Vector Error Correction Model

Definition

A vector error correction model (VECM) is a multivariate time-series model for a set of variables that are individually non-stationary — integrated of order one, $I(1)$ — but share one or more long-run equilibrium relationships (cointegration). It augments a vector autoregression (VAR) written in first differences with error-correction terms — the lagged deviations from the cointegrating relationships — so that whenever the system strays from equilibrium it is pulled back. VECM is the standard tool when one needs to capture both short-run dynamics and a binding long-run relationship at the same time, and in this wiki it is the leading framework for jointly forecasting related mortality populations without letting them diverge.

Key Ideas

Cointegration (Engle-Granger 1987): two or more $I(1)$ series are cointegrated if some linear combination of them is stationary, $I(0)$ . They then share a common stochastic trend and cannot drift arbitrarily far apart; the linear combination defines the long-run equilibrium.
The Granger representation theorem: if a set of variables is cointegrated, an error-correction representation exists (and conversely). This is what licenses writing the system as a VAR-in-differences plus correction terms rather than as a levels VAR.
Error-correction terms and adjustment speeds: the lagged equilibrium deviation enters each equation; its coefficient (the adjustment coefficient $\alpha$ ) governs how fast that variable moves back toward equilibrium. The signs and magnitudes of the $\alpha$ 's encode which series does the "correcting."
Cointegration rank ( $r$ ): the number of independent long-run relationships. It is a critical, sensitivity-driving modeling choice — $r = 0$ collapses the VECM to a VAR in pure differences (no long-run tie), while higher $r$ imposes more equilibrium structure.
VAR vs. VECM: a VAR in levels ignores the long-run constraint and can produce implausible divergence; a VAR in differences discards the equilibrium information entirely; a VECM keeps both the short-run dynamics and the long-run anchor.
Non-divergence by construction: because cointegrated series cannot separate without bound, a VECM is the natural model wherever divergence is implausible — most prominently, the mortality of closely related populations.

How It Works

For an $n$ -vector $Y_t$ , the VECM is:

$\Delta Y_t = \mu + \Pi Y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta Y_{t-i} + \varepsilon_t$

where the long-run matrix factors as $\Pi = \alpha \beta'$ with reduced rank $r < n$ : the columns of $\beta$ are the cointegrating vectors (so $\beta' Y_{t-1}$ is the vector of equilibrium deviations) and $\alpha$ holds the adjustment coefficients. The $\Gamma_i$ capture short-run dynamics.

Estimation is typically by Johansen maximum likelihood (which also supplies likelihood-ratio tests for the cointegration rank) or the Engle-Granger two-step procedure. Forecasts are generated by simulating the system forward; the error-correction terms continually pull deviating components back toward the shared trend, so forecast fans stay bounded relative to one another.

Why It Matters

Multi-population mortality forecasting: VECM is the leading method for jointly forecasting two or more related populations (e.g., a general population vs. insured lives) so that their mortality cannot diverge — grounding non-divergence in the cointegration structure of the Lee-Carter Model time factors rather than in an ad hoc "dominant population" assumption. It outperforms random-walk-plus-AR(1) and plain VAR alternatives on fit and out-of-sample tracking. See Multi-Population Mortality Modeling.
Longevity basis risk and capital: because the model is tail-driven, the choice among VECM, RWAR, and VAR can change Solvency II risk margins substantially even when the models agree on best-estimate forecasts. See Risk Margin.
Subpopulation mortality in social insurance: VECM is the appropriate tool for jointly projecting SSA-covered-worker and Disability Insurance–beneficiary mortality, or a widening earnings–mortality gradient, where single-population Lee-Carter would impose the wrong common drift. See SSA Mortality Forecasting.

Open Questions

How should the cointegration rank $r$ be selected robustly in the short, noisy samples typical of demographic data?
When subpopulation differentials are widening rather than mean-reverting, is a cointegration (non-divergence) assumption appropriate at all, or does it understate basis risk? See Risk Margin.