メインコンテンツへ移動

Macro by Mark の Cookie.ログイン、言語、安全性のために必須 Cookie を使用します。許可いただいた場合は、設定の保存、分析、Meta(Facebook)ピクセルによる広告計測のために任意の Cookie も使用します。詳しくはCookie ポリシーをご覧ください。

Macro by MarkMacro by Mark
  • ホーム
読み込み中…
読み込み中…
← モデル
Macro概念政策モデル

VECM Model

Vector error-correction model - for cointegrated macro series where long-run equilibrium relationships pull the system back.

概要グラフ数式参考資料

How do you model short-run dynamics among variables that share long-run equilibrium relationships?

Background

Clive Granger's 1981 insight was that two individually nonstationary time series can move together in a way that keeps their linear combination stationary. He called this cointegration. If log GDP and log consumption are both I(1) but their difference (the log of the consumption-to-GDP ratio) is I(0), the two series are cointegrated with cointegrating vector (1,−1)(1, -1)(1,−1), and a VAR in first differences throws away the equilibrium relationship - the levels contain information that differencing destroys. The Vector Error Correction Model (VECM) solves this by embedding the long-run equilibrium directly into a first-difference VAR as an error correction term. Granger shared the 2003 Nobel Prize with Robert Engle for this work.

Soren Johansen (1988, 1991) developed the maximum likelihood framework that made cointegration estimation practical for systems of n variables. His procedure jointly estimates the number of cointegrating relationships (the cointegrating rank r), the cointegrating vectors (the long-run equilibrium equations), and the short-run adjustment coefficients (the speed at which each variable corrects deviations from equilibrium). The Johansen trace and maximum eigenvalue tests for the cointegrating rank became the standard diagnostic in applied macroeconometrics.

The VECM representation is: Δyt=αβ′yt−1+Γ1Δyt−1+…+Γp−1Δyt−p+1+ut\Delta y_t = \alpha \beta' y_{t-1} + \Gamma_1 \Delta y_{t-1} + \ldots + \Gamma_{p-1} \Delta y_{t-p+1} + u_tΔyt​=αβ′yt−1​+Γ1​Δyt−1​+…+Γp−1​Δyt−p+1​+ut​. Here β′yt−1\beta' y_{t-1}β′yt−1​ is the r×1r \times 1r×1 vector of cointegrating residuals (deviations from equilibrium), α\alphaα is the n×rn \times rn×r loading matrix (how fast each variable adjusts), and the Γ\GammaΓ matrices capture short-run dynamics in differences. The key structural restriction: the long-run impact matrix Π=αβ′\Pi = \alpha \beta'Π=αβ′ has reduced rank r<nr < nr<n. This rank restriction distinguishes the VECM from a VAR in differences (r=0)(r = 0)(r=0) and a VAR in levels (r=n)(r = n)(r=n).

Central banks use VECMs to model systems where economic theory predicts long-run relationships: the Fisher equation linking nominal rates to inflation, purchasing power parity tying exchange rates to price levels, money demand equations relating money supply to income and interest rates. The Bank of England's medium-term macro model includes VECM components. The Riksbank and Norges Bank have published VECM-based analyses of the monetary transmission mechanism. In academic work, King, Plosser, Stock, and Watson (1991) used VECMs to test the permanent-transitory decomposition implied by real business cycle theory.

Connections to explore

Math ↗The structure and its conditions, step by step.References ↗Academic sources and guidance for using the model.

How the Parts Fit Together

α\boldsymbol{\alpha}α

Loading (adjustment) matrix

The n×rn \times rn×r matrix of adjustment speeds. Column jjj of α\alphaα tells you how each variable responds to a unit deviation from the jjj-th cointegrating relationship.

The input is an n x T matrix of I(1) variables observed at uniform frequency. The practitioner must determine the cointegrating rank r: how many independent long-run equilibrium relationships exist among the n variables. The Johansen trace test and maximum eigenvalue test provide statistical guidance, but economic theory often has the stronger say - a system with n variables has at most n - 1 cointegrating relationships, and the actual number depends on how many common stochastic trends drive the data.

The model separates into two layers. The long-run layer is β′yt−1\beta' y_{t-1}β′yt−1​: an r×1r \times 1r×1 vector where each element is a linear combination of the nnn variables that is stationary. These are the equilibrium relationships. The short-run layer consists of the Γ\GammaΓ matrices on lagged differences, capturing transitory dynamics that wash out in the long run. The loading matrix α\alphaα connects the two layers: it tells you which variables adjust when the system is out of equilibrium and how fast. A zero row in α\alphaα means that variable does not respond to the equilibrium error - it is weakly exogenous with respect to the long-run parameters.

Estimation uses Johansen's reduced-rank regression (concentrated MLE). First, regress Δyt\Delta y_tΔyt​ and yt−1y_{t-1}yt−1​ on lagged differences to concentrate out the short-run parameters. Then solve a generalized eigenvalue problem on the residual moment matrices. The rrr largest eigenvalues give the cointegrating vectors (eigenvectors of β\betaβ), and the corresponding α\alphaα is recovered by OLS. The trace test statistic is −T∑i=r+1nlog⁡(1−λ^i)-T \sum_{i=r+1}^{n} \log(1 - \hat{\lambda}_i)−T∑i=r+1n​log(1−λ^i​); under the null of rank rrr, this follows a non-standard distribution tabulated by Johansen. Asymptotic critical values depend on the deterministic specification (constant inside or outside the cointegrating space, trend inside or outside).

Assumptions

Testable

Variables are I(1)

Each variable in the system is integrated of order 1: nonstationary in levels but stationary in first differences. The Johansen framework extends to I(2) systems but the standard implementation assumes I(1).

If violated:Including I(0) variables inflates the apparent cointegrating rank. Including I(2) variables causes the Johansen asymptotic distributions to be incorrect. Pre-test with ADF or KPSS unit root tests.

Testable

Cointegrating rank is correctly specified

The number of cointegrating relationships r is known or correctly estimated by the trace/max eigenvalue tests. The rank determines the dimension of the long-run equilibrium space.

If violated:Over-specifying r (too many cointegrating vectors) forces spurious stationary combinations into the equilibrium space, biasing short-run dynamics. Under-specifying r omits genuine long-run information, losing forecast accuracy at medium-to-long horizons.

Testable

Linearity of the cointegrating relationship

The equilibrium relationship β′yt\beta' y_tβ′yt​ is linear. No threshold cointegration, no smooth transition between regimes.

If violated:Transaction costs, policy regime switches, and other nonlinearities can make the equilibrium relationship piecewise linear or nonlinear. Linear VECM adjustment speeds will be biased if the true adjustment is asymmetric.

Testable

Gaussian innovations

ut∼N(0,Σ)u_t \sim N(0, \Sigma)ut​∼N(0,Σ). Required for the Johansen MLE to achieve its asymptotic distribution. The trace test has some robustness to non-normality but loses power under heavy tails.

If violated:Non-normal innovations affect the size of the trace and max eigenvalue tests. Bootstrap versions of the Johansen test (Cavaliere, Rahbek, Taylor 2012) are recommended when residuals show excess kurtosis or skewness.

Testable

No structural breaks in the cointegrating relationship

The cointegrating vectors β\betaβ and the adjustment speeds α\alphaα are constant over the sample period.

If violated:A structural break in β\betaβ means the long-run equilibrium has shifted - the old cointegrating vector is no longer valid. Johansen tests have low power against the alternative of a break in β\betaβ. Gregory-Hansen (1996) tests allow for a single break in the cointegrating relationship.

Testable

Sufficient lag length

The number of lagged differences p-1 is large enough to whiten the residuals. Selected by information criteria applied to the unrestricted VAR in levels before imposing the rank restriction.

If violated:Too few lags leave serial correlation in the residuals, invalidating the asymptotic distributions. Too many lags waste degrees of freedom. Standard practice: select p for the VAR in levels, then use p-1 lagged differences in the VECM.

Applications

The Bank of England uses VECM specifications to model the sterling exchange rate against a basket of currencies, anchored by the purchasing power parity cointegrating relationship. The error correction term measures the real exchange rate misalignment - how far sterling has deviated from its PPP-implied level. The loading coefficient on the exchange rate equation measures how fast sterling corrects: estimates typically suggest a half-life of 3-5 years for PPP deviations, consistent with the Rogoff (1996) consensus.

Money demand estimation is a canonical VECM application. The long-run demand for real money balances is a function of real income and the opportunity cost of holding money (the interest rate). The cointegrating vector recovers the income elasticity and interest semi-elasticity of money demand. The adjustment speed α\alphaα tells the central bank how quickly money holdings return to equilibrium after a shock. The ECB's monetary analysis pillar historically relied on VECM-based money demand equations to assess whether M3 growth was consistent with price stability.

Term structure models use VECMs to capture the comovement among interest rates at different maturities. If the expectations hypothesis holds, the spread between the 10-year and 1-year rate is stationary even though both rates are I(1). The cointegrating vector is the yield spread; the error correction term captures the tendency of the spread to revert to its long-run mean. The Bank for International Settlements uses VECM-based term structure analysis for monitoring global financial conditions.

VECMs fail when the cointegrating relationship is unstable, when the sample is too short for reliable rank determination (T<80T < 80T<80 for systems with n>4n > 4n>4), or when the adjustment to equilibrium is nonlinear. Exchange rate cointegration at short horizons (quarterly data, T<80T < 80T<80) is notoriously fragile: the PPP relationship holds over decades but the adjustment speed is so slow that finite-sample tests cannot distinguish it from a unit root. In these cases, a BVAR with sum-of-coefficients priors may be more robust.

Extensions and open questions

Variants

  • Two-step Engle-Granger VECM: estimate the cointegrating vector by OLS on the levels equation, then use the residual as the error correction term in a single-equation framework. Simpler but less efficient than Johansen's system approach.
  • Structural VECM (SVECM): imposes identifying restrictions on the contemporaneous impact matrix to recover structural shocks, analogous to structural VAR identification but within the cointegrated system.
  • Threshold VECM: allows the adjustment speed α\alphaα to differ depending on the size or sign of the equilibrium error. Captures asymmetric adjustment - e.g., prices adjust faster when above equilibrium than below.
  • Panel VECM: pools cointegrating relationships across cross-sectional units (countries, firms) to increase power. Pedroni (1999) and Westerlund (2007) provide panel cointegration tests.
  • Bayesian VECM (Villani 2005, Econometric Theory): places priors on the cointegrating space and the adjustment speeds, avoiding the discrete rank-testing step. The prior can express uncertainty about the rank itself.

Open questions

  • Rank determination in small samples remains unreliable. The Johansen trace test has notoriously low power when the largest eigenvalue is close to zero, which is common when adjustment to equilibrium is slow. Information criteria for rank selection (Cheng and Phillips 2009) are an alternative but not widely adopted.
  • Cointegration under structural breaks: if the cointegrating vector shifts (e.g., a regime change in monetary policy), standard Johansen tests can either over- or under-estimate the rank. Sequential break-date estimation within the VECM framework lacks a fully satisfactory asymptotic theory.
  • Nonlinear cointegration: many economic relationships (PPP, the term structure, money demand) show nonlinear mean reversion. The linear VECM imposes symmetric constant-speed adjustment. Smooth transition VECMs and threshold VECMs relax this but at the cost of many additional parameters and weak identification of the threshold.

Keep going

Read next

Continue with a related concept, model, or example.

Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
ヘルプ
プライバシーポリシー利用規約倫理とコンプライアンス

© 2026 Mark Jayson Nation LLC