, monthly consensus reviews, and quarterly structural-model reviews.

A forecast calendar should specify cutoffs and responsibilities. For example:

The calendar is a control mechanism. Late submissions should be visible rather than silently incorporated, and emergency changes should have a separate exception route.

6. Forecast Accuracy, Bias, and Value

No single metric captures forecast quality in all circumstances. Let e_t = A_t − F_t denote the error for period t, where actual and forecast values use identical units.

Mean absolute error is:

where n is the number of evaluated periods. MAE is expressed in the original units and is easy to interpret, but it is not directly comparable across products with very different volumes.

Mean absolute percentage error is commonly written as:

It requires A_t ≠ 0 and becomes unstable when actual demand is zero or very small. For portfolios containing zeros, weighted or scaled measures are often more appropriate.

A bias measure can be expressed as mean error:

Positive ME indicates systematic under-forecasting under this sign convention; negative ME indicates over-forecasting. Bias should be examined alongside absolute error because positive and negative errors can cancel in an average.

Weighted absolute percentage error can be expressed as:

WAPE is meaningful when the denominator is positive and is often more stable than item-level MAPE for aggregated demand. However, it can be dominated by high-volume items and may conceal poor performance on strategic low-volume products.

Forecast value added

Forecast-value-added (FVA) analysis asks whether each process step improves the forecast relative to a defined benchmark. The benchmark is often the statistical baseline or a simple naïve method. For a lower-is-better error measure E:

A positive result means the step reduced error; a negative result means it worsened error. The comparison must use equivalent forecast origins, horizons, products, and actuals. FVA is not a judgment about whether a person is valuable. It evaluates whether a particular intervention, on average, improved forecast outcomes.

An adjustment may have negative FVA but still be justified if it addresses a rare, high-impact event that historical testing cannot represent. Therefore, FVA should be interpreted with event classification and business impact, not used as an automatic basis for eliminating all human judgment.

Worked example: baseline and adjustment evaluation

Assume actual demand for four weeks is 100, 120, 80, and 100 units. The baseline forecasts are 90, 110, 100, and 90 units. The errors A_t − F_t are 10, 10, −20, and 10; absolute errors are 10, 10, 20, and 10. Thus:

MAE = (10 + 10 + 20 + 10)/4 = 12.5 units.

A collaborative adjustment produces forecasts of 100, 115, 90, and 105 units. Absolute errors are 0, 5, 10, and 5, so:

MAE = (0 + 5 + 10 + 5)/4 = 5 units.

The adjustment has an FVA of 12.5 − 5 = 7.5 units per forecast period under MAE. The result supports the adjustment process for this sample, although a larger historical evaluation is required before concluding that the process consistently adds value.

7. Organizational Frameworks and Decision Rights

Forecasting is often organized through a cross-functional planning forum, frequently connected to sales and operations planning or integrated business planning. The forum should not be a meeting that merely exchanges numbers. It should be a decision system with defined inputs, escalation rules, and outputs.

A practical framework includes four layers:

1. Data and technology layer: transactional data, master data, models, workflows, and audit trails.
2. Analytical layer: baseline forecasts, diagnostics, scenarios, and uncertainty estimates.
3. Collaborative layer: commercial, operational, financial, and market intelligence.
4. Governance layer: ownership, approval, policy, performance management, and escalation.

The process owner coordinates the cycle but should not necessarily own every input. A RACI matrix—responsible, accountable, consulted, and informed—can clarify roles. For example, analytics may be responsible for the baseline, marketing accountable for promotion assumptions, operations consulted on constraints, and finance informed of the financial translation.

Centralization versus decentralization

A centralized forecasting team can improve methodological consistency, tool governance, and career development. A decentralized structure may provide stronger local market knowledge and faster response. Many organizations adopt a federated model: common standards, data definitions, and technology are centralized, while local teams contribute market intelligence within controlled rules.

The appropriate design depends on product complexity, geographic diversity, data maturity, and the cost of coordination. Centralization does not automatically produce objectivity; decentralization does not automatically produce responsiveness. Both require transparent decision rights and performance evidence.

8. Technology, Automation, and Artificial Intelligence

Technology can automate extraction, model execution, exception detection, workflow routing, versioning, and dashboarding. Automation is particularly valuable for repetitive work because it allows analysts to focus on diagnosis and decisions. However, automation can also scale bad definitions and data defects.

Machine-learning methods may improve forecasts where large, rich, and stable data sets contain useful explanatory signals. They require careful time-based validation, protection against data leakage, monitoring for drift, and a clear comparison with simple benchmarks. A complex model should earn its place through demonstrated incremental value, not novelty.

Generative AI may assist with summarizing forecast drivers, identifying anomalies, drafting commentary, or querying documented assumptions. It should not independently publish forecasts or invent causal explanations. Human accountability remains necessary for data authorization, model validation, commercial confidentiality, and decisions with material financial or service consequences.

A model-risk framework should document the model purpose, training data, assumptions, validation method, known limitations, monitoring thresholds, and retirement conditions. Explainability is especially important when forecasts affect customer allocation, employee capacity, or supplier commitments.

9. Case Study: NorthStar Home Appliances

Background

NorthStar Home Appliances, an illustrative mid-sized manufacturer, sells refrigerators through retailers in three regions. Its monthly forecast was produced by finance, adjusted informally by sales, and then sent to operations. The company experienced excess inventory in standard models and shortages during promotions.

Problem

An investigation found three process weaknesses. First, sales shipments were used as a demand measure even when stockouts occurred. Second, sales adjustments were entered without rationale or expiry dates. Third, no one compared the approved forecast with the statistical baseline. The organization debated forecast accuracy but could not identify which stage created error.

Analysis

NorthStar defined the forecast object as unconstrained retailer demand, measured in units by SKU-region-month. The horizon was eighteen months, with a monthly rolling cycle. Data ownership was assigned to sales operations for customer orders, supply planning for stockout flags, marketing for promotion events, and analytics for model execution.

The analytics team created a seasonal-naïve baseline and a model incorporating trend and promotion indicators. Both were evaluated using rolling-origin backtesting: models were trained only on information that would have been available at each historical forecast date. This avoided using future information accidentally.

The company introduced an adjustment log requiring the reason, amount, affected periods, owner, confidence, and expiry date. A monthly consensus meeting reviewed only material exceptions rather than every SKU. Forecast-value-added analysis compared baseline, sales-adjusted, and consensus forecasts using MAE and bias.

Approach and solution

NorthStar separated three measures: unconstrained demand, fulfilled shipments, and the approved operational plan. Stockout periods were flagged, and the demand model used available order and substitution evidence rather than blindly treating lost sales as zero demand. Promotions were represented as explicit events with start and end dates.

A twelve-month rolling horizon was published monthly, while a six-week operational view was refreshed weekly. The approval forum could reject an unsupported adjustment and request evidence, but it could also retain a judgmental adjustment for a documented major event.

Outcome

After two quarters, NorthStar reported lower aggregate absolute error and less systematic over-forecasting in standard products. More importantly, it could identify that many sales overrides reduced performance, while a smaller number of promotion-related overrides materially improved it. The organization removed routine manual adjustments and redirected sales participation toward event information and exception review.

Learning points

The principal lesson is organizational: improvement came not from selecting a sophisticated model alone, but from defining demand correctly, assigning ownership, separating baseline from judgment, and measuring each intervention. A forecast process creates value when it makes better decisions more likely, not merely when it produces more detailed reports.

10. Challenges and Limitations

Forecasting processes face recurring difficulties. Structural breaks—such as regulation, disruption, or a new competitor—can invalidate historical relationships. New products have little or no history and require analogues, market assumptions, or scenario ranges. Intermittent demand makes percentage metrics unstable. Hierarchical forecasts may not add up unless reconciliation rules are applied.

Human judgment can incorporate information absent from historical data, but it is vulnerable to optimism, anchoring, political pressure, and double counting. A manager may increase a forecast because a promotion is planned even though the promotion effect is already included in the model. Structured adjustment logs and independent challenge reduce this risk.

Targets should also be distinguished from forecasts. A target expresses an aspiration or commitment; a forecast expresses the most likely outcome conditional on current information. Using a target as if it were an unbiased forecast can conceal risk and distort inventory or capacity decisions.

Finally, accuracy metrics can create dysfunctional behavior. Teams may avoid forecasting low-volume items, manipulate timing, or prefer measures that hide unfavorable errors. Metrics should therefore be linked to decision quality, service, inventory, revenue, and process adherence rather than treated as isolated performance rankings.

Best-Practice Principles

1. Start with the decision and define the forecast object precisely.
2. Establish authoritative data definitions and named owners.
3. Generate a reproducible statistical baseline before judgmental intervention.
4. Require evidence, timing, ownership, and expiry for adjustments.
5. Use a rolling horizon with a stable, published calendar.
6. Evaluate both error and bias, using metrics appropriate to the data.
7. Apply FVA to process steps, not as a simplistic judgment of individuals.
8. Separate demand, supply constraints, targets, and financial plans.
9. Automate repeatable work but retain human accountability for material decisions.
10. Treat forecasting as a learning cycle supported by auditability and experimentation.

Chapter Summary

Forecasting is an end-to-end organizational process linking data, statistical analysis, expert knowledge, governance, and operational decisions. Effective practice begins with a clear definition of the decision, demand measure, hierarchy, horizon, and assumptions. Data owners must control definitions and quality, while analysts produce a reproducible baseline before collaborative review.

A complete cycle includes problem definition, data preparation, baseline generation, diagnostics, structured collaboration, reconciliation, approval, publication, monitoring, and continuous improvement. Rolling forecasts maintain a forward-looking horizon but must be governed by an appropriate update cadence. Accuracy, bias, and business consequences should be evaluated together. Forecast-value-added analysis provides a disciplined way to test whether interventions improve results relative to a benchmark.

Organizational frameworks require explicit decision rights, cross-functional participation, audit trails, and separation of forecasts from targets and constrained plans. Technology and AI can improve speed and insight, but neither substitutes for sound definitions, validation, accountability, and ethical governance. The central managerial principle is that forecasting value arises from better decisions and learning, not from numerical sophistication alone.

Key Terms

Baseline forecast: A reproducible forecast generated before managerial adjustments.

Bias: A systematic tendency to over-forecast or under-forecast.

Consensus forecast: The approved cross-functional view after structured review.

Demand censoring: Distortion of observed demand because supply was unavailable.

Forecast object: The quantity, event, or measure being predicted.

Forecast horizon: The future period covered by a forecast.

Forecast-value-added (FVA): The change in forecast error attributable to a process step relative to a benchmark.

Rolling forecast: A forecast that continually extends into a new future period as time advances.

Data owner: The person or function accountable for a data element’s definition, quality, access, and use.

Exception management: Focusing analytical and managerial attention on unusual or material cases.

Hierarchy: The product, geographic, customer, or channel structure used to aggregate and disaggregate forecasts.

MAE: Mean absolute error, expressed in the original units of the forecast.

MAPE: Mean absolute percentage error, subject to instability when actual values are zero or small.

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

Forecast adjustment: A documented change to a baseline forecast based on additional information or judgment.

Forecast calendar: The schedule of data extraction, modeling, review, approval, and publication.

Unconstrained demand: Demand expected if sufficient supply were available.

Model drift: A decline in model performance as relationships or data conditions change.

Target: An aspirational or committed outcome, not necessarily the most likely forecast.

Review Questions

Short Answer Questions

1. Why should forecasting begin with the decision being supported rather than with model selection?
2. What is the difference between a demand forecast and a constrained supply plan?
3. Why are data ownership and a data dictionary important?
4. Define a rolling forecast and explain how it differs from an annual fixed forecast.
5. What does a positive forecast-value-added result indicate when lower error is better?

Long Answer Questions

1. Explain the stages of an end-to-end forecasting process and the governance controls required at each stage.
2. Compare centralized, decentralized, and federated forecasting structures. Under what conditions might each be appropriate?
3. Discuss the strengths and limitations of MAE, MAPE, ME, and WAPE for evaluating a product portfolio.
4. Explain how collaborative judgment can improve a statistical baseline and describe the main risks of manual overrides.
5. Design a forecast calendar and RACI framework for a company with monthly consensus planning and weekly operational updates.

Analytical / Application Questions

1. A retailer uses shipments as demand history. Several products experienced stockouts during a major promotion. Diagnose the likely forecasting distortion and propose a data and modeling response.
2. A baseline has MAE of 18 units, while a reviewed forecast has MAE of 15 units over the same evaluation set. Calculate the FVA and explain why the result is not sufficient by itself to justify every future override.
3. Design a monitoring dashboard for a rolling forecast process. Specify the measures, segmentation, thresholds, and actions you would use to distinguish model problems from data, supply, or execution problems.

Multiple Choice Questions with Answers

Question 1

Which activity should occur first in a well-designed forecasting process?

A. Select the most sophisticated available algorithm

B. Define the decision, forecast object, horizon, and users

C. Ask sales managers to submit their preferred numbers

D. Publish the forecast to the planning system

Correct Answer: B. Define the decision, forecast object, horizon, and users

Question 2

What is the principal purpose of a statistical baseline?

A. To replace all managerial judgment permanently

B. To provide a reproducible reference before collaborative adjustments

C. To convert a target into an unbiased forecast

D. To guarantee that forecast error is zero

Correct Answer: B. To provide a reproducible reference before collaborative adjustments

Question 3

A positive mean error defined as ME = (1/n)∑(A_t − F_t) indicates which condition?

A. Systematic over-forecasting

B. Systematic under-forecasting

C. No forecast error

D. High forecast volatility only

Correct Answer: B. Systematic under-forecasting

Question 4

Why can MAPE be problematic for a portfolio containing zero-demand periods?

A. It is expressed in original units

B. It cannot compare different products

C. Division by zero or very small actuals makes it unstable

D. It measures bias rather than absolute error

Correct Answer: C. Division by zero or very small actuals makes it unstable

Question 5

If a baseline MAE is 20 units and an adjusted forecast MAE is 16 units, what is the FVA using FVA = E_before − E_after?

A. −36 units

B. −4 units

C. 4 units

D. 36 units

Correct Answer: C. 4 units

Question 6

Which statement best describes a rolling forecast?

A. It is prepared once and never revised

B. It always covers a constant forward-looking window as time advances

C. It contains only historical actuals

D. It is identical to a sales target

Correct Answer: B. It always covers a constant forward-looking window as time advances

Question 7

Which adjustment is best supported in a collaborative forecasting process?

A. An unexplained increase because the number feels low

B. A change linked to a documented promotion with dates and an owner

C. A permanent override based on one unusual observation

D. An adjustment made after the forecast has been published without version control

Correct Answer: B. A change linked to a documented promotion with dates and an owner

Question 8

What is a major risk of using shipments as a demand measure during stockouts?

A. Demand may be understated because fulfilled shipments are constrained

B. Seasonality is automatically removed

C. The forecast horizon becomes longer

D. The data becomes more granular than required

Correct Answer: A. Demand may be understated because fulfilled shipments are constrained

Question 9

What does forecast-value-added analysis primarily evaluate?

A. Whether each process intervention improves forecast performance relative to a benchmark

B. Whether employees attend consensus meetings

C. Whether the forecast equals the financial target

D. Whether the most complex model is being used

Correct Answer: A. Whether each process intervention improves forecast performance relative to a benchmark

Question 10

Which governance practice is most appropriate when generative AI is used to summarize forecast drivers?

A. Allow it to publish forecasts without review

B. Permit unrestricted entry of confidential customer data

C. Require human accountability, authorized data use, and validation of explanations

D. Replace all documented assumptions with generated text

Correct Answer: C. Require human accountability, authorized data use, and validation of explanations