Causal and Explanatory Forecasting

Introduction and Context

Demand forecasting is often introduced as the extrapolation of historical demand patterns. That approach is useful when the future resembles the past, but it becomes unreliable when demand is shaped by changing prices, promotions, economic conditions, weather, competitor actions, or special events. Causal and explanatory forecasting addresses this problem by relating demand to variables that plausibly influence it.

The central idea is not simply to identify variables that move with demand. It is to develop a defensible representation of how demand responds to controllable and uncontrollable drivers. This distinction is important. A forecasting model may predict accurately without identifying a causal relationship. Conversely, a model that is causally meaningful may forecast poorly if important variables are unavailable or future values are uncertain.

In supply chain management, causal forecasting supports pricing, promotion planning, inventory positioning, capacity decisions, assortment design, workforce planning, and scenario analysis. It is especially valuable when managers ask questions such as:

• How much additional demand will a 5% price reduction generate?
• What is the likely effect of a promotional display?
• How will a heatwave affect beverage sales?
• What happens to orders if unemployment increases?
• How long after a marketing campaign should demand be expected to rise?

These are explanatory questions as well as forecasting questions. They require attention to model specification, data quality, timing, confounding, multicollinearity, and the limits of statistical inference.

Learning Objectives

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

1. Explain the purpose and logic of causal and explanatory forecasting.
2. Specify and interpret regression models for demand.
3. Incorporate price, promotion, elasticity, events, weather, leading indicators, and macroeconomic variables.
4. Distinguish prediction from causation and recognize confounding and reverse causality.
5. Diagnose multicollinearity, lag effects, omitted-variable bias, and endogeneity.
6. Evaluate models using both statistical and operational criteria.
7. Apply causal forecasting responsibly in supply chain decisions.

1. The Logic of Explanatory Forecasting

A demand observation can be represented as:

where Y_t is demand in period t, X_t contains explanatory variables that may be controllable or externally observed, Z_t contains additional influences such as seasonality or trend, and ε_t is unexplained variation. The function f(·) may be linear or nonlinear.

A basic multiple regression model is:

Here, β₀ is the intercept, β_j is the expected change in demand associated with a one-unit change in X_j, holding the other included variables constant, and ε_t is the error term. The units of each coefficient are units of demand divided by units of the corresponding explanatory variable.

The phrase “holding other variables constant” is essential. It means that a coefficient describes a conditional association within the model. It does not automatically prove that changing X_j will cause the predicted change in demand.

Explanatory forecasting differs from purely time-series forecasting in emphasis. A time-series model may use past demand, moving averages, exponential smoothing, or autoregressive terms. A causal model uses drivers such as price, advertising, temperature, income, or holidays. Hybrid models combine both approaches:

The lagged demand term can capture persistence, while the other variables represent demand drivers. However, including lagged demand can make interpretation more difficult because it may absorb effects of omitted variables and create dynamic rather than immediate coefficients.

2. Model Specification for Supply Chain Demand

2.1 The dependent variable

The dependent variable should be defined precisely. It may be unit sales, revenue, orders, shipment quantity, store-level demand, or lost-sales-adjusted demand. Sales are not always equal to demand. If inventory is unavailable, observed sales may be censored below true demand. A regression of sales on price may then produce misleading conclusions because stockouts often occur during high-demand periods.

The forecasting horizon also matters. A daily model for fresh food has different drivers and error structures from a monthly model for industrial equipment. Variables should be measured at compatible time and geographical levels.

2.2 Prices and promotions

Price is frequently a major demand driver. A simple linear price model is:

where Q_t is quantity demanded and P_t is price. Under ordinary downward-sloping demand, β₁ < 0. The linear form implies a constant absolute change in quantity for each price unit, which may be unrealistic over a broad price range.

A log-log model is often useful for elasticity:

In this specification, β₁ is the price elasticity of demand. It is approximately the percentage change in quantity associated with a 1% change in price, holding other variables constant. If β₁ = -1.4, a 1% price increase is associated with approximately a 1.4% decrease in quantity.

Elasticity is not necessarily constant. A model can include nonlinear terms, price bands, or interactions with customer segment, channel, or promotion status. For example:

The interaction term allows price sensitivity to differ during promotions.

Promotions can be represented using indicator variables. If Promo_t = 1 during a promotion and 0 otherwise, β₂ estimates the average difference associated with the promotion, conditional on the other variables. A promotion variable may need to distinguish discount depth, display, advertising, coupons, digital placement, and duration. Treating every promotion as identical can conceal important differences.

Promotion effects may include:

• baseline lift during the promotion;
• forward buying before an anticipated price increase;
• post-promotion dips caused by inventory already purchased;
• cannibalization of related products; and
• substitution from competing brands.

2.3 Events and calendar effects

Holidays, sporting events, school terms, pay cycles, product launches, and local festivals can be modeled using indicator variables or event-specific measures. A holiday dummy may capture an average shift, while the number of days before a holiday can represent a buildup pattern.

Calendar variables should be designed carefully. For example, a retailer may need separate variables for Christmas week, the week before Christmas, public holidays, and payday weekends. A single “holiday” variable may be too crude. Event variables should also avoid look-ahead bias: a forecast may use an event known in advance, but not an event classification created retrospectively.

2.4 Weather variables

Weather can be an important short-term driver for beverages, apparel, energy, travel, construction materials, and emergency supplies. Relevant measures include temperature, precipitation, snowfall, humidity, wind, and extreme-weather indicators.

The relationship may be nonlinear. Ice cream demand may rise with temperature up to a point, whereas heating demand may increase at low temperatures. A quadratic specification is:

The marginal effect of temperature is:

This derivative shows that the effect changes with the temperature level. The quadratic form should be used only when supported by data and subject-matter reasoning; otherwise it can generate implausible predictions outside the observed range.

Weather forecasts themselves are uncertain. A demand forecast based on a weather forecast should account for that uncertainty, especially at longer horizons. Scenario forecasts using warm, normal, and cool conditions may be more honest than a single point estimate.

2.5 Macroeconomic variables and leading indicators

Macroeconomic variables include income, unemployment, interest rates, inflation, consumer confidence, industrial production, housing starts, and exchange rates. Their relevance depends on the product and customer base. Industrial production may lead demand for components, while consumer confidence may be relevant for discretionary retail purchases.

A leading indicator is a variable whose movement tends to precede movement in the target demand series. A lagged explanatory model may be written as:

The lag must reflect the operational mechanism. If a change in housing starts affects appliance orders after six months, a one-month lag is not theoretically appropriate merely because it improves in-sample fit.

Distributed-lag models represent effects spread across several periods:

The total association over the included periods is β₁ + β₂ + β₃, assuming the coefficients are interpreted on a compatible scale. Distributed lags require sufficient data and care because adjacent lagged variables are often highly correlated.

3. Regression Estimation and Interpretation

Ordinary least squares estimates coefficients by minimizing the sum of squared residuals:

where Ŷ_t is the fitted value and n is the number of observations. The residual is e_t = Y_t - Ŷ_t.

The standard interpretation of an estimated coefficient β̂_j requires assumptions including a correctly specified conditional mean, no perfect multicollinearity, and an error term with conditional mean zero:

For conventional standard errors, additional assumptions concerning independence, constant variance, and an appropriate error distribution may be required. Time-series demand commonly violates independence through autocorrelation and may exhibit changing variance. Robust or time-series-specific standard errors, dynamic models, or validation procedures may therefore be necessary.

A model should be evaluated using out-of-sample performance, not only statistical significance or in-sample R². Common forecast measures include:

MAE is expressed in demand units and is easy to interpret. RMSE gives greater weight to large errors. Percentage measures such as MAPE can be problematic when actual demand is zero or close to zero. Forecast evaluation should use rolling-origin or time-based validation so that future information does not enter the training sample.

4. Prediction Is Not Causation

A predictive relationship identifies information useful for estimating an outcome. A causal relationship claims that an intervention or change in one variable would alter the outcome, all else relevant being appropriately controlled. These claims are not equivalent.

For example, umbrella sales may rise when rain rises. Rain plausibly causes umbrella demand. But ice cream sales and drowning incidents may both rise in summer. Their correlation does not mean ice cream causes drowning; temperature and outdoor activity are confounding influences.

4.1 Confounding

A confounder affects both the explanatory variable and demand. Suppose a retailer discounts a product when demand is expected to be weak. A simple regression may find that lower prices are associated with lower sales, even though the discount increased sales relative to what would otherwise have occurred. Expected demand is a confounder and the pricing decision is endogenous.

Confounding can be reduced by:

• including relevant observed variables;
• using randomized experiments;
• exploiting natural experiments or policy changes;
• using instrumental variables when valid instruments exist;
• applying fixed effects for stable differences across stores or products; and
• designing difference-in-differences analyses with credible comparison groups.

No statistical technique automatically solves confounding. The identifying assumptions must be defended using institutional knowledge and research design.

4.2 Reverse causality and simultaneity

Demand may influence price rather than merely respond to price. A capacity-constrained supplier may raise prices when demand is strong. In that case, price and demand are jointly determined. Similar simultaneity can arise when advertising budgets increase in response to weak sales or when inventory allocation responds to expected demand.

4.3 Omitted-variable bias

If a relevant variable is omitted and correlated with an included variable, the estimated coefficient can be biased. For example, if both price and advertising change during promotions but advertising is omitted, the price coefficient may partly capture advertising effects.

4.4 Selection and measurement problems

Observed demand may reflect only customers who saw an offer, stores selected for a test, or products that remained in stock. Measurement error in prices, promotion dates, weather, or economic indicators can weaken estimates and sometimes create bias. Data governance should therefore document definitions, sources, revisions, missing values, and data transformations.

5. Multicollinearity and Model Stability

Multicollinearity occurs when explanatory variables are strongly related. Price, discount percentage, promotion display, advertising, and coupon distribution may occur together. Seasonal indicators may also be correlated with weather and holidays.

Multicollinearity does not necessarily bias ordinary least squares coefficients when the exogeneity assumption holds, but it increases coefficient variance and makes individual effects difficult to separate. Symptoms include unstable coefficients across samples, large standard errors, unexpected signs, and sensitivity to adding or removing variables.

Useful responses include:

• combining variables when they represent one underlying construct;
• selecting variables using operational and theoretical reasoning;
• collecting data in periods with independent variation;
• using regularization for prediction; and
• reporting joint effects or scenario effects rather than overinterpreting individual coefficients.

Ridge regression minimizes squared error plus a penalty on coefficient size:

where λ ≥ 0 controls the penalty. Ridge can improve predictive stability but does not make coefficients causal. Lasso uses an absolute-value penalty and may set some coefficients exactly to zero. Regularization is a prediction tool, not a substitute for identification.

6. Lag Effects, Dynamics, and Seasonality

Demand responses often occur with delay. A marketing campaign may create immediate awareness, a later purchase, and a post-campaign decline. A price change may influence customer search and replenishment at different speeds.

Lagged effects should be chosen from process knowledge, customer behavior, and empirical testing. A model with too few lags misses effects; a model with too many lags can overfit and create multicollinearity. Constraints such as gradually declining lag weights can improve interpretability.

Seasonality should be represented explicitly when it is not already captured by other variables. Monthly or weekly indicators, Fourier terms, or seasonal indices may be used. Care is required because a holiday variable and a seasonal dummy can explain the same variation. Identifiability is a modeling issue: a model must contain enough independent information to distinguish effects.

Autocorrelated residuals indicate that the model has left systematic time dependence unexplained. Residual diagnostics should examine residual plots, autocorrelation functions, forecast errors by segment, and errors around promotions or events. A model that performs well on average may still fail systematically during the decisions that matter most.

7. A Practical Modeling Workflow

A disciplined workflow is more important than selecting a sophisticated algorithm.

1. Define the decision and horizon. Specify whether the forecast supports replenishment, pricing, capacity, or budgeting.
2. Define the demand measure. Separate true demand, sales, orders, and shipments; identify stockouts and cancellations.
3. Construct a causal map. List plausible drivers, mechanisms, timing, and variables affected by management decisions.
4. Assemble and audit data. Align time zones, units, product hierarchies, prices, promotion dates, weather stations, and economic releases.
5. Explore relationships. Use plots, summaries, segment comparisons, and lag analysis; do not treat correlation as proof.
6. Specify candidate models. Include baseline time-series models and explanatory alternatives.
7. Estimate and diagnose. Examine residuals, coefficient signs, collinearity, autocorrelation, heteroscedasticity, and influential observations.
8. Validate temporally. Use rolling or expanding windows and compare performance against operational benchmarks.
9. Conduct scenarios. Test realistic price, promotion, weather, and economic assumptions.
10. Deploy with monitoring. Track forecast error, drift, coefficient stability, stockouts, and changes in customer behavior.

The best model is not necessarily the one with the lowest historical error. It must also be available at the forecast origin, understandable to decision makers, robust under plausible changes, and consistent with operational constraints.

8. Illustrative Worked Example: Price Elasticity

Suppose a retailer estimates the following log-log relationship for a product:

Assume the model is estimated from adequate data, promotion status is correctly measured, and the coefficient is interpreted conditionally. The price elasticity is −1.2. Therefore, a 1% price increase is associated with approximately a 1.2% decrease in quantity, holding promotion status constant.

If price falls by 5%, the approximate percentage change in quantity is:

This is an approximation based on percentage changes. It does not establish that a price cut will cause exactly a 6% increase, because the estimate may be confounded by promotion selection, competitor prices, stock availability, or changing customer composition. It also does not determine profit. If unit margin falls substantially, revenue or contribution margin may decline despite higher quantity.

The promotion coefficient is interpreted in log-quantity units. The exact multiplicative quantity factor associated with the promotion, under this model, is:

Thus, conditional on the model and holding price constant, promotion is associated with approximately 28.4% higher quantity. This interpretation assumes the promotion coefficient applies uniformly and that the log transformation is appropriate.

9. Illustrative Case: FreshBrew Beverages

The following is a fictional illustrative case.

Background

FreshBrew operates 420 convenience stores and sells cold coffee, bottled water, and hot beverages. Management currently uses last year’s weekly sales adjusted for seasonality. Forecast error increases during heatwaves, major sports events, and discount campaigns. Stockouts occur frequently for cold coffee during unusually warm weeks.

Problem

The commercial team wants to know whether a planned 10% price reduction will increase contribution margin. Supply chain managers also need forecasts four weeks ahead, when weather forecasts and event schedules are available but realized weather is not.

Analysis

The analytics team defines weekly store-level unit demand and flags stockout weeks. It assembles price, discount depth, display placement, digital advertising, temperature, precipitation, sports-event attendance, holiday indicators, and local unemployment. Store and product fixed effects are included to account for stable differences in location and product preference. A distributed-lag temperature term is tested because weather affects both immediate consumption and replenishment visits.

The team discovers that discounts are more common in stores where managers anticipate weak demand. A simple regression therefore understates the effect of discounts. To address this, the team conducts a staggered pilot in which eligible stores are randomly assigned different discount schedules, while maintaining inventory availability. The pilot also records competitor prices and display compliance.

Approach and solution

The operational forecasting system combines:

• a seasonal baseline;
• price and promotion variables;
• nonlinear temperature effects;
• event indicators;
• lagged weather variables;
• store and product effects; and
• forecast scenarios for warm, normal, and cool weather.

The pricing decision uses the experimentally estimated promotion lift rather than the observational coefficient. Forecasts are reconciled across store, region, and network levels so that local plans remain consistent with procurement and distribution capacity.

Outcome

The pilot shows that the discount increases unit demand, but the lift is smaller when displays are absent and when a competitor runs a deeper promotion. During warm weeks, demand rises sharply, but the incremental demand is concentrated in high-traffic stores. The company therefore applies the promotion selectively, increases safety stock for the highest-risk stores, and uses weather scenarios rather than one deterministic forecast.

Learning points

1. Sales data can understate demand when stockouts occur.
2. Observational price coefficients may be biased by managerial targeting.
3. Promotion effects depend on execution and competitive context.
4. Weather should be modeled with nonlinearities and forecast uncertainty.
5. Causal evidence is most valuable when decisions involve intervention.
6. Forecasting and causal analysis can share data but require different claims and validation standards.

10. Technology, AI, and Governance

Modern forecasting platforms can combine regression, gradient-boosting models, probabilistic forecasting, and machine learning. These tools may improve prediction by detecting nonlinearities and interactions. However, predictive importance is not causal importance. A model may rely on a proxy for a decision variable, encode historical managerial bias, or fail when policy changes alter the data-generating process.

Good governance requires versioned data, documented feature definitions, reproducible transformations, access controls, model approval, and monitoring for drift. Forecast users should know whether a variable is observed at the forecast origin or becomes available only later. Explanations should distinguish model attribution—what helped the model predict—from an estimated intervention effect.

Ethically, pricing and allocation models should be reviewed for discriminatory outcomes, unfair treatment of customer groups, and exploitative use of sensitive information. Regulatory obligations vary by jurisdiction, but transparency, purpose limitation, data minimization, and auditability are broadly responsible practices. Managers should avoid presenting uncertain causal estimates as facts.

11. Challenges and Limitations

Causal forecasting faces several persistent limitations:

• Structural change: Relationships estimated from historical data may change after new competitors, technologies, regulations, or consumer habits appear.
• Extrapolation: A model may be reliable within the observed range but unsafe for extreme prices or unprecedented weather.
• Data revisions: Macroeconomic indicators may be revised after the forecast was made, creating accidental look-ahead bias in evaluation.
• Granularity: Store-level data may be noisy, while aggregate data can hide substitution and heterogeneous responses.
• Interference: One store’s promotion can affect nearby stores, violating the assumption that units are independent.
• Capacity constraints: Predicted demand may exceed available inventory, transport, labor, or production capacity.
• Endogenous decisions: Prices, advertising, assortment, and inventory are chosen by managers in response to information that may not be recorded.
• Uncertainty: Point forecasts conceal uncertainty in coefficients, future drivers, and model form.

Best practice is to report prediction intervals or scenarios, test sensitivity to model choices, compare with simpler benchmarks, and involve domain experts in interpreting unusual results.

Chapter Summary

Causal and explanatory forecasting relates demand to variables such as price, promotion, events, weather, leading indicators, and macroeconomic conditions. Multiple regression provides a flexible framework, but its coefficients are conditional associations unless stronger identification assumptions or experimental designs support causal interpretation. Log-log models provide elasticity estimates; dummy variables represent events and promotions; nonlinear and distributed-lag models represent changing and delayed effects.

Reliable practice requires accurate demand measurement, attention to stockouts, temporal alignment, validation without look-ahead bias, and diagnosis of multicollinearity, autocorrelation, heteroscedasticity, confounding, omitted variables, and endogeneity. Forecasting systems should combine statistical performance with operational usefulness, interpretability, robustness, and governance. Prediction and causation are related but distinct: a variable can improve prediction without representing a valid intervention effect.

Key Terms

Causal forecasting: Forecasting that uses variables believed to influence demand.

Explanatory variable: A variable used to explain or predict the dependent variable.

Multiple regression: A regression model with more than one explanatory variable.

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

Price elasticity: Percentage change in quantity associated with a percentage change in price.

Promotion lift: Incremental demand associated with a promotion relative to a suitable baseline.

Leading indicator: A variable whose movement tends to precede movement in the forecast target.

Lag effect: An influence that occurs after a delay.

Distributed lag: A model allowing an explanatory variable to affect several future periods.

Confounder: A variable related to both an explanatory variable and the outcome.

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

Reverse causality: A situation in which the outcome also influences the explanatory variable.

Multicollinearity: Strong correlation among explanatory variables.

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

Fixed effect: A model component controlling for stable differences across entities or periods.

Out-of-sample validation: Evaluation using observations not used to estimate the model.

MAE: Mean absolute error, the average absolute forecast error.

RMSE: Root mean squared error, an error measure that penalizes large errors more heavily.

Look-ahead bias: Using information that would not have been available when the forecast was made.

Intervention effect: The change in an outcome attributable to a specified action or treatment.

Review Questions

Short Answer Questions

1. What is the difference between a time-series forecast and a causal forecast?
2. How is a coefficient interpreted in a multiple regression model?
3. What does a price elasticity of −1.5 mean?
4. Why can stockouts make observed sales an unreliable measure of demand?
5. Define multicollinearity and state one consequence.

Long Answer Questions

1. Explain how price, promotion, and competitor actions can be incorporated into a demand model. Discuss the risks of interpreting their coefficients causally.
2. Describe how weather and event variables should be modeled for a retail supply chain.
3. Explain confounding, reverse causality, and omitted-variable bias using supply chain examples.
4. Compare MAE, RMSE, and percentage-based forecast measures, including their limitations.
5. Develop a complete workflow for building and governing an explanatory demand forecasting system.

Analytical / Application Questions

1. A retailer observes that discounted products sell less than full-price products. Explain at least three reasons this relationship might occur and design a study to estimate the true discount effect.
2. A monthly model includes price, promotion, advertising expenditure, holiday indicators, temperature, and three lags of temperature. Identify possible multicollinearity and lag-interpretation problems, and propose diagnostic or modeling responses.
3. A forecast is needed four weeks ahead, but future temperature is uncertain. Design a scenario-based forecasting and inventory decision process that distinguishes uncertainty in weather from uncertainty in model coefficients.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from causation?

A. Prediction always requires a randomized experiment.

B. A predictive relationship necessarily identifies an intervention effect.

C. Prediction concerns forecast usefulness, whereas causation concerns the effect of changing a variable.

D. Causation can be established whenever the model has a high R².

Correct Answer: C. Prediction concerns forecast usefulness, whereas causation concerns the effect of changing a variable.

Question 2

In a log-log demand model, the coefficient on log price is −1.2. What is its usual interpretation?

A. A one-unit price increase reduces demand by 1.2 units.

B. A 1% price increase is associated with approximately a 1.2% demand decrease.

C. A 1.2% price increase reduces demand by 1%.

D. Price explains 1.2% of all demand variation.

Correct Answer: B. A 1% price increase is associated with approximately a 1.2% demand decrease.

Question 3

Which situation is most likely to create confounding in estimating a discount effect?

A. Discounts are assigned randomly across stores.

B. Discounts are offered only when managers expect weak demand.

C. Prices are recorded in the same currency.

D. Demand is measured in units rather than revenue.

Correct Answer: B. Discounts are offered only when managers expect weak demand.

Question 4

What is a primary consequence of multicollinearity?

A. It always biases coefficients upward.

B. It guarantees that forecasts are unbiased.

C. It increases uncertainty and makes individual effects difficult to separate.

D. It eliminates the need for out-of-sample validation.

Correct Answer: C. It increases uncertainty and makes individual effects difficult to separate.

Question 5

If a weather variable has coefficients on both current and lagged values, the model is an example of a:

A. Distributed-lag model

B. Purely seasonal index

C. Randomized block design

D. Naive benchmark model

Correct Answer: A. Distributed-lag model

Question 6

Which forecast measure is expressed in the original units of demand?

A. MAE

B. MAPE only

C. Price elasticity

D. A standardized coefficient only

Correct Answer: A. MAE

Question 7

Why can MAPE be problematic for products with zero demand in some periods?

A. It cannot be calculated when the actual value is zero.

B. It always overweights large errors.

C. It is measured in price units.

D. It requires a causal experiment.

Correct Answer: A. It cannot be calculated when the actual value is zero.

Question 8

In a quadratic temperature model, the marginal effect of temperature is:

A. β₁

B. β₂Temp_t²

C. β₁ + 2β₂Temp_t

D. β₀ + β₁ + β₂

Correct Answer: C. β₁ + 2β₂Temp_t

Question 9

Which practice prevents look-ahead bias in forecast evaluation?

A. Using revised economic data unavailable at the forecast date

B. Randomly mixing past and future observations before splitting the data

C. Using time-based or rolling-origin validation

D. Selecting variables using the complete future dataset

Correct Answer: C. Using time-based or rolling-origin validation

Question 10

A model has excellent historical fit but fails after a major competitor enters the market. This most directly illustrates:

A. Structural change

B. Unit consistency

C. Perfect randomization

D. Constant elasticity

Correct Answer: A. Structural change