Vector Error Correction Model (VECM)

The Error Correction Term is the defining mechanism of VECM, measuring the deviation from the long-run equilibrium in the previous period.

• Mechanism: The ECT is the lagged residual from the cointegrating regression (e.g., ECT_{t-1} = Y_{t-1} - βX_{t-1}).
• Speed of Adjustment: The coefficient α on the ECT dictates how quickly the system corrects a disequilibrium. A negative α (between -1 and 0) ensures convergence.
• Economic Logic: If two assets are cointegrated, a positive ECT implies Y is overvalued relative to X, triggering a short-sale signal in Y and a long signal in X.

VECM separates the data-generating process into two distinct components to isolate permanent trends from transient noise.

• Long-Run Matrix (Π): This matrix is decomposed as Π = αβ', where β contains the cointegrating vectors defining the stationary equilibrium relations, and α contains the loading weights for adjustment speed.
• Short-Run Dynamics (Γ): The Γ_i matrices capture the transitory effects of lagged differenced variables, representing the immediate momentum and mean-reversion tendencies.
• Deterministic Terms: Constants and trends can be restricted to lie inside the cointegrating space (restricted trend) or outside it (unrestricted drift).

Unlike a standard VAR in differences, a VECM imposes long-run neutrality constraints on the impulse response functions (IRFs).

• Persistence: Shocks to a cointegrated system have transitory effects on the equilibrium error; the IRF of the cointegrating relation must decay to zero.
• Permanent Components: The model identifies k - r common stochastic trends driving the permanent shifts in the system.
• Forecasting Superiority: By enforcing the long-run equilibrium, VECM forecasts do not diverge over long horizons, unlike unrestricted VARs which accumulate compounding drift errors.

VECM is the mathematical backbone of statistical arbitrage, specifically pairs trading, where a spread between two assets is assumed mean-reverting.

• Spread Calculation: The cointegrating vector β defines the hedge ratio. If β = (1, -1.5), the trader buys 1 share of Asset A and shorts 1.5 shares of Asset B.
• Signal Generation: When the spread (ECT) exceeds 2 standard deviations, the VECM predicts a correction. The α coefficient determines the expected half-life of this reversion.
• Risk Management: A breakdown in cointegration (structural break) invalidates the VECM; traders monitor recursive eigenvalue stability tests to exit positions.

FEVD in a VECM context quantifies how much of the forecast error variance of a variable is attributable to shocks to the permanent vs. transitory components.

• Permanent Shocks: Shocks to the common stochastic trends explain the long-horizon variance of the levels of the variables.
• Transitory Shocks: Shocks to the cointegrating relations explain the short-horizon variance and have zero long-run impact.
• Application: In macroeconomics, this distinguishes whether inflation volatility is driven by supply-side (permanent) or demand-side (transitory) disturbances.

The statistical bedrock upon which VECM is built. Two or more non-stationary time series are cointegrated if a linear combination of them is stationary, implying a long-run equilibrium relationship. Without cointegration, a VECM is misspecified and reduces to a standard VAR in first differences. The Engle-Granger two-step method and the Johansen test are the primary frameworks for detecting this property before model estimation.

The unrestricted parent model of VECM. A VAR models each variable as a linear function of its own past values and the past values of all other variables in the system. VECM is a restricted VAR that explicitly models the error correction term (ECT) to capture the speed of adjustment back to equilibrium. If variables are cointegrated, a VAR in first differences alone is misspecified due to omitted variable bias.

A primary diagnostic tool for analyzing estimated VECM dynamics. An IRF traces the time path of a one-unit exogenous shock to one variable's error term on all endogenous variables in the system. In a VECM context, IRFs reveal whether shocks have transitory or permanent effects and how quickly the system reverts to its long-run equilibrium after a perturbation.

A fundamental prerequisite for standard statistical inference. A stationary process has a constant mean, variance, and autocorrelation structure over time. VECM is specifically designed for I(1) non-stationary variables that become stationary after first differencing. Failing to account for unit roots leads to spurious regression, where unrelated trending variables appear highly correlated.

A predictive, not structural, causality concept. A variable X Granger-causes Y if past values of X contain statistically significant information that improves the forecast of Y beyond using past values of Y alone. In a VECM framework, Granger causality can operate through two distinct channels: the short-run lagged differences and the long-run error correction term, allowing for richer causal inference.

A critical violation of the classical linear regression assumptions where an explanatory variable is correlated with the error term. In VECM systems, endogeneity often arises from simultaneity bias—where variables mutually influence each other. The model addresses this by treating all variables as jointly endogenous, but structural identification for causal interpretation requires additional restrictions via Cholesky decompositions or sign restrictions.