Threshold factor models for high-dimensional time series
Xialu Liu, Rong Chen(2020)
- We consider a threshold factor model for high-dimensional time series in which the dynamics of the time series is assumed to switch between different regimes according to the value of a threshold variable.
- Professor George C. Tiao is one of the pioneers in the field of jointly modeling multiple time series and has made significant contributions
- In many applications it is often observed that the loading matrix of a factor model may vary.
- we formally introduce a threshold factor model, propose an estimation procedure for the loading spaces and the number of factors based on eigen-analysis of the cross moment matrices of the observed process, develop an objective function for the identification of the threshold value and the threshold variable, and investigate their theoretical properties.
Notation
y t = A x t + ε t \mathbf{y}_{t}=\mathbf{A} \mathbf{x}_{t}+\boldsymbol{\varepsilon}_{t} yt=Axt+εt
- x t = ( x t , 1 , x t , 2 , … , x t , k 0 ) ′ \mathbf{x}_{t}=\left(x_{t, 1}, x_{t, 2}, \ldots,x_{t,k_{0}}\right)^{\prime} xt=(xt,1,xt,2,…,xt,k0)′ is a set of unobserved (latent) factor time series with dimension k 0 k_{0} k0
- y t \mathbf{y}_{t} yt be an observed p × 1 p \times 1 p×1 time series t = 1 , … , n t=1, \ldots, n t=1,…,n
- A is the loading matrix of the common factors,
- ε t \boldsymbol{\varepsilon}_{t} εt is an error process or an idiosyncratic component.
- z t z_{t} zt is a partially known threshold variable
- Q i \mathbf{Q}_{i} Qi are orthonormal matrices.
Model
y t = A x t + ε t \mathbf{y}_{t}=\mathbf{A} \mathbf{x}_{t}+\boldsymbol{\varepsilon}_{t} yt=Axt+εt
y t = { A 1 x t + ε t , 1 z t < r 0 , A 2 x t + ε t , 2 z t ⩾ r 0 , and ε t , i ∼ N ( 0 , Σ t , i ) , i = 1 , 2 \mathbf{y}_{t}=\left\{\begin{array}{ll}\mathbf{A}_{1} \mathbf{x}_{t}+\boldsymbol{\varepsilon}_{t, 1} & z_{t}<r_{0}, \\ \mathbf{A}_{2} \mathbf{x}_{t}+\boldsymbol{\varepsilon}_{t, 2} & z_{t} \geqslant r_{0},\end{array} \text { and } \boldsymbol{\varepsilon}_{t, i} \sim N\left(\mathbf{0}, \boldsymbol{\Sigma}_{t, i}\right), \quad i=1,2\right. yt={A1xt+εt,1A2xt+εt,2zt<r0,zt⩾r0, and εt,i∼N(0,Σt,i),i=1,2
Q i ′ Q i = I k 0 , and A i = Q i Γ i , i = 1 , 2 \mathbf{Q}_{i}^{\prime} \mathbf{Q}_{i}=\mathbf{I}_{k_{0}}, \text { and } \mathbf{A}_{i}=\mathbf{Q}_{i} \Gamma_{i}, \quad i=1,2 Qi′Qi=Ik0, and Ai=QiΓi,i=1,2
R t = ∑ i = 1 2 Γ i x t I t , i \mathbf{R}_{t}=\sum_{i=1}^{2} \Gamma_{i} \mathbf{x}_{t} I_{t, i} Rt=i=1∑2ΓixtIt,i
y t = ∑ i = 1 2 ( Q i R t + ε t , i ) I t , i \mathbf{y}_{t}=\sum_{i=1}^{2}\left(\mathbf{Q}_{i} \mathbf{R}_{t}+\boldsymbol{\varepsilon}_{t, i}\right) I_{t, i} yt=i=1∑2(QiRt+εt,i)It,i
Their aim is to estimate the loading spaces M ( A k ) , i = 1 , 2 \mathcal{M}\left(\boldsymbol{A}_{k}\right), i=1,2 M(Ak),i=1,2, the number of factors k 0 k_{0} k0, and the threshold value r 0 r_{0} r0, given the threshold variable.