Demand Forecasting and Operations Analytics

Introduction

Demand forecasting is the systematic estimation of future customer demand for a product, service, or resource. It connects market information with operational decisions such as purchasing, capacity planning, workforce scheduling, inventory control, production planning, transport allocation, and financial budgeting. In supply chain management, a forecast is not a prediction of what must happen; it is an informed estimate subject to uncertainty.

Operations analytics extends forecasting by using data, models, visualisation, performance measures, and structured decision processes. Its purpose is not merely to produce a number but to improve the quality, speed, transparency, and accountability of decisions. A technically sophisticated forecast can still be operationally useless if it is based on poor data, ignores supply constraints, or is not communicated to the people who must act on it.

Forecasting may be performed at several levels: strategic forecasts over years for facility and network design; tactical forecasts over months for aggregate capacity and procurement; and operational forecasts over days or weeks for replenishment, scheduling, and transport. Forecasts may also be produced at different levels of aggregation, such as total company demand, product family, stock-keeping unit (SKU), customer, region, or channel.

Learning Objectives

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

1. Explain the role, scope, and limitations of demand forecasting in operations.
2. Distinguish qualitative, quantitative, causal, time-series, and judgmental methods.
3. Calculate and interpret moving-average and exponential-smoothing forecasts.
4. Identify trend and seasonal patterns and understand their operational implications.
5. Measure forecast error using appropriate accuracy and bias indicators.
6. Use visualisation and sensitivity analysis to support evidence-based decisions.
7. Evaluate forecast governance, data quality, technology, and ethical issues.
8. Recommend a practical forecasting process for a supply-chain organisation.

1. The Nature of Demand and Forecasting

Demand is the quantity customers are willing and able to purchase under specified conditions. It should be distinguished from sales. Sales may be lower than true demand because of stockouts, credit restrictions, delivery failures, or lost orders. Conversely, sales may be temporarily inflated by promotions, forward buying, or abnormal orders. A forecasting process must therefore identify whether the available data represent unconstrained demand or merely fulfilled transactions.

Demand has several dimensions:

• Level: the underlying average magnitude of demand.
• Trend: a persistent upward or downward movement.
• Seasonality: a recurring pattern linked to a calendar or known cycle.
• Cyclical movement: longer-term fluctuations associated with economic or industry conditions.
• Irregular variation: random or unusual effects that are difficult to predict.
• Intermittency: demand occurring sporadically with many zero observations.

A useful conceptual representation is:

where Y_t is observed demand in period t, T_t is the underlying level and trend component, S_t is a seasonal component, and I_t is irregular variation. This additive form is most suitable when seasonal fluctuations have approximately constant absolute size. When seasonal variation grows in proportion to the level, a multiplicative representation is more appropriate:

The decomposition is a model, not a law. Its usefulness depends on the data-generating process, the forecast horizon, and the stability of the environment.

Forecasting is also a hierarchical activity. A top-level forecast for total demand may be accurate while individual SKU forecasts are poor, or the reverse. Organisations should define whether forecasts must be coherent: for example, the sum of regional forecasts should equal the national forecast. Reconciliation methods can enforce such consistency, but they do not automatically improve the underlying information.

2. The Forecasting Process

A professional forecasting process normally follows these stages:

1. Define the decision. Specify what is being forecast, the unit of measure, the time bucket, the horizon, and the decision it supports.
2. Collect and validate data. Check completeness, units, time stamps, duplicated records, missing values, outliers, promotions, stockouts, and changes in product codes.
3. Explore the pattern. Plot the series and examine level, trend, seasonality, variability, and structural breaks.
4. Select a method. Compare suitable qualitative and quantitative methods rather than assuming that a complex model is superior.
5. Generate a baseline forecast. A transparent benchmark, such as the naïve forecast, provides a reference for judging more elaborate methods.
6. Incorporate legitimate information. Include promotions, price changes, planned contracts, market intelligence, and known events, while documenting assumptions.
7. Validate and measure error. Use a holdout period or rolling-origin evaluation that resembles actual forecasting conditions.
8. Review collaboratively. Sales, marketing, procurement, operations, finance, and customers may contribute information, but overrides must be evidence-based.
9. Decide and monitor. Translate the forecast into inventory, capacity, purchasing, and service decisions; then monitor accuracy and bias.
10. Learn and improve. Investigate forecast failures and revise data, parameters, processes, or assumptions.

The forecast itself should be accompanied by a confidence or prediction interval where possible. A point forecast without an indication of uncertainty can encourage false precision. The width of an interval normally increases with the forecast horizon because more future disturbances can occur.

3. Qualitative Forecasting Methods

Qualitative methods are particularly valuable when historical data are unavailable, unreliable, or not representative of the future. They are also useful for long-range strategic questions, new products, regulatory changes, technological disruption, and markets where expert knowledge contains information not yet visible in the data.

Expert judgement

Managers or specialists estimate future demand using experience and contextual knowledge. This method is fast and inexpensive but is vulnerable to optimism, anchoring, availability bias, political pressure, and inconsistent assumptions. Judgement should therefore be documented, compared with a quantitative baseline, and reviewed after the outcome is known.

Sales-force composite

Sales representatives estimate demand for their territories or accounts. They may possess current information about customer projects and competitor activity. However, estimates can be distorted by incentives, sandbagging, or uneven market knowledge. Training, standard definitions, and independent review improve reliability.

Market research and customer surveys

Surveys, interviews, test marketing, conjoint analysis, and observation can identify preferences and purchase intentions. Stated intention is not identical to actual behaviour, so survey design, sampling, response bias, and the difference between willingness and ability to pay must be considered.

Delphi method

The Delphi method obtains successive, anonymous estimates from a panel of experts. A coordinator summarises the responses and circulates them for another round. Anonymity reduces the influence of status and dominant personalities. Consensus, however, should not be confused with truth; a minority view may contain the most important warning.

Scenario analysis

Scenario analysis constructs internally consistent alternatives such as high growth, base case, and disruption. It is not a probability forecast unless probabilities are explicitly assigned. Its value lies in preparing decisions that remain robust under different futures.

Qualitative evidence should be traceable. An organisation should record who supplied an assumption, what evidence supported it, its expected timing, and how it will be tested. This creates accountability without eliminating professional judgement.

4. Quantitative Time-Series Forecasting

Quantitative time-series methods use historical observations of the variable being forecast. They are most useful when the future is expected to exhibit patterns related to the past. Their principal limitation is that structural change, new competitors, policy shifts, or major disruptions can invalidate historical relationships.

4.1 Naïve and seasonal-naïve benchmarks

The simplest forecast for the next period equals the most recent observation:

where F_t+1 is the forecast for the next period and Y_t is actual demand in the current period. For seasonal demand, a seasonal-naïve forecast uses the observation from the same season in the previous cycle:

where h is the forecast horizon and m is the number of periods in a seasonal cycle. These benchmarks require little computation and are essential for evaluating whether a more complex method adds value.

4.2 Simple moving average

A moving average forecasts the next period using the arithmetic mean of the most recent n observations:

Here, n is the window length. The method assumes that recent observations are relevant and that no strong trend or changing seasonal pattern dominates the series. A short window reacts quickly but is more volatile; a long window is smoother but slower to respond.

Worked example. Weekly demand for a component was 120, 135, 150, and 145 units in the most recent four weeks. A three-week moving-average forecast for week 5 is:

In practice the organisation may round to 143 units, 144 units, or retain decimals until the final planning decision. The rounding rule should be consistent with the item and the decision context. The oldest observation, 120, is excluded because the window contains only the three latest observations.

4.3 Weighted moving average

A weighted moving average assigns different weights to recent observations:

with the conditions:

The weights are dimensionless, and the forecast has the same units as demand. Greater weight on recent demand improves responsiveness when conditions change, but it can amplify random noise. Weights should be selected using validation or a clear operational rationale, not convenience alone.

4.4 Simple exponential smoothing

Simple exponential smoothing is suitable for a series with a relatively stable level and no systematic trend or seasonality. The recursive equation is:

where 0 ≤ α ≤ 1 is the smoothing constant, Y_t is actual demand, and F_t is the forecast for period t. Both terms have demand units, while α is unitless. An initial forecast must be specified, often as the first observation or an average of an initial set of observations.

A high α gives more weight to recent demand and responds quickly; a low α produces greater smoothing. The method can also be written as an error-correction equation:

The quantity Y_t − F_t is the forecast error for period t. This form makes the updating logic especially clear.

Worked example. Suppose the current forecast is 200 units, actual demand is 220 units, and α = 0.30. Then:

The forecast rises by only 6 units because exponential smoothing deliberately moderates the effect of one observation.

4.5 Trend-adjusted and seasonal methods

Simple exponential smoothing should not be used uncritically when a clear trend exists. Holt’s method estimates both level and trend:

Here, L_t is the estimated level, B_t is the estimated trend, α and β lie between 0 and 1, and h is the number of periods ahead. The additive trend form assumes that the trend increment is approximately constant. Damped-trend variants reduce the long-range trend effect when indefinite extrapolation would be unrealistic.

When seasonality is present, a seasonal exponential-smoothing model such as Holt–Winters can represent level, trend, and seasonal indices. Additive seasonality is suitable when the seasonal effect is approximately constant in size; multiplicative seasonality is suitable when it is proportional to the level. Parameter estimation and initialisation should be performed carefully, preferably through software that supports out-of-sample validation.

4.6 Causal and regression models

A causal model explains demand using explanatory variables such as price, advertising, income, weather, competitor activity, or number of active customers. A simple linear model is:

where Y_t is demand, X_t is an explanatory variable, β₀ is the intercept, β₁ is the expected change in demand for a one-unit change in X_t, and ε_t is the unexplained error. Causal interpretation requires more than correlation: omitted variables, reverse causality, price endogeneity, and common trends can produce misleading relationships.

Regression is useful when the explanatory variables are known or can themselves be forecast for the decision horizon. A model that relies on future advertising expenditure or competitor price must include a process for obtaining those inputs. Otherwise, apparent model accuracy during development may not be achievable in practice.

5. Trend and Seasonal Analysis

Visualisation is the first analytical tool. A line chart of demand over time can reveal changes that summary statistics conceal. Analysts should inspect the original series, a time plot with operational events, and—where relevant—seasonal subseries or year-over-year comparisons.

Trend may be estimated by fitting a time variable, smoothing, or comparing moving averages. A simple trend equation is:

where Ŷ_t is fitted demand, t is a time index, a is the intercept, and b is the estimated change per period. The slope has demand units per time period. A fitted trend is descriptive within the observed range; extrapolating far beyond that range is risky.

Seasonal indices quantify the typical relative or absolute effect of a season. Under a multiplicative model, an index above 1 indicates demand above the underlying level, while an index below 1 indicates demand below it. For a complete cycle, indices are commonly normalised so their average is 1. Under an additive model, seasonal effects are expressed in demand units and typically sum to zero across a complete cycle. The choice must match the behaviour of variability.

Seasonality must be distinguished from a one-time event. A festival that occurs annually may create seasonality; a one-off product launch does not. Calendar effects such as the number of working days, moving holidays, and weather should be represented explicitly where they materially affect demand.

6. Measuring Forecast Error and Bias

Let the forecast error for period t be:

A positive error means actual demand exceeded the forecast (under-forecasting); a negative error means the forecast exceeded actual demand (over-forecasting). The sign convention must be stated because some organisations define error in the reverse direction.

Important measures include:

Here, n is the number of evaluated periods, Y_t is actual demand, and F_t is forecast demand. Units must be considered: MAE and RMSE use demand units, MSE uses squared demand units, and percentage measures are dimensionless.

MAPE can be severely misleading for low-volume or zero-demand items. Alternatives include MAE, WAPE, or scaled errors. A scaled error compares the model error with the error of a naïve benchmark, but its exact form must be defined consistently. No single metric is universally best. A service-critical item may justify a metric linked to stockout risk, while a portfolio may require volume-weighted performance.

Worked error example. Actual demand over three periods is 100, 120, and 80 units; forecasts are 90, 130, and 100 units. Errors are 10, −10, and −20 units.

The negative ME indicates under-forecasting under the stated convention. MAE shows typical absolute error, while RMSE is higher because it gives extra influence to the 20-unit error. With only three observations, these estimates are illustrative rather than statistically stable.

Forecast accuracy should be assessed on data not used to fit or tune the method. A single holdout period may be unrepresentative; rolling-origin evaluation repeatedly trains on earlier data and tests on subsequent observations. Accuracy should also be segmented by product, customer, horizon, and demand pattern. Aggregate accuracy can conceal systematic failure in strategically important items.

7. Data Visualisation and Operations Analytics

Effective visualisation supports understanding before modelling. Useful displays include:

• time-series line charts with promotions, stockouts, and disruptions annotated;
• actual-versus-forecast plots;
• error histograms and error-over-time charts;
• seasonal profiles by month, weekday, or hour;
• scatter plots of demand against price, weather, or marketing expenditure;
• Pareto charts ranking items by volume, value, error, or service impact;
• heat maps showing demand by product and location;
• control-style dashboards showing forecast bias, inventory exposure, and service outcomes.

A dashboard should distinguish leading indicators, such as forecast bias and changes in customer orders, from lagging outcomes, such as stockouts and obsolete inventory. Visual design should use clear units, consistent time scales, labelled axes, accessible colours, and meaningful comparisons. Dual axes can create misleading impressions and should be used sparingly.

Analytics maturity develops from descriptive analysis (what happened), to diagnostic analysis (why it happened), predictive analysis (what may happen), and prescriptive analysis (what should be done). Forecasting is predictive, but prescriptive decisions also require costs, capacity, constraints, service targets, and risk preferences.

8. Sensitivity Analysis and Evidence-Based Decisions

A forecast is an input to a decision, not the decision itself. Sensitivity analysis examines how a recommendation changes when important assumptions vary. For example, a planner may test demand growth of 2%, 5%, and 8%, alternative lead times, different service targets, or a promotion that increases demand but reduces margin.

If a decision outcome is represented by Z = f(X₁, X₂, …, X_k), sensitivity analysis changes one or more inputs X_i and observes the resulting change in Z. A local sensitivity measure can be expressed as:

where ∆Z is the change in outcome and ∆X_i is the corresponding change in an input. This is an approximate slope, not a universal causal effect. If inputs interact, one-at-a-time changes may miss important combinations; scenario analysis or simulation is then more appropriate.

Evidence-based decision making combines four forms of evidence: data, model output, professional expertise, and stakeholder or customer knowledge. A disciplined decision record should state the base forecast, assumptions, alternatives considered, expected operational consequences, risks, trigger conditions, and review date. This avoids both blind reliance on models and unstructured managerial override.

9. Applied Case: FreshBite Distribution (Illustrative)

Background

FreshBite Distribution supplies chilled snack boxes to supermarkets. Weekly demand has increased, but promotions and occasional stockouts make sales data difficult to interpret. The operations team must set a four-week replenishment plan for a high-volume product.

Problem

The latest four weeks of recorded sales were 1,020, 1,080, 1,150, and 1,120 boxes. Investigation found that week 1 had a two-day stockout, so recorded sales understated unconstrained demand. A supermarket promotion occurred in week 3 and will not repeat in the next four weeks. The sales team expects modest growth because two new stores will open, but it has not provided a documented estimate.

Analysis

A four-week moving average would be:

However, the calculation treats the stockout week and promotional week as ordinary observations. The analyst therefore creates a baseline using the three unaffected or more representative observations, after estimating that normal demand in week 1 would have been approximately 1,070 boxes. The adjusted average becomes:

The week-3 promotion is excluded from the baseline but retained as an event record for future promotional planning. The team compares this baseline with simple exponential smoothing and a sales estimate. Exponential smoothing gives a similar result, but the parameter is selected using rolling validation rather than chosen because it produces a preferred answer.

Approach and solution

The team adopts 1,090 boxes as the statistical baseline, then develops scenarios:

The base estimate is calculated as 1,090 × 1.04 = 1,133.6, rounded to 1,134 boxes. The replenishment decision also considers shelf-life, supplier lead time, transport capacity, and the cost of waste. Rather than simply ordering the high scenario, FreshBite uses a review trigger: if the first two weeks of orders exceed the base plan by a defined threshold, the forecast is revised and an expedited supplier option is activated.

Outcome

This approach prevents the stockout from depressing the baseline and prevents the one-time promotion from being treated as recurring demand. The scenario range exposes the risk of new-store uncertainty. The trigger creates flexibility without committing immediately to maximum inventory. After four weeks, the team compares actual demand, forecast error, waste, and service performance, then updates the model.

Learning points

1. Data cleansing requires operational knowledge, not only statistical software.
2. Sales history may not equal true demand.
3. Events should be coded and treated explicitly.
4. A forecast range and response plan can be more useful than one precise number.
5. Judgement is valuable when documented, testable, and combined with a quantitative baseline.

10. Governance, Technology, and Ethical Issues

Forecast governance defines ownership, data standards, approval rights, override rules, performance reporting, and escalation. A forecast should have a named owner, a documented version, and a clear cut-off date. Changes made after the cut-off should be logged rather than silently replacing the original forecast.

Artificial intelligence and machine learning can identify nonlinear relationships, high-dimensional patterns, and interactions across products and locations. Their benefits depend on adequate data, stable definitions, appropriate validation, and operational integration. Complex models may fail during regime changes, produce unstable results for sparse items, or be difficult to explain. A strong practice is to compare them with transparent benchmarks and to use explainability, monitoring, and human review for consequential decisions.

Ethical issues include privacy in customer-level data, discriminatory allocation of scarce products, manipulation of forecasts to influence budgets, and opacity in automated recommendations. Personal data should be minimised, protected, and used for legitimate purposes. Forecasts affecting employment, customer access, or essential goods require heightened oversight. Organisations should not present probabilistic output as certainty.

Contemporary developments include probabilistic forecasting, demand sensing using near-real-time signals, collaborative planning, forecast reconciliation across hierarchies, and digital twins for scenario testing. These techniques are valuable when they improve decisions, not merely when they increase technical sophistication. Resilience planning increasingly combines forecasts with stress tests for disruption, supplier failure, extreme weather, and abrupt demand shifts.

11. Best-Practice Checklist

A capable operations analyst should:

• define demand, horizon, frequency, and unit before modelling;
• separate sales from lost or constrained demand where possible;
• visualise data and annotate known events;
• establish naïve or seasonal-naïve benchmarks;
• select methods according to pattern, horizon, and data quality;
• validate using future-like holdout or rolling-origin tests;
• report both accuracy and bias;
• avoid MAPE when actual demand is zero or very small;
• distinguish recurring seasonality from one-time events;
• document judgemental overrides and their evidence;
• use scenarios and sensitivity analysis for uncertain assumptions;
• connect forecast performance to service, inventory, cost, and waste;
• monitor model drift and review parameters regularly;
• protect data and maintain an auditable governance process.

Chapter Summary

Demand forecasting estimates future requirements under uncertainty and supports purchasing, capacity, inventory, scheduling, and service decisions. Demand may contain level, trend, seasonality, cyclical movement, irregular variation, and intermittency. Qualitative approaches—such as expert judgement, Delphi panels, market research, and scenarios—are important when data are limited or structural change is expected. Quantitative approaches include naïve forecasts, moving averages, exponential smoothing, trend and seasonal models, and causal regression.

Moving averages smooth recent observations; weighted averages adjust responsiveness; exponential smoothing updates forecasts through a smoothing constant. Trend and seasonal models are appropriate only when their assumptions fit the data. Forecast error should be measured with clearly defined conventions using metrics such as ME, MAE, RMSE, and WAPE, recognising the limitations of percentage measures for zero or low demand. Visualisation reveals patterns and anomalies, while sensitivity and scenario analysis show how decisions change under uncertainty.

Effective forecasting is a governed business process rather than a purely statistical exercise. Data quality, event identification, cross-functional judgement, validation, ethical use, technology, and continuous monitoring determine whether forecasts create operational value. The strongest decisions combine transparent models with contextual expertise and explicit response plans.

Key Terms

Demand forecasting: Estimating future customer requirements for a specified item, time, and market.

Operations analytics: Use of data, models, and analysis to improve operational decisions.

Time series: Observations recorded sequentially over time.

Level: The underlying average magnitude of a demand series.

Trend: Persistent upward or downward movement in demand.

Seasonality: A recurring pattern associated with a calendar or cycle.

Intermittent demand: Demand with sporadic occurrences and many zero periods.

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

Moving average: The mean of a fixed number of recent observations.

Weighted moving average: A moving average assigning different weights to observations.

Exponential smoothing: A recursive forecasting method that weights recent information more strongly.

Smoothing constant: The parameter controlling responsiveness in exponential smoothing.

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

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

MAE: Mean absolute error, measured in the original demand units.

RMSE: Root mean squared error, which penalises large errors.

MAPE: Mean absolute percentage error, problematic when actual demand is zero or very small.

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

Sensitivity analysis: Examination of how an outcome changes when assumptions or inputs vary.

Forecast governance: Policies and controls governing data, methods, ownership, overrides, and review.

Review Questions

Short Answer Questions

1. What is the difference between sales and unconstrained demand?
2. State two conditions under which qualitative forecasting is especially useful.
3. What does a high smoothing constant do in simple exponential smoothing?
4. Why can MAPE be unsuitable for intermittent demand?
5. What is the purpose of a naïve forecast benchmark?

Long Answer Questions

1. Explain the stages of a professional forecasting process, from defining the decision to monitoring results.
2. Compare moving averages and exponential smoothing in terms of responsiveness, data requirements, and limitations.
3. Explain additive and multiplicative seasonality and describe when each representation is appropriate.
4. Discuss how forecast bias and forecast accuracy differ, and explain why both should be monitored.
5. Evaluate the role of qualitative judgement in a quantitative forecasting system.

Analytical / Application Questions

1. A retailer records demand of 240, 260, 250, and 280 units in four successive weeks. Calculate the three-week moving-average forecast for week 5 and explain one limitation of the result.
2. Actual demand is 500 units and the forecast is 460 units. Calculate the forecast error using e_t = Y_t − F_t and interpret its sign.
3. A promotion, a stockout, and a new store opening occur in the same demand history. Design a forecasting and sensitivity-analysis approach that prevents these events from being treated as identical ordinary observations.

Multiple Choice Questions with Answers

Question 1

Which forecast is the standard naïve forecast for period t+1?

A. The average of all historical observations

B. The actual demand in period t

C. The maximum historical demand

D. The forecast error in period t

Correct Answer: B. The actual demand in period t

Question 2

A three-period moving average uses demands of 90, 120, and 150 units. What is the forecast for the next period?

A. 110 units

B. 115 units

C. 120 units

D. 130 units

Correct Answer: C. 120 units

Question 3

In simple exponential smoothing, what is the effect of increasing α, while keeping the data unchanged?

A. The forecast responds more strongly to recent demand

B. The forecast becomes independent of recent demand

C. Seasonality is automatically removed

D. Forecast errors become exactly zero

Correct Answer: A. The forecast responds more strongly to recent demand

Question 4

Actual demand is 80 units and forecast demand is 95 units. Using e_t = Y_t − F_t, what is the error?

A. 15 units, indicating under-forecasting

B. −15 units, indicating over-forecasting

C. 175 units, indicating under-forecasting

D. −175 units, indicating over-forecasting

Correct Answer: B. −15 units, indicating over-forecasting

Question 5

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

A. MAE

B. MSE

C. MAPE

D. WAPE

Correct Answer: A. MAE

Question 6

Why is MAPE potentially unsuitable when actual demand is zero?

A. It always produces a negative value

B. Its denominator becomes zero

C. It ignores forecast values

D. It measures only bias

Correct Answer: B. Its denominator becomes zero

Question 7

Which method is most directly designed to collect successive anonymous estimates from experts?

A. Delphi method

B. Seasonal naïve method

C. Simple moving average

D. Linear regression

Correct Answer: A. Delphi method

Question 8

Under a multiplicative seasonal model, a seasonal index of 1.20 generally indicates that demand is:

A. 20% below the underlying level

B. 20 demand units above the underlying level

C. 20% above the underlying level

D. Unrelated to the underlying level

Correct Answer: C. 20% above the underlying level

Question 9

What is the principal purpose of rolling-origin evaluation?

A. To remove all outliers automatically

B. To test forecasting performance under repeated future-like conditions

C. To guarantee that the forecast is unbiased

D. To increase the smoothing constant

Correct Answer: B. To test forecasting performance under repeated future-like conditions

Question 10

Which practice best supports evidence-based forecast overrides?

A. Replacing the statistical forecast without recording the reason

B. Using the highest estimate to ensure capacity is never short

C. Documenting the evidence, assumption, owner, and review date

D. Ignoring forecasts whenever expert opinion differs

Correct Answer: C. Documenting the evidence, assumption, owner, and review date