Multivariate and Causal Forecasting

Introduction

Many business variables evolve together. Sales may respond to advertising, prices, promotions, weather, and competitor actions; interest rates may influence borrowing, investment, exchange rates, and housing demand; and inventory may respond to demand forecasts, replenishment policies, and supplier lead times. A univariate time-series model can forecast one variable using its own history, but it may omit information contained in related series.

Multivariate forecasting models address this problem by modeling several time series jointly. Causal forecasting goes further: it asks whether a variable or intervention produces changes in another variable, rather than merely improving prediction. This distinction is fundamental. A variable can be useful for forecasting without being a cause, because both variables may respond to a third factor, because one may be a proxy, or because information flows in the opposite direction.

This chapter develops vector autoregression (VAR), Granger causality, cointegration, vector error-correction models (VECM), leading indicators, intervention analysis, and the assumptions required for causal interpretation. The emphasis is on sound modeling, practical interpretation, and responsible use in business analytics.

Learning Objectives

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

1. Explain why multiple time series may need to be modeled jointly.
2. Specify and interpret a VAR model.
3. Conduct and interpret Granger-causality tests without treating them as proof of structural causation.
4. Distinguish stationarity from nonstationarity and explain spurious regression.
5. Explain cointegration and formulate a VECM.
6. Identify and evaluate leading indicators.
7. Design an intervention analysis for a policy, campaign, disruption, or event.
8. State the assumptions needed for causal conclusions.
9. Evaluate forecasting and causal models using appropriate diagnostics and governance practices.

1. From Univariate to Multivariate Forecasting

Let y_t denote a variable observed at time t. A univariate autoregressive model of order p, written AR(p), is commonly expressed as:

Here, c is an intercept, φ_i are autoregressive coefficients, and ε_t is an innovation or error with conditional mean zero. A multivariate model allows past values of several variables to explain each current value. For example, sales may depend on past sales, advertising expenditure, and price.

A multivariate model is useful when:

• variables influence one another over time;
• a forecast of one series improves when other series are observed;
• management decisions concern a system rather than one isolated outcome;
• shocks propagate across markets, products, or departments;
• long-run equilibrium relationships are economically meaningful.

The central danger is overinterpretation. Correlated dynamics do not automatically establish causal mechanisms. A forecasting model is judged primarily by predictive performance, whereas a causal model requires defensible assumptions about interventions, confounding, timing, measurement, and the data-generating process.

2. Vector Autoregression

2.1 Basic formulation

Let y_t be a K×1 vector containing K jointly observed variables. A VAR of order p is:

where:

• c is a K×1 intercept vector;
• A_i is a K×K coefficient matrix for lag i;
• ε_t is a K×1 innovation vector;
• E(ε_t)=0 and the contemporaneous covariance matrix is Σ_ε = E(ε_tε_t′).

For two variables, sales [[MATH(s_t)]] and advertising [[MATH(a_t)]], a VAR(1) may be written:

The coefficient [[MATH(a₁₂)]] measures the association between lagged advertising and current sales, conditional on lagged sales and the second equation’s system variables. It is not automatically the effect of an advertising intervention.

2.2 Interpretation and forecasting

A VAR treats each included variable as potentially endogenous: its current value is modeled using past values of all variables in the system. Forecasts are generated recursively. First, estimate the one-step-ahead value using observed lags; then use that prediction as an input for later horizons.

A VAR can be expressed compactly using the lag operator [[MATH(L)]], where [[MATH(Ly_t = y_{t−1})]]:

A stable VAR requires the roots of [[MATH|A(z)|=0]] to lie outside the unit circle, equivalently the eigenvalues of the companion-form transition matrix must have modulus less than one. Stability implies that shocks have diminishing effects over time. If this condition fails, forecasts and impulse responses require special treatment.

2.3 Specification choices

Important decisions include variable selection, lag length, deterministic terms, frequency, transformations, and sample period. Information criteria such as AIC, BIC, or HQ can guide lag selection, but the choice should also reflect business timing and residual diagnostics. Seasonal dummies, trends, intervention indicators, or exogenous variables may be included when justified.

A large VAR can quickly consume degrees of freedom. With K variables and p lags, each equation contains approximately Kp lag coefficients, excluding deterministic terms. Parsimony is therefore important. Regularized VARs, Bayesian VARs, factor models, and shrinkage methods can help when the system is large, but their assumptions must be documented.

2.4 Diagnostics

After estimating a VAR, analysts should examine:

• residual autocorrelation, using multivariate portmanteau or related tests;
• residual heteroskedasticity and outliers;
• stability of the estimated system;
• normality when inference relies on it;
• sensitivity to lag length, sample period, and variable transformations;
• out-of-sample forecast accuracy.

Forecast measures should match the decision. Common measures include mean absolute error, MAE = (1/n)∑|y_t−ŷ_t|, and root mean squared error, RMSE = √[(1/n)∑(y_t−ŷ_t)²]. MAE is in the units of the outcome and treats errors linearly; RMSE gives greater weight to large errors. Comparisons should use the same forecast horizon, evaluation window, and information set.

3. Granger Causality: Predictive Precedence

A variable x_t is said to Granger-cause y_t relative to an information set if past values of x_t improve prediction of y_t after past values of y_t and other included information have been considered.

In a VAR, the null hypothesis that x does not Granger-cause y is often tested by restricting all coefficients on lagged x in the y equation to zero:

An F test, Wald test, or likelihood-ratio test may be used, depending on the estimation framework and assumptions. Rejection indicates predictive content, not necessarily a real-world causal effect. For example, a newspaper’s publication time may predict daily stock movements because both reflect an underlying information cycle; the publication itself need not cause the movement.

Granger tests are sensitive to omitted variables, lag length, structural breaks, nonstationarity, contemporaneous feedback, and measurement frequency. Testing many pairs also creates multiple-testing concerns. Economic theory and institutional knowledge should guide the information set and hypotheses.

4. Stationarity, Unit Roots, and Spurious Regression

A weakly stationary series has a constant mean, constant finite variance, and autocovariances that depend on lag length rather than calendar time. Many business series—prices, cumulative sales, employment levels, and money balances—are nonstationary. A common nonstationary process is a unit-root process:

Its variance grows with time, so ordinary regression of one unrelated trending series on another can produce high R² and apparently significant coefficients even when no meaningful relationship exists. This is called spurious regression.

Analysts may difference a series:

or use growth rates such as 100×(ln y_t−ln y_t−1). Differencing can improve stationarity, but it may remove economically meaningful long-run information. Unit-root tests, including augmented Dickey–Fuller-type tests, have low power in some settings and should be interpreted alongside plots, institutional knowledge, and alternative specifications.

5. Cointegration and the Vector Error-Correction Model

5.1 Concept

Two or more nonstationary series are cointegrated if a linear combination of them is stationary. If sales and household income each trend over time, their difference may nevertheless remain bounded because economic forces maintain a long-run relationship.

For a vector y_t, cointegration exists when there is a nonzero vector β such that:

Here, I(1) denotes a series that becomes stationary after first differencing, and I(0) denotes stationarity in levels. The vector β defines a long-run equilibrium relation; its scale is not unique, so a normalization is required for interpretation.

5.2 Error correction

A VECM represents both short-run changes and correction toward long-run equilibrium. For a VAR(p) with cointegration rank r, one representation is:

where:

β′y_t−1 contains r equilibrium errors, and α contains adjustment coefficients. Γ_i capture short-run effects. If the equilibrium error is positive and the relevant element of α is negative, the variable tends to decrease in the next period, moving the system back toward equilibrium.

The Johansen approach estimates the cointegration rank using the rank of Π. The cases are: rank zero, implying no cointegration and a model in differences; full rank, implying stationarity in levels; and intermediate rank, implying one or more cointegrating relations. Determining rank requires judgment because results depend on deterministic terms, lag length, sample size, breaks, and test assumptions.

5.3 Practical interpretation

Suppose a normalized equilibrium relation is:

If z_t is stationary, sales and income are linked in the long run. A VECM does not claim that income mechanically causes all sales changes. It says that deviations from the estimated equilibrium help predict subsequent adjustments, conditional on the model and information set.

6. Leading Indicators and Dynamic Information

A leading indicator changes before the target variable and provides useful information about its future movement. Examples include new orders leading shipments, building permits leading construction activity, job vacancies leading employment, or web searches leading product demand.

A useful indicator should have:

• a defensible timing mechanism;
• reliable and timely measurement;
• a stable relationship, or a relationship whose changes can be monitored;
• incremental predictive value beyond the target’s own history;
• acceptable revision and availability characteristics.

Lead-lag correlation alone is insufficient. Analysts should conduct rolling-origin validation and compare a benchmark model with and without the indicator. Real-time forecasting must use only information available at the forecast date; otherwise, data revisions and publication delays create look-ahead bias.

A leading indicator can fail when behavior changes, policy alters the mechanism, the indicator becomes widely anticipated, or a structural break occurs. Indicator selection should therefore combine statistical evidence with domain theory.

7. Intervention Analysis

Intervention analysis evaluates whether a known event is associated with a change in a time series after accounting for its prior dynamics. Events may include a price campaign, product launch, strike, regulatory change, cyberattack, store closure, or supply disruption.

Let I_t be an intervention indicator. A simple pulse and step specification are:

A dynamic intervention model may be written:

A pulse represents a temporary shock; a step represents a permanent level shift. A gradual effect may use distributed lags or a transfer-function structure. The coefficients θ_j measure conditional associations with the intervention, not automatically causal effects.

Causal interpretation is more credible when the intervention date is externally determined, no other major event coincides with it, pre-intervention trends are stable, outcomes are measured consistently, and a credible comparison group is available. Interrupted time-series designs can be strengthened with a control series through difference-in-differences. A basic specification is:

Here, α_i are unit effects, λ_t are time effects, and τ estimates the treatment effect under the parallel-trends assumption: absent treatment, treated and comparison units would have followed similar trends.

8. Forecasting Is Not Causation

The distinction can be summarized as follows:

Causal claims concern counterfactuals: what would have happened to the same unit at the same time under a different intervention. Time-series prediction does not observe this counterfactual by itself.

Key causal assumptions include:

1. Consistency: the observed outcome under the received treatment equals the potential outcome for that treatment.
2. No unmeasured confounding, when using adjustment: conditional on measured covariates, treatment assignment is as good as random for the comparison being made.
3. Positivity: each relevant group has a nonzero probability of receiving each treatment level.
4. Correct temporal ordering: causes precede effects, and variables are measured at appropriate frequencies.
5. Stable measurement and treatment definition: the intervention and outcome are not redefined in ways that invalidate comparisons.
6. No problematic interference: one unit’s treatment does not unpredictably alter another unit’s outcome, unless the model incorporates spillovers.

Directed acyclic graphs can clarify assumed relationships. Analysts should distinguish confounders, mediators, colliders, and proxies. Conditioning on a collider can create bias, even when the variables appear relevant for prediction.

9. Worked Example: Advertising and Weekly Sales

Consider 104 weeks of observations for a retailer: sales, advertising expenditure, price, and competitor price. The business wants to forecast sales and assess whether a new advertising campaign caused an increase.

Step 1: Forecasting model

The analyst first plots the variables, checks missing values, records publication timing, and uses logarithms for strongly right-skewed positive variables. A VAR with two lags is selected using information criteria and residual diagnostics. The sales equation includes lagged sales, advertising, prices, and competitor price; the advertising equation is also estimated because advertising may respond to expected demand.

Suppose a joint test rejects the null that lagged advertising coefficients in the sales equation are zero, while the reverse test is not rejected. Advertising therefore has incremental predictive content for sales in this sample. This is a Granger-predictive result, not proof that advertising caused sales.

Step 2: Long-run assessment

Sales and prices appear to trend, so the analyst tests for unit roots. If sales and income or market size are integrated and cointegrated, a VECM is preferable to an unrestricted VAR in differences because it preserves the equilibrium relation. If advertising is a stationary promotional-flow variable, it may enter as an exogenous or short-run explanatory variable, subject to the design.

Step 3: Campaign intervention

The campaign begins in week 70. A pulse indicator is inappropriate if the campaign continues; a step indicator may capture a permanent level change, while a distributed-lag structure can capture gradual effects. The analyst also identifies a comparable region where the campaign was not introduced and estimates an interrupted time-series or difference-in-differences model.

The causal interpretation depends on whether the treated region and comparison region had similar pre-campaign trends, whether competitor promotions changed at the same time, and whether management selected the region because it expected unusually high growth. If campaign placement was targeted toward rising demand, ordinary regression may overstate the campaign effect.

Step 4: Evaluation and decision

Rolling-origin forecasts are compared with a seasonal naïve benchmark. The campaign effect is reported with uncertainty intervals, alternative control groups, placebo intervention dates, and sensitivity to lag length. Management receives two separate outputs: a forecast distribution for inventory planning and a conditional estimate of campaign impact for budget allocation. Separating these outputs prevents a strong forecast relationship from being presented as a proven causal return on advertising.

10. Implementation Workflow

A defensible multivariate project generally follows this sequence:

1. Define the decision, target, forecast horizon, and causal question separately.
2. Establish the information set available at each forecast date.
3. Map business timing and plausible mechanisms before selecting variables.
4. Inspect levels, logarithms, differences, seasonality, breaks, outliers, and missingness.
5. Test or assess stationarity and cointegration using complementary evidence.
6. Select a parsimonious VAR, VECM, intervention model, or alternative design.
7. Diagnose residuals, stability, parameter changes, and forecast errors.
8. Evaluate using rolling or expanding windows and realistic benchmarks.
9. For causal claims, state the counterfactual and identification assumptions explicitly.
10. Conduct placebo, sensitivity, subgroup, and alternative-specification analyses.
11. Document data revisions, model versions, exclusions, and decision thresholds.
12. Monitor drift and retrain only under a governed process.

11. Limitations, Ethics, and Governance

Multivariate models can amplify errors because misspecification in one equation may affect the whole system. Structural breaks can invalidate historical relationships. High-dimensional models may overfit. Revisions to macroeconomic data can create a false impression of historical forecasting performance. Simultaneous decisions—such as setting advertising in response to expected sales—can produce endogeneity.

Causal analytics also raises ethical concerns. Models used for pricing, credit, workforce planning, or marketing can distribute benefits and burdens unevenly. Analysts should assess whether variables encode protected characteristics or access inequalities, whether intervention effects differ across groups, and whether automated decisions are explainable and contestable. Privacy, data minimization, access control, retention, and regulatory requirements must be incorporated into the data lifecycle.

AI-assisted forecasting can support feature engineering, anomaly detection, model comparison, and natural-language reporting. It does not remove the need for temporal validation, leakage controls, causal design, or human accountability. Large language models may generate plausible but unsupported causal narratives; all such claims require statistical and domain verification.

Contemporary Developments and Best Practices

Modern forecasting increasingly combines classical econometrics with machine learning. Dynamic factor models summarize many correlated indicators; Bayesian VARs shrink unstable estimates toward sensible priors; state-space models allow time-varying parameters; and causal machine-learning methods estimate heterogeneous treatment effects. These tools are valuable when their assumptions match the problem, but complexity should not replace identification or interpretability.

Best practice is to maintain separate model layers: a transparent benchmark, a production forecast model, and a causal evaluation design. Report point forecasts together with prediction intervals, and report causal estimates together with confidence or credible intervals and assumptions. Use real-time data vintages when evaluating historical decisions. Reassess relationships after major policy, technology, market, or behavioral changes.

Chapter Summary

Multivariate forecasting models several time series jointly and can exploit cross-series information. A VAR represents each variable as a function of lags of all variables in the system, while stability, parsimony, and diagnostic testing determine whether its forecasts are credible. Granger causality means predictive improvement from past values; it does not establish structural causation.

Nonstationary series can produce spurious regression. Cointegration identifies stationary long-run combinations of nonstationary variables, and a VECM combines short-run changes with adjustment toward equilibrium. Leading indicators are useful only when their timing, availability, stability, and incremental predictive value are validated. Intervention analysis models the effect of pulses, steps, or gradual events, but causal interpretation requires a credible counterfactual and assumptions such as appropriate timing, no serious confounding, and—where relevant—parallel trends.

Professional analysts separate the forecasting question from the causal question, use realistic information sets, validate out of sample, test robustness, monitor structural change, and communicate uncertainty. Ethical governance is essential when forecasts or causal estimates influence people, markets, access, or resource allocation.

Key Terms

Vector autoregression (VAR): A system in which each variable depends on lags of all included variables.

Vector error-correction model (VECM): A model combining short-run differences with adjustment toward long-run equilibrium.

Granger causality: Incremental predictive content of one variable’s past for another variable.

Stationarity: A time-series property involving stable distributional features over time.

Unit root: A stochastic trend causing nonstationary behavior.

Differencing: Subtracting a prior observation from the current observation.

Cointegration: A stationary linear combination of individually nonstationary series.

Cointegrating vector: Coefficients defining a long-run equilibrium relation.

Adjustment coefficient: A VECM parameter describing response to disequilibrium.

Leading indicator: A variable that changes before and helps predict a target.

Intervention analysis: Modeling a known event’s association with a time-series change.

Pulse intervention: An indicator equal to one at a specific time only.

Step intervention: An indicator that switches on from an intervention date onward.

Spurious regression: Apparently significant regression caused by unrelated nonstationary trends.

Impulse response: The dynamic response of system variables to a specified shock.

Structural break: A change in the data-generating process or parameter relationships.

Counterfactual: The outcome that would occur under an alternative treatment or intervention.

Confounder: A variable affecting both treatment and outcome that can bias causal estimates.

Look-ahead bias: Use of information unavailable at the time a forecast would have been made.

Review Questions

Short Answer Questions

1. What distinguishes a VAR from a univariate autoregressive model?
2. What does rejection of a Granger-causality null hypothesis imply?
3. Define cointegration and explain why it matters for modeling nonstationary series.
4. How do pulse and step interventions differ?
5. Why is out-of-sample validation important in multivariate forecasting?

Long Answer Questions

1. Explain the structure, interpretation, stability condition, and diagnostic requirements of a VAR.
2. Compare a VAR in differences with a VECM, including the role of long-run equilibrium.
3. Discuss why Granger causality should not be interpreted as proof of structural causation.
4. Explain the assumptions required for a credible interrupted time-series or difference-in-differences analysis.
5. Design a complete workflow for forecasting demand using sales, price, advertising, and economic indicators.

Analytical / Application Questions

1. A firm finds that advertising Granger-causes sales, but advertising spending rises before expected demand peaks. Explain at least two reasons why the result may not identify the causal effect of advertising and propose an improved design.
2. Two nonstationary variables have a stable long-run relationship, but short-run deviations are persistent. Explain why a VECM may be appropriate and interpret the sign of an adjustment coefficient.
3. A policy begins in one region while a similar region remains untreated. Describe the tests, plots, placebo checks, and assumptions you would use before claiming a causal policy effect.

Multiple Choice Questions with Answers

Question 1

What is the primary meaning of Granger causality?

A. The first variable is a proven structural cause of the second

B. Past values of one variable improve prediction of another, conditional on the model

C. The two variables have identical variances

D. The variables must be cointegrated

Correct Answer: B. Past values of one variable improve prediction of another, conditional on the model

Question 2

In a VAR, each equation generally includes:

A. Only lags of its own dependent variable

B. Only contemporaneous values of other variables

C. Lags of all variables included in the system

D. No error term

Correct Answer: C. Lags of all variables included in the system

Question 3

Cointegration occurs when:

A. All variables are stationary in levels

B. A linear combination of nonstationary series is stationary

C. Two variables have zero correlation

D. Every variable has a unit root and no stable combination

Correct Answer: B. A linear combination of nonstationary series is stationary

Question 4

What does a step intervention usually represent?

A. A one-time observation error

B. A permanent level change beginning at a specified date

C. A seasonal cycle with no intervention

D. A forecast interval

Correct Answer: B. A permanent level change beginning at a specified date

Question 5

Why can regression of unrelated trending series be misleading?

A. It necessarily produces negative coefficients

B. It can produce apparently significant relationships through common trends

C. It eliminates all autocorrelation

D. It guarantees stationarity

Correct Answer: B. It can produce apparently significant relationships through common trends

Question 6

In a VECM, β′yₜ typically represents:

A. The forecast error variance

B. The short-run innovation vector

C. The long-run equilibrium error

D. The number of observations

Correct Answer: C. The long-run equilibrium error

Question 7

Which condition is central to a difference-in-differences causal interpretation?

A. The outcome must have zero variance

B. The treated and comparison groups must have identical levels at every date

C. Parallel trends in the absence of treatment

D. The treatment must be randomly timed by the analyst

Correct Answer: C. Parallel trends in the absence of treatment

Question 8

Which practice prevents look-ahead bias?

A. Using revised future data in historical forecasts

B. Using only information available at each forecast date

C. Selecting variables after observing the test-period outcomes

D. Evaluating only in-sample fit

Correct Answer: B. Using only information available at each forecast date

Question 9

A pulse intervention indicator equals one:

A. At every date after the intervention

B. Only at the intervention date

C. Only before the intervention

D. Whenever the outcome is stationary

Correct Answer: B. Only at the intervention date

Question 10

A strong predictive relationship between two variables by itself establishes:

A. A causal effect under all conditions

B. No useful information

C. Predictive content, but not necessarily causation

D. Cointegration in levels

Correct Answer: C. Predictive content, but not necessarily causation