Causal Forecasting, Regression and Econometric Models

Introduction

Time-series methods describe how demand has behaved over time. Causal forecasting goes further: it explains demand through variables that managers can observe, influence, or anticipate. Price, advertising, promotions, distribution, income, interest rates, weather, competitor actions, holidays, and policy changes may all help explain sales. This makes causal forecasting especially valuable for pricing, budgeting, capacity planning, inventory control, and strategic decision-making.

A causal model is not automatically a causal explanation. Regression measures conditional associations in observed data; credible causal interpretation requires a defensible theory, appropriate timing, adequate controls, and protection against confounding, reverse causality, and structural change. This chapter develops multiple regression, dynamic regression, leading-indicator models, and selected econometric tools for operations managers.

Learning Objectives

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

1. Formulate a multiple regression model for demand forecasting.
2. Interpret coefficients, elasticities, interactions, dummy variables, and transformations.
3. Develop dynamic regression models containing lags, trend, seasonality, and autocorrelated errors.
4. Evaluate leading indicators and distinguish predictive usefulness from causal validity.
5. Diagnose multicollinearity, endogeneity, heteroskedasticity, autocorrelation, and structural change.
6. Use interventions, instrumental variables, panel data, and quasi-experimental designs appropriately.
7. Translate model results into operational and managerial decisions.

1. From Associational to Causal Forecasting

Let Y_t denote demand in period t. A purely extrapolative forecast uses past demand, such as a moving average or an autoregressive model. A causal forecast uses explanatory variables X_1t, X_2t, …, X_kt:

Here, β₀ is the intercept, β_j is the partial effect of predictor X_j when other included predictors are held constant, and ε_t contains omitted influences and random variation. The units of β_j are units of demand per unit of X_j.

The phrase “holding other variables constant” is a mathematical interpretation, not proof that such a comparison is operationally or causally possible. For example, price may be reduced precisely when weak demand is expected. A negative price coefficient could therefore reflect both the demand response to price and management’s reaction to weak demand.

A useful distinction is:

• Forecasting question: Given information available at time t, what will demand be at t+h?
• Causal question: What would demand have been if a specified intervention, such as a price change, had not occurred?

A model can forecast well without identifying a causal effect, and a credible causal model may forecast poorly if future explanatory variables are unavailable or poorly predicted.

2. Multiple Regression for Demand

2.1 Model specification

A practical demand model may include price, promotion, distribution, calendar effects, and economic variables:

Q_t is quantity demanded; P_t is price; A_t is advertising or promotional expenditure; D_t is distribution or availability; I_t is an economic indicator such as disposable income; and S_t is a vector of seasonal or event indicators. The vector γ contains their coefficients.

The ordinary least squares (OLS) estimator chooses coefficients that minimize the sum of squared residuals:

In matrix notation, when X'X is nonsingular:

X is the design matrix, Y is the vector of observed demands, and β̂ is the estimated coefficient vector. OLS is appropriate as a descriptive or predictive estimator under broad conditions, but unbiased causal interpretation requires stronger assumptions.

2.2 Core assumptions

The classical linear model commonly assumes:

1. Correct conditional mean: E(ε_t|X)=0. The included predictors contain sufficient information for the conditional mean being estimated.
2. No perfect multicollinearity: no predictor is an exact linear combination of other predictors.
3. Independent or appropriately modeled errors: observations are not treated as independent when serial dependence is present.
4. Constant conditional variance for conventional standard errors: Var(ε_t|X)=σ².
5. Normality is not required for unbiased coefficients, but is useful for exact small-sample tests under the classical model.

In time-series demand data, independent errors and constant variance often fail. Coefficients may remain useful, but inference and forecasting intervals must then be adjusted.

2.3 Interpretation of coefficients

In a level-level model, β₁ is the expected change in demand units associated with a one-unit increase in price, conditional on other variables. If quantity is measured in cases per week and price in dollars per case, β₁ has units of cases per week per dollar.

A log transformation often gives more useful interpretations:

In a log-log model, β₁ is an elasticity: approximately the percentage change in demand associated with a 1% price change. A negative price elasticity is normally consistent with a downward-sloping demand relationship, although the estimate may be distorted by endogeneity.

In a log-level model:

A one-unit increase in P_t is associated approximately with a 100β₁ percent change in demand when β₁ is small. More exactly, the percentage change is 100(e^β₁−1) percent, assuming the model is correctly specified.

2.4 Promotions, events, and interactions

A binary promotion indicator M_t can represent whether a promotion is active:

For a level-dependent model, β₂ is the expected difference between promotion and non-promotion periods at the same price and other covariate values. In a log-demand model, the exact percentage difference is 100(e^β₂−1) percent.

An interaction allows the effect of price to differ during promotion:

The price slope outside promotion is β₁, while during promotion it is β₁+β₃. Omitting a meaningful interaction can produce misleading average effects.

Events such as festivals, product launches, strikes, or extreme weather should be coded using carefully defined indicators or event windows. The model should distinguish a one-day shock from a permanent level shift.

3. Dynamic Regression and Time-Series Structure

Demand often responds with delay. Advertising today may influence sales for several weeks; a price reduction may generate immediate purchases and later stock depletion. Dynamic regression incorporates lags:

The short-run effect of a one-unit increase in X at lag j is β_j. If all lag coefficients represent the response to the same temporary change, the cumulative effect is:

This sum is meaningful only when the lag structure, units, and intervention interpretation are compatible.

3.1 Distributed-lag models

A distributed-lag model estimates separate or constrained responses across lags. A polynomial distributed lag imposes a smooth pattern, reducing parameters. A geometric lag model assumes effects decline at a constant rate. Such restrictions can improve precision but may introduce bias if the imposed shape is wrong.

3.2 Autoregressive distributed-lag models

A model may include its own past demand:

If |φ|<1 and the process is stable, the long-run effect of a permanent one-unit increase in X is:

The formula assumes a stable, correctly specified model and a constant effect over the relevant horizon. For example, if φ=0.60 and β=4 units, the long-run effect is 4÷(1−0.60)=10 units. The immediate effect is 4 units; persistence adds the remaining 6 units over time.

Including lagged demand can capture habit persistence and omitted dynamics, but it complicates causal interpretation. If a lagged dependent variable is correlated with the error, especially in short panels or after structural changes, standard estimators may be biased.

3.3 Seasonality and trend

Trend can be represented by a time index t, while monthly or weekly seasonality can be represented by dummy variables. With an intercept, one category must be omitted as the reference category to avoid perfect multicollinearity. Fourier terms may provide a parsimonious smooth seasonal representation:

Here m is the seasonal cycle length in observations. Seasonal controls should not absorb the very event effect that management wishes to estimate.

4. Leading Indicators and Forecast Timing

A leading indicator changes before the target demand measure and may improve forecasts. Examples include search volume, retailer orders, building permits, consumer confidence, freight bookings, and supplier purchase orders. A useful indicator must be available before the forecast decision and must add information beyond the demand series’ own history.

A predictive model can be written as:

Z_t is the candidate indicator, W_t contains other information available at time t, and h is the forecast horizon. The indicator should be aligned by information-release date, not merely by the date printed in a database. Publication delays and revisions can create look-ahead bias.

A leading indicator is not necessarily causal. Search activity may predict demand because both respond to an unobserved event. For operational forecasting, predictive performance may be sufficient. For policy, pricing, or investment decisions, causal evidence is usually required.

5. Diagnosing Model Problems

5.1 Multicollinearity

Multicollinearity occurs when predictors contain overlapping information. Price and promotion, advertising and distribution, or GDP and consumer income may move together. It generally inflates coefficient variance and makes individual effects unstable, though it does not necessarily bias OLS when exogeneity holds.

The variance inflation factor for predictor j is:

R_j² is obtained by regressing predictor j on the other predictors. A high VIF is a warning, not a universal pass–fail rule. Remedies include combining variables, removing redundant predictors, using theory to select controls, collecting more informative data, or applying regularization for prediction. Dropping a variable solely because it is statistically insignificant can create omitted-variable bias.

5.2 Endogeneity and reverse causality

A predictor is endogenous when it is correlated with the error term:

In demand analysis, price is often set in response to expected demand, creating simultaneity. Promotion may be targeted to products expected to perform poorly. Omitted competitor actions may affect both price and sales.

Consequences include biased and inconsistent causal coefficients. Possible strategies are randomized experiments, natural experiments, fixed effects, lagged variables when theoretically defensible, instrumental variables, control functions, and structural models. A lag is not automatically an instrument: past price may still correlate with persistent unobserved demand.

5.3 Instrumental variables

An instrument Z for endogenous X should satisfy:

1. Relevance: Cov(Z,X) ≠ 0.
2. Exclusion: Z affects demand only through X.
3. Independence or exogeneity: Z is unrelated to unobserved demand shocks.

Two-stage least squares first estimates:

and then substitutes the fitted value of X into the demand equation. Weak instruments can create severe finite-sample problems. Instruments must therefore be defended institutionally and economically, not selected merely because they produce a desired coefficient.

5.4 Autocorrelation and heteroskedasticity

Serially correlated residuals imply that errors in adjacent periods are related. This can make conventional standard errors too small and can degrade forecasts. Residual plots, autocorrelation functions, and tests such as Breusch–Godfrey can help diagnose the issue. Remedies include dynamic terms, explicit error models, feasible generalized least squares, or heteroskedasticity-and-autocorrelation-consistent standard errors.

Heteroskedasticity occurs when error variance changes with scale, product, season, or demand level. Robust standard errors improve inference under broad variance patterns, but they do not repair omitted variables or endogeneity. Transformations such as logarithms may stabilize variance when economically appropriate.

5.5 Nonstationarity and spurious regression

Two trending series may show a high R² and significant coefficients even when they are unrelated. Stationarity concerns whether a series’ probabilistic behavior is stable over time. Differencing can remove stochastic trends, but may also eliminate economically meaningful long-run information. If nonstationary series are cointegrated, an error-correction model can represent short-run adjustment toward a long-run relationship:

Here (Y_t−1−θX_t−1) is the lagged disequilibrium term; a negative λ commonly indicates adjustment toward equilibrium. Cointegration analysis requires adequate sample size, sound economic theory, and careful testing.

6. Structural Change and Model Stability

Relationships change because of competitor entry, regulation, technology, supply disruption, inflation, channel migration, or consumer behavior. A coefficient estimated from pre-disruption data may be inappropriate after the disruption.

A break can be modeled with an intervention indicator B_t:

A slope change requires an interaction:

Before the break, the slope is β₁; after it, the slope is β₁+γ. Rolling estimation, recursive residuals, Chow-type tests, and formal change-point methods can support diagnosis. However, a statistical break should be interpreted alongside operational knowledge.

Forecast evaluation must be time-ordered. Randomly shuffling time-series observations leaks future information. Use rolling-origin or expanding-window validation, and compare against simple benchmarks such as seasonal-naïve forecasts. Assess both accuracy and decision value.

Common metrics include:

MAE is in demand units; RMSE penalizes large errors more strongly. Percentage metrics can be unstable when actual demand is near zero.

7. Worked Example: Price and Promotion

Suppose a retailer estimates weekly demand in thousands of units using:

where P is price in dollars and M equals 1 during promotion and 0 otherwise. At P=5 without promotion, predicted demand is:

Thus, predicted demand is 80 thousand units. At the same price during promotion:

The model attributes an additional 18 thousand units to promotion, conditional on the assumptions and included variables.

The price coefficient implies eight thousand fewer units for a one-dollar increase, holding promotion status constant. At the non-promotion point, the point price elasticity is:

This is a local elasticity, not necessarily valid at every price. It also should not be interpreted as a causal elasticity unless price endogeneity and other identification concerns have been addressed.

8. Illustrative Case: FreshBrew Coffee Pods

Background

FreshBrew sells coffee pods through supermarkets and an online channel. Weekly demand varies with price, digital advertising, temporary discounts, holidays, and household income. Management wants a twelve-week forecast and an estimate of whether discounts create incremental volume or merely shift purchases forward.

Problem

The commercial team proposes regressing weekly sales on the discount rate and advertising expenditure. Analysts observe that discounts are more frequent before weak weeks, and advertising rises around holidays. A simple regression would therefore risk reverse causality and confounding.

Analysis

The analytics team first creates an information-timed data set: only variables known before each forecast origin are retained. It includes lagged sales, regular price, discount indicator, advertising distributed across current and two lagged weeks, holiday indicators, distribution coverage, and a competitor-price index. Product and channel fixed effects are used in a panel model, while standard errors are clustered at the product level.

The team checks correlations and VIF values, plots residuals, tests for serial dependence, and performs rolling-origin validation. Because discounts are targeted, the team does not present the discount coefficient as a causal effect. Instead, it uses a controlled regional pilot in which eligible stores are assigned different discount schedules. Store assignment is stratified by historical demand and region.

Approach and solution

The forecasting model uses dynamic regression for operational planning. The pilot estimates incremental lift by comparing treated and control stores while controlling for pre-treatment demand. A difference-in-differences specification is considered only after checking that treatment and control stores have approximately parallel pre-intervention trends:

α_i represents store effects, δ_t represents common time effects, and τ is the estimated average treatment effect under the design assumptions.

The company also tests whether online advertising affects supermarket demand through channel substitution. It reports separate direct-channel effects rather than assuming that total sales equal the sum of independently estimated channel effects.

Outcome

The dynamic model improves twelve-week forecast accuracy over a seasonal-naïve benchmark, particularly when future promotional calendars are known. The pilot indicates positive incremental volume, but the effect is smaller than the raw association in historical data. Inventory planners use the forecast distribution rather than a single point estimate, and pricing managers use the experimental estimate for decisions about discount depth.

Learning points

1. Forecast variables must be available at the time a forecast is made.
2. Historical promotion coefficients can be biased by targeting.
3. A pilot can answer a causal question more credibly than observational regression.
4. Forecasting and causal estimation may require related but different models.
5. Channel effects, stockouts, and purchase timing must be considered before declaring promotional success.

9. Managerial, Ethical, and Governance Considerations

Causal forecasts affect prices, staffing, supplier commitments, and customer access. Managers should document variable definitions, data vintages, transformations, exclusions, model versions, and approval responsibilities. Forecasts should communicate uncertainty, not conceal it behind spurious precision.

Data governance includes access control, retention rules, audit trails, and protection of personal information. Search, loyalty, and location data may create privacy risks even when individual identifiers are removed. Fairness concerns arise when models use proxies for income, neighborhood, age, or other protected characteristics to allocate inventory or vary prices. Governance should require purpose limitation, lawful data use, bias assessment, human review for high-impact decisions, and monitoring after deployment.

AI and automated machine-learning systems can discover nonlinearities and interactions, but predictive importance is not causal evidence. Flexible models may overfit, extrapolate poorly after structural breaks, and obscure decision logic. A strong practice is to combine domain theory and experimental design with machine learning for residual prediction, scenario generation, or heterogeneous-effect exploration. Out-of-time validation and model monitoring remain essential.

10. Best-Practice Workflow

1. Define the decision, horizon, unit of analysis, and causal or predictive objective.
2. Establish a conceptual demand theory before selecting variables.
3. Build an information-timed data set and prevent look-ahead bias.
4. Explore levels, transformations, missingness, outliers, seasonality, and breaks.
5. Fit a transparent benchmark before adding complexity.
6. Add lags, interactions, events, and economic variables only when justified.
7. Diagnose multicollinearity, endogeneity, autocorrelation, heteroskedasticity, and stability.
8. Use experiments or credible quasi-experiments for causal claims.
9. Validate with rolling or expanding time windows and compare decision costs.
10. Report assumptions, uncertainty, limitations, and monitoring triggers.

Chapter Summary

Causal forecasting links demand to price, promotion, economic conditions, events, and other explanatory variables. Multiple regression estimates partial relationships, while transformations and interactions improve interpretation when their assumptions are appropriate. Dynamic regression represents delayed responses, persistence, seasonality, and autocorrelated errors; distributed-lag coefficients can be combined to obtain cumulative effects, and stable autoregressive models support long-run multipliers.

Leading indicators can improve forecasts when they are timely and add information, but predictive usefulness does not establish causality. Multicollinearity reduces precision, endogeneity threatens causal interpretation, and autocorrelation, heteroskedasticity, nonstationarity, and structural change can invalidate naïve inference or degrade forecasts. Experiments, instrumental variables, fixed effects, difference-in-differences, and error-correction models are specialized tools whose assumptions must be defended. Professional practice combines statistical diagnostics, operational knowledge, time-ordered validation, ethical governance, and clear communication of uncertainty.

Key Terms

Causal forecasting: Forecasting that uses explanatory variables and seeks to understand intervention-relevant relationships.

Multiple regression: A model relating one dependent variable to several predictors.

Ordinary least squares: An estimator minimizing the sum of squared residuals.

Partial effect: The association between a predictor and outcome conditional on included variables.

Elasticity: Percentage responsiveness of an outcome to a percentage change in a predictor.

Interaction: A term allowing one variable’s effect to depend on another variable.

Dummy variable: A coded indicator, commonly 0 or 1, representing a category or event.

Dynamic regression: Regression containing lags, autoregressive terms, or both.

Distributed lag: A model assigning effects to current and past values of a predictor.

Leading indicator: A variable available before and predictive of a future target.

Multicollinearity: Overlap or strong linear dependence among predictors.

Endogeneity: Correlation between a predictor and the model error term.

Instrumental variable: An external variable used to identify an endogenous regressor under relevance and exclusion assumptions.

Heteroskedasticity: Nonconstant conditional error variance.

Autocorrelation: Dependence among errors across time.

Stationarity: Stability of a series’ probabilistic behavior over time.

Cointegration: A stable long-run relationship among nonstationary variables.

Structural break: A change in a model’s underlying relationship over time.

Difference-in-differences: A quasi-experimental estimator comparing changes across treated and control groups.

Rolling-origin validation: Time-ordered evaluation using successive historical forecast origins.

Review Questions

Short Answer Questions

1. Distinguish a predictive forecasting question from a causal question.
2. What does a regression coefficient mean in a level-level model?
3. Why can a lagged predictor be useful in demand forecasting?
4. Define multicollinearity and state one consequence.
5. Why must forecast validation preserve temporal order?

Long Answer Questions

1. Explain the assumptions of OLS and discuss which are especially vulnerable in demand time series.
2. Compare level-level, log-level, and log-log demand models, including coefficient interpretation.
3. Explain endogeneity in pricing and compare experiments, instrumental variables, and fixed effects as responses.
4. Develop a dynamic regression framework for a product affected by advertising carryover and seasonality.
5. Discuss how structural breaks and nonstationarity can make historical causal estimates unreliable.

Analytical / Application Questions

1. A model is Q̂=200−10P+25M, where demand is in thousands, price is in dollars, and M is a promotion indicator. Calculate demand at P=8 with and without promotion, and interpret the promotion coefficient.
2. An autoregressive distributed-lag model has φ=0.75 and a current-period effect β=5. Calculate the long-run effect of a permanent one-unit increase in the predictor, assuming stability.
3. A retailer observes that discounted products have high sales, but discounts are scheduled when managers expect weak demand. Explain why the raw comparison is not causal and propose a credible evaluation design.

Multiple Choice Questions with Answers

Question 1

In a level-level demand model, what does a price coefficient of −6 mean?

A. A 1% price increase reduces demand by 6%.

B. A one-unit price increase is associated with six fewer demand units, holding other variables constant.

C. Demand is six times price.

D. Price explains 6% of demand.

Correct Answer: B. A one-unit price increase is associated with six fewer demand units, holding other variables constant.

Question 2

In a log-log model, the coefficient on price is most naturally interpreted as:

A. A demand elasticity.

B. A demand intercept.

C. A seasonal index.

D. A variance estimate.

Correct Answer: A. A demand elasticity.

Question 3

If an autoregressive distributed-lag model has φ = 0.5 and β = 3, what is the long-run effect of a permanent one-unit increase in X?

A. 1.5 units

B. 3 units

C. 6 units

D. 9 units

Correct Answer: C. 6 units

Question 4

Which condition is required for a valid instrumental variable?

A. It must be perfectly correlated with the error term.

B. It must affect the outcome through every possible pathway.

C. It must be relevant to the endogenous predictor and satisfy the exclusion condition.

D. It must be identical to the dependent variable.

Correct Answer: C. It must be relevant to the endogenous predictor and satisfy the exclusion condition.

Question 5

A high VIF primarily signals that:

A. The dependent variable is nonstationary.

B. Predictors contain substantial overlapping information.

C. The model has no intercept.

D. Forecast errors are necessarily unbiased.

Correct Answer: B. Predictors contain substantial overlapping information.

Question 6

Why can historical promotion coefficients overstate causal promotional lift?

A. Promotions may be targeted when weak demand is expected.

B. Promotion indicators always create perfect multicollinearity.

C. Promotion variables cannot be included in regression.

D. Demand is always independent of promotion timing.

Correct Answer: A. Promotions may be targeted when weak demand is expected.

Question 7

Which validation approach is most appropriate for a time-series forecasting model?

A. Randomly shuffle all observations before splitting.

B. Use only the largest historical demand value.

C. Use rolling-origin or expanding-window evaluation.

D. Evaluate on the same observations used for fitting only.

Correct Answer: C. Use rolling-origin or expanding-window evaluation.

Question 8

In a model with an interaction P×M, the price slope during promotion is:

A. The intercept only.

B. The coefficient of promotion alone.

C. The price coefficient plus the interaction coefficient.

D. Always zero.

Correct Answer: C. The price coefficient plus the interaction coefficient.

Question 9

What is a leading indicator?

A. A variable observed only after the target is realized.

B. A variable that changes before and helps predict the target.

C. A coefficient with a positive sign.

D. A residual with zero variance.

Correct Answer: B. A variable that changes before and helps predict the target.

Question 10

If Cov(X_t,ε_t)≠0, the predictor is described as:

A. Exogenous.

B. Stationary.

C. Endogenous.

D. Seasonally adjusted.

Correct Answer: C. Endogenous.