跳到主要内容

Macro by Mark 的 Cookie.我们使用必要 Cookie 处理登录、语言和安全。经您同意,我们还会将可选 Cookie 用于保存偏好、分析,以及通过 Meta(Facebook)Pixel 衡量广告效果。请阅读我们的Cookie 政策。

Macro by MarkMacro by Mark
  • 首页
加载中…
加载中…
← 模型
Macro概念政策模型

Dynamic factor model

Extracts a small number of latent factors from many macro indicators and lets those factors drive a current-quarter read.

概述图表数学参考资料

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.

Static Factor Model and Principal Components Estimation

Start with the static representation. The n×1n \times 1n×1 vector of observed variables xtx_txt​ is decomposed into a common component Λft\Lambda f_tΛft​ and an idiosyncratic component ete_tet​. The factors ftf_tft​ are r×1r \times 1r×1, the loadings Λ\LambdaΛ are n×rn \times rn×r, and the idiosyncratic errors ete_tet​ are n×1n \times 1n×1. The key identifying restriction: the common and idiosyncratic components are orthogonal.

Principal components estimation does not require specifying factor dynamics or the distribution of the errors. Stack observations into a T×nT \times nT×n matrix XXX. The first rrr principal components of XXX are consistent estimators of the factor space as both nnn and TTT grow. The eigendecomposition of the n×nn \times nn×n sample covariance matrix Σ^=X′X/T\hat{\Sigma} = X'X / TΣ^=X′X/T yields eigenvectors v1,…,vnv_1, \ldots, v_nv1​,…,vn​ with eigenvalues λ1≥…≥λn\lambda_1 \geq \ldots \geq \lambda_nλ1​≥…≥λn​. The estimated loadings are Λ^=n[v1,…,vr]\hat{\Lambda} = \sqrt{n} [v_1, \ldots, v_r]Λ^=n​[v1​,…,vr​], and the estimated factors are f^t=Λ^′xt/n\hat{f}_t = \hat{\Lambda}' x_t / nf^​t​=Λ^′xt​/n.

xt=Λft+et,E[ftet′]=0\mathbf{x}_t = \boldsymbol{\Lambda} \mathbf{f}_t + \mathbf{e}_t, \quad E[\mathbf{f}_t \mathbf{e}_t'] = \mathbf{0}xt​=Λft​+et​,E[ft​et​′]=0
Λ\boldsymbol{\Lambda}Λft\mathbf{f}_tft​et\mathbf{e}_tet​

The static factor model. Each variable is a linear combination of r common factors plus idiosyncratic noise.

Σx=ΛΣfΛ′+Σe\boldsymbol{\Sigma}_x = \boldsymbol{\Lambda} \boldsymbol{\Sigma}_f \boldsymbol{\Lambda}' + \boldsymbol{\Sigma}_eΣx​=ΛΣf​Λ′+Σe​
Λ\boldsymbol{\Lambda}Λ

The covariance decomposition. The first rrr eigenvalues of Σx\Sigma_xΣx​ capture the common component.

Λ^=n⋅[v^1,…,v^r]\hat{\boldsymbol{\Lambda}} = \sqrt{n} \cdot [\hat{\mathbf{v}}_1, \ldots, \hat{\mathbf{v}}_r]Λ^=n​⋅[v^1​,…,v^r​]
nnn

The PCA estimate of the loadings. The n\sqrt{n}n​ normalization ensures Λ′Λ/n=Ir\Lambda'\Lambda/n = I_rΛ′Λ/n=Ir​.

f^t=1nΛ^′xt\hat{\mathbf{f}}_t = \frac{1}{n} \hat{\boldsymbol{\Lambda}}' \mathbf{x}_tf^t​=n1​Λ^′xt​
nnn

The estimated factors are weighted averages of the observed variables, with weights given by the loadings.

IC(r)=ln⁡(1nT∑i=1n∑t=1T(xit−λ^i′f^t)2)+r⋅g(n,T)IC(r) = \ln\left(\frac{1}{nT}\sum_{i=1}^{n}\sum_{t=1}^{T}(x_{it} - \hat{\lambda}_i' \hat{f}_t)^2\right) + r \cdot g(n, T)IC(r)=ln(nT1​i=1∑n​t=1∑T​(xit​−λ^i′​f^​t​)2)+r⋅g(n,T)
rrrnnn

Information criterion for selecting the number of factors. Minimize IC⁡(r)\operatorname{IC}(r)IC(r) over r=0,1,…,rmax⁡r = 0, 1, \ldots, r_{\max}r=0,1,…,rmax​.

Standardize the panel

PCA is scale-dependent; variables with larger variance dominate the first component unless standardized
x~it=xit−xˉisi,si=sample std dev of variable i\tilde{x}_{it} = \frac{x_{it} - \bar{x}_i}{s_i}, \quad s_i = \text{sample std dev of variable } ix~it​=si​xit​−xˉi​​,si​=sample std dev of variable i

Compute the sample covariance matrix

The eigendecomposition of this matrix identifies the factor space
Σ^=1TX~′X~,X~ is T×n\hat{\boldsymbol{\Sigma}} = \frac{1}{T} \tilde{\mathbf{X}}' \tilde{\mathbf{X}}, \quad \tilde{\mathbf{X}} \text{ is } T \times nΣ^=T1​X~′X~,X~ is T×n
nnn

Eigendecompose the covariance matrix

The largest eigenvalues correspond to the common factors; the rest to idiosyncratic variation
Σ^=∑j=1nλ^jv^jv^j′,λ^1≥λ^2≥⋯≥λ^n\hat{\boldsymbol{\Sigma}} = \sum_{j=1}^{n} \hat{\lambda}_j \hat{\mathbf{v}}_j \hat{\mathbf{v}}_j', \quad \hat{\lambda}_1 \geq \hat{\lambda}_2 \geq \cdots \geq \hat{\lambda}_nΣ^=j=1∑n​λ^j​v^j​v^j​′,λ^1​≥λ^2​≥⋯≥λ^n​
nnnλ^j\hat{\lambda}_jλ^j​v^j\hat{\mathbf{v}}_jv^j​

Select the number of factors r

Too few factors omit common variation; too many incorporate noise
r^=arg⁡min⁡r≤rmax⁡IC(r)\hat{r} = \arg\min_{r \leq r_{\max}} IC(r)r^=argr≤rmax​min​IC(r)
rrr

Extract the estimated factors and loadings

The first r eigenvectors span the factor space
Λ^=n[v^1,…,v^r],f^t=1nΛ^′x~t\hat{\boldsymbol{\Lambda}} = \sqrt{n}[\hat{\mathbf{v}}_1, \ldots, \hat{\mathbf{v}}_r], \quad \hat{\mathbf{f}}_t = \frac{1}{n}\hat{\boldsymbol{\Lambda}}'\tilde{\mathbf{x}}_tΛ^=n​[v^1​,…,v^r​],f^t​=n1​Λ^′x~t​
nnn

Verify consistency conditions

The PCA estimator is consistent as min⁡(n,T)→∞\min(n,T) \to \inftymin(n,T)→∞ under the approximate factor model
1T∑t=1T∥f^t−Hft∥2→p0 as n,T→∞\frac{1}{T}\sum_{t=1}^{T}\|\hat{\mathbf{f}}_t - \mathbf{H}\mathbf{f}_t\|^2 \xrightarrow{p} 0 \text{ as } n, T \to \inftyT1​t=1∑T​∥f^t​−Hft​∥2p​0 as n,T→∞
H\mathbf{H}Hft\mathbf{f}_tft​nnn

Rotation indeterminacy

PCA identifies the factor space but not the individual factors. For any orthogonal r×rr \times rr×r matrix H,ΛHH, \Lambda HH,ΛH and H′ftH' f_tH′ft​ give the same fit. Naming factors (calling the first one a 'demand factor') requires additional restrictions or structural identification beyond PCA.

Notations

λ^j\hat{\lambda}_jλ^j​

Sample covariance eigenvalue

v^j\hat{\mathbf{v}}_jv^j​

Eigenvector

H\mathbf{H}H

Factor rotation matrix

About these symbols

Sample covariance eigenvalue: jjj-th largest eigenvalue of the sample covariance matrix (not to be confused with factor loadings Λ\LambdaΛ)

Eigenvector: Eigenvector corresponding to the j-th eigenvalue

Factor rotation matrix: Rotation matrix: factors are identified only up to this orthogonal transformation

    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​

    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​

    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.

    Empirical model

    ARDL / Bounds Testing

    a contrasting take on Dynamic Factor Model

    Contrast it →
    Empirical model

    BVAR

    a contrasting take on Dynamic Factor Model

    Contrast it →

    继续学习

    概念

    • Output and income→
    • Business cycles→
    • Labor market structure→

    政策

    • Monetary policy→
    • Fiscal policy→

    模型

    • Bridge model→
    • MIDAS→
    • VAR→
    Macro by Mark
    帮助
    隐私政策使用条款伦理与合规

    © 2026 Mark Jayson Nation LLC