Multivariate Time Series and Long-Run Relationships

Introduction

Business variables rarely evolve in isolation. Sales depend on prices, promotions, competitor activity, income, inventory, and macroeconomic conditions. Interest rates influence exchange rates, investment, housing demand, and credit losses. Energy consumption responds to weather, production, prices, and operating schedules. A univariate forecast may be useful, but it ignores information contained in related series.

Multivariate time-series analysis models several variables jointly. This permits analysts to study dynamic feedback, predictive content, temporary shocks, and long-run equilibrium relationships. The principal tools in this chapter are vector autoregression (VAR), Granger predictability tests, impulse-response functions (IRFs), forecast-error variance decomposition (FEVD), cointegration, and vector error-correction models (VECMs).

The central modeling challenge is to distinguish short-run dynamics from long-run relationships. Two series may wander substantially over time yet remain linked by an economically meaningful equilibrium. Modeling their changes alone can discard this information; modeling their levels without testing for nonstationarity can produce misleading regression results.

Learning Objectives

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

1. Explain why cross-series dynamics matter in business forecasting.
2. Specify, estimate, and interpret a VAR model.
3. Distinguish correlation, contemporaneous association, and Granger predictability.
4. Interpret impulse responses and forecast-error variance decompositions.
5. Explain unit roots, cointegration, and spurious regression.
6. Formulate and interpret a VECM and its error-correction mechanism.
7. Select between VAR, VECM, and alternative forecasting approaches.
8. Evaluate whether multivariate information improves operational or strategic forecasts.

1. Multivariate Forecasting as a Dynamic System

Let yₜ be an m × 1 vector containing m variables observed at time t, such as sales, price, and advertising expenditure. A multivariate model represents each variable as a function of its own history and the histories of other variables. In contrast with a collection of separate univariate models, a system model permits feedback: price may affect sales, while sales may influence subsequent pricing decisions.

A useful distinction is:

• Contemporaneous association: variables move together in the same period.
• Lagged dependence: past values of one series help explain another.
• Granger predictability: past values of one series improve prediction of another, conditional on the information set and model specification.
• Structural causality: an intervention or theoretically identified shock produces an effect. Granger predictability does not, by itself, establish this stronger claim.

The modeling objective must be defined before estimation. A forecasting system may prioritize predictive accuracy, whereas a policy or managerial analysis may prioritize interpretable responses to shocks. These goals require different identification assumptions.

2. Vector Autoregression

2.1 VAR(p) formulation

A reduced-form VAR with p lags is

Here, yₜ is the m × 1 vector of endogenous variables; c is an m × 1 intercept vector; Aᵢ is an m × m coefficient matrix for lag i; and εₜ is an m × 1 innovation vector. The innovations have conditional mean zero, E(εₜ | information available before t) = 0, and covariance matrix Σₑ. In a reduced-form VAR, contemporaneous relationships are contained in the covariance matrix and are not automatically given a causal interpretation.

For two variables, yₜ = (salesₜ, priceₜ)′, a VAR(1) can be written as

The coefficient a₁₂ measures the conditional association between lagged price and current sales, not a guaranteed causal price effect. Other variables, omitted lags, reverse decision-making, and measurement error may affect interpretation.

A deterministic component can include an intercept, trend, seasonal dummies, intervention indicators, or exogenous variables. With exogenous regressors xₜ, one may write

where B maps known or modeled external predictors into the system.

2.2 Stationarity and stability

A VAR in levels is stable or covariance-stationary when shocks have transitory effects and the roots of the determinant equation

lie outside the unit circle, where Iₘ is the m × m identity matrix. Equivalently, the eigenvalues of the companion-form transition matrix have modulus less than one. Under stability, the process has constant unconditional means and variances, and it admits a moving-average representation.

For a stable VAR, the moving-average representation is

where Ψ₀ = Iₘ and the Ψⱼ matrices describe the effect at horizon j of innovations. The long-run cumulative response to a one-unit innovation, when the sum converges, is related to ∑ⱼ₌₀∞ Ψⱼ. Stability is essential for standard interpretation; a model with unit roots requires special treatment.

2.3 Lag selection

Too few lags leave serial correlation in residuals and distort dynamic relationships. Too many lags consume degrees of freedom and may increase forecast variance. Candidate lag orders are commonly compared using information criteria such as

where Σ̂ₑ(p) is the estimated innovation covariance matrix, N is the effective sample size, and k is the number of estimated parameters per system adjustment as defined by the implementation. AIC typically penalizes complexity less than BIC. Selection should also consider residual diagnostics, business plausibility, sample size, and forecast performance.

3. Granger Predictability

Suppose a VAR contains series xₜ and yₜ. x is said to Granger-predict y if lagged x contains information that improves prediction of y after accounting for lagged y and other included variables. The null hypothesis that x does not Granger-predict y is commonly tested by restricting all coefficients on lagged x in the y equation to zero:

An F test or Wald test may be used, depending on estimation and assumptions. Rejection means that the restricted model loses predictive information under the chosen sample, variables, lags, and horizon. It does not prove that x causes y, nor that an intervention in x will produce the estimated response.

In systems containing integrated or cointegrated variables, testing must respect the integration structure. A levels VAR or VECM framework may be required; testing a relationship inappropriately transformed into differences can omit long-run information.

4. Impulse-Response Analysis

An impulse-response function traces the predicted response of each variable to a specified innovation over future horizons. In a reduced-form VAR, innovations may be correlated, so a response to “one innovation in x” is not uniquely defined unless the shocks are orthogonalized or otherwise structurally identified.

A common approach is Cholesky orthogonalization. If

where P is a lower-triangular matrix, define orthogonal shocks uₜ by εₜ = Puₜ, with Var(uₜ) = Iₘ. The response at horizon h is then ΨₕP. The ordering of variables determines P and therefore affects the responses. Ordering should be justified by timing or institutional assumptions, not selected merely because it produces attractive results.

Alternative identification methods include:

• structural VAR restrictions based on economic theory;
• sign restrictions on responses;
• long-run restrictions;
• external instruments or proxy shocks;
• high-frequency identification around known events.

IRFs should be reported with confidence intervals, usually obtained through asymptotic, bootstrap, or Bayesian methods. A response that is statistically indistinguishable from zero should not be described as a reliable effect. The horizon, shock size, units, and identification scheme must always be stated.

5. Forecast-Error Variance Decomposition

FEVD partitions the uncertainty of a forecast into contributions associated with different shocks. If the h-step forecast error for variable i is represented using orthogonal shocks, the contribution of shock j is

where eᵢ is a selection vector with one in position i, Θₛ is the response matrix at horizon s, and the denominator is the total h-step forecast-error variance for variable i. Under orthogonal shocks, contributions sum to one for each variable and horizon. FEVD is informative about sources of forecast uncertainty, but it is sensitive to identification, ordering, variable selection, and horizon.

Forecast variance shares should not be confused with economic importance. A shock may account for a small share of forecast variance but still have large consequences in a particular scenario; conversely, a statistically dominant source of variance may be difficult for managers to influence.

6. Nonstationarity and Cointegration

6.1 Unit roots

A scalar process is often called integrated of order one, I(1), when it is nonstationary in levels but becomes stationary after first differencing. A basic example is

where uₜ is stationary white noise. Its first difference, ∆xₜ = xₜ − xₜ₋₁, equals uₜ and is stationary.

Unit-root tests, such as the augmented Dickey–Fuller test, test a null of a unit root against a stationary alternative. Their results can be sensitive to deterministic terms, lag choice, structural breaks, and sample size. A failure to reject is not proof of a unit root. Analysts should combine tests, plots, domain knowledge, and break analysis.

6.2 Spurious regression

Regressing one unrelated trending I(1) series on another can produce a high R² and apparently significant coefficients even when no meaningful relationship exists. This is spurious regression. Differencing may resolve the problem when only short-run dynamics matter, but it may also remove a valid long-run relationship.

6.3 Definition of cointegration

The m-vector yₜ is cointegrated if its components are individually I(1), but at least one nonzero linear combination is stationary. For a cointegrating vector β,

is I(0), even though the individual elements of yₜ are I(1). The stationary zₜ is often interpreted as an equilibrium error or spread. Cointegration does not require variables to have identical units; however, β must produce a meaningful linear combination, and unit consistency should be checked when interpreting it.

For example, if price and unit cost are both measured in currency per unit, their difference may represent a stable markup. If revenue and advertising expenditure are measured in different monetary scales, the coefficients must be interpreted with care and may depend on normalization.

7. The Vector Error-Correction Model

A VAR(p) can be reparameterized as a VECM:

where

and, for j = 1, …, p − 1,

The matrix Π contains long-run information. If its rank is r, with 0 < r < m, it can be factorized as

where β is an m × r matrix of cointegrating vectors and α is an m × r matrix of adjustment coefficients. β′yₜ₋₁ contains equilibrium errors; α determines how each variable responds to disequilibrium.

For one cointegrating relationship, a component of the model may be written as

If α₁ is negative, sales growth tends to reduce a positive equilibrium error, assuming the error is defined as sales − β₀ − β₁price. Sign interpretation depends on normalization. The equation does not imply that sales alone restores equilibrium; the adjustment may occur through several variables.

7.1 Cointegration rank

The rank r of Π determines the model:

Johansen’s trace and maximum-eigenvalue procedures estimate rank in a system, allowing multiple cointegrating vectors. Results depend on deterministic-term specification, lag length, sample size, and structural stability. The number of cointegrating vectors should be economically interpretable, not selected mechanically.

7.2 Identification and normalization

A cointegrating vector is identified only up to scale: multiplying β by any nonzero constant and α by its reciprocal leaves Π unchanged. A normalization, such as setting the coefficient on sales to one, is therefore a presentation choice. Economic restrictions may be required to identify a uniquely interpretable relationship. Weak-exogeneity restrictions can test whether a variable responds to long-run disequilibrium.

8. A Professional Modeling Workflow

A defensible multivariate forecasting project typically follows these stages:

1. Define the decision and forecast target. Specify variables, frequency, horizon, forecast origin, and whether forecasts are conditional on future prices, promotions, or macroeconomic inputs.
2. Construct and audit the data. Align timestamps, handle missing values, document revisions, inspect outliers, and ensure consistent units. Avoid using information unavailable at the forecast origin.
3. Explore the system. Plot levels and differences, inspect correlations and cross-correlations, and identify seasonality, breaks, and calendar effects.
4. Assess integration. Use unit-root tests with suitable intercepts, trends, and lag terms. Consider structural-break tests and domain knowledge.
5. Select lag length. Compare information criteria and residual diagnostics while preserving a realistic validation sample.
6. Test cointegration when variables are I(1). Use a suitable system method and deterministic specification.
7. Estimate the appropriate model. Use a level VAR for stationary variables, a VECM for cointegrated I(1) variables, or a differenced VAR when no cointegration is supported.
8. Diagnose residuals. Check serial correlation, heteroskedasticity, non-normality, stability, and outlier influence. Non-normal residuals may affect small-sample inference even if forecasts remain usable.
9. Evaluate forecasts out of sample. Compare against sensible benchmarks, including seasonal naïve, drift, univariate ARIMA or exponential smoothing, and simpler regressions.
10. Interpret dynamics cautiously. Use IRFs, FEVD, and Granger tests only with explicit assumptions and uncertainty intervals.
11. Stress-test the model. Examine rolling windows, alternative lag orders, variable orderings, break periods, and plausible scenarios.
12. Deploy with monitoring. Track forecast errors, data revisions, parameter drift, and changes in business processes.

For a forecast ŷₜ₊h and observed value yₜ₊h, common evaluation measures are

and

where H is the number of forecast observations. MAE is in the units of the target; RMSE gives greater weight to large errors. Comparisons should use the same forecast origins and information availability. For multiple horizons or series, report scale-aware measures and business-weighted costs where appropriate.

9. Illustrative Case: Retail Demand, Price, and Inventory

Background

A national retailer forecasts weekly demand for a product category. Management suspects that demand, price, and inventory are dynamically connected. Price reductions may increase demand, while low inventory may constrain observed sales. Persistent inflation also causes prices and revenue to trend upward.

Problem

The existing model forecasts demand from its own lags and seasonal indicators. It performs adequately at a one-week horizon but misses demand surges after price changes and produces biased forecasts during prolonged inflation. Managers want to know whether a multivariate model improves forecasts and whether inventory affects the interpretation of observed sales.

Analysis and approach

The team defines yₜ = (log demandₜ, log priceₜ, log inventoryₜ)′ and includes promotion indicators and holiday dummies as exogenous variables. Logs make coefficients approximately interpretable as elasticities and reduce scale differences, but they require positive observations and careful treatment of stockouts.

Plots show that all three log-level series are persistent. Unit-root testing is inconclusive in a short sample, so the team examines both a differenced VAR and a VECM. Economic reasoning suggests that demand and price may share a long-run relationship, while inventory may respond to demand and replenishment policies. Johansen testing supports one cointegrating vector under the selected deterministic specification.

The estimated VECM includes two weekly lags in differences. The cointegrating relation is normalized on demand:

The demand equation has a negative adjustment coefficient. Thus, when demand is above its long-run relation with price and inventory, subsequent demand growth tends to be lower, conditional on the other terms. Price and inventory equations also contain adjustment coefficients, indicating that equilibrium restoration is shared across the system rather than imposed entirely on demand.

Granger tests indicate that lagged price and inventory jointly improve demand prediction. However, the team avoids claiming that the tests establish causal price elasticity. Price changes are partly managerial responses to expected demand, and inventory is affected by supply constraints. For causal promotion decisions, the firm would need experimental or structurally identified evidence.

IRFs are estimated using an economically justified ordering in which replenishment information is available before weekly pricing decisions, while alternative orderings are used as sensitivity checks. A price innovation produces a temporary demand response, but the magnitude is uncertain and differs across identification schemes. FEVD shows that inventory and price explain a meaningful share of medium-horizon demand forecast variance, whereas short-horizon uncertainty remains dominated by demand-specific shocks.

Outcome

Rolling-origin evaluation compares the VECM with the existing univariate model, a differenced VAR, and a seasonal naïve benchmark. The VECM improves MAE at two- to eight-week horizons, especially during periods of persistent price and inventory movements, but offers little advantage at the one-week horizon. The firm therefore uses the VECM for replenishment planning and the simpler model for rapid daily operational updates.

Learning points

1. Multivariate information improves forecasts only when related variables are timely, measured reliably, and predictive out of sample.
2. Stockouts create censoring: observed sales may understate unconstrained demand. A VECM cannot automatically solve this measurement problem.
3. Cointegration supports a long-run interpretation but does not prove causal mechanisms.
4. Forecast gains should be evaluated by horizon and decision cost, not by in-sample fit alone.
5. Alternative identification schemes and rolling validation are essential when managerial actions influence the explanatory variables.

10. Limitations, Governance, and Contemporary Practice

Structural breaks and regime change

Pandemics, regulation, technology adoption, mergers, and pricing-policy changes can alter coefficients and cointegrating relationships. A historically stable VECM may fail after a regime change. Rolling estimation, break tests, time-varying parameter models, and intervention variables can help, but none guarantees stability.

High dimensionality

A VAR with m variables and p lags contains approximately pm² slope coefficients before deterministic terms. Parameter growth can become severe. Bayesian VARs, shrinkage, factor-augmented VARs, sparse methods, and carefully selected variables can improve estimation. Complexity should be justified by forecast gains and interpretability.

Exogeneity and real-time information

Future exogenous variables may not be known when a forecast is produced. A model that uses realized future promotions or macroeconomic values can suffer from look-ahead bias. They must be forecast separately, supplied as a genuine scenario, or excluded.

Ethical and regulatory considerations

Forecasts influence staffing, credit access, inventory allocation, and prices. Analysts should document data provenance, transformations, missing-value treatment, model versions, and uncertainty. Sensitive variables may create discriminatory outcomes through direct use or proxy effects. Governance should include access controls, reproducible pipelines, monitoring for drift, human review of high-impact decisions, and clear communication that predictive relationships are not automatically causal.

Machine learning and hybrid systems

Modern systems may combine VAR or VECM structure with gradient boosting, recurrent neural networks, temporal convolutional networks, or transformers. Machine-learning models can capture nonlinearities and interactions, while econometric models offer explicit long-run relationships and interpretable shocks. Hybrid modeling is most defensible when it is evaluated against strong baselines, uses time-respecting validation, prevents leakage, and preserves uncertainty estimates. A more complex model is justified only when it produces reliable incremental value.

Best-Practice Checklist

• Match the model to the forecast decision and data-generating context.
• Distinguish forecasting from causal analysis.
• Test integration and cointegration before interpreting level relationships.
• Use sufficient lags to remove residual autocorrelation without overfitting.
• Justify deterministic terms and structural identification.
• Report uncertainty for IRFs and forecast comparisons.
• Validate recursively out of sample using real-time information sets.
• Compare against univariate and naïve benchmarks.
• Investigate breaks, stockouts, revisions, and measurement changes.
• Monitor performance after deployment and revisit the model when business mechanisms change.

Chapter Summary

Multivariate time-series models treat business variables as a dynamic system in which own lags, cross-series lags, shocks, and long-run restrictions interact. A VAR models stationary variables jointly; Granger tests assess whether lagged information improves prediction; IRFs trace responses to identified shocks; and FEVD attributes forecast uncertainty to those shocks. These tools describe dynamic dependence but do not automatically establish structural causality.

When variables are I(1), differencing may remove nonstationarity but can discard equilibrium information. Cointegration exists when a linear combination of nonstationary variables is stationary. A VECM represents both short-run changes and correction toward long-run equilibria through the decomposition Π = αβ′. Appropriate rank testing, normalization, diagnostics, identification, and out-of-sample validation are essential.

Multivariate information improves business forecasts only when it is timely, stable, and genuinely predictive beyond strong benchmarks. Professional practice therefore combines econometric theory, domain knowledge, data governance, forecast evaluation, and continuous monitoring.

Key Terms

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

VECM: A vector error-correction model combining differenced dynamics with long-run equilibrium adjustment.

Granger predictability: Improvement in conditional prediction from including past values of another series.

Impulse-response function: A sequence of responses to a specified innovation over future horizons.

Forecast-error variance decomposition: A partition of forecast uncertainty among identified shocks.

Stationarity: A property in which relevant probabilistic characteristics do not change over time.

Unit root: A dynamic root on the unit circle that generally produces nonstationarity.

Integrated process: A process requiring differencing a specified number of times to become stationary.

Cointegration: Stationary linear dependence among individually nonstationary series.

Cointegrating vector: A vector β for which β′yₜ is stationary.

Equilibrium error: The deviation of a system from its long-run cointegrating relationship.

Adjustment coefficient: An element of α describing how a variable responds to disequilibrium.

Cointegration rank: The number of independent cointegrating relationships.

Reduced-form innovation: An unexplained system disturbance that may be contemporaneously correlated with other disturbances.

Structural shock: An economically interpreted disturbance identified using additional assumptions.

Cholesky identification: Orthogonalization based on a triangular factorization of the innovation covariance matrix.

Spurious regression: Apparently significant regression among unrelated nonstationary series.

Forecast origin: The information date at which a forecast is produced.

Structural break: A change in parameters or data-generating behavior over time.

Review Questions

Short Answer Questions

1. What is the difference between contemporaneous association and Granger predictability?
2. Why can a VAR in levels be inappropriate for nonstationary variables?
3. Define cointegration and explain why it matters for forecasting.
4. What do the matrices α and β represent in a VECM?
5. Why does Cholesky ordering affect an impulse-response analysis?

Long Answer Questions

1. Explain the formulation, stability condition, lag selection, and interpretation of a VAR model.
2. Discuss how IRFs and FEVD complement one another and identify their limitations.
3. Derive the conceptual transition from a VAR in levels to a VECM and explain the role of Π.
4. Design a complete workflow for evaluating whether multivariate information improves a business forecast.
5. Critically assess why Granger predictability should not be interpreted as structural causality.

Analytical / Application Questions

1. A retailer has weekly sales, price, and inventory series that appear trending. Propose tests and model-selection steps before estimating a forecasting system.
2. A VECM shows a negative adjustment coefficient in the sales equation but no significant adjustment in the price equation. Explain the economic interpretation and the assumptions required for that conclusion.
3. A VAR improves in-sample fit but performs worse than a seasonal naïve model in rolling forecasts. Diagnose possible reasons and recommend a managerial course of action.

Multiple Choice Questions with Answers

Question 1

Which statement best describes Granger predictability?

A. It proves that one variable causes another.

B. It shows that past values improve conditional prediction.

C. It requires contemporaneous correlation to be zero.

D. It applies only to stationary economic variables.

Correct Answer: B. It shows that past values improve conditional prediction.

Question 2

In a stable VAR, the eigenvalues of the companion transition matrix must have:

A. Modulus greater than one

B. Modulus equal to one

C. Modulus less than one

D. Positive real parts only

Correct Answer: C. Modulus less than one

Question 3

If individually I(1) variables have a stationary linear combination, they are:

A. White noise

B. Cointegrated

C. Strictly exogenous

D. Seasonally adjusted

Correct Answer: B. Cointegrated

Question 4

In the VECM decomposition Π = αβ′, β primarily contains:

A. Short-run lag coefficients

B. Innovation variances

C. Cointegrating vectors

D. Forecast-error measures

Correct Answer: C. Cointegrating vectors

Question 5

Why can Cholesky impulse responses depend on variable ordering?

A. Ordering changes the sample size.

B. The triangular factorization assigns contemporaneous relationships according to the order.

C. Ordering removes all lagged effects.

D. The method requires variables to be normally distributed.

Correct Answer: B. The triangular factorization assigns contemporaneous relationships according to the order.

Question 6

Which measure gives greater weight to large forecast errors?

A. MAE

B. Median absolute error

C. RMSE

D. Mean error sign

Correct Answer: C. RMSE

Question 7

A rank of zero for Π in a system of I(1) variables generally indicates:

A. m independent cointegrating relationships

B. No cointegrating relationships

C. That all variables are stationary

D. That residuals are necessarily autocorrelated

Correct Answer: B. No cointegrating relationships

Question 8

What is the principal danger of regressing one unrelated trending series on another?

A. Perfect multicollinearity

B. Spurious regression

C. Automatic overdifferencing

D. Guaranteed forecast improvement

Correct Answer: B. Spurious regression

Question 9

FEVD is primarily used to:

A. Estimate missing observations

B. Partition forecast uncertainty across identified shocks

C. Select the seasonal frequency

D. Prove structural causality

Correct Answer: B. Partition forecast uncertainty across identified shocks

Question 10

Which practice best prevents look-ahead bias?

A. Using realized future promotions as regressors

B. Evaluating forecasts with information available at each forecast origin

C. Selecting variables from the complete future sample without validation

D. Replacing missing future values with actual observations

Correct Answer: B. Evaluating forecasts with information available at each forecast origin