5. Causal, Econometric and Machine-Learning Forecasting

Introduction and context

Demand forecasting in logistics is not only an exercise in extending historical sales patterns. Managers often need to know why demand changes, how it will respond to prices or promotions, whether an economic indicator provides advance warning, and which analytical method is sufficiently accurate and explainable for operational decisions. These questions move forecasting from descriptive extrapolation toward causal, econometric and machine-learning approaches.

A time-series model may identify that demand is usually higher in December, but it does not by itself establish whether the increase is caused by seasonal consumption, promotional activity, changes in income, or a competitor’s stock-out. Causal and econometric models attempt to relate demand to explanatory variables. Machine-learning models can discover complex, nonlinear relationships among many variables, often improving predictive performance. However, predictive accuracy does not automatically imply causal validity. A model that forecasts well may still provide poor guidance when managers change prices, advertising or service levels.

For supply-chain professionals, the central task is therefore not to select the most sophisticated algorithm. It is to match the method to the decision, the available data, the forecasting horizon, the cost of error and the required level of explanation.

Learning objectives

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

1. Distinguish descriptive, predictive and causal forecasting questions.

2. Formulate and interpret regression and econometric demand models.

3. Use leading indicators and price–promotion variables appropriately.

4. Explain endogeneity, multicollinearity, autocorrelation and omitted-variable bias.

5. Compare regression with decision trees, ensembles and neural networks.

6. Design time-aware validation procedures and manage overfitting.

7. Evaluate forecasts using both statistical and business measures.

8. Apply explainability, governance and ethical controls to advanced forecasting.

9. Select an appropriate modelling approach for logistics decisions.

From extrapolation to explanation

Forecasting methods can be placed on a continuum. Univariate methods use only the history of the variable being forecast. Causal and econometric models introduce external drivers such as price, income, weather, promotions and competitor activity. Machine-learning methods use algorithms capable of representing nonlinearities and interactions among large sets of predictors.

Three questions should be separated. A descriptive question asks what pattern has occurred: for example, whether demand rises before a festival. A predictive question asks what is likely to happen next week. A causal question asks what would happen if the price were reduced by 5 per cent while other relevant conditions were held constant. The third question is the most demanding because it concerns a counterfactual decision rather than a passive prediction.

This distinction is essential in logistics. A model may predict high demand during a promotion because the promotion is scheduled precisely when demand is already expected to be high. Treating the observed association as the effect of the promotion can lead to over-ordering and incorrect evaluation of marketing expenditure.

Regression as a managerial forecasting language

A basic multiple regression model can be written as:

D_t = β₀ + β₁P_t + β₂A_t + β₃S_t + β₄I_t + ε_t

where D_t is demand in period t, P_t is price, A_t represents advertising or promotion, S_t represents seasonal or calendar variables, I_t is an economic or market indicator, β parameters measure relationships, and ε_t captures unexplained influences.

The intercept β₀ is the expected value of demand when all numerical predictors equal zero, although that condition may not be economically meaningful. A coefficient such as β₁ estimates the expected change in demand associated with a one-unit change in price, conditional on the other variables in the model. If price is measured in currency units and β₁ is negative, the model is consistent with a downward-sloping demand relationship.

Logarithmic specifications are frequently more useful for commercial interpretation. In a log–log model,

ln(D_t) = β₀ + β₁ln(P_t) + β₂ln(I_t) + ε_t,

β₁ is an estimated price elasticity: a 1 per cent increase in price is associated, approximately, with a β₁ per cent change in demand. A log-level model, in which ln(D) is explained by price, interprets a coefficient approximately as a proportional demand change for a unit price change. Transformations should be selected for interpretability and statistical suitability, not simply because they are conventional.

Categorical variables such as region, product family, weekday or promotion type can be represented through dummy variables. Interaction terms allow the effect of one variable to depend on another. For instance, the effect of a price discount may be larger during a holiday period:

D_t = β₀ + β₁Discount_t + β₂Holiday_t + β₃(Discount_t × Holiday_t) + ε_t.

Here β₃ estimates the additional discount effect during the holiday, subject to the assumptions of the model.

Regression coefficients must not be interpreted mechanically. A statistically significant coefficient may be commercially negligible, while a practically important relationship may be estimated imprecisely because the data are limited or highly variable. Confidence intervals, effect sizes, residual diagnostics and business plausibility should accompany coefficient interpretation.

Econometric issues in demand data

Ordinary least squares is useful, but demand data often violate its simplest assumptions. The following problems are particularly important in supply-chain forecasting.

Omitted-variable bias occurs when a relevant driver is excluded and is correlated with an included predictor. If a promotion is more likely to occur when managers expect strong demand, a model omitting baseline demand conditions may overstate the promotion effect.

Endogeneity arises when a predictor is related to the error term. Price is often endogenous: firms may lower prices when demand is weak, so the observed price–demand relationship reflects both consumer response and managerial reaction. Instrumental variables, natural experiments, randomized tests or carefully designed panel models may be required when causal interpretation matters.

Multicollinearity occurs when predictors move together, such as advertising expenditure and promotion intensity. It can make individual coefficients unstable even when the model predicts reasonably well. Correlation matrices, variance-inflation diagnostics, domain knowledge and regularization can help, but removing a variable should not be done solely because it is correlated with another if both represent distinct managerial decisions.

Autocorrelation means that errors are related across time. A demand shock in one week may affect replenishment and sales in the next week. Ignoring autocorrelation can produce misleading standard errors and inefficient estimates. Lagged demand, autoregressive errors, dynamic regression or appropriate time-series methods may be used.

Non-stationarity occurs when the statistical properties of a series change over time, often because of growth, inflation, market expansion or structural change. Regressing unrelated trending variables on one another can create a spurious relationship. Differencing, growth rates, trend terms, cointegration methods or structural-break analysis may be appropriate.

Heteroscedasticity occurs when error variance changes with the level of demand. Forecast uncertainty may be much greater for a high-volume item than for a low-volume item. Log transformations, weighted estimation or robust standard errors can improve analysis.

Model diagnostics should include residual plots, out-of-sample errors, stability checks and assessment of whether the variables would be available at the time the forecast is issued. A model containing final sales, post-period returns or revised economic data may appear accurate in testing but be unusable in practice because of information leakage.

Leading indicators and dynamic relationships

A leading indicator changes before the target variable and may provide advance information. Examples include housing starts for building materials, vehicle registrations for spare parts, construction permits for industrial equipment, search activity for selected consumer products, and customer quotations for project-based logistics.

The usefulness of a leading indicator depends on more than correlation. It should have a defensible operational relationship, a stable timing advantage, acceptable data quality and availability before the forecast cut-off. A contemporaneous indicator may explain demand without improving advance warning.

Lagged regression can represent timing explicitly:

D_t = β₀ + β₁I_{t-1} + β₂D_{t-1} + β₃S_t + ε_t.

Distributed-lag models include several lags of an indicator and estimate whether its effect is immediate, delayed or persistent. Distributed lags are valuable when a market signal influences orders gradually. Excessive lags, however, increase dimensionality and can create unstable coefficients.

Cross-correlation analysis can suggest candidate lead–lag relationships, but it does not prove causality. Analysts should examine operational mechanisms and test relationships on periods not used to select the lags. When many indicators and lags are searched, false discoveries become likely.

Price, promotion and cannibalisation models

Price and promotion models are central to commercial forecasting. A useful model may distinguish regular price, discount depth, display placement, advertising, coupon availability, promotion duration and competitor price. A simple elasticity model is:

ln(D_t) = β₀ + β₁ln(P_t) + β₂Promotion_t + β₃ln(CompetitorPrice_t) + β₄Season_t + ε_t.

The coefficient on competitor price may indicate substitution, while a promotion variable estimates incremental demand relative to a defined baseline. Baseline definition is crucial: promotional uplift should be measured against the demand that would have occurred without the promotion, not against the previous week alone.

Promotion effects can include forward buying, post-promotion dips, stockpiling and channel substitution. A discount may increase a product’s sales while reducing sales of another product in the same portfolio. This is cannibalisation. Models that forecast each item independently can therefore produce an unrealistic aggregate plan.

Price response may also be nonlinear. A small discount may have little effect until a psychological threshold is crossed; a very deep discount may generate stockpiling rather than sustainable consumption. Splines, piecewise terms, interaction variables or machine-learning models can represent these effects, but greater flexibility must be controlled through validation and business constraints.

Machine learning for demand forecasting

Machine learning differs less in its objective than in its capacity to represent relationships. It can use lagged demand, rolling statistics, calendar features, price, promotion, inventory position, weather, web traffic and operational variables. The target may be point demand, a quantile, a classification such as stock-out risk, or a complete predictive distribution.

Decision trees recursively split the data into regions, such as high or low price and promotional or non-promotional periods. They are intuitive and represent interactions, but individual trees can be unstable and prone to overfitting.

Random forests average many trees trained with data and feature variation. They usually reduce variance and can model nonlinear relationships without requiring a particular functional form. They may be less transparent than a single tree and can behave poorly when extrapolating beyond the range of the training data.

Gradient-boosted trees build trees sequentially, each concentrating on residual errors from earlier trees. They often perform strongly on structured tabular data. Learning rate, tree depth, number of estimators, subsampling and regularization require careful tuning. Excessive tuning on one validation set can itself overfit.

Neural networks learn layered transformations of inputs. Feed-forward networks can model complex relationships, while recurrent networks and temporal architectures can represent sequential information. Neural networks require adequate data, careful scaling, regularization and substantial monitoring. In many supply-chain applications with moderate-sized tabular data, a well-engineered boosted-tree model may be more effective and easier to govern than a deep architecture.

Machine learning is not automatically causal. Feature importance indicates predictive contribution, not the effect of an intervention. If a model uses a variable that is a consequence of a promotion, it may forecast accurately but cannot estimate the promotion’s incremental impact. Causal machine learning, uplift modelling and treatment-effect estimation address such questions, but they still require suitable identification assumptions and data design.

Validation, overfitting and forecast evaluation

Random train–test splitting is often inappropriate for time-dependent demand because it allows future information to influence the past. A better design uses rolling-origin evaluation. The model is trained on an initial period, tested on the following period, then the training window is advanced and the process is repeated. This simulates repeated operational forecasting.

Validation must reproduce the real information set. Features should be calculated as they would have been known at forecast issuance, and promotion plans should reflect whether they were actually committed or merely proposed. Hyperparameter tuning should use a separate validation structure, with a final untouched test period used once for confirmation.

Common measures include mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), weighted absolute percentage error and bias. MAPE is unstable when actual demand is zero or near zero. RMSE penalizes large errors more strongly than MAE. In logistics, a cost-weighted measure may be more useful because under-forecasting a critical spare part can be more damaging than over-forecasting a low-value item.

Forecast evaluation should also consider service level, inventory cost, expediting, waste, capacity utilization and revenue. A model with a slightly higher statistical error may produce a better supply-chain outcome if it reduces costly shortages without excessive inventory. Bias should be evaluated by segment, location and product life-cycle stage rather than only in aggregate.

Overfitting occurs when a model learns noise or idiosyncrasies of the training data. Warning signs include excellent training accuracy, deteriorating validation performance, unstable feature importance and implausible response curves. Controls include simpler feature sets, regularization, early stopping, cross-validation appropriate to time, pruning, dropout for neural networks, and explicit business constraints.

Explainability and model governance

Explainability has at least two levels. Global explanation asks how the model behaves overall; local explanation asks why it produced a particular forecast. Partial-dependence plots, accumulated-local-effects plots, permutation importance and SHAP-style contribution analysis can assist interpretation. These tools should be treated as diagnostic aids rather than unquestionable causal evidence.

Governance requires a documented purpose, data lineage, variable definitions, forecast cut-off, model version, validation results, approval authority and monitoring plan. Important controls include detection of data leakage, review of unusual predictions, monitoring for drift, challenger models and periodic back-testing. Human override should be recorded with a reason and later evaluated; otherwise overrides can conceal model weakness or introduce unmeasured bias.

Ethically, organizations should avoid using protected or intrusive personal information without legitimate purpose and appropriate consent. Forecasts affecting employee allocation, customer access or essential goods should be reviewed for unfair outcomes. Commercial confidentiality, cybersecurity and compliance with data-protection requirements also matter when external data and automated systems are used.

An illustrative case: forecasting promotional demand for a regional grocery distributor

Background. A fictional regional grocery distributor supplies 420 stores. It experiences frequent shortages of snack products during promotional weeks and excess stock immediately afterward. The planning team currently uses last year’s weekly sales adjusted by a planner’s judgement.

Situation. The commercial department wants to estimate the incremental effect of discount depth, display position and advertising. The supply team also wants a four-week forecast. Historical data contain weekly sales, regular and promotional price, store, product, display, advertising, inventory availability and calendar variables. Several weeks contain stock-outs, so recorded sales may be lower than true demand.

Analysis. Analysts first separate periods with severe stock-outs and mark them as censored observations rather than treating low sales as normal demand. They create a baseline regression with product, store and seasonal effects, then add discount depth, display and advertising. A promotion indicator is insufficient because a 5 per cent and a 30 per cent discount are unlikely to have equal effects. Product-level models reveal strong substitution between two similar snack flavours, while aggregate portfolio demand is more stable.

The team finds that promotion weeks are more common before holidays, creating confounding between promotion and seasonality. It therefore includes holiday interactions and tests the model using rolling-origin validation. A boosted-tree challenger is trained with the same information cut-off and lagged features. The tree model produces lower MAE, especially for unusual combinations of price and display, but its predicted response to extreme discounts is implausible because such discounts were rare in the historical data.

Approach. The distributor uses the regression model for price and promotion scenario analysis because its coefficients and interactions can be reviewed by commercial managers. The boosted-tree model supports short-term baseline forecasting, while a reconciliation step ensures that product forecasts are consistent with the portfolio forecast. For future promotions, the organization introduces controlled tests in selected stores, enabling better estimation of incremental effects. Prediction intervals are used to set safety stock rather than relying only on a point forecast.

Outcome. The combined approach improves forecast discipline without assuming that the machine-learning model has discovered causal promotion effects. Planners receive a forecast, an explanation of major drivers, an uncertainty range and a warning when the proposed promotion lies outside historical experience.

Learning points. First, observed sales are not always demand when inventory is unavailable. Second, causal promotion analysis requires attention to how promotions are assigned. Third, statistical accuracy and managerial interpretability serve different purposes. Finally, model deployment should include operational constraints, uncertainty and monitoring rather than a one-time algorithm selection.

Selecting an approach in practice

Regression is often preferred when the organization needs interpretable price, promotion or economic effects and has a moderate number of well-defined variables. Econometric designs are appropriate when decisions depend on credible counterfactuals, policy evaluation or market response. Tree ensembles are attractive for nonlinear tabular prediction with many interactions. Neural networks become more compelling when data are large, high-dimensional or sequential, such as transactions, text, images or dense sensor streams.

A practical architecture may be an ensemble rather than a single method: a causal model for intervention scenarios, a machine-learning model for residual patterns, and a judgmental or rules-based layer for known events. Combining forecasts can reduce dependence on one model, but combination should be validated and governed. No model should be accepted solely because it is technically advanced.

Common mistakes and good practice

Frequent errors include using future information, interpreting correlation as causation, ignoring stock-outs, evaluating only average accuracy, tuning on the test set, creating too many lagged variables, extrapolating beyond observed prices, and allowing unrecorded overrides. Another mistake is to optimize forecast accuracy while disregarding the inventory policy that converts forecasts into orders.

Good practice begins with a clearly defined decision and forecast horizon. It continues with a documented information set, a defensible feature-generation process, time-aware validation, segment-level error analysis and comparison against simple benchmarks. Models should be stress-tested under plausible price, promotion, supply and market scenarios. Deployment should include drift monitoring, retraining criteria, exception handling and accountable human review.

Current developments

Forecasting systems increasingly combine probabilistic prediction with optimization, allowing uncertainty to flow into inventory and capacity decisions. Automated machine-learning tools can search models and features, but automation does not remove the need for leakage controls or domain review. Causal representation learning, uplift modelling, hierarchical machine learning, probabilistic neural networks and foundation models for time series are active areas of development. Their value will depend on reliable data, transparent evaluation and the ability to operate within supply-chain constraints.

Chapter Summary

Causal, econometric and machine-learning forecasting extend demand planning beyond historical extrapolation. Regression models express relationships among demand, price, promotion, seasonality and external indicators, while econometric analysis draws attention to endogeneity, omitted variables, autocorrelation, non-stationarity and identification. Leading indicators are useful when they provide a stable and operationally meaningful timing advantage. Price–promotion models must distinguish baseline demand, incremental uplift, cannibalisation and post-promotion effects.

Decision trees and ensembles capture nonlinearities and interactions; neural networks can represent complex sequential or high-dimensional patterns. Their predictive strength does not make them causal, and their use requires time-aware validation, leakage prevention, regularization and explainability. The appropriate method depends on the decision, data, cost of error and governance requirements. Strong professional practice combines statistical evidence, operational knowledge, uncertainty estimates and ongoing monitoring.

Key Terms

Causal forecasting: Forecasting that estimates the effect of an intervention or explanatory driver on a target.

Econometric model: A statistical model grounded in economic theory and designed to estimate relationships among variables.

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

Elasticity: The percentage change in one variable associated with a percentage change in another.

Leading indicator: A variable whose movement precedes movement in the forecast target.

Distributed lag: A model representing the effects of a predictor across several time lags.

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

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

Autocorrelation: Dependence between errors or observations across time.

Non-stationarity: Change in a series’ statistical properties over time.

Heteroscedasticity: Non-constant variance of model errors.

Cannibalisation: Loss of sales of one product caused by increased sales of another related product.

Decision tree: A recursive rule-based model that partitions observations into regions.

Random forest: An ensemble averaging many varied decision trees.

Gradient boosting: A sequential ensemble method that fits new learners to prior errors.

Overfitting: Learning noise or sample-specific patterns that do not generalize.

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

Rolling-origin evaluation: Repeated time-ordered testing that simulates operational forecasting.

Explainability: The capacity to communicate how a model generates or supports a prediction.

Forecast bias: Systematic tendency to over-forecast or under-forecast.

Review Questions

Short Answer Questions

1. Distinguish predictive forecasting from causal forecasting.

2. What is the managerial interpretation of a price elasticity?

3. Why can a stock-out make recorded sales an unreliable measure of demand?

4. Define a leading indicator and state two criteria for judging its usefulness.

5. What is data leakage in time-dependent forecasting?

Long Answer Questions

1. Explain how omitted variables, endogeneity and multicollinearity can affect a price–promotion model.

2. Discuss the construction and interpretation of a multiple regression model for logistics demand.

3. Compare decision trees, random forests, gradient boosting and neural networks for supply-chain forecasting.

4. Explain rolling-origin validation and why random train–test splitting may be inappropriate.

5. Discuss how forecast accuracy, explainability and governance should be balanced in model selection.

Analytical / Application Questions

1. A retailer lowers prices whenever it expects weak demand. Explain why a simple regression of sales on price may give a misleading price effect, and propose two ways to improve identification.

2. A machine-learning model has excellent training accuracy but poor performance during unusual promotions. Diagnose the likely problem and recommend a validation and modelling strategy.

3. A distributor must choose between an interpretable elasticity model and a more accurate boosted-tree model for a proposed price change. Design a decision framework that uses both models responsibly.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a causal forecast from a predictive forecast?

A. A causal forecast uses only historical demand values.

B. A causal forecast estimates what would happen under an intervention or change in a driver.

C. A causal forecast always has lower forecast error.

D. A causal forecast does not require explanatory variables.

Correct Answer: B. A causal forecast estimates what would happen under an intervention or change in a driver.

Question 2

In a log–log demand model, a price coefficient of -1.4 is most appropriately interpreted as:

A. A one-unit price increase reduces demand by 1.4 units.

B. A 1.4 per cent increase in price causes demand to increase by 1 per cent.

C. A 1 per cent increase in price is associated with approximately a 1.4 per cent decrease in demand.

D. Price explains 1.4 per cent of all demand variation.

Correct Answer: C. A 1 per cent increase in price is associated with approximately a 1.4 per cent decrease in demand.

Question 3

A firm cuts prices mainly when it expects demand to be weak. This creates a major risk of:

A. Endogeneity

B. Heteroscedasticity only

C. Seasonal differencing

D. Underfitting due to regularization

Correct Answer: A. Endogeneity

Question 4

Which feature makes a variable a useful leading indicator?

A. It is measured after the forecast period ends.

B. It is correlated with demand only in the training sample.

C. It provides stable information before the target changes.

D. It is always more accurate than the demand series itself.

Correct Answer: C. It provides stable information before the target changes.

Question 5

Why is rolling-origin evaluation generally preferable to random splitting for weekly demand data?

A. It eliminates all forecast error.

B. It preserves the temporal order and simulates future forecasting.

C. It guarantees causal estimates.

D. It automatically removes multicollinearity.

Correct Answer: B. It preserves the temporal order and simulates future forecasting.

Question 6

Which statement about machine-learning feature importance is correct?

A. It always measures the causal effect of a feature.

B. It shows predictive contribution but does not by itself establish causality.

C. It is meaningful only for linear regression.

D. It proves that the feature should be manipulated.

Correct Answer: B. It shows predictive contribution but does not by itself establish causality.

Question 7

A promotion increases sales of one product but reduces sales of a similar product. This is an example of:

A. Autocorrelation

B. Cannibalisation

C. Heteroscedasticity

D. Non-stationarity

Correct Answer: B. Cannibalisation

Question 8

Which situation is a clear example of data leakage?

A. Using last week’s sales as a lagged predictor.

B. Using a promotion plan known when the forecast is issued.

C. Using final revised sales data that were unavailable at forecast issuance.

D. Using a calendar holiday variable.

Correct Answer: C. Using final revised sales data that were unavailable at forecast issuance.

Question 9

Which model is generally most suitable for estimating an interpretable price elasticity?

A. A carefully specified econometric regression

B. An unpruned single decision tree only

C. A neural network without feature controls

D. A random forecast generated from historical noise

Correct Answer: A. A carefully specified econometric regression

Question 10

A boosted-tree model predicts implausible demand for a discount never observed in training data. The principal concern is:

A. The model may extrapolate poorly beyond the observed price range.

B. The model must be causal because it is nonlinear.

C. The model has no possibility of overfitting.

D. The model has automatically corrected for stock-outs.

Correct Answer: A. The model may extrapolate poorly beyond the observed price range.