Causal, Regression and Econometric Forecasting

Introduction

Forecasting is often presented as an exercise in extrapolating historical patterns. In operations management, however, important changes in demand, cost, capacity, lead time and service requirements are frequently caused by identifiable drivers. A promotion can increase sales, a price change can reduce demand, a fuel-price increase can raise distribution cost, and a recession can alter industrial orders. Forecasting methods that explicitly represent such drivers are therefore valuable because they support both prediction and managerial explanation.

Regression and econometric forecasting relate a dependent variable to one or more explanatory variables. A causal forecast goes further: it interprets the explanatory variables as drivers that would plausibly change the outcome under a defined intervention. This distinction is fundamental. A variable may improve predictive accuracy without being a genuine cause. For example, umbrella sales may predict traffic congestion, although umbrellas do not cause congestion; rainfall is a common cause of both.

This chapter develops multiple regression, dynamic regression and econometric forecasting for operational applications. It addresses variable selection, interaction effects, multicollinearity, autocorrelation, heteroscedasticity and causal interpretation. The emphasis is not merely on fitting equations, but on building models that are statistically defensible, operationally useful and transparent about uncertainty.

Learning Objectives

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

1. Distinguish predictive regression from causal analysis.
2. Formulate and interpret multiple linear regression models.
3. Incorporate categorical variables, nonlinear effects, interactions and lagged drivers.
4. Select variables using theory, diagnostics and out-of-sample evidence.
5. Diagnose multicollinearity, autocorrelation and heteroscedasticity.
6. Construct and evaluate dynamic regression forecasts.
7. Explain the assumptions required for causal interpretation.
8. Apply regression forecasting to an operational decision and communicate its limitations.

1. Regression Forecasting as a Driver-Based Framework

Let y_t denote the quantity to be forecast at time t, and let x_1t, x_2t, …, x_kt represent observed or planned drivers. A multiple regression model is commonly written as:

Here, β₀ is the intercept, β_j is the partial effect associated with driver x_j, and ε_t is the unexplained error. The model is linear in its unknown coefficients, although variables themselves may be transformed or combined.

The conditional mean is:

where X_t is the vector of drivers at time t. The model is appropriate when the conditional mean is reasonably represented by the specified functional form, observations are sufficiently informative, and the error structure is handled appropriately. Linearity does not require the raw relationship between every variable and the outcome to be a straight line; polynomial, logarithmic and interaction terms can represent nonlinear relationships.

A forecast made for a future period uses future driver values. If those values are unknown, they must themselves be forecast, planned, or represented through scenarios. Consequently, forecast uncertainty includes both model uncertainty and driver uncertainty.

Regression versus causal forecasting

A predictive interpretation asks: “How accurately does the model estimate future demand?” A causal interpretation asks: “What would happen to demand if a specified driver were changed, holding the relevant alternatives constant?” The second question requires stronger assumptions and a credible identification strategy.

A coefficient is not automatically causal merely because it is statistically significant. Omitted variables, reverse causality, measurement error, selection bias and simultaneous decisions can all distort estimates. Regression is a language for expressing relationships; causal claims require design, institutional knowledge and assumptions that make the causal comparison credible.

2. Building a Multiple Regression Model

2.1 Define the forecasting unit and timing

The analyst should first define the unit of observation: store-week, product-day, supplier-month or plant-quarter. The time at which each variable becomes available must be recorded. A model that uses a promotion variable known only after the sales period has ended will produce leakage and unrealistic forecasts.

The dependent variable should reflect the decision. For example, gross demand may be preferable to observed sales when stockouts censor demand. If the outcome is a count, proportion or strictly positive cost, a generalized linear model or transformation may be more suitable than ordinary least squares.

2.2 Interpretation of coefficients

In a model with numerical variables, β_j is the expected change in y associated with a one-unit increase in x_j, holding the other included variables constant. This is a partial association. Units matter: if demand is measured in units and advertising expenditure in thousands of currency units, the coefficient measures units per thousand currency units.

For a log-level model,

100β₁ approximately represents the percentage change in y associated with a one-unit increase in x, for small coefficients. The exact percentage change is 100(e^β₁-1). This interpretation requires positive y and a meaningful logarithmic scale.

For a log-log model,

β₁ is an elasticity: a 1% change in x is associated with approximately a β₁% change in y.

2.3 Estimation by ordinary least squares

Ordinary least squares (OLS) selects coefficients that minimize the sum of squared residuals:

where ŷ_t is the fitted value and e_t = y_t − ŷ_t is the residual. In matrix notation, with a full-rank design matrix X,

The inverse exists only when the columns of X are linearly independent. Exact linear dependence is perfect multicollinearity and prevents unique OLS estimation.

For unbiased coefficient estimation under the classical linear model, important conditions include a correctly specified conditional mean, random or appropriately structured sampling, no perfect multicollinearity and E(ε|X)=0. For conventional standard errors, additional assumptions concerning error variance and dependence are required. Forecasting can remain useful when some classical assumptions fail, but inference and model correction must then be reconsidered.

2.4 Goodness of fit and forecast accuracy

The coefficient of determination is:

where SST = ∑(y_t − ȳ)². R² is the proportion of in-sample variation explained by the fitted model relative to a mean-only benchmark. It does not establish causality and may increase when irrelevant variables are added.

Adjusted R² penalizes the number of predictors:

where n is the sample size and k is the number of predictors excluding the intercept. For forecasting, out-of-sample measures are usually more important. Common measures include:

MAE is in the units of the outcome and treats errors linearly. RMSE gives greater weight to large errors. MAPE, (100/n)∑|e_t/y_t|, is problematic when actual values are zero or close to zero and can penalize positive and negative errors asymmetrically. Metrics should therefore be selected according to the operational cost of error.

3. Representing Operational and Economic Drivers

Categorical variables

A categorical variable such as region or promotion type is represented using indicator variables. If a factor has m categories, a model with an intercept normally includes m−1 indicators. The omitted category is the reference group. Including all m indicators together with an intercept creates perfect multicollinearity.

Seasonality and calendar effects

Month, weekday, holiday and quarter indicators can represent recurring calendar effects. They should not be used mechanically: a holiday effect may change over time, and a calendar variable may be a proxy for an unobserved operational process. Fourier terms, such as sine and cosine functions, can represent smooth seasonality, but their periods must correspond to the data frequency and business cycle.

Nonlinear effects

A linear price effect may be inadequate if demand falls sharply after a threshold. A quadratic specification is:

The marginal association is ∂E(y|x)/∂x = β₁ + 2β₂x. The turning point, when meaningful and within the observed range, is x* = −β₁/(2β₂). Extrapolating a polynomial far beyond the data can be unstable, so splines or theory-based transformations may be safer.

Interaction effects

An interaction allows the effect of one driver to depend on another:

The effect of x is β₁ + β₃z. Thus, β₁ is the effect of x only when z=0. Centering z at a meaningful value can make the main effect easier to interpret, although centering does not eliminate the substantive interaction. Interaction terms should be hypothesized before estimation where possible, because searching many interactions increases the risk of spurious findings.

4. Variable Selection and Model Specification

Variable selection should combine operational theory, data quality, timing, statistical evidence and validation performance. A purely automated procedure may select variables that exploit accidental historical patterns and may omit strategically important variables whose effects are weak in a particular sample.

A disciplined process is:

1. Define the decision and forecasting horizon.
2. Create a causal or operational process map.
3. Screen variables for availability, leakage, missingness and measurement quality.
4. Fit a theory-informed baseline model.
5. Add nonlinearities, interactions and lags only when justified.
6. Compare candidate models using rolling-origin validation.
7. Examine residuals and stability across time, products and locations.
8. Prefer the simplest model that meets accuracy and governance requirements.

AIC and BIC can compare in-sample likelihood-based models; BIC imposes a stronger penalty for complexity. They should not replace time-respecting validation. Stepwise selection based solely on p-values produces unstable models, biased coefficient estimates and overstated significance. Regularization can help when there are many correlated predictors. Ridge regression minimizes squared error plus λ∑β_j², while lasso minimizes squared error plus λ∑|β_j|. The tuning parameter λ≥0 is selected by validation. Regularization improves prediction in many settings, but coefficients should not automatically be interpreted as causal effects.

5. Multicollinearity

Multicollinearity occurs when predictors contain overlapping information. Advertising expenditure, impressions and reach may be highly correlated; temperature and season may also move together. It does not necessarily bias OLS coefficients when exogeneity holds, but it increases variance, makes estimates sensitive to small data changes and complicates interpretation.

The variance inflation factor for predictor x_j is:

where R_j² comes from regressing x_j on the other predictors. A high VIF is a warning, not a universal proof of unacceptable modeling. Remedies include combining measures using domain knowledge, removing redundant variables, collecting more informative data, centering variables involved in interactions, or using ridge regression. Dropping a variable solely because it is correlated with another may create omitted-variable bias if the removed variable is substantively important.

6. Dynamic Regression and Econometric Structure

Operational outcomes often depend on their own history. A dynamic regression can be written as:

Lagged drivers can also be included:

The contemporaneous coefficient β₀ in the lag expression should not be confused with the intercept; notation must be defined carefully in implementation. A distributed-lag model distinguishes immediate from delayed effects. If the driver is a promotion, lag terms may represent stock-up purchases and post-promotion depletion.

When the dependent variable is included as a lagged predictor, the coefficient on a driver is often a short-run effect. Under a stable model with one lag, y_t = α + βx_t + φy_t−1 + ε_t, the long-run level effect of a permanent one-unit increase in x is:

This requires stability, commonly |φ|<1, and a meaningful constant relationship over the relevant horizon. The expression should not be used when the process is explosive, structurally changing or improperly differenced.

Autocorrelation

Autocorrelation means that errors are related across time. Positive autocorrelation often produces runs of under- and over-forecasting. OLS coefficients may remain unbiased under strict exogeneity, but conventional standard errors are generally unreliable, and forecasts may be inefficient.

A residual autocorrelation function, residual time plot and tests such as Durbin–Watson or Breusch–Godfrey can provide evidence. Durbin–Watson is approximately 2 when first-order residual autocorrelation is absent, but it is not a universal diagnostic and has limitations with lagged dependent variables.

Remedies include adding omitted lags or seasonality, modeling errors with ARIMA or autoregressive structures, using generalized least squares, or using heteroscedasticity-and-autocorrelation-consistent standard errors for inference. The remedy should address the data-generating process rather than simply suppress a diagnostic statistic.

Heteroscedasticity

Heteroscedasticity occurs when error variance changes with predictors or time. For example, demand errors may be larger for high-volume stores. OLS coefficient estimates can remain unbiased under exogeneity, but usual standard errors are invalid and prediction intervals may be poorly calibrated.

Residual-versus-fitted plots, scale-location plots and tests such as Breusch–Pagan can be useful. Remedies include transforming the outcome, weighted least squares, modeling variance explicitly, or using heteroscedasticity-robust standard errors. Robust standard errors improve inference but do not automatically improve point forecasts or correct a misspecified mean equation.

7. Econometric and Causal Interpretation

A causal effect is defined relative to a potential outcome: the outcome that would be observed under treatment or exposure and the outcome that would be observed for the same unit under an alternative condition. Only one is observed, creating the fundamental problem of causal inference.

A regression coefficient may be interpreted causally only under assumptions such as:

• Consistency: the recorded treatment corresponds to the intervention of interest.
• No unmeasured confounding: after conditioning on appropriate covariates, treatment assignment is independent of potential outcomes.
• Positivity: relevant treatment levels occur for comparable observations.
• Correct temporal ordering: causes precede effects.
• Stable measurement and interference assumptions: one unit’s treatment does not unexpectedly alter another unit’s outcome, unless modeled.

Operational decisions often create endogeneity. Managers may increase inventory when they anticipate high demand, making inventory appear positively associated with demand even if additional inventory does not cause demand. Price may be lowered when demand is weak, producing a misleading negative relationship. Instrumental variables, natural experiments, difference-in-differences, randomized tests and regression discontinuity designs can sometimes address such problems, but each requires design-specific assumptions.

A useful causal diagram distinguishes treatment, outcome, confounders, mediators and colliders. Conditioning on a collider can create bias. Analysts should therefore avoid selecting controls merely because they are correlated with the outcome; controls should be chosen based on the causal question and timing.

Forecasting and causal estimation can use different models. A variable that is highly predictive but occurs after the decision cannot support a pre-decision causal claim. Conversely, a valid causal driver may add little short-term predictive power if its variation is small.

8. Worked Example: Regional Distribution Demand

Consider a distributor forecasting weekly cases demanded by regional warehouses. An illustrative model, estimated on historical data, is:

where ŷ_t is forecast weekly demand in cases, P_t is price in currency units per case, A_t is advertising expenditure in thousands of currency units, H_t is a holiday indicator equal to 1 during a designated holiday week and 0 otherwise, and ŷ_t−1 is the previous week’s demand forecast or observed demand, depending on the implementation. The coefficient definitions and data-generation procedure must be fixed before operational use.

For a non-holiday week with price 20, advertising expenditure 10 (thousand currency units), and previous-week demand 4,000 cases:

The point forecast is therefore approximately 1,644 cases. The price coefficient suggests that, holding included variables constant, a one-currency-unit increase in price is associated with 18 fewer cases per week. This is not automatically a causal price elasticity. The advertising coefficient implies 0.35 additional cases for each additional thousand currency units, which appears operationally small and should be checked for unit definition, scaling and statistical uncertainty.

If the holiday indicator changes from 0 to 1, the forecast increases by 75 cases, conditional on the other variables. If the price rises by 2 currency units, the linear model reduces the forecast by 36 cases. Such calculations are valid only within the observed price range and assuming no interaction or nonlinear price effect.

Suppose residual plots reveal larger errors in high-demand weeks and a residual autocorrelation of 0.4 at lag one. The analyst should not merely report the original regression. A revised approach might model the residual process, consider a log or variance-stabilizing transformation, and use rolling-origin validation. Forecast intervals should widen when demand volatility is high. If future advertising expenditure is not known, the analyst should produce forecasts under planned, conservative and aggressive advertising scenarios rather than silently treating the historical value as known.

9. Illustrative Case: FreshMart Promotion Planning

Background. FreshMart, a supermarket chain, must decide weekly orders for a perishable product. It has store-level sales, price, promotion type, weather, local events, inventory availability and competitor-price data for three years.

Problem. A simple seasonal average underestimates demand during promotions and creates stockouts. Management proposes a regression using promotion status, price, temperature and last week’s sales.

Analysis. The analytics team first checks whether sales are censored by stockouts. It identifies weeks in which inventory reached zero and flags them rather than treating recorded sales as unconstrained demand. It then divides data chronologically into training, validation and test periods. A baseline model includes weekday, month and store effects. A dynamic model adds lagged demand and promotion variables. Interaction terms allow promotion effectiveness to differ by store format.

The team finds that price and promotion are strongly correlated because promotions usually involve discounts. Rather than interpreting both coefficients mechanically, it defines the estimand: the expected sales change from a specified promotion at a specified price. It reports VIF values, coefficient intervals and validation errors. Residual diagnostics show autocorrelation during holiday periods, so holiday lags and an error model are considered. Robust or clustered inference is used where appropriate, but forecast performance is assessed through future-period simulation.

Approach and solution. The final forecasting system uses a dynamic regression with store effects, promotion-price interactions, lagged demand, holiday indicators and an explicit stockout treatment. It generates scenario forecasts for each proposed promotion. A separate randomized promotion test is recommended to estimate incremental causal lift, because historical promotions were targeted to stores expected to have high demand.

Outcome. In the illustrative case, the model reduces validation MAE relative to the seasonal average and improves order planning, but management is warned that improved forecast accuracy does not prove that promotions cause the estimated lift. The randomized test is used for causal measurement, while the regression remains the operational forecasting engine.

Learning points. The case demonstrates the importance of defining the outcome, preventing leakage, recognizing censored demand, separating prediction from causation, modeling dynamics and validating against future-like data.

10. Implementation, Governance and Contemporary Practice

A production forecasting process should maintain a data dictionary, version-controlled feature definitions, documented model assumptions, reproducible training data and an audit trail of forecasts and overrides. Forecasts should be monitored for bias, drift, degradation by segment and changes in driver availability. A model that performs well overall may systematically underforecast rural stores or overforecast low-volume products.

Machine-learning methods such as boosted trees can capture nonlinearities and interactions, while neural networks may help with large, high-frequency data. However, greater predictive flexibility does not guarantee causal validity or operational robustness. Interpretable regression remains valuable when managers need to understand driver effects, run scenarios and defend decisions. Explainability tools describe model behavior but do not establish causality.

Governance issues include privacy, fairness, access control, data provenance and appropriate use of personally identifiable information. Demand forecasts can influence staffing, supplier allocation and customer availability; systematic errors may therefore create unequal service outcomes. Human review should focus on exceptional conditions and documented override rules, not on arbitrary adjustment of inconvenient forecasts.

Challenges and Limitations

Regression forecasting is limited by structural breaks, omitted variables, poor measurement, changing customer behavior and uncertain future drivers. Coefficients estimated during stable supply conditions may fail during disruption. Extrapolation beyond the historical range is especially risky. Aggregation can hide segment-specific effects, while disaggregation can create noisy estimates.

A statistically significant coefficient may be operationally negligible. A high R² may coexist with poor future forecasts. Conversely, a model with modest in-sample fit may be valuable if it is stable, interpretable and accurate at the decision horizon. Forecast intervals should be communicated alongside point forecasts, and scenario analysis should be used when economic or operational drivers are uncertain.

Best-Practice Checklist

• Define the forecast horizon, unit and decision before selecting variables.
• Use only information available at the forecast origin.
• Separate descriptive, predictive and causal claims.
• Start with a theory-informed baseline.
• Represent seasonality, nonlinearities and interactions deliberately.
• Examine multicollinearity and coefficient stability.
• Diagnose autocorrelation and heteroscedasticity.
• Use rolling-origin validation for time-dependent data.
• Forecast uncertain drivers through scenarios or linked models.
• Report uncertainty, limitations and data-quality concerns.
• Monitor performance and fairness after deployment.
• Use experiments or credible quasi-experiments for causal claims.

Chapter Summary

Regression forecasting links an operational outcome to measurable drivers and is especially useful when managers can observe or plan prices, promotions, capacity, economic indicators and calendar effects. Multiple regression estimates partial relationships, while transformations, indicators and interactions allow richer functional forms. Dynamic regression incorporates lags of outcomes and drivers and can distinguish short-run from long-run effects under stability conditions.

Good modeling requires more than maximizing in-sample fit. Variable selection should combine theory, timing, data quality and rolling validation. Multicollinearity increases uncertainty; autocorrelation affects time-series inference and efficiency; heteroscedasticity makes conventional standard errors unreliable. Each problem requires a remedy suited to the underlying data process.

Causal interpretation demands stronger assumptions than prediction. Confounding, reverse causality, selection and post-treatment controls can invalidate a causal claim. Regression forecasting should therefore be paired with experiments or credible quasi-experimental designs when the managerial question concerns intervention effects. In practice, a sound forecasting system combines statistical diagnostics, operational knowledge, transparent governance, scenario analysis and continuous monitoring.

Key Terms

Causal effect: The change in an outcome attributable to a defined intervention under a specified counterfactual.

Conditional mean: The expected value of an outcome given specified predictors.

Dynamic regression: Regression that includes lagged outcomes or lagged explanatory variables.

Endogeneity: A condition in which a predictor is related to the model error, undermining causal interpretation.

Heteroscedasticity: Non-constant error variance across observations or predictor values.

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

Lag: A previous-period value of a variable.

Long-run effect: The eventual effect of a persistent driver change in a stable dynamic model.

Multicollinearity: Strong linear dependence or overlap among predictors.

Ordinary least squares: Estimation by minimizing the sum of squared residuals.

Partial effect: The modeled association of one predictor holding other included predictors constant.

Prediction interval: An interval representing uncertainty about a future individual outcome.

Regularization: Penalization used to control model complexity and improve generalization.

Residual: The difference between an observed and fitted outcome.

Rolling-origin validation: Time-respecting evaluation that repeatedly trains on earlier data and tests on later data.

Scenario forecast: A forecast conditional on specified future driver assumptions.

Short-run effect: The immediate or near-term response to a driver change.

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

Variable leakage: Use of information unavailable at the time a forecast would have been made.

Variance inflation factor: A diagnostic measuring how predictor correlation inflates coefficient variance.

Review Questions

Short Answer Questions

1. What distinguishes a predictive regression from a causal regression?
2. Why are m−1 indicator variables normally used for a categorical variable with m categories and an intercept?
3. What does a significant interaction coefficient imply?
4. State two consequences of autocorrelation in regression forecasting.
5. Why can a high R² fail to guarantee good future forecasts?

Long Answer Questions

1. Explain the assumptions underlying OLS estimation and distinguish assumptions needed for unbiased coefficients from those needed for conventional standard errors.
2. Discuss a complete variable-selection process for an operational regression forecast, including theory, leakage prevention and time-based validation.
3. Explain multicollinearity, its diagnosis and at least three possible responses.
4. Develop and interpret a dynamic regression model containing lagged demand and lagged promotional effects.
5. Compare predictive regression, randomized experimentation and quasi-experimental methods for estimating promotion impact.

Analytical / Application Questions

1. A retailer’s model includes price, discount percentage and promotion status. Explain why these variables may be collinear, how you would diagnose the problem and how you would define an interpretable promotion estimand.
2. A forecast model has low training error but shows long runs of positive residuals during economic downturns. Diagnose possible causes and propose a validation and model-revision plan.
3. An operations manager wants to claim that higher inventory causes higher sales because inventory and sales are positively related. Explain the likely endogeneity problem and design a more credible analysis.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes forecasting from causal inference?

A. Forecasting always requires randomized data.

B. Causal inference asks what would happen under a defined intervention, whereas forecasting focuses on future prediction.

C. Causal inference never uses regression.

D. Forecasting coefficients are automatically causal.

Correct Answer: B. Causal inference asks what would happen under a defined intervention, whereas forecasting focuses on future prediction.

Question 2

In a multiple regression, the coefficient on a numerical predictor is interpreted as the expected change in the outcome when that predictor increases by one unit, holding what constant?

A. The sample size

B. The residual variance only

C. The other included predictors

D. The forecast horizon only

Correct Answer: C. The other included predictors

Question 3

What is the principal purpose of rolling-origin validation?

A. To guarantee causal identification

B. To respect temporal ordering when evaluating forecasts

C. To eliminate multicollinearity exactly

D. To increase the in-sample R²

Correct Answer: B. To respect temporal ordering when evaluating forecasts

Question 4

If a model contains an intercept and a categorical variable with four categories, how many indicator variables are normally included?

A. 1

B. 2

C. 3

D. 4

Correct Answer: C. 3

Question 5

For the model y_t = α + βx_t + φy_t−1 + ε_t, what is the stable long-run effect of a permanent one-unit increase in x?

A. βφ

B. β/(1−φ)

C. β+φ

D. β/(1+φ)

Correct Answer: B. β/(1−φ)

Question 6

Which condition is associated with perfect multicollinearity?

A. Residuals have unequal variance

B. Predictors are exactly linearly dependent

C. The outcome is measured in units

D. The sample contains a seasonal pattern

Correct Answer: B. Predictors are exactly linearly dependent

Question 7

What is a common consequence of heteroscedasticity when ordinary standard errors are used?

A. The dependent variable becomes categorical

B. The usual standard errors may be unreliable

C. All OLS coefficients become undefined

D. Autocorrelation is necessarily removed

Correct Answer: B. The usual standard errors may be unreliable

Question 8

Why can historical price-demand regression fail to identify a causal price effect?

A. Price is always measured without error

B. Managers may change price in response to anticipated demand

C. Demand cannot be forecast using regression

D. A price variable cannot have units

Correct Answer: B. Managers may change price in response to anticipated demand

Question 9

In a model with an interaction term β₃xz, the marginal effect of x is:

A. β₁+β₃z

B. β₁+β₂x

C. β₃/(x+z)

D. β₁β₃z

Correct Answer: A. β₁+β₃z

Question 10

Which evaluation measure is expressed in the same units as the forecasted outcome?

A. MAE

B. MAPE expressed as a percentage

C. R²

D. VIF

Correct Answer: A. MAE