Structural Vector Error Correction Model (SVECM)

• SVECM is a multivariate time series framework that captures both short-run dynamics and long-run equilibrium by imposing theory-driven structural restrictions.
• It decomposes observed dynamics into interpretable components through techniques like impulse response functions and variance decompositions.
• The model is extensively applied in macroeconomics and environmental studies to infer causal relationships and validate equilibrium via cointegration analysis.

A Structural Vector Error Correction Model (SVECM) is a multivariate time series framework that integrates both short-term dynamics and long-term equilibrium relationships, while simultaneously enabling structural identification of underlying shocks. SVECMs extend classical Vector Error Correction Models (VECMs) by embedding structural assumptions—typically based on economic, physical, or domain-specific theory—into the innovation structure. This facilitates causal inference and the decomposition of observed dynamics into interpretable structural components, frequently through imposing contemporary restrictions or exploiting higher-order identification strategies. Applications span empirical macroeconomics, environmental sciences, and other fields characterized by cointegrated, non-stationary systems subjected to identifiable structural shocks (Nakano et al., 11 Jan 2026).

1. Model Specification

The SVECM builds upon the reduced-form VECM, which models the joint dynamics of integrated ($I(1)$) variables subject to cointegration, by incorporating restrictions that map reduced-form residuals to structurally meaningful innovations. The standard reduced-form VECM is defined as: $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$ where $y_t$ is a $K$-dimensional vector of $I(1)$ variables, $\Delta y_t$ denotes first differences, $\Pi = \alpha\beta'$ (with rank $r<K$) encapsulates long-run adjustment via loading coefficients $\alpha$ and cointegrating vectors $\beta$, $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$0 are short-run parameter matrices, $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$1 captures deterministic components, and $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$2 are reduced-form residuals.

The structural form then specifies

$$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$3

with $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$4 a contemporaneous impact matrix and $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$5 orthogonalized structural shocks. The identification of $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$6 requires external restrictions, grounded in theory or empirical regularities. In the paleoclimate context, $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$7 (Antarctic temperature anomaly, log atmospheric CO$$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$8, insolation), and the VECM is estimated with three lags and no extraneous deterministic terms beyond the cointegration constant (Nakano et al., 11 Jan 2026).

2. Structural Identification and Restriction Schemes

Structural identification in SVECM frameworks is achieved by imposing (often physically or economically motivated) a priori restrictions on the contemporaneous effects encoded in $$\Delta y_t = \Pi y_{t-1} + \sum_{i=1}^{p-1} \Gamma_i \Delta y_{t-i} + D_t + \varepsilon_t,$$9. For the trivariate climate SVECM, $y_t$0 is parameterized as follows: $y_t$1 with zero restrictions such as $y_t$2 (no instantaneous CO$y_t$3 Temperature feedback, reflecting oceanic inertia), $y_t$4 (insolation, $y_t$5, is strictly exogenous), and allowance for rapid T$y_t$6CO$y_t$7 interaction ($y_t$8) (Nakano et al., 11 Jan 2026). These constraints ensure structural shocks are uniquely recovered from reduced-form innovations, enabling credible causal inference. Such design is validated against cointegration and exogeneity tests, with the chosen exclusion restrictions justified by domain theory.

3. Cointegration Analysis and Long-Run Relationships

Cointegration establishes statistically robust, long-run equilibrium relationships among non-stationary variables modeled by SVECMs. For the paleoclimate application, the Johansen method identifies cointegration rank $y_t$9, implying one binding equilibrium: $K$0 A 1% increase in CO$K$1 ($K$2) mandates a temperature increase of approximately $K$3°C to preserve equilibrium. The magnitude of the long-run coefficient corresponds to an Earth System Sensitivity (ESS) of $K$4C per CO$K$5 doubling, with the statistical significance validated through trace tests and $K$6-statistics on the cointegrating vector ($K$7, $K$8) (Nakano et al., 11 Jan 2026). The model further demonstrates robustness to lag order and deterministic term specification.

4. Short-Run Dynamics and Impulse Response Analysis

Short-run adjustment is governed by the loading vector $K$9, quantifying how deviations from equilibrium dissipate over time. For temperature anomalies, the estimated loading of $I(1)$0 ($I(1)$1) denotes mean reversion at approximately $I(1)$2 per century. CO$I(1)$3 levels actively adjust as well, while insolation remains strictly exogenous and non-adjusting.

Structural impulse response functions (IRFs) provide dynamic profiles of system variables following exogenous shocks:

• A structural CO$I(1)$4 shock induces a gradual, persistent temperature rise, asymptoting to $I(1)$5C per $I(1)$6 CO$I(1)$7 (over $I(1)$8 centuries).
• A structural temperature shock yields an immediate $I(1)$91.1% $\Delta y_t$0 jump, decaying with a half-life of $\Delta y_t$1200 years.

IRFs are calculated via the companion form representation, exploiting the identified $\Delta y_t$2 and reduced-form dynamics (Nakano et al., 11 Jan 2026).

5. Forecast Error Variance Decomposition (FEVD) and Causal Attribution

FEVD quantifies the contribution of specific structural shocks to the forecast error variance of each endogenous variable at varying horizons. For temperature,

$\Delta y_t$3

where $\Delta y_t$4 selects temperature, and $\Delta y_t$5, $\Delta y_t$6 are matrices mapping shocks to outcomes. In the long run ($\Delta y_t$7), structural CO$\Delta y_t$8 shocks account for approximately $\Delta y_t$9 of temperature's forecast error variance, confirming CO$\Pi = \alpha\beta'$0 as a primary contributor to long-term climate variability in the system (Nakano et al., 11 Jan 2026).

6. Empirical Implementation, Robustness, and Extensions

Empirical implementation consists of sequentially verifying integration order via unit-root tests (all series $\Pi = \alpha\beta'$1), applying Engle–Granger and Johansen procedures for cointegration (confirming $\Pi = \alpha\beta'$2), and performing estimation with lag order selected by the AIC (here, $\Pi = \alpha\beta'$3). The estimated long-run relationship and dynamic properties exhibit strong robustness to lag specification and deterministic trend inclusion. Sensitivity analyses reveal that modifying structural restrictions, such as imposing $\Pi = \alpha\beta'$4 to exclude T$\Pi = \alpha\beta'$5CO$\Pi = \alpha\beta'$6 feedback, inflates the ESS estimate by approximately $\Pi = \alpha\beta'$7C, underscoring the implications of identification assumptions (Nakano et al., 11 Jan 2026).

SVECMs thus provide a comprehensive, replicable framework for decomposing non-stationary systems into interpretable equilibrium paths and short-run causal mechanisms. Applications extend wherever cointegrated dynamics and theory-motivated structural inquiries coalesce.