Judgmental Forecasting and Forecast Combination

Introduction and Context

Forecasting is often presented as a statistical task, but many operational decisions concern futures for which historical data are incomplete, unstable, or structurally irrelevant. A new product has no sales history; a competitor may enter unexpectedly; a regulation may change demand; and a sales team may possess information about customer negotiations that has not yet appeared in the data. Judgmental forecasting addresses these situations by using informed human assessment. Forecast combination integrates such assessments with statistical models, market intelligence, and external forecasts.

Judgment is not simply a substitute for mathematics. It is a source of information that must be elicited, documented, tested, and governed. Human experts can recognize structural changes and contextual signals, but they are also vulnerable to anchoring, overconfidence, political pressure, and groupthink. The central managerial challenge is therefore to design a process in which expert knowledge adds information without merely adding noise or organizational bias.

Learning Objectives

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

1. Explain the purpose, strengths, and limitations of judgmental forecasting.
2. Apply expert panels, the Delphi method, sales-force composites, analogies, and scenario methods.
3. Diagnose cognitive, motivational, and organizational biases in forecasting.
4. Distinguish forecasts from assumptions, targets, and decisions.
5. Combine statistical, judgmental, and external forecasts systematically.
6. Evaluate forecast processes using accuracy, calibration, transparency, and decision usefulness.
7. Design a governed forecasting process for an operational setting.

1. The Nature of Judgmental Forecasting

A forecast is an estimate of an uncertain future outcome, usually conditional on information available at a specified time. A target is a desired outcome, while a plan describes actions intended to achieve an outcome. Confusing these concepts produces biased forecasts. For example, “sales must grow by 10 percent” is a target; it is not evidence that sales will grow by 10 percent.

Judgmental forecasting uses structured or unstructured assessments by individuals or groups with relevant knowledge. Its inputs may include:

• customer conversations and contract pipelines;
• knowledge of competitor actions;
• regulatory and technological developments;
• operational constraints;
• product-launch information;
• local market knowledge; and
• interpretation of unusual or unprecedented events.

Judgment is especially valuable when the environment is changing, the forecast horizon is long, the product is new, or data are sparse. It is less defensible when experts merely reproduce a recent number without explaining what information they add.

A useful distinction is between judgmental adjustment and judgmental replacement. In adjustment, a statistical forecast is produced first and experts modify it for documented reasons. In replacement, experts produce the entire forecast. Adjustment is usually easier to audit because the baseline and the reasons for departure remain visible.

2. Major Judgmental Methods

2.1 Expert panels

An expert panel brings together specialists from functions such as marketing, operations, finance, procurement, engineering, and customer service. The panel discusses evidence and produces a forecast, often with a range and confidence assessment.

A well-designed panel should specify:

1. the forecast variable, horizon, units, and segmentation;
2. the information supplied to participants;
3. the evidence required for an adjustment;
4. the process for resolving disagreement; and
5. the method for recording assumptions and uncertainty.

Panels can exploit diverse knowledge, but face social conformity, status effects, dominant personalities, and political bargaining. A chair should distinguish evidence from advocacy and should prevent the group from treating consensus as proof.

2.2 The Delphi method

Delphi is a structured, usually anonymous, multiround elicitation process. Experts independently provide estimates and rationales. A coordinator summarizes the responses, often reporting the median, range, and reasons for disagreement. Experts then revise their estimates in subsequent rounds.

Anonymity reduces status pressure and allows minority views to survive. Delphi is appropriate for long-range technological, policy, or market questions where no single expert possesses complete knowledge. It is not automatically superior to a meeting: poorly selected experts, vague questions, or artificial pressure toward consensus can still undermine results.

Delphi should not require convergence. Persistent disagreement may reveal genuinely different assumptions. The final output can therefore be a distribution or a set of conditional scenarios rather than one number.

2.3 Sales-force composite forecasting

In a sales-force composite, sales representatives estimate future sales for their territories, accounts, or products. These estimates are aggregated, commonly after review by regional and product managers.

Salespeople possess direct customer knowledge, but incentives can distort estimates. A representative may understate demand to obtain an easier quota or overstate it to secure inventory and management attention. The process should therefore distinguish the sales forecast from the sales target and should compare estimates with historical accuracy by territory and product.

A basic aggregate is:

where f_i,t is the forecast from salesperson or territory i for period t, and F_t is the total forecast. This sum is appropriate when territory forecasts are expressed in the same units and refer to mutually exclusive demand. Double-counting occurs if several representatives claim the same account or opportunity.

2.4 Analogy methods

An analogy forecasts a new situation using the historical development of a comparable product, market, technology, or customer segment. The analogy may be quantitative, such as aligning adoption curves, or qualitative, such as comparing the sequence of regulatory and competitive events.

A valid analogy requires more than superficial similarity. Analysts should assess:

• similarity of customer need;
• price and economic conditions;
• distribution and adoption mechanisms;
• competitive structure;
• technology maturity;
• scale and market boundaries; and
• differences in timing and institutional context.

Analogy is particularly useful for new products, but it can fail when the supposedly comparable product benefited from conditions that no longer exist.

2.5 Scenario methods

Scenario planning develops several coherent descriptions of how the future could unfold. A scenario is not a probability-free forecast and should not be confused with a wish list. It is a structured combination of assumptions, drivers, events, and consequences.

A practical scenario process is:

1. define the decision and horizon;
2. identify external drivers and critical uncertainties;
3. construct a small number of internally consistent scenarios;
4. translate each scenario into operational demand implications;
5. identify leading indicators; and
6. prepare contingent actions.

Scenarios are valuable when uncertainty is structural rather than merely random. They support capacity flexibility, inventory policies, workforce plans, and trigger-based decisions. Their limitation is that organizations may produce attractive narratives without assigning indicators or actions.

3. Cognitive and Organizational Biases

Bias is a systematic tendency that can affect estimates, information selection, or interpretation. It does not imply bad faith; however, unmanaged bias can create persistent forecast error.

Forecast governance should also address organizational bias. A finance department may prefer conservative numbers to protect budgets; sales may prefer high numbers to obtain inventory; operations may prefer low numbers to reduce workload. These are legitimate interests, but they should not be hidden inside a supposedly objective forecast.

4. Eliciting and Documenting Expert Judgment

A disciplined elicitation process has five stages.

Stage 1: Define the question

Specify the measure, population, time period, horizon, aggregation level, and information cutoff. “What will demand be?” is inadequate. “What is the unconstrained unit demand for Product A in Region North during April, excluding promotional uplift?” is substantially better.

Stage 2: Establish the baseline

Produce a transparent statistical or naïve benchmark where possible. The baseline prevents experts from treating an unsupported number as the starting point and allows the value of judgment to be assessed.

Stage 3: Obtain independent assessments

Ask each expert for a point estimate, a plausible range, key assumptions, and confidence. Independent estimates should precede group discussion to reduce anchoring and conformity.

Stage 4: Reconcile differences

Discuss why estimates differ. Differences may reflect alternative assumptions rather than disagreement about facts. Record the assumptions and identify which are testable.

Stage 5: Audit the result

Store the original baseline, adjustments, rationale, participants, information set, and final forecast. After the period ends, compare the forecast with actual demand and evaluate both accuracy and reasoning.

For a point forecast F_t and actual value A_t, a common absolute error measure is:

where AE_t is absolute error in the same units as demand. Mean absolute error over n periods is:

MAE is interpretable and treats positive and negative errors symmetrically. It should be supplemented by bias measures because a method can have modest average absolute error while consistently overforecasting or underforecasting. One simple mean error is:

Here, positive values indicate that actual demand tends to exceed forecasts. Percentage measures require caution when actual demand is zero or close to zero.

For probabilistic forecasts, calibration matters. If experts state that an interval has 80 percent confidence, approximately 80 percent of comparable realized outcomes should fall within such intervals over a sufficiently large evaluation set. This is a property of repeated forecasts, not a guarantee for one observation.

5. Forecast Combination

Forecast combination uses two or more forecasts to produce a final forecast. Components may include a time-series model, an expert adjustment, a sales-force estimate, a supplier forecast, and an industry outlook.

5.1 Simple weighted combination

The general weighted combination is:

where F_t^C is the combined forecast, F_j,t is component forecast j, w_j is its weight, and m is the number of forecasts. For a convex combination, weights satisfy w_j ≥ 0 and ∑ⱼ w_j = 1. Under these conditions, the combined forecast lies between the smallest and largest component forecasts.

Weights may be equal, judgmentally assigned, or estimated from historical performance. Equal weighting is a strong benchmark because it is transparent and often robust when estimation samples are short or component forecasts are highly correlated.

5.2 Error-based weighting

Suppose component j has historical mean squared error:

A simple inverse-error weighting rule is:

This rule requires positive, comparable MSE values and a representative evaluation period. It can be unstable when one forecast happens to perform exceptionally well in a short sample. It also ignores dependence among forecast errors. Two forecasts based on the same data may appear to provide two independent signals when they do not.

5.3 Combining a baseline with a judgmental adjustment

A particularly useful structure is:

where F_t^B is the baseline forecast and J_t is a documented judgmental adjustment. The adjustment should have a reason, such as a confirmed promotion, a lost contract, or a capacity constraint. It should not be used merely to make the number agree with a budget.

5.4 Worked example

An appliance manufacturer produces three April forecasts for a region:

• statistical model: 10,000 units;
• sales-force composite: 11,200 units; and
• external market forecast: 10,600 units.

Management assigns weights of 0.50, 0.30, and 0.20 respectively. The weights are nonnegative and sum to one. Therefore:

The result is not necessarily more accurate simply because it is a combination. Its credibility depends on the independence, relevance, and historical performance of the components. If the sales-force composite already incorporates the statistical forecast, the effective information content may be lower than the formula suggests.

6. Advanced Combination Issues

Correlated forecast errors

If two forecasts make similar errors, combining them provides less diversification. In statistical terms, the variance of a weighted combination depends not only on each component’s variance but also on covariances among errors. For two forecasts with errors e₁ and e₂, weights w and 1−w, and error variances and covariance defined conventionally:

The formula applies when the forecasts are combined linearly and the error moments are meaningful for the evaluation population. Positive covariance reduces the benefit of combination; negative covariance can improve diversification, although negative relationships may not persist.

Hierarchical reconciliation

Forecasts are often generated at product, region, customer, and total-company levels. Their sums may be inconsistent. A governance process should specify whether the organization forecasts at the lowest level and aggregates upward, forecasts the total and allocates downward, or uses a reconciliation method. Reconciliation should preserve business meaning and avoid creating apparent precision.

Uncertainty ranges

A final point estimate should be accompanied by a range where decisions are sensitive to uncertainty. Ranges may be generated from statistical prediction intervals, expert quantiles, or scenario outcomes. The method and confidence interpretation must be stated. A range is not credible merely because its endpoints are wide.

7. Illustrative Case: Northstar Medical Devices

Background

Northstar Medical Devices is preparing a six-month demand plan for a newly launched portable diagnostic device. Historical sales are limited to eight weeks, and a regulatory approval has recently expanded the eligible customer base. Hospitals are also facing budget uncertainty.

Problem

The statistical model predicts 4,800 units for the next quarter. Regional sales managers propose 6,200 units based on customer discussions. An industry analyst predicts market growth that would imply 5,500 units. Operations must decide whether to produce 5,000, 5,500, or 6,000 units, while avoiding excessive inventory of a product with a short technology cycle.

Analysis

The forecasting team identifies three distinct information sources: the model captures early sales patterns; sales managers know named opportunities but may have quota incentives; and the analyst provides an external market perspective but may use a broad market definition. The team requests account-level evidence from sales, verifies which opportunities have procurement approval, and constructs two scenarios:

• Moderate adoption: delayed hospital budgets and gradual conversion, producing approximately 5,000 units.
• Accelerated adoption: regulatory approval translates quickly into orders, producing approximately 6,000 units.

The team also identifies leading indicators: signed purchase orders, approved capital budgets, distributor inventory, and weekly conversion rates.

Approach and solution

The baseline remains 4,800 units. After evidence review, the team combines the statistical forecast, an evidence-adjusted sales forecast of 5,700 units, and the external forecast of 5,500 units with weights 0.40, 0.35, and 0.25:

Rather than produce exactly 5,290 units, operations adopts a flexible plan: initial production of 5,000 units, supplier capacity reserved for an additional 1,000, and a review after four weeks. The review is triggered if signed orders exceed 4,600 units or if conversion rates fall below a defined threshold.

Outcome

The process produces a transparent central forecast, an explicit range, and an operational response to uncertainty. It does not claim that the combined number is objectively correct. Its value lies in integrating evidence, exposing assumptions, and linking forecast uncertainty to reversible decisions.

Learning points

1. New-product forecasting requires analogies, expert knowledge, and scenarios because history is limited.
2. Sales information is useful only when opportunities are defined and double-counting is controlled.
3. Combination weights should reflect evidence and be reviewed after outcomes are known.
4. A forecast should be connected to triggers and capacity options, not treated as a single irreversible commitment.

8. Technology, AI, and Governance

Digital planning platforms can collect estimates, preserve version histories, compare forecasts with actuals, and calculate accuracy by person, product, and horizon. Natural-language systems can summarize customer notes or identify assumptions, but automated summaries may reproduce biased or incomplete source material.

Machine-learning systems can combine structured and unstructured data, yet they do not eliminate judgment. Human review remains necessary when relationships change, data are generated by strategic actors, or decisions affect customers and employees. Organizations should document data provenance, access rights, model version, overrides, and reasons for material adjustments.

AI-assisted forecasting should meet four governance principles:

• traceability: users can identify the data and logic behind a recommendation;
• contestability: experts can challenge an output with evidence;
• fairness: systematic disadvantage to regions, customers, or groups is monitored;
• accountability: a named decision owner remains responsible for the forecast process.

Confidential customer information must be protected, and forecasts should not be used to conceal discriminatory allocation rules or to manipulate stakeholders.

9. Best-Practice Forecasting Process

A mature process typically follows these principles:

1. Start with a transparent statistical baseline.
2. Define the forecast separately from targets and budgets.
3. Elicit independent judgment before group discussion.
4. Require explicit assumptions and supporting evidence.
5. Use ranges and scenarios when structural uncertainty is high.
6. Weight forecasts using validated historical performance, not organizational rank.
7. Examine correlation among forecast errors.
8. Track accuracy, bias, calibration, and decision outcomes.
9. Review forecasts at the horizon and aggregation level at which decisions are made.
10. Establish override thresholds and escalation rules.
11. Preserve an audit trail of forecasts, changes, and rationales.
12. Treat forecast performance as a learning process rather than a basis for blaming individuals.

Challenges and Limitations

Judgmental forecasting is expensive in time and coordination. Experts may not understand probability, may provide overly narrow ranges, or may lack incentives to reveal uncertainty. Consensus can conceal disagreement, while excessive documentation can turn the process into bureaucracy. Historical accuracy is also imperfect: an expert who performed well in a stable market may not be reliable during disruption.

Combination methods can create false precision. A weighted average may hide incompatible definitions, different horizons, or overlapping information. Forecast accuracy metrics can also encourage undesirable behavior, such as avoiding difficult products or manipulating the timing of forecasts. Evaluation should therefore consider forecast quality, decision consequences, responsiveness, and process integrity.

Chapter Summary

Judgmental forecasting adds contextual and forward-looking information when historical data are inadequate or structural change is important. Expert panels, Delphi, sales-force composites, analogies, and scenarios differ in their mechanisms and risks. Effective use requires precise questions, independent elicitation, explicit assumptions, documented adjustments, and post-outcome evaluation.

Human judgment is affected by anchoring, availability, confirmation, overconfidence, groupthink, status effects, and strategic incentives. Governance should separate forecasts from targets, protect dissent, and maintain an audit trail. Forecast combination can integrate statistical, judgmental, and external forecasts through transparent weighted averages or baseline-plus-adjustment processes. Weights should reflect validated performance and information diversity, while correlated errors and inconsistent definitions must be examined. The final forecast should support decisions through ranges, scenarios, triggers, and flexible operating plans.

Key Terms

Judgmental forecasting: Forecasting based primarily on structured human assessment.

Expert panel: A group of specialists that develops a forecast through organized deliberation.

Delphi method: An anonymous, multiround expert-elicitation process.

Sales-force composite: An aggregate forecast prepared from sales representatives’ estimates.

Analogy method: Forecasting a new situation using a comparable historical case.

Scenario planning: Constructing coherent alternative futures to guide decisions.

Baseline forecast: A transparent reference forecast produced before judgmental adjustment.

Judgmental adjustment: A documented modification to a baseline forecast.

Forecast combination: Integration of two or more forecasts into one output.

Convex combination: A weighted combination with nonnegative weights that sum to one.

Forecast bias: A systematic tendency to overforecast or underforecast.

Anchoring: Excessive influence of an initial estimate on subsequent judgment.

Overconfidence: Excessive certainty relative to actual forecasting performance.

Groupthink: Suppression of dissent in pursuit of apparent consensus.

Calibration: Agreement between stated uncertainty and observed frequencies.

Mean absolute error: The average absolute difference between actual and forecast values.

Scenario trigger: An observable condition that initiates a planned response.

Forecast governance: Policies and controls for producing, reviewing, and using forecasts.

Information diversity: The extent to which forecast components contribute distinct evidence.

Forecast override: A deliberate change to a model output supported by stated reasoning.

Review Questions

Short Answer Questions

1. Distinguish a forecast from a target and a plan.
2. What problem does anonymity address in the Delphi method?
3. State two conditions under which judgmental forecasting is especially useful.
4. Define anchoring and give one control response.
5. Why can equal-weight forecast combination be a useful benchmark?

Long Answer Questions

1. Compare expert panels, Delphi, sales-force composites, analogy methods, and scenario planning.
2. Explain how cognitive and organizational biases can affect operational forecasts.
3. Describe a complete process for eliciting and documenting expert judgment.
4. Explain the advantages and limitations of combining statistical and judgmental forecasts.
5. Discuss how forecast governance should change when AI tools are introduced.

Analytical / Application Questions

1. A statistical forecast is 8,000 units, a sales-force forecast is 9,200 units, and an external forecast is 8,600 units. Using weights of 0.50, 0.30, and 0.20 respectively, calculate the combined forecast and explain two assumptions required for meaningful combination.
2. Design a Delphi study to forecast adoption of a new industrial technology over five years. Specify participant selection, rounds, information supplied, treatment of disagreement, and evaluation.
3. A sales team consistently forecasts 20 percent above actual demand because forecasts influence inventory allocation. Diagnose the likely organizational incentives and design a governance response that preserves useful customer information without rewarding inflation.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a forecast from a target?

A. A forecast describes a desired outcome, while a target describes uncertainty.

B. A forecast estimates a likely outcome, while a target expresses a desired outcome.

C. A forecast is always statistical, while a target is always judgmental.

D. A forecast is confidential, while a target must be public.

Correct Answer: B. A forecast estimates a likely outcome, while a target expresses a desired outcome.

Question 2

What is the defining feature of the Delphi method?

A. A single senior manager makes the final estimate.

B. Sales representatives submit territory estimates.

C. Experts provide anonymous estimates over multiple rounds.

D. A statistical model is replaced by a scenario narrative.

Correct Answer: C. Experts provide anonymous estimates over multiple rounds.

Question 3

A baseline forecast is 10,000 units and a documented judgmental adjustment is −600 units. What is the adjusted forecast?

A. 9,400 units

B. 9,600 units

C. 10,600 units

D. 16,000 units

Correct Answer: A. 9,400 units

Question 4

Which bias occurs when an initial estimate excessively influences subsequent judgments?

A. Availability bias

B. Anchoring bias

C. Confirmation bias

D. Escalation of commitment

Correct Answer: B. Anchoring bias

Question 5

Three forecasts are 10,000, 11,200, and 10,600 units, with weights 0.50, 0.30, and 0.20. What is the weighted combination?

A. 10,200 units

B. 10,480 units

C. 10,600 units

D. 10,933 units

Correct Answer: B. 10,480 units

Question 6

Why may combining two forecasts provide little benefit?

A. Their errors may be highly correlated.

B. Their units may be identical.

C. Their weights may sum to one.

D. Both forecasts may include uncertainty.

Correct Answer: A. Their errors may be highly correlated.

Question 7

Which practice most directly reduces groupthink?

A. Allowing the most senior participant to speak first

B. Requiring consensus before estimates are recorded

C. Eliciting independent estimates before discussion

D. Removing all external information

Correct Answer: C. Eliciting independent estimates before discussion

Question 8

What does a positive mean error defined as actual minus forecast generally indicate?

A. A tendency to underforecast

B. A tendency to overforecast

C. Perfect calibration

D. Zero absolute error

Correct Answer: A. A tendency to underforecast

Question 9

Which feature makes a scenario useful for operations?

A. It presents only the most optimistic future.

B. It identifies assumptions, indicators, and contingent actions.

C. It eliminates the need for a baseline forecast.

D. It guarantees the probability of each outcome.

Correct Answer: B. It identifies assumptions, indicators, and contingent actions.

Question 10

A convex forecast combination requires weights that are:

A. All negative and sum to zero

B. Nonnegative and sum to one

C. Greater than one and unequal

D. Based only on managerial seniority

Correct Answer: B. Nonnegative and sum to one