Skip to main content

Cookies on Macro by Mark.We use essential cookies for sign-in, locale, and security. With your permission, we also use optional cookies for preferences, for analytics, and for advertising measurement with the Meta (Facebook) Pixel. Read our cookie policy.

Macro by MarkMacro by Mark
  • Home
Loading…
Loading…
← Models
MacroConceptsPolicyModels

Network spillover Model

Network-based decomposition of forecast-error variance into directional spillovers across sectors or markets.

OverviewGraphMathReferences

Understand the mathematics

Follow the same equations and assumptions one section at a time or in a continuous reading.

Hover or focus a symbol defined in this model to find its exact matches. Select it to keep the highlight; Escape clears the selection.

VAR Estimation and Generalized Forecast-Error Variance Decomposition

The Diebold-Yilmaz framework starts with a standard VAR and moves to its Wold (infinite-order moving average) representation. The generalized FEVD, due to Pesaran and Shin (1998), allocates forecast-error variance to each variable's innovations without depending on the ordering of variables in the system.

This order-invariance is critical when there is no theoretical ordering - as is typically the case when the variables are equity markets or sovereign spreads for 20 countries.

yt=c+A1yt−1+⋯+Apyt−p+et,et∼(0,Σ)y_t = c + A_1 y_{t-1} + \cdots + A_p y_{t-p} + e_t, \quad e_t \sim (0, \Sigma)yt​=c+A1​yt−1​+⋯+Ap​yt−p​+et​,et​∼(0,Σ)
Σ\SigmaΣ

Standard VAR(ppp) in NNN variables. ccc is an N×1N \times 1N×1 intercept, AjA_jAj​ are N×NN \times NN×N coefficient matrices, Σ\SigmaΣ is the N×NN \times NN×N innovation covariance.

yt=μ+∑h=0∞Φhet−h,Φ0=INy_t = \mu + \sum_{h=0}^{\infty} \Phi_h e_{t-h}, \quad \Phi_0 = I_Nyt​=μ+h=0∑∞​Φh​et−h​,Φ0​=IN​
Φh\Phi_hΦh​

Wold MA representation. The matrices Φh\Phi_hΦh​ can be computed recursively from the VAR coefficients: Φh=∑j=1min⁡(h,p)AjΦh−j\Phi_h = \sum_{j=1}^{\min(h,p)} A_j \Phi_{h-j}Φh​=∑j=1min(h,p)​Aj​Φh−j​.

θ~ijH=σjj−1∑h=0H−1(ei′ΦhΣej)2∑h=0H−1ei′ΦhΣΦh′ei\tilde{\theta}^H_{ij} = \frac{\sigma_{jj}^{-1} \sum_{h=0}^{H-1} (e_i' \Phi_h \Sigma e_j)^2}{\sum_{h=0}^{H-1} e_i' \Phi_h \Sigma \Phi_h' e_i}θ~ijH​=∑h=0H−1​ei​′Φh​ΣΦh​′ei​σjj−1​∑h=0H−1​(ei​′Φh​Σej​)2​
θ~ijH\tilde{\theta}^H_{ij}θ~ijH​HHHeie_iei​Φh\Phi_hΦh​Σ\SigmaΣ

Unnormalized generalized FEVD. The numerator measures how much of iii's HHH-step forecast-error variance is driven by jjj's innovations, accounting for contemporaneous correlation through Σ\SigmaΣ. The denominator is iii's total HHH-step forecast-error variance.

1

The forecast error is a linear combination of future innovations, weighted by the MA coefficients.
The H-step forecast error for variable i:yi,t+H−y^i,t+H∣t=∑h=0H−1ei′Φhet+H−h\text{The } H\text{-step forecast error for variable } i: \quad y_{i,t+H} - \hat{y}_{i,t+H|t} = \sum_{h=0}^{H-1} e_i' \Phi_h e_{t+H-h}The H-step forecast error for variable i:yi,t+H​−y^​i,t+H∣t​=h=0∑H−1​ei​′Φh​et+H−h​
HHHeie_iei​Φh\Phi_hΦh​

2

The total forecast-error variance sums the contributions from all horizons, accounting for innovation covariance.
Variance: Var(yi,t+H−y^i,t+H∣t)=∑h=0H−1ei′ΦhΣΦh′ei\text{Variance: } \text{Var}(y_{i,t+H} - \hat{y}_{i,t+H|t}) = \sum_{h=0}^{H-1} e_i' \Phi_h \Sigma \Phi_h' e_iVariance: Var(yi,t+H​−y^​i,t+H∣t​)=h=0∑H−1​ei​′Φh​ΣΦh​′ei​
HHHeie_iei​Φh\Phi_hΦh​Σ\SigmaΣ

3

The generalized approach uses the innovation to variable jjj and accounts for the fact that other variables' innovations may be correlated. The shock is the expected change in ete_tet​ given a one-standard-deviation shock to ej,te_{j,t}ej,t​, under the joint distribution.
Contribution of variable j’s innovation: shock δj=σjj−1/2Σej\text{Contribution of variable } j\text{'s innovation: shock } \delta_j = \sigma_{jj}^{-1/2} \Sigma e_jContribution of variable j’s innovation: shock δj​=σjj−1/2​Σej​
Σ\SigmaΣ

4

The squared impulse of the generalized shock to variable jjj on variable iii, accumulated over HHH horizons.
Numerator: σjj−1∑h=0H−1(ei′ΦhΣej)2\text{Numerator: } \sigma_{jj}^{-1} \sum_{h=0}^{H-1} (e_i' \Phi_h \Sigma e_j)^2Numerator: σjj−1​h=0∑H−1​(ei​′Φh​Σej​)2
HHHeie_iei​Φh\Phi_hΦh​Σ\SigmaΣ

5

Row-normalization ensures each row sums to 1. This is a convenience, not a structural result - the unnormalized rows do not sum to 1 because the generalized shocks are not orthogonal.
θijH=θ~ijH  /  ∑k=1Nθ~ikH\theta^H_{ij} = \tilde{\theta}^H_{ij} \;/\; \sum_{k=1}^N \tilde{\theta}^H_{ik}θijH​=θ~ijH​/k=1∑N​θ~ikH​
θijH\theta^H_{ij}θijH​θ~ijH\tilde{\theta}^H_{ij}θ~ijH​NNNHHH

6

The full connectedness table. Diagonal: own variance shares. Off-diagonal: pairwise spillovers. Each row sums to 1 by construction. Column sums vary and reflect each variable's importance as a source of shocks.
Spillover table: ΘH=[θijH]N×N\text{Spillover table: } \Theta^H = [\theta^H_{ij}]_{N \times N}Spillover table: ΘH=[θijH​]N×N​
HHHθijH\theta^H_{ij}θijH​

GFEVD vs. Cholesky FEVD

The Cholesky FEVD orthogonalizes innovations using the Cholesky factor of Σ\SigmaΣ, making the result dependent on variable ordering. With N=10N = 10N=10 variables, there are 10!=3,628,80010! = 3{,}628{,}80010!=3,628,800 possible orderings, each giving a different decomposition. The GFEVD avoids this by not orthogonalizing, but at the cost of non-orthogonal shocks whose contributions do not sum to exactly 1 before normalization.

Notations

Φh\Phi_hΦh​

Moving-average coefficient matrix

Σ\SigmaΣ

Innovation covariance

eie_iei​

Basis vector

σjj\sigma_{jj}σjj​

Innovation variance

θ~ijH\tilde{\theta}^H_{ij}θ~ijH​

Unnormalized variance share

θijH\theta^H_{ij}θijH​

Normalized variance share

About these symbols

Moving-average coefficient matrix: N×NN \times NN×N MA coefficient matrix at horizon hhh

Innovation covariance: N×NN \times NN×N innovation covariance

Basis vector: iii-th standard basis vector (N×1N \times 1N×1)

Innovation variance: variance of variable jjj's innovation (diagonal of Σ\SigmaΣ)

Unnormalized variance share: unnormalized GFEVD element

Normalized variance share: row-normalized GFEVD element

    Transition checks

    Sampled dynamics
    zt=G(zt−1;θ)+ηt,yt=Hzt+εtz_t=G(z_{t-1};\theta)+\eta_t,\qquad y_t=H z_t+\varepsilon_tzt​=G(zt−1​;θ)+ηt​,yt​=Hzt​+εt​
    HHH

    The state is updated at observed dates. A line between dates is a display convention unless the model specifies behavior inside the interval.

    Read adjacent points as adjacent sample calculations. Smoothness is a property of the estimator, not of the chart line.

    Do not infer within-period behavior from a sampled estimator.

    Variable timing

    Latent state

    zt=G(zt−1;θ)+ηtz_t=G(z_{t-1};\theta)+\eta_tzt​=G(zt−1​;θ)+ηt​

    The state updates at sample dates. The visible line is only as smooth as the estimated state sequence.

    Observation equation

    yt=Hzt+εty_t=Hz_t+\varepsilon_tyt​=Hzt​+εt​
    HHH

    Observed series can move differently from latent states because measurement noise and loadings matter.

    Keep going

    Read next

    Continue with a related concept, model, or example.

    Agent-based model

    Supply-Chain / Network ABM

    a contrasting take on Network Spillover (Diebold-Yilmaz)

    Contrast it →
    Theory model

    IS-LM Model

    a well-connected theoretical model two steps from here

    New angle: theoretical model →
    Macro by Mark
    Help
    Privacy PolicyTerms of UseEthics and Compliance

    © 2026 Mark Jayson Nation LLC