Forecast Error Variance Decomposition (FEVD)

• FEVD is a statistical tool that decomposes forecast error variance into contributions from individual shocks, clarifying dynamic spillover effects.
• It employs classical VAR models and advanced nonlinear, Bayesian, or sparse frameworks to measure time-dependent influences in system dynamics.
• Extensions like Hermite FEVD and Bayesian methods enhance uncertainty quantification and interpretability in high-dimensional financial and macroeconomic applications.

Forecast Error Variance Decomposition (FEVD) is a fundamental analytic tool in multivariate time series analysis, providing a principled breakdown of a system’s $h$-step forecast error variance into contributions attributable to shocks in individual system components. FEVD plays a central role in econometric inference on dynamic spillovers, systemic risk, financial connectedness, and structural transmission mechanisms. Initially developed for linear Vector Autoregressive (VAR) models, modern research generalizes FEVD to nonlinear, high-dimensional, and sparse settings and embeds it within Bayesian and network-theoretic frameworks (Deng et al., 2022, Lin et al., 26 Jul 2025, Lee, 14 Mar 2025, Ouyang et al., 7 Jan 2026).

1. Mathematical Formulation and Classical FEVD

Consider a stationary VAR($p$) process: $$y_t = c + \sum_{\ell=1}^p\Phi_\ell\, y_{t-\ell} + \varepsilon_t, \qquad \varepsilon_t \sim (0, \Sigma)$$ The process admits a vector moving-average (VMA) representation: $$y_t = \mu + \sum_{k=0}^\infty \Psi_k \varepsilon_{t-k}$$ The $h$-step forecast error is $$e_{t+h} = y_{t+h} - \mathbb{E}[y_{t+h}| \mathcal{F}_t] = \sum_{s=0}^{h-1} \Psi_s \varepsilon_{t+h-s}$$, and its variance is $$\mathrm{Var}(e_{t+h}) = \sum_{s=0}^{h-1} \Psi_s \Sigma \Psi_s^\top$$.

The FEVD quantifies, for each target variable $i$ and source $j$, the proportion of $i$’s $h$-step forecast error variance attributable to innovations in $j$: $$\theta_{ij}^{(h)} = \frac{ \displaystyle\sum_{s=0}^{h-1} [e_i^\top \Psi_s \Sigma e_j]^2 } { \displaystyle\sum_{s=0}^{h-1} e_i^\top \Psi_s \Sigma \Psi_s^\top e_i }$$ where $e_i$ is the $i$th unit selection vector. By construction $\sum_j \theta_{ij}^{(h)} = 1$ (up to rounding).

In the "generalized" FEVD (GFEVD) (Ouyang et al., 7 Jan 2026), which accommodates correlated innovations and is invariant to variable ordering, the numerator becomes $$\sigma_{jj}^{-1} \sum_{s=0}^{h-1} [e_i^\top \Psi_s \Sigma e_j]^2$$ with $\sigma_{jj}$ denoting the $j$th diagonal of $\Sigma$.

This decomposition respects both the ordering of innovation arrival and the pathway of signal transmission.

2. Horizon Dependence and Dynamic Interpretation

The forecast horizon $h$ in FEVD parameterizes the temporal depth of variance decomposition.

• For $h=1$, FEVD recovers immediate (“Granger-causal”) effects.
• For intermediate $h$, both direct and indirect spillovers—mediated by intermediary nodes and higher-order lags—are captured.
• As $h\to h^*$, under system stationarity, MA coefficients $\Psi_s$ decay and FEVD estimates converge, representing the system’s long-run equilibrium variance shares (Deng et al., 2022, Lin et al., 26 Jul 2025).

Empirical analyses frequently determine an “equilibrium” horizon $h^*$ by monitoring the maximum across edges of the incremental change in directed FEVD weights, stopping when changes fall below a numerical threshold:

$$\max_{j,k} |\bar{s}_h^{k \to j} - \bar{s}_{h-1}^{k \to j}| < \varepsilon$$

where $\bar{s}_h^{k \to j}$ denotes a (possibly posterior-averaged) normalized FEVD edge (Deng et al., 2022).

Choice of $h$ is nontrivial: small $h$ omits indirect flows; large $h$ amplifies estimation error; equilibrium $h^*$ balances these considerations (Lin et al., 26 Jul 2025).

3. Extensions: Nonlinear, Bayesian, and Sparse FEVD

Recent work extends FEVD beyond the linear-Gaussian VAR:

• Nonlinear FEVD (HFEVD): For nonlinear SVARs, the Hermite FEVD (HFEVD) (Lee, 14 Mar 2025) uses Hermite polynomial expansions to express the conditional variance of the $h$-step forecast error in terms of component-wise, horizon-wise, and degree-of-nonlinearity contributions:

$$\mathrm{Var}[Y_{t+h} | Y_t] = \sum_{K \neq 0} \left( c_K(Y_t)\, \prod_{i} k_i! \right)^2$$

Multi-index $K$ indexes both time and innovation component, separating marginal and interaction contributions. HFEVD generalizes linear FEVD’s history invariance, allowing risk allocation in nonlinear state evolution.

• Bayesian FEVD and Uncertainty Quantification: Bayesian Spillover Graphs (BSG) (Deng et al., 2022) employ conjugate priors on the VAR coefficients, sample from the joint posterior, and compute FEVD for each posterior draw. Posterior means yield estimated edge weights, while credible intervals are constructed from empirical quantiles across samples. This allows honest quantification of posterior uncertainty in all FEVD-based network measures.
• Sparse FEVD for Network Parsimony: High-dimensional FEVD networks are typically dense, impeding interpretability and economic application. Sparse FEVD estimation (Ouyang et al., 7 Jan 2026) recasts the decomposition as a regression problem and selects active spillover channels by penalized information criteria. The procedure retains links with largest variance shares, shrinking negligible influences to zero, and tunes penalization by pseudo-out-of-sample forecast performance.

4. Algorithmic and Computational Aspects

Standard FEVD computation proceeds as follows (Lin et al., 26 Jul 2025, Deng et al., 2022):

1. Estimate the VAR(p) (frequentist: OLS, Bayesian: posterior sampling).
2. Compute the MA (impulse-response) coefficients $\{\Psi_s\}_{s=0}^{h-1}$ recursively.
3. Estimate the covariance $\Sigma$ of innovations.
4. Compute, for each $i,j$:
   • Numerator: $$N_{ij}(h) = \sum_{s=0}^{h-1} [e_i^\top \Psi_s \Sigma e_j]^2$$
   • Denominator: $$D_i(h) = \sum_{s=0}^{h-1} e_i^\top \Psi_s \Sigma \Psi_s^\top e_i$$
   • Proportion: $\theta_{ij}(h) = N_{ij}(h) / D_i(h)$
5. Optionally, for network analysis or portfolio construction, transform FEVD shares into edge weights or costs (e.g., $C_{i \to j} = 1 - \theta_{j \leftarrow i}(h)$), symmetrize as needed, and apply graph algorithms (Lin et al., 26 Jul 2025).

In Bayesian settings, average shares and construct credible intervals over posterior samples (Deng et al., 2022). For sparse FEVD, retain largest off-diagonal elements according to an information criterion penalty (Ouyang et al., 7 Jan 2026).

Numerical safeguards (e.g., positive-definite covariance adaptations, stability checks for denominators, clipping small values) are essential in large or near-singular systems (Lin et al., 26 Jul 2025).

5. Applications in Finance and Macroeconomics

FEVD underpins several applied domains:

• Financial Connectedness and Spillover Networks: FEVD quantifies the share of uncertainty transmitted from one financial instrument or market to another, providing a natural basis for directed network construction. Practitioners use these weights to extract market backbone structures (e.g., via Minimum Spanning Tree algorithms) and to identify systemic “source” and “sink” nodes (Lin et al., 26 Jul 2025, Deng et al., 2022, Ouyang et al., 7 Jan 2026).
• Dynamic Portfolio Allocation: By identifying stocks or assets central in the FEVD-derived network, adaptive portfolios can be constructed that exploit interdependence and shock propagation, often outperforming naive allocation schemes (Lin et al., 26 Jul 2025).
• Systemic Risk and Macroeconomic Transmission: In macroeconomics, FEVD reveals the main transmitters and receivers of structural shocks (e.g., fiscal, monetary, volatility). Nonlinear HFEVD can further isolate risk components due to nonlinearities or interactions (Lee, 14 Mar 2025).
• Network Sparsification for Interpretation: Sparse FEVD frameworks yield interpretable, economically plausible networks by retaining only the most statistically and economically significant transmission routes (Ouyang et al., 7 Jan 2026).

6. Limitations, Extensions, and Comparisons

The classical FEVD assumes linearity and Gaussianity, which may fail in systems with regime-switching, nonlinear feedback, or heavy-tailed innovations. The HFEVD framework directly isolates nonlinear and interaction effects and reveals when linear analysis misallocates forecast risk (Lee, 14 Mar 2025).

Dense FEVD-linked networks can obscure meaningful spillover channels; sparse estimation provides drastic improvement in structural interpretability without compromising forecast accuracy (Ouyang et al., 7 Jan 2026). However, larger forecast horizons $h$ can increase estimation error and introduce indirect links that may not be economically meaningful.

In the statistical learning literature, multi-step forecast error decompositions parallel FEVD at the model-uncertainty level, partitioning expected error into irreducible noise, structural approximation (bias), and estimation variance. In this analogy, the “irreducible noise” term corresponds to the FEVD share from unpredictable shocks, model approximation gap to omitted structural effects, and estimation variance to uncertainty in estimated impulse weights (Green et al., 14 Nov 2025). This connection underscores FEVD’s foundational role in both system-theoretic and predictive modeling paradigms.

7. Summary Table: Representative FEVD Variants

FEVD Variant	Key Features	Relevant References
Classical	Linear VAR; VMA form; historical invariance	(Lin et al., 26 Jul 2025, Deng et al., 2022)
Generalized	Correlated shocks; order-invariance	(Ouyang et al., 7 Jan 2026, Deng et al., 2022)
Hermite (HFEVD)	Nonlinear SVARs; horizon, component, and nonlinearity separation	(Lee, 14 Mar 2025)
Bayesian	Full posterior uncertainty; credible intervals; BSG	(Deng et al., 2022)
Sparse	Information-criterion selection; interpretable networks	(Ouyang et al., 7 Jan 2026)

All modern FEVD implementations rest on the rigorous partitioning of forecast error variance, but differ in assumptions, inferential approach, and suitability for complex system diagnostics.