4. Demand Estimation, Causal Inference and Forecasting

Introduction and context

Managerial economics becomes operational when decision makers must estimate how customers respond to price, advertising, product features, service quality, competitors, and broader economic conditions. A firm may need to decide whether to reduce price, expand capacity, discontinue a product, target a promotion, or enter a market. Each decision requires evidence about demand, but evidence can answer different questions. A descriptive analysis may show that sales are higher in regions receiving more advertising. A forecasting model may predict next quarter’s sales. A causal analysis asks a more demanding question: what would sales have been if advertising had not increased?

These distinctions are central to managerial judgment. Correlation describes systematic association; prediction concerns accuracy about future or unobserved outcomes; causal inference concerns the effect of an intervention relative to a credible counterfactual. Confusing these purposes can produce costly decisions. Regression may estimate an economically meaningful relationship, but the coefficient is not automatically causal. A highly accurate forecasting model may provide little insight into which policy will increase demand. Conversely, a carefully identified causal estimate may be valuable for policy but insufficient for accurate prediction in a changing market.

This chapter develops an integrated framework for demand estimation, causal inference, and forecasting. It considers regression, experiments, quasi-experiments, instrumental variables, panel data, time-series methods, and machine learning. Particular attention is given to identification, uncertainty, data quality, validation, managerial interpretation, and ethical governance.

Learning objectives

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

1. Distinguish correlation, prediction, and causation in managerial analysis.

2. Formulate demand questions using economic theory and counterfactual reasoning.

3. Specify and interpret regression models while recognizing threats to identification.

4. Evaluate randomized experiments, quasi-experiments, instrumental variables, and panel-data designs.

5. Select and validate appropriate time-series forecasting methods.

6. Explain how machine learning complements, but does not replace, economic reasoning.

7. Assess uncertainty, data quality, model risk, ethical concerns, and organizational implications.

8. Translate empirical findings into disciplined managerial decisions.

From an economic question to an estimand

A sound analysis begins with the decision rather than with the dataset. Suppose an online retailer is considering a 10 percent price reduction. Several questions are possible: How much will expected demand change? What will happen to contribution margin? Which customers will respond? Will competitors react? Will the effect persist? These are distinct estimands—precisely defined quantities that the analysis seeks to learn.

A simple demand function can be written as

Q_it = f(P_it, Y_it, A_it, X_it, u_it),

where Q is quantity demanded, P is price, Y represents income or purchasing power, A is advertising or promotion, X includes product and market conditions, and u captures unobserved influences. A commonly estimated log-linear model is

ln Q_it = β₀ + β₁ ln P_it + β₂ ln Y_it + β₃ ln A_it + γ'X_it + ε_it.

In this formulation, β₁ is a price elasticity if the specification is appropriate. Yet observed price is often chosen in response to expected demand. Managers may lower prices when demand is weak and raise prices when demand is strong. Consequently, price can be correlated with the error term, producing endogeneity and biased ordinary least squares estimates.

The counterfactual framework clarifies the problem. For unit i, let Y_i(1) denote the outcome if a treatment occurs—for example, exposure to a promotion—and Y_i(0) the outcome without it. The individual treatment effect is Y_i(1) − Y_i(0), but both potential outcomes cannot normally be observed for the same unit at the same time. The average treatment effect is E[Y(1) − Y(0)]. Causal research attempts to estimate this missing counterfactual using design assumptions, comparison groups, or institutional variation.

Correlation, prediction, and causation

Correlation is a property of observed variables. It can be useful for diagnosis and description, but it does not establish that changing one variable will change another. A positive relationship between advertising and sales may reflect effective advertising, but it may also reflect the fact that managers advertise more when they anticipate strong demand.

Prediction evaluates how well a model estimates outcomes not used in fitting it. A model can predict sales effectively by using variables that are causally irrelevant but stable proxies for demand. Prediction is therefore judged primarily by out-of-sample performance, whereas causal inference depends on assumptions about treatment assignment, confounding, measurement, and the counterfactual.

Causation concerns an intervention or change. A causal claim requires a defensible explanation of why the comparison identifies the effect of interest. Random assignment is powerful because, in expectation, it makes treatment independent of both observed and unobserved characteristics. In nonexperimental settings, researchers must justify alternatives such as conditional independence, parallel trends, exclusion restrictions, or continuity at a threshold.

This distinction matters strategically. If a chief marketing officer wants to allocate next month’s inventory, prediction may be the immediate priority. If the officer wants to know whether a campaign should be expanded, causal evidence is more important. In many decisions both are needed: causal analysis estimates the effect of an intervention, while forecasting predicts the market environment in which the intervention will operate.

Regression as a language for demand analysis

Regression provides a flexible framework for describing conditional relationships, testing hypotheses, and estimating effects under identifying assumptions. A multiple regression controls for observed factors that may otherwise obscure a relationship. For example, a store-level sales model might include price, promotions, local income, competitor price, seasonality, store fixed effects, and week fixed effects.

The coefficient on price represents the partial association between price and sales, holding included variables constant. It should not be called a causal elasticity unless the remaining variation in price is plausibly exogenous, or a valid research design addresses endogeneity. Omitted-variable bias occurs when an excluded determinant of sales is correlated with an included regressor. Reverse causality occurs when the outcome influences the explanatory variable. Simultaneity occurs when price and quantity are jointly determined by supply and demand.

Other concerns include multicollinearity, which inflates standard errors when predictors move together; heteroskedasticity, which makes conventional standard errors unreliable; serial correlation in time-ordered observations; functional-form misspecification; and measurement error. Robust or clustered standard errors can improve inference, but they do not repair a biased coefficient caused by endogeneity. Statistical significance also does not establish economic significance. A small estimated effect may be precisely measured but commercially irrelevant, while a large effect may be uncertain because the sample is limited.

A useful regression workflow includes: defining the decision and estimand; constructing variables using economic logic; inspecting distributions and missingness; specifying a baseline model; testing alternative specifications; examining residuals and influential observations; reporting confidence intervals; and assessing sensitivity to plausible confounders. Researchers should pre-specify major hypotheses where feasible and distinguish exploratory analysis from confirmatory testing.

Experimental evidence: the cleanest route to causality

A randomized controlled trial assigns units—customers, stores, regions, or time periods—to treatment and control conditions. For example, an online retailer may randomly display a discount to some eligible customers while withholding it from others. The intention-to-treat effect compares average outcomes according to assignment, regardless of whether every treated customer uses the discount.

Experiments require attention to implementation. Randomization must be preserved; treatment and control groups must not contaminate one another; outcomes must be measured consistently; and the observation period must capture delayed or repeated effects. Interference occurs when one unit’s treatment affects another unit, such as when treated customers share a promotion. Noncompliance arises when assigned treatment is not actually received. Attrition can create bias if dropout differs systematically between groups.

A/B tests are valuable but not universally sufficient. Short experiments may miss long-run brand effects, customer learning, inventory constraints, or competitor responses. Repeated experimentation can also create ethical and operational concerns, especially when prices or access differ across customers. Cluster randomization may be preferable when individual-level treatment would create spillovers, although it generally requires more observations to achieve precision.

Quasi-experimental designs and institutional variation

When randomization is impossible, quasi-experimental methods seek variation that approximates random assignment. Their credibility depends on explicit assumptions and diagnostic evidence.

A difference-in-differences design compares the change over time in a treated group with the change in a comparison group. If a regulation affects stores in one jurisdiction but not another, the estimator is credible when treated and control groups would have followed parallel trends absent the intervention. Researchers should inspect pre-treatment trends, consider event-study estimates, and assess whether other changes occurred simultaneously.

A regression-discontinuity design exploits a treatment rule based on a cutoff, such as eligibility for a subsidy when a score exceeds a threshold. Units just above and below the cutoff may be comparable. Validity requires that units cannot precisely manipulate the assignment variable and that no other determinants change discontinuously at the threshold. The estimate is generally local to observations near the cutoff.

A controlled interrupted time series examines whether an outcome changes in level or trend after an intervention, ideally relative to an unaffected series. Synthetic control methods construct a weighted combination of comparison units to approximate the treated unit’s pre-intervention path. Matching can improve comparability on observed characteristics, but matching alone does not eliminate unobserved confounding.

These methods are not interchangeable recipes. Each identifies a particular effect under particular assumptions. A doctoral-level analysis should state the identifying assumption, explain why it is plausible, test implications where possible, and acknowledge the population to which the estimate generalizes.

Instrumental variables and panel data

Instrumental variables are useful when a key explanatory variable is endogenous. An instrument Z must satisfy two central conditions. First, relevance: Z must affect the endogenous variable, such as price. Second, exclusion: Z must affect demand only through that endogenous variable, not through another pathway. The resulting two-stage least squares estimator is often interpreted as a local average treatment effect for units whose treatment responds to the instrument, provided additional assumptions hold.

Potential instruments include a cost shock affecting prices but not demand directly, or administrative allocation rules unrelated to customers’ latent preferences. Instruments must be defended substantively, not merely statistically. A strong first-stage relationship does not prove the exclusion restriction. Weak instruments can produce unstable estimates and misleading inference.

Panel data observe multiple units over time. Store, customer, or firm fixed effects remove time-invariant unobserved characteristics. Time fixed effects capture shocks common to all units, such as macroeconomic conditions. A fixed-effects model may be written as

Y_it = α_i + λ_t + βX_it + ε_it,

where α_i represents unit effects and λ_t represents period effects. Fixed effects do not control for unobserved factors that vary over time and correlate with X. They also do not automatically establish causality. Dynamic effects, anticipation, treatment timing, and serial correlation must be considered. Clustered inference is often appropriate when errors are correlated within units over time.

Forecasting demand through time

Forecasting asks what is likely to happen, conditional on information available at the forecast date. A useful forecasting system begins with a clear horizon, frequency, decision context, and loss function. A model for daily replenishment may prioritize short-horizon accuracy, whereas a capacity-planning model may need interpretable long-run scenarios.

Time-series methods include naïve forecasts, moving averages, exponential smoothing, trend models, seasonal decomposition, and autoregressive integrated moving average models. A seasonal naïve forecast, for instance, predicts this month’s demand using demand in the same month of the previous seasonal cycle. Such simple benchmarks are essential: a complex method that cannot outperform a reasonable benchmark may not justify its cost or opacity.

Forecasting data must be split chronologically. Randomly mixing future observations into the training set creates look-ahead bias. Rolling-origin evaluation repeatedly trains on an expanding or moving historical window and evaluates the next period. Relevant metrics include mean absolute error, root mean squared error, mean absolute percentage error, and weighted business losses. MAPE can be problematic when actual demand is zero or near zero; metric choice should reflect the decision cost of over- and under-forecasting.

Forecast intervals are as important as point forecasts. They communicate uncertainty arising from residual variation, parameter estimation, future shocks, and possible structural change. Scenario analysis can supplement statistical intervals when managers face discontinuities such as a new competitor, policy change, or supply disruption.

Machine learning and the economics of prediction

Machine-learning methods—regularized regression, random forests, gradient boosting, neural networks, and related algorithms—can model nonlinearities and interactions across large feature sets. They are particularly useful for high-dimensional prediction, customer-level personalization, demand sensing, and anomaly detection.

The central discipline remains validation. Training error is not evidence of generalization. Models should be evaluated on held-out or rolling future data, with leakage prevented in feature construction. Hyperparameters should be selected within the training process, and performance should be compared with economically meaningful benchmarks. Model drift occurs when customer behavior, product mix, measurement systems, or market conditions change.

Machine learning does not automatically identify causal effects. A model may accurately predict which customers will purchase without revealing whether a discount causes incremental purchases. Uplift modeling and causal forests attempt to estimate heterogeneous treatment effects, but their conclusions still depend on treatment assignment and identification assumptions. Explainability, fairness assessment, monitoring, and human review are especially important when models influence prices, credit, employment, or access to services.

An integrated analytical workflow

A practical workflow links economic reasoning with statistical discipline. First, articulate the managerial decision and whether it requires description, prediction, causal inference, or all three. Second, define the unit of analysis, time horizon, outcome, treatment, and target population. Third, audit the data: provenance, definitions, missing values, duplicates, timing, measurement changes, and access restrictions. Fourth, use descriptive analysis to understand distributions, trends, seasonality, outliers, and potential confounding.

Fifth, select a design appropriate to the question. Use randomization where feasible; otherwise consider panel, difference-in-differences, discontinuity, instrumental-variable, or controlled time-series approaches. Sixth, estimate a transparent baseline before adding complexity. Seventh, validate using methods consistent with deployment: out-of-sample temporal validation for forecasting and robustness or falsification tests for causal designs. Eighth, quantify uncertainty and conduct sensitivity analysis. Finally, translate estimates into financial consequences, implementation requirements, and monitoring indicators.

Illustrative case: pricing a regional subscription service

Background

The fictional company Northstar Learning sells monthly digital subscriptions to households in 40 metropolitan markets. Managers observed that markets with higher introductory discounts often recorded more sign-ups. The commercial team proposed a permanent price reduction, claiming that the data demonstrated strong price sensitivity.

Situation and problem

The observed discount was not randomly assigned. Regional managers offered deeper discounts in markets where they expected weak demand and where competitors were active. A regression of sign-ups on discount depth therefore produced a negative and unstable relationship. The board needed evidence about the incremental effect of a discount, its effect on retention, and the profitability of expansion.

Analysis and approach

The analytics team first defined the estimand as the change in 90-day contribution margin caused by offering a 15 percent discount to eligible new customers. It audited customer records, discovered inconsistent definitions of activation, and separated acquisition from retention outcomes.

Northstar then ran a market-clustered randomized experiment in 20 comparable markets for eight weeks. Markets were assigned either the existing offer or the 15 percent discount, while the remaining markets were monitored for external shocks. The team pre-specified primary outcomes, measured assignment compliance, and checked baseline balance. Because market-level assignment limited statistical power, the analysis used cluster-robust inference and reported confidence intervals rather than relying only on a significance threshold.

The experiment showed that the discount increased initial sign-ups but reduced average revenue per subscriber. The 90-day margin effect was positive in some customer segments and negative in others. A separate retention analysis indicated that discount-sensitive customers were more likely to cancel when the introductory period ended. The company used the experimental estimate to target the offer rather than adopting a universal price cut.

Outcome and learning points

Northstar introduced the discount only in segments with positive estimated incremental margin, maintained a control holdout for monitoring, and reviewed results quarterly. The case illustrates several principles. First, an observed price–sales relationship was not a causal elasticity. Second, the relevant outcome was contribution margin over an appropriate horizon, not sign-ups alone. Third, experimentation can reveal heterogeneous effects. Fourth, data definitions and implementation quality are as important as model selection. Finally, a credible estimate still requires managerial judgment about scalability, competitor response, and long-run brand effects.

Challenges, risks, and professional responsibilities

Demand data are often affected by selection bias, missing transactions, changing product definitions, stockouts, returns, cancellations, and strategic managerial behavior. A sales decline may reflect lower demand, unavailable inventory, poorer placement, or a tracking failure. Analysts should distinguish zero sales from missing sales and avoid treating operational constraints as consumer preferences.

Common mistakes include controlling for post-treatment variables, using future information in a forecast, interpreting fixed effects as a complete solution to causality, searching across specifications until a desired result appears, ignoring multiple testing, and presenting a point estimate without its uncertainty. Overfitting and excessive model complexity can make a system fragile and difficult to govern.

Ethical practice requires transparency about data sources, consent and privacy, fairness across customer groups, and the consequences of automated decisions. Differential prices may be economically efficient in some settings but discriminatory or exploitative in others. Personal data should be minimized, protected, and used for legitimate purposes. Governance should assign responsibility for model approval, monitoring, documentation, drift detection, and human escalation. Regulatory requirements vary by jurisdiction, but sound professional practice requires that analytical decisions be explainable, auditable, and contestable where stakeholders may be materially affected.

Emerging directions

Several developments are reshaping demand analytics. Privacy-preserving measurement is becoming more important as individual tracking becomes restricted. Causal machine learning is expanding analysis of heterogeneous treatment effects, although its validity remains dependent on research design. Interference-aware experiments are increasingly relevant for platforms where users affect one another. Probabilistic forecasting and decision-focused loss functions are improving inventory and capacity decisions. Digital experimentation can also be integrated with structural economic models that represent substitution, strategic behavior, and market equilibrium.

These developments do not eliminate the need for theory. The more automated the analytical system, the more important it becomes to specify the decision, understand the data-generating process, test assumptions, and monitor consequences. Advanced methods should be adopted when they improve decisions, not merely because they are technically sophisticated.

Best practices for doctoral managers and analysts

• Begin with a clearly defined estimand and decision.
• Separate descriptive, predictive, and causal objectives.
• Use economic theory to guide variables, mechanisms, and plausible confounders.
• Prefer randomization when feasible and document assumptions when it is not.
• Establish simple benchmarks before adopting complex models.
• Use temporal or design-consistent validation and prevent leakage.
• Report effect sizes, uncertainty, practical significance, and external validity.
• Test robustness through alternative specifications, placebo tests, pre-trends, and sensitivity analysis.
• Treat data provenance and measurement quality as analytical issues, not administrative details.
• Monitor deployed models for drift, fairness, performance, and unintended effects.

Chapter Summary

Demand analysis requires more than fitting a regression to observed sales. Correlation, prediction, and causation answer different questions and require different standards of evidence. Regression is useful for conditional relationships, but endogeneity, omitted variables, measurement error, and misspecification can undermine causal interpretation. Experiments provide strong identification when implemented correctly, while quasi-experimental designs use institutional variation under explicit assumptions. Instrumental variables address endogeneity only when relevance and exclusion are credible; panel data control for time-invariant unit characteristics but do not solve all confounding.

Forecasting depends on temporal structure, appropriate benchmarks, rolling or chronological validation, and explicit uncertainty. Machine learning can improve prediction and reveal nonlinear patterns, but it does not itself establish causality. Effective managerial analytics combines economic theory, credible research design, data-quality controls, validation, ethical governance, and translation into financial and operational decisions.

Key Terms

Estimand: The precisely defined quantity a study seeks to estimate.

Counterfactual: The outcome that would have occurred under an alternative treatment or condition.

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

Omitted-variable bias: Bias caused by excluding a determinant related to an included explanatory variable.

Price elasticity: The percentage change in quantity associated with a one-percent change in price.

Ordinary least squares: A regression method that minimizes the sum of squared residuals.

Randomized controlled trial: An experiment assigning units to treatment conditions by chance.

Intention-to-treat effect: The effect of assignment to treatment, regardless of treatment compliance.

Difference-in-differences: A design comparing changes over time between treated and comparison groups.

Parallel trends: The assumption that treated and comparison groups would have evolved similarly without treatment.

Regression discontinuity: A design exploiting treatment assignment at a known cutoff.

Instrumental variable: A variable that affects treatment but influences the outcome only through treatment under the exclusion restriction.

Fixed effects: Unit- or time-specific effects included to account for systematic unobserved factors.

Forecast: An estimate of a future outcome using information available at a specified time.

Look-ahead bias: Forecasting error caused by using information unavailable at the forecast date.

Rolling-origin evaluation: Repeatedly training on past data and testing on the next time period.

Overfitting: Fitting noise or idiosyncratic patterns that fail to generalize.

Model drift: Deterioration caused by changes in data, behavior, or the environment.

External validity: The extent to which findings generalize beyond the study context.

Review Questions

Short Answer Questions

1. How do correlation, prediction, and causation differ?

2. What is an estimand, and why should it be defined before model selection?

3. Explain omitted-variable bias in a demand regression.

4. What is the purpose of rolling-origin evaluation?

5. State the relevance and exclusion requirements for an instrumental variable.

Long Answer Questions

1. Explain how economic theory and counterfactual reasoning should guide demand estimation.

2. Compare randomized experiments with difference-in-differences and regression-discontinuity designs.

3. Discuss the principal threats to causal interpretation in a multiple regression model.

4. Explain how forecasting models should be selected, validated, and communicated to managers.

5. Evaluate the contribution and limitations of machine learning in managerial demand analysis.

Analytical / Application Questions

1. A retailer observes higher sales in stores receiving more advertising. Develop two alternative explanations for this relationship and propose a design that could estimate the causal advertising effect.

2. A subscription firm’s price reduction increases sign-ups but decreases 90-day contribution margin. How should management evaluate the policy, and what additional evidence is required before scaling it?

3. Design an empirical strategy to estimate the effect of a regulation introduced in some regions but not others. State the identifying assumption, relevant diagnostics, and threats to validity.

22. Multiple Choice Questions with Answers

Question 1

A retailer observes that stores with greater discounts have weaker sales. Which explanation most directly illustrates endogeneity?

A. Discounts are measured in percentages rather than currency

B. Managers offer larger discounts when they anticipate weak demand

C. Sales data contain seasonal variation

D. Customers prefer products with higher prices

Correct Answer: B. Managers offer larger discounts when they anticipate weak demand

Question 2

What is the defining purpose of a causal estimand?

A. To maximize in-sample fit

B. To describe the correlation between two variables

C. To define the effect relative to a counterfactual outcome

D. To eliminate the need for data validation

Correct Answer: C. To define the effect relative to a counterfactual outcome

Question 3

In a difference-in-differences design, the central identifying assumption is usually that treated and comparison groups would have had:

A. Identical outcome levels after treatment

B. Parallel trends absent the intervention

C. The same sample size in every period

D. No observable differences before treatment

Correct Answer: B. Parallel trends absent the intervention

Question 4

Which condition is part of the exclusion restriction for an instrumental variable?

A. The instrument must be correlated with the outcome through every available pathway

B. The instrument must affect the endogenous treatment

C. The instrument must affect the outcome only through the endogenous treatment

D. The instrument must be randomly generated by the analyst

Correct Answer: C. The instrument must affect the outcome only through the endogenous treatment

Question 5

Why is randomly splitting time-series observations into training and test sets problematic?

A. It always increases model bias

B. It can allow future information to enter model development

C. It prevents calculation of forecast intervals

D. It makes seasonal patterns impossible to estimate

Correct Answer: B. It can allow future information to enter model development

Question 6

A model predicts customer purchases accurately but cannot show whether a discount causes additional purchases. This illustrates the distinction between:

A. Prediction and causation

B. Correlation and measurement

C. Forecasting and data cleaning

D. Elasticity and seasonality

Correct Answer: A. Prediction and causation

Question 7

Which result would most strongly support the validity of a regression-discontinuity design?

A. A large treatment effect far from the cutoff

B. Evidence that units precisely manipulate the assignment score

C. Continuity of other determinants around the cutoff

D. A high in-sample R-squared

Correct Answer: C. Continuity of other determinants around the cutoff

Question 8

Why might fixed effects fail to establish a causal effect?

A. They cannot remove time-invariant unit characteristics

B. They do not control for unobserved factors that change over time

C. They prevent the use of panel data

D. They always create omitted-variable bias

Correct Answer: B. They do not control for unobserved factors that change over time

Question 9

Which practice best addresses overfitting in a demand prediction system?

A. Increasing model complexity until training error is zero

B. Evaluating performance on properly held-out future data

C. Removing all economic variables

D. Reporting only statistically significant predictors

Correct Answer: B. Evaluating performance on properly held-out future data

Question 10

A discount raises sign-ups but lowers long-run contribution margin. Which managerial conclusion is most defensible?

A. The discount must be rejected because sign-ups are irrelevant

B. The discount should be adopted universally because acquisition increased

C. The policy should be evaluated using the relevant horizon, margin, and heterogeneous effects

D. The regression coefficient alone determines the optimal policy

Correct Answer: C. The policy should be evaluated using the relevant horizon, margin, and heterogeneous effects