Dem& Forecasting Analytics

Chapter 6
Causal, Econometric and Judgmental Forecasting

Post Graduate Program in Logistics
Dem& Forecasting Analytics   •   Post Graduate

Causal, Econometric and Judgmental Forecasting

Introduction

Forecasting in logistics is not merely the extrapolation of historical demand. Demand changes because prices change, promotions are introduced, competitors act, income and industrial activity vary, weather disrupts operations, and customers anticipate future conditions. A forecast that ignores these drivers may describe the past accurately while predicting the future poorly.

Causal, econometric and judgmental methods address this problem by asking why demand changes, not only how it has changed. Causal forecasting estimates relationships between demand and explanatory variables. Econometric forecasting places these relationships within a broader economic and statistical system. Judgmental forecasting incorporates information that may not yet appear in historical data, such as a competitor’s planned launch, a regulatory change or a customer’s confidential expansion plan.

The central managerial challenge is integration. Human insight can improve a statistical forecast, but it can also introduce anchoring, optimism, political pressure and overreaction. Effective forecasting therefore combines quantitative evidence with structured judgment, documented assumptions and formal bias controls.

Learning Objectives

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

  1. Distinguish causal, econometric and judgmental forecasting.
  2. Specify and interpret regression models for logistics demand.
  3. Evaluate leading indicators, price effects, promotions and elasticity.
  4. Construct scenarios and assess forecast sensitivity.
  5. Design market research, Delphi and expert-judgment processes.
  6. Identify bias, endogeneity, multicollinearity and structural change.
  7. Combine statistical forecasts and human adjustments responsibly.
  8. Translate forecast evidence into inventory, capacity and transport decisions.

1. From Time-Series Extrapolation to Causal Explanation

A time-series model uses information contained in the demand series itself. A causal model adds variables believed to influence demand. For example, parcel volume may depend on online sales, delivery price, promotional intensity, fuel prices and business activity.

The distinction is not that causal models prove causation automatically. A regression coefficient is causal only under credible assumptions about omitted variables, reverse causality, measurement and model specification. Regression is therefore a tool for estimating conditional relationships; causal interpretation requires research design and domain knowledge.

A general demand model can be written as:

Dt = f(Pt, At, It, St, Wt, Ct, εt)

Here, Dt is demand in period t; Pt is price; At is advertising or promotion; It is income or industrial activity; St represents service availability; Wt represents weather or other external conditions; Ct represents competitor conditions; and εt is the unexplained component. The function f may be linear, logarithmic, nonlinear or dynamic.

A useful causal model must specify the decision context. A retailer forecasting baseline demand may use observed price and promotion. A pricing manager estimating the effect of a future price change must consider that prices are often set in response to expected demand. This simultaneity can make simple regression misleading.

2. Regression for Demand Forecasting

2.1 Multiple linear regression

The standard multiple regression model is:

Dt = β0 + β1X1t + β2X2t + ⋯ + βkXkt + εt

Dt is the dependent variable, Xjt are explanatory variables, β0 is the intercept, βj is the partial effect of variable j holding other included variables constant, and εt is the error term. The coefficients are usually estimated by ordinary least squares (OLS), which minimizes the sum of squared residuals:

min Σt=1n(Dt − D̂t)2

OLS interpretation requires assumptions including a correctly specified conditional mean, no perfect multicollinearity, exogeneity of regressors for unbiased causal interpretation, and an appropriate treatment of error dependence. Forecasting may remain useful when some causal assumptions fail, but interpretation and extrapolation then require caution.

For logistics, explanatory variables may include:

  • calendar effects: month, weekday, holidays and trading days;
  • prices, discounts and freight charges;
  • promotion type, display, advertising and campaign reach;
  • customer orders, production or construction activity;
  • competitor availability and market share;
  • weather, strikes, congestion and service reliability;
  • lagged demand and inventory availability.

2.2 Functional forms

The functional form should reflect the decision problem and data-generating process.

A level-level model assumes a unit change in X changes demand by a constant number of units. A log-level model uses ln(Dt) as the dependent variable; approximately, a one-unit change in X changes demand by 100β percent when β is small. A log-log model is especially useful for elasticity:

ln(Dt) = β0 + β1 ln(Pt) + β2 ln(At) + εt

In this model, β1 is the price elasticity of demand, subject to the model’s assumptions. If β₁ = −1.4, a 1% increase in price is associated with an approximate 1.4% decrease in demand, holding included variables constant.

Dummy variables represent categories. For example, a promotion indicator equals 1 during a promotion and 0 otherwise. With several categories, one category must be omitted as the reference group to avoid perfect multicollinearity.

Lagged variables capture delayed responses:

Dt = β0 + β1Pt + β2Pt−1 + β3At + εt

The coefficient on current price measures the immediate association, while the coefficient on lagged price captures delayed response. Dynamic models should also consider inventory constraints and customer order timing, since observed sales may be less than unconstrained demand.

2.3 Diagnostics and validation

A model should be evaluated both statistically and operationally. Important diagnostics include residual plots, outlier analysis, autocorrelation, changing variance, multicollinearity and stability over time. A high R² does not guarantee useful forecasts; a model can explain historical variation while failing after a market change.

Forecast validation should use time-ordered data. Randomly splitting observations can leak future information into the training set. A rolling-origin procedure estimates the model using an initial window, forecasts the next period, expands the window and repeats the process.

Common accuracy measures include:

MAE = (1/n)Σ|Dt − D̂t|
RMSE = √[(1/n)Σ(Dt − D̂t)2]

MAE is in demand units and is easy to interpret. RMSE penalizes large errors more heavily. Percentage measures such as MAPE can be unstable or undefined when actual demand is zero or near zero. Forecast evaluation should therefore reflect the logistics cost of errors, including stockouts, excess inventory, expediting and lost service.

3. Econometric Forecasting and Identification

Econometric forecasting combines economic theory, statistical estimation and institutional knowledge. It is particularly useful when demand is influenced by macroeconomic conditions, prices, income, interest rates, exchange rates, industrial output or policy.

3.1 Endogeneity

A major threat is endogeneity: an explanatory variable is correlated with the error term. Price is often endogenous because firms reduce prices when they expect weak demand. In a simple model:

Dt = β0 + β1Pt + εt

if managers choose Pt partly in response to unobserved εt, OLS may produce a biased estimate of β₁. The coefficient may not represent the true demand response.

Possible remedies include randomized price experiments, natural experiments, instrumental variables, fixed effects, difference-in-differences and carefully designed panel data. An instrumental variable must be correlated with price but affect demand only through price, conditional on the model. This exclusion condition is difficult and must be defended substantively, not merely statistically.

3.2 Panel and structural models

Panel data combine observations across customers, products, locations or time. A fixed-effects model controls for unobserved characteristics that are stable within a unit:

Dit = αi + βXit + γt + εit

Here, i identifies a customer, product or location; αi captures time-invariant unit effects; γt captures common time effects; and β measures the relationship between changing X and demand after these controls.

Structural models explicitly represent decisions such as consumer choice, inventory availability and substitution among products. They require stronger assumptions but can support policy simulation, such as estimating the effect of a new delivery fee or a competitor’s entry.

3.3 Leading indicators

A leading indicator changes before the target variable and may improve forecast timeliness. Examples include new building permits for construction-material logistics, purchasing-manager indices for industrial freight, online search activity for consumer demand, and order-book data for contract logistics.

A leading indicator should not be selected solely because it correlates historically with demand. Analysts should test timing, stability, revisions, data availability and the possibility that the relationship is spurious. A cross-correlation or lag model may be useful, but economic logic remains essential.

A practical indicator assessment asks:

  1. Does the indicator precede the target at the required decision horizon?
  2. Is the relationship stable across cycles and regions?
  3. Are revisions known and available in real time?
  4. Can management act on the signal?
  5. Does it add predictive value beyond existing variables?

4. Price, Promotion and Elasticity

4.1 Price elasticity

Point price elasticity is defined as:

Ep = (∂D/∂P)(P/D)

Ep is elasticity, D is demand and P is price. It is dimensionless. Demand is typically price-inelastic when |E_p| < 1 and elastic when |E_p| > 1, although interpretation depends on the market and relevant range.

For discrete changes, arc elasticity reduces dependence on the choice of base period:

Earc = [(D2 − D1)/((D1 + D2)/2)] ÷ [(P2 − P1)/((P1 + P2)/2)]

This approximation is appropriate for finite changes when the two observations are comparable and other important drivers are controlled.

Elasticity estimates are useful for revenue and capacity planning, but price effects may differ by customer segment, route, service level and competitive context. A price increase can reduce volume while increasing revenue, but the result depends on the size of the elasticity and cost structure.

4.2 Promotions and cannibalisation

Promotion variables should distinguish discount depth, duration, placement, advertising support and product availability. A promotion may increase total category demand, shift demand from a later period, or transfer customers from another product. The relevant quantity for logistics is often incremental demand, not observed demand during the campaign.

A model can include interactions:

Dt = β0 + β1Pt + β2Promot + β3(Pt × Promot) + εt

The interaction term allows the price response during promotion to differ from the ordinary price response. Interpretation must be based on combined marginal effects rather than reading β₁ or β₂ in isolation.

5. Scenario Analysis and Sensitivity

A single forecast conceals uncertainty. Scenario analysis creates internally consistent descriptions of possible futures, such as base, high-demand and disruption scenarios. Scenarios are not probabilities unless probabilities are explicitly assigned.

A scenario should specify assumptions, mechanisms, indicators and management actions. For example, a high-demand scenario may combine stronger industrial output, stable fuel prices and a competitor service failure. An adverse scenario may combine recession, a price increase and port disruption.

Sensitivity analysis changes one assumption while holding others constant. It identifies variables to which the decision is most exposed. Scenario analysis changes several related assumptions consistently. Both should distinguish:

  • forecast uncertainty: imperfect knowledge of future conditions;
  • parameter uncertainty: imperfect knowledge of model coefficients;
  • model uncertainty: uncertainty about the appropriate structure;
  • operational uncertainty: uncertainty in execution, capacity or lead time.

Scenario outputs should be translated into decisions: required fleet capacity, warehouse labor, safety stock, supplier commitments and contingency transport. A forecast is valuable when it improves decisions, not merely when it minimizes an abstract error metric.

6. Judgmental Forecasting

Judgment is essential when historical data are scarce, the environment is changing, or important information is qualitative. Typical applications include new products, network redesign, major customer wins, regulatory changes and disruptions.

6.1 Sources of judgment

Judgment may come from sales teams, procurement specialists, operations managers, customers, industry analysts and senior executives. Each source has a different information set and incentive. Sales teams may know customer intentions but may also have incentives to overstate pipeline. Operations managers understand capacity constraints but may anchor on recent workload.

Judgment should be elicited separately from advocacy. The question is not “What target do we want?” but “What outcome is supported by evidence, and what assumptions produce it?”

6.2 Delphi method

The Delphi method is a structured, usually anonymous, multi-round process. Experts independently provide estimates and rationales. A facilitator summarizes the group response, shares controlled feedback and requests revised estimates. The process continues until convergence or until the reasons for disagreement are sufficiently clear.

Its advantages include reduced domination by senior participants, explicit reasoning and the ability to incorporate dispersed expertise. Its limitations include panel-selection bias, conformity in later rounds, facilitator influence and false consensus. A strong Delphi design preserves minority views and reports dispersion, not only the median.

6.3 Expert judgment and structured elicitation

A practical elicitation protocol should:

  1. define the forecast variable, horizon and unit;
  2. provide a common information pack;
  3. request a point estimate and plausible range;
  4. ask for assumptions and evidence;
  5. elicit probabilities or confidence where appropriate;
  6. record judgments before group discussion;
  7. compare judgments with the statistical baseline;
  8. document the final adjustment and its owner;
  9. review accuracy and calibration after the outcome is known.

Ranges should be interpreted carefully. A 90% interval is meaningful only if experts understand the coverage question and are reasonably calibrated. Excessively narrow ranges often indicate overconfidence.

6.4 Market research

Market research can reveal intended purchase, substitution, service preferences and reactions to price. Methods include surveys, interviews, focus groups, conjoint analysis, test markets and experiments. Stated intention is not identical to realized demand, so research evidence should be triangulated with behavioral data.

Sampling, nonresponse, question wording, hypothetical bias and social desirability can distort results. Research should identify the target population, sampling frame, field period and weighting method. For logistics, research should also examine operational feasibility: a customer may prefer same-day delivery but be unwilling to pay for it.

7. Combining Quantitative and Human Forecasts

A disciplined combination process begins with an unconstrained statistical baseline. Forecasters then review exceptions, new information and model limitations. Adjustments should be explicit, quantified and traceable.

A simple weighted combination is:

Fc = wFq + (1 − w)Fj, 0 ≤ w ≤ 1

Fc is the combined forecast, Fq is the quantitative forecast, Fj is the judgmental forecast and w is the weight. Weights may be chosen using historical validation rather than status or hierarchy. If judgment adds value only in special situations, a conditional adjustment or override policy may be better than a permanent fixed weight.

Forecast reconciliation is important when product, customer, regional and network forecasts differ. A bottom-up forecast may reflect operational detail but double-count accounts. A top-down forecast may be coherent at the market level but miss local constraints. Reconciliation should identify the source of differences and preserve an auditable hierarchy.

Bias controls include:

  • blind or independent initial estimates;
  • statistical baselines shown before discussion;
  • documented override reasons;
  • limits on unexplained adjustments;
  • post-period review of overrides;
  • separation of forecasting from target setting;
  • calibration training and accuracy dashboards.

A forecast should not be changed merely to meet a budget or incentive target. Governance must distinguish an objective forecast from a desired plan.

8. Illustrative Case: NorthStar Parcel Logistics

Background

NorthStar delivers parcels for online retailers in three metropolitan regions. It must set weekly line-haul capacity and temporary labor six weeks ahead. Historical volume is available for 156 weeks. The company also has online retail sales, delivery price, promotion intensity, public holidays and a purchasing-manager index.

Problem

The existing forecast uses a seasonal moving average. It performs adequately in ordinary weeks but underestimates promotional peaks and overestimates volume during industrial slowdowns. Managers frequently apply undocumented increases to the forecast.

Analysis

The analytics team specifies a model with weekly demand as the dependent variable and seasonal indicators, retail sales, price, promotion intensity and the lagged purchasing-manager index as predictors. It uses rolling-origin validation and evaluates MAE separately for ordinary and promotional weeks. The team checks whether observed parcel volume is constrained by sorting capacity; constrained weeks are flagged because sales may understate true demand.

The price coefficient is interpreted cautiously because prices change when managers anticipate demand. Promotional effects are estimated using campaign indicators and post-campaign variables to test for demand pull-forward. Regional fixed effects are used in a panel model to control for stable differences among metropolitan areas.

Approach and solution

The organization adopts a three-stage process. First, the model produces a baseline and prediction interval. Second, regional managers independently submit evidence-based adjustments using a standard form. Third, a cross-functional panel reviews only material exceptions, separating operational constraints from demand assumptions. A Delphi round is used for a planned competitor entry because no historical analogue is reliable.

Scenarios are developed for normal, strong promotion and competitor disruption. Each scenario is linked to labor, vehicle and subcontractor actions. The final forecast combines the validated baseline with documented judgment rather than replacing it.

Outcome

In an illustrative six-month evaluation, the new process reduces MAE in promotional weeks and decreases unexplained overrides. More importantly, managers can explain why capacity was increased and which assumptions would trigger a revision. The result is not certainty; it is a more transparent decision system with faster response to new evidence.

Learning points

  1. Promotional demand requires variables beyond ordinary seasonality.
  2. Capacity constraints must be distinguished from low demand.
  3. A leading indicator is useful only when its timing and stability are tested.
  4. Judgment adds value when it contributes new information, not when it merely changes a number.
  5. Forecast governance is part of analytical quality.

9. Limitations, Ethics and Governance

Causal forecasting can fail under structural breaks, omitted variables, measurement error and policy changes. Historical relationships may not survive a new competitor, technology or regulation. Models trained on aggregate data may conceal differences among small customers or regions.

Ethically, forecasts can influence staffing, service access, supplier contracts and customer treatment. Analysts should avoid discriminatory proxies, protect commercially sensitive data and communicate uncertainty honestly. Automated systems should retain human accountability, versioned data, reproducible code and an audit trail of overrides.

Privacy rules may restrict the use of customer-level browsing, location or transaction data. Data minimization, access control, retention limits and purpose limitation are appropriate governance principles. Forecast users should know whether a number is observed, modeled, judgmentally adjusted or scenario-based.

Artificial intelligence can improve feature discovery, text extraction from market reports and probabilistic forecasting. However, machine-learning models do not eliminate endogeneity, bias or data leakage. Explainability, temporal validation, drift monitoring and human review are essential. Generative systems should not be treated as independent experts without evidence, calibration and accountability.

10. Best-Practice Implementation Framework

A mature causal and judgmental forecasting system follows a repeatable cycle:

  1. Define the decision: specify horizon, granularity, cost of errors and decision owner.
  2. Map causal drivers: distinguish controllable variables, external indicators and constraints.
  3. Build a benchmark: retain a simple seasonal or naïve model for comparison.
  4. Estimate and diagnose: test functional form, lags, residual behavior and stability.
  5. Validate temporally: use rolling-origin evaluation and relevant business loss measures.
  6. Elicit judgment structurally: collect independent estimates, assumptions and ranges.
  7. Create scenarios: connect assumptions to operational responses.
  8. Combine transparently: record weights, overrides, reasons and uncertainty.
  9. Monitor drift: track forecast error, bias, override performance and driver changes.
  10. Learn after outcomes: update models, panel membership and elicitation procedures.

The most important principle is proportionality. A high-value, volatile lane may justify a detailed econometric model and weekly expert review. A stable, low-value SKU may require only a simple benchmark and exception monitoring.

Chapter Summary

Causal forecasting explains demand through prices, promotions, economic activity, competition, weather, service conditions and other drivers. Regression provides a flexible framework, but coefficients have causal meaning only when assumptions concerning exogeneity, specification and measurement are credible. Econometric methods extend this analysis through leading indicators, panel data, fixed effects, experiments and structural models.

Price elasticity quantifies responsiveness, while promotion analysis must distinguish incremental demand, substitution and demand pull-forward. Scenario analysis exposes uncertainty and links alternative futures to capacity and inventory actions. Judgmental methods, including market research, Delphi and structured expert elicitation, are valuable when data are limited or the environment is changing, but they require controls against anchoring, overconfidence, groupthink and target-driven adjustments.

The strongest forecasting practice combines a validated quantitative baseline with documented, evidence-based human insight. It monitors forecast accuracy, bias, structural change and the performance of overrides. Ethical governance requires privacy protection, transparency, auditability and accountability for both automated and human decisions.

Key Terms

Causal forecasting: Forecasting demand using variables believed to influence the outcome.

Econometric forecasting: Forecasting that combines economic theory, statistical estimation and observed data.

Judgmental forecasting: Forecasting based on structured expert or managerial knowledge.

Endogeneity: Correlation between an explanatory variable and the model error term.

Exogeneity: A condition in which explanatory variables are appropriately unrelated to the unexplained error for the intended inference.

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

Leading indicator: A variable that changes before the target variable.

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

Dummy variable: A binary indicator used to represent a category or event.

Interaction term: A model term allowing the effect of one variable to depend on another.

Panel data: Observations across multiple units and time periods.

Fixed effects: Unit-specific or time-specific effects used to control for stable unobserved factors.

Delphi method: A structured, usually anonymous, multi-round expert-elicitation process.

Scenario analysis: Evaluation of outcomes under internally consistent alternative assumptions.

Sensitivity analysis: Assessment of how results change when an assumption changes.

Forecast override: A deliberate adjustment to a model-generated forecast.

Calibration: Agreement between stated confidence or probabilities and observed frequencies.

Cannibalisation: Loss of demand from one product or period caused by an offer affecting another.

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

Rolling-origin validation: Time-ordered evaluation using repeated expanding or moving estimation windows.

Review Questions

Short Answer Questions

  1. What distinguishes a causal forecast from a purely time-series forecast?
  2. Define price elasticity and explain why it is dimensionless.
  3. What is a leading indicator? Give one logistics example.
  4. State two advantages and two limitations of the Delphi method.
  5. Why should forecast overrides be documented?

Long Answer Questions

  1. Explain the assumptions and limitations of multiple linear regression for logistics demand forecasting.
  2. Discuss how endogeneity can distort estimated price effects and describe possible remedies.
  3. Compare market research, Delphi and structured expert judgment.
  4. Explain how scenario analysis supports capacity and inventory decisions.
  5. Design a governance process for combining a statistical baseline with managerial judgment.

Analytical / Application Questions

  1. A parcel operator observes that price reductions coincide with weak demand. Explain why a simple regression may produce a misleading price coefficient and propose a credible identification strategy.
  2. A promotion raises weekly shipments by 18%, but shipments fall 12% below normal in the following week. Analyze whether the promotion generated incremental demand or pulled demand forward, and identify additional data required.
  3. A model has lower RMSE but higher stockout costs than a judgmentally adjusted model. Recommend how the organization should evaluate and combine the forecasts, considering asymmetric logistics costs.

Multiple Choice Questions with Answers

Question 1

Which condition is most important for interpreting a regression coefficient as a causal effect?

A. The model has the highest possible R²

B. The explanatory variable is exogenous under a credible research design

C. The dependent variable has a positive mean

D. The sample contains exactly 100 observations

Correct Answer: B. The explanatory variable is exogenous under a credible research design

Question 2

In a log-log demand model, a price coefficient of −1.4 is interpreted approximately as:

A. A 1% price increase is associated with a 1.4% demand increase

B. A 1.4-unit price increase causes a 1% demand decrease

C. A 1% price increase is associated with a 1.4% demand decrease

D. Price has no relationship with demand

Correct Answer: C. A 1% price increase is associated with a 1.4% demand decrease

Question 3

Which measure is expressed in the same units as demand?

A. MAE

B. Price elasticity

C. MAPE

D. Correlation coefficient

Correct Answer: A. MAE

Question 4

What is a leading indicator?

A. A variable that changes after the target

B. A variable that precedes the target and may improve forecast timing

C. A variable with zero variance

D. A variable used only for accounting

Correct Answer: B. A variable that precedes the target and may improve forecast timing

Question 5

What is a principal purpose of the Delphi method?

A. To guarantee unanimous expert opinions

B. To replace all quantitative data

C. To obtain structured expert estimates while reducing dominance effects

D. To maximize the influence of the most senior participant

Correct Answer: C. To obtain structured expert estimates while reducing dominance effects

Question 6

If price elasticity is −0.6, demand is conventionally described as:

A. Price elastic

B. Unit elastic

C. Price inelastic

D. Perfectly elastic

Correct Answer: C. Price inelastic

Question 7

Which example best represents demand pull-forward?

A. A promotion increases both campaign-week and total-quarter demand equally

B. A promotion increases campaign-week demand but reduces demand shortly afterward

C. A promotion has no effect on shipment timing

D. A competitor exits and market demand disappears

Correct Answer: B. A promotion increases campaign-week demand but reduces demand shortly afterward

Question 8

Why is random train-test splitting potentially inappropriate for time-series forecasting?

A. It always reduces the sample size to zero

B. It can allow future information to influence model estimation

C. It prevents calculation of MAE

D. It makes all variables categorical

Correct Answer: B. It can allow future information to influence model estimation

Question 9

In the combination formula Fc = wFq + (1 − w)Fj, what does w represent?

A. The forecast horizon

B. The error variance only

C. The weight assigned to the quantitative forecast

D. The number of judgmental experts

Correct Answer: C. The weight assigned to the quantitative forecast

Question 10

Which practice best controls politically motivated forecast adjustments?

A. Allowing unlimited undocumented overrides

B. Separating forecasts from targets and recording override reasons

C. Using only the most optimistic estimate

D. Suppressing the statistical baseline

Correct Answer: B. Separating forecasts from targets and recording override reasons
Dem& Forecasting Analytics
Chapter 6 of 14