Qualitative, Judgmental and Causal Forecasting

Introduction and context

Demand forecasting is not a choice between “human judgment” and “data.” In practice, supply-chain planners combine historical information, market intelligence, managerial experience and causal evidence. This is especially important when a product is new, demand is changing, historical data are distorted, or external drivers such as price, promotion, income and weather materially affect sales.

This chapter examines two broad families of methods. Qualitative and judgmental methods use structured knowledge from customers, salespeople, executives, experts and analogous markets. Causal methods explain demand through variables that influence it, using regression or econometric models. The central managerial challenge is to extract useful judgment without allowing optimism, anchoring, politics or untested assumptions to contaminate the forecast.

A sound forecast is therefore more than a number. It is a documented statement of expected demand, the assumptions supporting it, the uncertainty surrounding it and the actions that should follow.

Learning objectives

After studying this chapter, the reader should be able to:

1. Explain the purpose, strengths and limitations of qualitative and judgmental forecasting.
2. Apply market research, sales-force composites, executive opinion, Delphi and analogy methods.
3. Distinguish independent, dependent and consensus-based judgment processes.
4. Develop and interpret regression and econometric forecasts using price, promotion and macroeconomic drivers.
5. Assess model assumptions, multicollinearity, endogeneity, omitted variables and forecast error.
6. Design a structured forecast-assumption process that reduces judgmental bias.
7. Translate a forecast into supply-chain decisions while communicating uncertainty responsibly.

1. The role of judgment in supply-chain planning

Judgment is most valuable when the future differs from the past. Examples include a product launch, a regulatory change, a competitor withdrawal, a major promotion, a channel migration or a disruption in supply. Historical time-series methods may extrapolate an obsolete pattern, whereas informed experts can recognize a structural change.

Judgment is also necessary when data are sparse or unreliable. However, experience is not automatically accurate. A salesperson may understand a customer but overstate an opportunity to secure inventory. An executive may impose an ambitious growth target. A buyer may underestimate demand after a recent stockout. Forecast governance must therefore separate evidence from advocacy.

A useful distinction is:

• Forecast judgment: an informed estimate of future demand.
• Business target: a desired level of sales or profit.
• Supply plan: the operational response to expected demand, constrained by capacity, inventory and lead time.

These may be reconciled in a planning process, but they should not be treated as interchangeable.

2. Qualitative and judgmental forecasting methods

2.1 Market research and customer-intention studies

Market research obtains information directly from customers or prospective customers. Common approaches include surveys, interviews, focus groups, concept tests, conjoint studies, pilot launches and retailer panels.

A customer-intention survey may ask whether a respondent is likely to purchase a product. The resulting intentions should not be interpreted as actual demand without adjustment. Respondents may express interest without having a budget, may overstate socially desirable behavior or may not face the purchase decision in realistic conditions.

A practical market-research forecast normally involves:

1. Defining the population, market boundary and forecast horizon.
2. Designing a representative sample or explicitly identifying a purposive sample.
3. Measuring awareness, purchase intention, expected frequency, price sensitivity and switching behavior.
4. Calibrating stated intentions with historical conversion rates or comparable launches.
5. Segmenting results by customer type, channel, geography and use case.
6. Documenting sampling error, nonresponse and interpretation limits.

Market research is particularly useful for product design, segmentation and identifying causal mechanisms. It is less reliable when respondents have little experience with the product or when the purchase decision is routine and unconscious.

2.2 Sales-force composite

In a sales-force composite, each salesperson estimates demand for assigned accounts, territories or products. These estimates are aggregated, reviewed and reconciled with capacity and market information.

The method benefits from the sales force’s close contact with customers. It can detect tenders, competitor activity, project pipelines and account-level changes earlier than centralized statistical systems. Its weaknesses include inconsistent estimation skills, strategic inflation or understatement, double counting and excessive focus on existing accounts.

Good practice includes:

• defining the unit being forecast, such as units, cases or revenue;
• providing a common forecast horizon and probability definitions;
• separating committed orders, probable opportunities and speculative opportunities;
• requiring evidence for material changes;
• using historical calibration by salesperson or account type;
• reviewing forecasts independently of compensation negotiations where possible.

A probability-weighted pipeline can be useful. If opportunity i has potential volume q_i and estimated probability p_i, its expected volume is:

Here, 0 ≤ p_i ≤ 1, and the opportunities should be sufficiently distinct to avoid double counting. This is an expected-value estimate, not a guarantee that fractional orders will occur.

2.3 Executive opinion and management consensus

Executive opinion uses the knowledge of senior managers in marketing, sales, finance, operations and strategy. It is efficient and useful when information is dispersed or a strategic decision must be made quickly. It can also be vulnerable to hierarchy effects, groupthink, anchoring on a budget and suppression of dissent.

A disciplined process begins with independent estimates and written assumptions. The group then examines differences, identifies evidence and records the reason for the final consensus. The chair should distinguish a genuine evidence-based adjustment from a target imposed by authority.

2.4 Delphi method

The Delphi method obtains forecasts from a panel of experts through multiple anonymous rounds. A coordinator aggregates responses, summarizes reasoning and returns the summary to participants for reconsideration. Anonymity reduces status pressure and allows participants to revise their views without public embarrassment.

A robust Delphi process requires:

• carefully selected experts with relevant, not merely prestigious, experience;
• a clearly specified variable and horizon;
• independent first-round responses;
• controlled feedback based on distributions and reasons;
• a stopping rule, such as stable median estimates or a specified number of rounds;
• documentation of disagreement rather than forcing artificial consensus.

The Delphi median is often preferable to a simple mean when estimates contain extreme values. Consensus is not proof of accuracy; a panel can share the same mistaken assumption.

2.5 Analogy forecasting

Analogy forecasting uses the adoption, demand or life cycle of a comparable product, market or technology. The analogy may be historical, geographic, product-based or behavioral.

The method requires explicit comparison of similarity and difference. Relevant dimensions include price, distribution, brand strength, customer need, regulation, technology, competitive intensity and economic conditions. A simple scaling relationship may be used when a comparable market is appropriately proportional:

Here, Q_new is forecast demand for the new situation, Q_ref is demand in the reference situation and s is a documented scaling factor. The relationship is only defensible when the scaling factor captures important differences and the reference case is genuinely comparable.

Analogy is useful for new products but can fail through superficial similarity. Two products may have similar features but entirely different purchase frequency, channel economics or switching costs.

3. Combining judgment without losing accountability

Forecast combination can improve robustness when methods contain different errors. Possible combinations include statistical baseline plus documented override, weighted expert estimates, scenario ranges and hierarchical reconciliation across product, region and channel.

A weighted judgmental combination can be written as:

F_j is forecast j, w_j is its weight and F is the combined forecast. Weights should reflect demonstrated performance, information quality and relevance—not organizational rank alone. If reliable error histories are available, weights may be informed by past accuracy; otherwise, equal weights can be a transparent starting point.

The planning team should preserve the baseline forecast, the override, the reason for the override, the owner, the date and the expected trigger for revision. This creates a learning loop in which judgment can be evaluated after actual demand is known.

4. Bias in judgmental forecasting

Important biases include:

• Anchoring: insufficient adjustment from a salient initial number, such as last year’s sales or a budget.
• Availability bias: overemphasis on memorable recent events.
• Confirmation bias: seeking evidence that supports an existing view.
• Overconfidence: expressing narrow ranges despite substantial uncertainty.
• Recency bias: assuming the latest trend will continue indefinitely.
• Optimism or strategic bias: overstating demand to obtain capacity or understating it to reduce accountability.
• Groupthink: suppressing dissent in a consensus meeting.
• Status-quo bias: failing to recognize a structural market change.

Bias controls should be procedural, not merely educational. Effective controls include independent estimates before discussion, reference-class forecasting, outside-view comparisons, premortems, explicit probability ranges, evidence requirements and post-forecast accuracy reviews.

A premortem asks: “Assume this forecast was materially wrong. What caused the error?” The answers are converted into scenarios, monitoring indicators and contingency actions.

5. Structured forecast assumptions

Every material forecast should have an assumption register. A useful register contains:

Assumptions should be classified as controllable, observable but uncontrollable, or speculative. They should also distinguish base, upside and downside cases. A scenario is not a probability distribution unless probabilities are explicitly assigned and justified.

The process should include a forecast meeting agenda: review baseline, inspect actual-versus-forecast performance, examine new evidence, challenge assumptions, agree adjustments, document dissent and assign follow-up actions. This creates traceability and prevents unexplained overrides.

6. Causal forecasting

Causal forecasting models demand as a function of explanatory variables. Typical drivers include price, promotions, distribution, advertising, competitor actions, income, employment, interest rates, weather and calendar effects.

A multiple linear regression model is commonly expressed as:

y_t is demand in period t, x_j,t is driver j, β₀ is the intercept, β_j is the partial effect of driver j holding the other included drivers constant, and ε_t is the unexplained error. Units must be consistent: if demand is cases and price is currency per case, the price coefficient has units of cases per currency unit.

The model is useful when the relationships are sufficiently stable, drivers are measured with reasonable accuracy and future driver values are available or can themselves be forecast. Regression estimates association conditional on included variables; it does not automatically establish causation.

6.1 Price and promotion effects

If price is included in a demand model, a negative price coefficient is often expected, but the sign should be tested rather than assumed. Promotion variables may represent discount depth, display, feature advertising, campaign intensity or a binary event indicator. A promotion coefficient can be misleading if promotions are targeted precisely when demand is expected to be weak or strong.

For percentage interpretation, a log-log model is often useful:

Here, Q_t is demand, P_t is price and I_t may be income or another positive driver. In this specification, β₁ is the price elasticity: approximately the percentage change in demand associated with a 1% change in price, holding included variables constant. Log models require positive values and careful treatment of zeros.

A semilog model such as ln(Q_t) = β₀ + β₁P_t + ε_t interprets β₁ as the change in log demand per unit of price, not directly as a percentage change in demand without transformation.

6.2 Worked regression example

An illustrative retailer estimates weekly demand for a packaged product using price and promotional display:

Q̂ is forecast weekly demand in cases, P is price in currency units per case, and D equals 1 when a display promotion is present and 0 otherwise. The equation is assumed to be estimated from comparable weeks and used within the observed range of price and promotion conditions.

At P = 10 with a display, D = 1:

Q̂ = 1,200 − 40(10) + 180(1) = 980 cases.

Without the display, demand would be 800 cases at the same price. The estimated incremental effect is 180 cases, conditional on the model and its assumptions. The calculation does not prove that every future display will produce exactly 180 additional cases.

6.3 Model evaluation and limitations

A regression model should be assessed using both statistical and managerial criteria. Important checks include:

• Out-of-sample accuracy: performance on periods not used for estimation.
• Residual analysis: whether errors show trend, seasonality or changing variance.
• Multicollinearity: whether drivers move together, making individual effects unstable.
• Autocorrelation: whether errors in adjacent periods are related.
• Heteroscedasticity: whether error variance changes with the level of demand.
• Structural breaks: whether relationships changed after a competitor entry, policy change or disruption.
• Endogeneity: whether a driver is jointly determined with demand.
• Omitted-variable bias: whether an absent factor distorts included coefficients.

A high R² does not guarantee good future forecasts. It measures in-sample variation explained under a particular specification, not managerial usefulness or causal truth. Forecast errors should be evaluated with measures appropriate to the data. Mean absolute error is:

where y_t is actual demand, ŷ_t is forecast demand and n is the number of evaluation periods. MAE is expressed in demand units and is easy to interpret. Percentage measures can be unstable when actual demand is zero or near zero.

7. Econometric forecasting and external drivers

Econometric forecasting extends regression by representing economic behavior, dynamic relationships and feedback. Demand may depend on lagged price, advertising carryover, income, employment, exchange rates, interest rates and competitor conditions.

A distributed-lag model may be written as:

Current and lagged price effects allow demand to respond over time. Advertising A_t may have delayed effects, and income I_t may capture macroeconomic purchasing power. The model should not include variables merely because they are available; each variable needs a plausible mechanism, reliable measurement and a known or forecastable future value.

Forecasting with causal models involves two uncertainties: uncertainty in the estimated coefficients and uncertainty in future driver values. Scenario analysis is therefore valuable. For example, the planning team may generate base, downside and upside forecasts using alternative price, promotion and income assumptions.

Causal models can also support what-if analysis, such as estimating the effect of a price change. Such analysis is credible only when the model reflects the decision context. Historical price variation caused by different conditions may not identify the effect of a new pricing policy.

8. Illustrative case: FreshBite chilled meals

Background

FreshBite, an illustrative regional manufacturer, sells chilled ready-to-eat meals through supermarkets. Demand has historically been forecast using last year’s sales and sales-manager overrides. The company is introducing a premium line while facing volatile food inflation and frequent promotional activity.

Problem

The baseline forecast underestimates demand during displays but overestimates demand after price increases. Sales managers submit highly variable estimates, and senior management routinely raises the forecast to support an aggressive growth plan. Production consequently alternates between overtime and excess finished inventory.

Analysis

The planning team first separates shipments from consumption because retailer inventory accumulation had inflated some historical observations. It then creates a weekly dataset containing unit sales, price, display status, distribution, competitor price, unemployment and holiday indicators. A regression model is estimated only over weeks with reliable availability data. The team tests lagged promotion effects and reviews residuals for seasonal patterns.

Parallel qualitative work includes interviews with key retailers, a structured sales-force composite and a two-round Delphi panel involving category managers, buyers and demand planners. Experts independently estimate premium-line adoption and document assumptions about trial, repeat purchase, competitor response and shelf availability.

Approach and solution

FreshBite adopts the following process:

1. The statistical model supplies the baseline forecast.
2. Sales estimates are collected independently by account before the consensus meeting.
3. Material overrides require evidence and an assumption-register entry.
4. Base, upside and downside scenarios are produced using alternative promotion and price plans.
5. A premortem identifies risks including competitor discounting, stockouts and lower repeat purchase.
6. Forecast performance is reviewed monthly by channel and product family.
7. Compensation discussions are separated from the operational forecast meeting.

Outcome

In the first planning cycle, the principal benefit is not a claim of perfect accuracy but improved transparency. Planners can explain why the forecast changes, distinguish promotional uplift from underlying demand and identify which assumptions require monitoring. Production uses the base case for capacity planning and retains defined contingencies for the upside and downside cases.

Learning points

The case demonstrates that judgment and causal analysis are complementary. Market intelligence explains why demand may change; regression quantifies historical relationships; structured governance controls bias; and scenario planning recognizes that driver forecasts are uncertain.

9. Technology, ethics and governance

Modern planning platforms can capture overrides, automate model scoring, monitor forecast drift and combine structured external data with internal sales information. Machine-learning systems may identify nonlinear effects or interactions, but they do not remove the need for causal reasoning, data governance and managerial accountability. A complex model can reproduce biased historical decisions or fail when market conditions change.

Organizations should govern access to customer data, respect consent and privacy requirements, protect commercially sensitive information and document the provenance of external datasets. Forecasts should not be used to discriminate unfairly among customers or suppliers. AI-generated recommendations should be reviewable, tested against baselines and accompanied by uncertainty indicators.

Best-practice checklist

• Define the forecast object, horizon, unit and decision use.
• Begin with a transparent baseline.
• Select qualitative methods according to the information need.
• Collect independent judgments before group discussion.
• Document assumptions, evidence, owner and review trigger.
• Distinguish demand from sales targets and constrained shipments.
• Test causal relationships rather than assuming them.
• Validate models out of sample and inspect residual behavior.
• Use scenarios when future drivers are uncertain.
• Measure forecast performance and recalibrate judgment systematically.
• Preserve dissent and prevent authority from substituting for evidence.
• Communicate ranges and risks, not only a single point estimate.

Chapter Summary

Qualitative and judgmental forecasting methods are essential when historical data are limited, structural change is occurring or relevant information is held by customers, salespeople and experts. Market research captures customer intentions and market mechanisms; sales-force composites exploit account knowledge; executive opinion provides strategic integration; Delphi controls status effects through anonymous iteration; and analogy supports new-product forecasting when comparable reference cases are carefully assessed.

Judgment must be governed because anchoring, optimism, groupthink, availability and strategic incentives can distort forecasts. Independent estimates, documented assumptions, premortems, evidence-based overrides and post-forecast evaluation improve accountability.

Causal and econometric models relate demand to price, promotions, distribution and macroeconomic conditions. Regression coefficients require careful interpretation, and causal claims can be undermined by multicollinearity, endogeneity, omitted variables, autocorrelation and structural breaks. Out-of-sample evaluation, scenario analysis and transparent assumptions are therefore central to professional practice. The strongest supply-chain process combines statistical baselines, informed judgment and explicit uncertainty.

Key Terms

Qualitative forecasting: Forecasting based primarily on expert, customer or market knowledge rather than historical numerical patterns.

Judgmental forecasting: Estimation of future demand using human assessment, often structured through a formal process.

Market research: Systematic collection and analysis of information about customers, markets and competitors.

Sales-force composite: Aggregation of forecasts prepared by salespeople for accounts or territories.

Executive opinion: Forecast produced or adjusted through senior-management assessment.

Delphi method: Anonymous, iterative expert forecasting with controlled feedback.

Analogy forecasting: Estimation based on a comparable product, market or historical adoption pattern.

Forecast override: A documented adjustment to a statistical or baseline forecast.

Assumption register: A record of assumptions, evidence, effects, owners and review triggers.

Anchoring: Excessive reliance on an initial or salient value.

Premortem: Structured exercise asking why a forecast might fail before implementation.

Causal forecasting: Forecasting demand from variables believed to influence it.

Regression: Statistical estimation of a relationship between a dependent variable and explanatory variables.

Elasticity: Percentage responsiveness of one variable to a percentage change in another.

Multicollinearity: Strong correlation among explanatory variables that makes individual effects difficult to estimate.

Endogeneity: A condition in which an explanatory variable is correlated with the model error.

Structural break: A change in the underlying relationship governing the data.

Scenario analysis: Evaluation of outcomes under alternative assumptions about future conditions.

Out-of-sample validation: Testing a forecasting method on observations not used to estimate it.

Forecast governance: Rules and processes that control forecast creation, adjustment, review and accountability.

Review Questions

Short Answer Questions

1. Why should a sales target be distinguished from a demand forecast?
2. State two strengths and two limitations of the Delphi method.
3. What information should be recorded for a material forecast override?
4. Define price elasticity in a log-log demand model.
5. Why does a high in-sample R² not guarantee a good forecast?

Long Answer Questions

1. Compare market research, sales-force composite, executive opinion, Delphi and analogy forecasting in terms of use, strengths and risks.
2. Explain how an organization can combine a statistical baseline with judgment while preserving accountability.
3. Discuss the principal sources of bias in executive and sales-force forecasting and propose procedural controls.
4. Explain the assumptions and limitations of a multiple regression demand model containing price, promotion and income.
5. Describe how scenario analysis and an assumption register improve supply-chain planning under uncertainty.

Analytical / Application Questions

1. A sales team submits forecasts that are consistently higher than actual demand, while the statistical baseline is accurate except during promotions. Design a revised forecasting and governance process.
2. Using Q̂ = 1,200 − 40P + 180D, calculate demand at P = 12 with promotion and without promotion. Interpret the difference and state one limitation of the interpretation.
3. A new product has no sales history, but a similar product launched in another country three years ago. Develop an analogy-based forecast approach, identifying the adjustments, assumptions, validation checks and risks.

Multiple Choice Questions with Answers

Question 1

Which method obtains forecasts through anonymous, iterative rounds of expert judgment?

A. Sales-force composite

B. Delphi method

C. Executive opinion

D. Market basket analysis

Correct Answer: B. Delphi method

Question 2

In a sales pipeline, an opportunity has potential volume 500 units and probability 0.4. Its expected volume is:

A. 125 units

B. 200 units

C. 500 units

D. 900 units

Correct Answer: B. 200 units

Question 3

Which practice most directly reduces anchoring before a consensus meeting?

A. Allowing the most senior executive to estimate first

B. Using last year’s forecast as the only starting point

C. Collecting independent estimates before discussion

D. Removing all assumptions from the forecast

Correct Answer: C. Collecting independent estimates before discussion

Question 4

In the model Q̂ = 1,200 − 40P + 180D, what is forecast demand when P = 10 and D = 1?

A. 800 cases

B. 980 cases

C. 1,020 cases

D. 1,380 cases

Correct Answer: B. 980 cases

Question 5

In a log-log demand model, the coefficient on log price is commonly interpreted as:

A. The unit change in demand for one currency of price

B. The price elasticity of demand

C. The forecast intercept

D. The residual standard deviation

Correct Answer: B. The price elasticity of demand

Question 6

Which issue occurs when explanatory variables are strongly correlated with one another?

A. Multicollinearity

B. Seasonal adjustment

C. Underfitting by definition

D. Sampling neutrality

Correct Answer: A. Multicollinearity

Question 7

What is the principal purpose of out-of-sample validation?

A. To increase the number of historical observations

B. To test performance on data not used for model estimation

C. To guarantee causal interpretation

D. To eliminate all forecast uncertainty

Correct Answer: B. To test performance on data not used for model estimation

Question 8

Which statement about an executive forecast is most appropriate?

A. It is automatically more accurate than a statistical forecast

B. It should be treated as a sales target

C. Its assumptions and evidence should be documented

D. It should never be revised

Correct Answer: C. Its assumptions and evidence should be documented

Question 9

Why can a historical promotion coefficient be misleading?

A. Promotions may be targeted in response to expected demand conditions

B. Promotion variables can never be measured

C. Price and promotion are always identical

D. Regression cannot include binary variables

Correct Answer: A. Promotions may be targeted in response to expected demand conditions

Question 10

Which item is a suitable entry in a forecast-assumption register?

A. An undocumented personal preference

B. An assumption, supporting evidence, owner and review trigger

C. Only the final forecast number

D. A target with no operational implication

Correct Answer: B. An assumption, supporting evidence, owner and review trigger