Skip to content

Summary Sheet

ARMA model forecast

(10.5B)(Zt3)=at(1-0.5B)(Z_t - 3) = a_t

Let Xt=Zt3X_t = Z_t - 3, then (10.5B)Xt=atAR(1)(1-0.5B)X_t = a_t\sim AR(1). And because the root B=2B=2 lies outside of the unit circle, the series is stationary. Rewrite as a MA()MA(\infty)

(10.5B)(1+ψ1B+ψ2B2+...)=1(1-0.5B)(1+\psi_1B + \psi_2B^2 + ...) = 1

ψ1=1/2\psi_1 = 1/2
ψ2=1/4\psi_2 = 1/4
ψk=2k\psi_k = 2^{-k}

Then, X^t(1)=02(i+l)ati,Z^t(l)=3+X^t(l)\hat X_t(1) = \sum_{0}^\infty 2^{-(i+l)}a_{t-i}, \hat Z_t(l) = 3 + \hat X_t(l)

et(l)=i=0l1ψiat+lie_t(l)=\sum_{i=0}^{l-1} \psi_i a_{t+l-i},
then var(et(l))=i=0l1ψi2at+li2=σ20l1ψi2var(e_t(l)) = \sum_{i=0}^{l-1} \psi_i^2 a_{t+l-i}^2 = \sigma^2 \sum_0^{l-1}\psi_i^2

(1B+0.25B2)(Zt1)=at(1-B+0.25B^2)(Z_t-1)=a_t

Let Xt=Zt1X_t = Z_t -1, (0.25B2B+1)=(0.5B1)2B=2(0.25B^2 - B + 1)=(0.5B-1)^2 \Rightarrow B=2 is a stationary AR(2)AR(2) model.

Xt=Xt1+14Xt2+atX_t = X_{t-1}+\frac{1}{4}X_{t-2} + a_t, then
Z^t(1)=1+Xt+14Xt1\hat Z_t(1) = 1+X_t + \frac{1}{4} X_{t-1}
Z^t(2)=1+Xt+14Xt1+14Xt=54Xt+14Xt1\hat Z_t(2) =1+ X_t + \frac{1}{4} X_{t-1}+ \frac{1}{4} X_t=\frac{5}{4}X_t + \frac{1}{4} X_{t-1}
Z^t(3)=1+54Xt+14Xt1+14(Xt+14Xt1)=32Xt+516Xt1\hat Z_t(3) = 1+\frac{5}{4} X_t + \frac{1}{4} X_{t-1}+ \frac{1}{4}(X_t+\frac{1}{4}X_{t-1})=\frac{3}{2}X_t + \frac{5}{16}X_{t-1} Z^t(4)=1+32Xt+516Xt1+516Xt+116Xt1=2916Xt+38Xt1\hat Z_t(4) = 1+\frac{3}{2}X_t + \frac{5}{16}X_{t-1}+\frac{5}{16}X_t + \frac{1}{16}X_{t-1} = \frac{29}{16}X_{t}+\frac{3}{8}X_{t-1}

(1B+0.25B2)(1+ψ1B+ψ2B2+...)=1(1-B+0.25B^2)(1+\psi_1B+\psi_2B^2+...) = 1
ψ11=0ψ1=1\psi_1 -1 = 0\Rightarrow \psi_1 = 1
ψ2ψ1+0.25=0ψ2=0.75\psi_2 -\psi_1 +0.25 = 0\Rightarrow \psi_2 = 0.75
ψ3ψ2+0.25ψ1=0ψ3=0.5\psi_3 -\psi_2 +0.25\psi_1 = 0\Rightarrow \psi_3 = 0.5

Therefore,
var(et(1))=σ2var(e_t(1)) = \sigma^2
var(et(2))=2σ2var(e_t(2)) = 2\sigma^2
var(et(3))=4116σ2var(e_t(3)) = \frac{41}{16}\sigma^2

(1B+0.25B2)(Zt+3)=(1+0.25B)at(1-B+0.25B^2)(Z_t + 3) = (1+0.25B)a_t

All B lies out of the unit circle, it's a stationary and invertible ARMA(2,1) model.

(1B+0.25B2)(1+ψ1B+ψ2B2+...)=1+0.25B(1-B+0.25B^2)(1+\psi_1B+\psi_2B^2 + ...) = 1+0.25B
ψ1=0.25+1=1.25\psi_1 = 0.25+1=1.25
ψ2=ψ1=1.25\psi_2 = \psi_1 = 1.25
ψ3=ψ20.25ψ1=1516\psi_3 = \psi_2 -0.25\psi_1 = \frac{15}{16}
ψ4=ψ30.25ψ2=58\psi_4 = \psi_3 - 0.25\psi_2 = \frac{5}{8}
ψ5=ψ40.25ψ3=2564\psi_5 = \psi_4 - 0.25\psi_3 = \frac{25}{64}

var(et(1))=σ2var(e_t(1)) = \sigma^2
var(et(2))=4116σ2var(e_t(2)) = \frac{41}{16}\sigma^2
var(et(3))=338σ2var(e_t(3)) = \frac{33}{8}\sigma^2

TFN

Consider a dynamic regression model yt=0kvixti+nty_t = \sum_0^k v_i x_{t-i}+n_t where both xt,ntx_t, n_t are stationary and invertible ARMA model given by ϕx(B)xt=θ(B)at,ϕn(B)nt=θn(B)et\phi_x(B)x_t = \theta(B)a_t, \phi_n(B)n_t = \theta_n(B)e_t and cov(et,as)=0cov(e_t,a_s)= 0

state the prewhitening process for how to identify the value of k

Apply ϕx(B)/θ(B)\phi_x(B)/\theta(B) on the regression model, we get

ϕx(B)θ(B)yt=0kviϕx(B)θ(B)xti+ϕx(B)θ(B)nt\frac{\phi_x(B)}{\theta(B)}y_t = \sum_0^k v_i \frac{\phi_x(B)}{\theta(B)} x_{t-i} + \frac{\phi_x(B)}{\theta(B)} n_t
y^t=v(B)at+ϵt\hat y_t = v(B)a_t + \epsilon_t

By having ϕn(B)=ϕx(B),θn(B)=θ(B)\phi_n(B)=\phi_x(B), \theta_n(B)=\theta(B), we have ϵe\epsilon \sim e so that

y^t=v(B)at+et\hat y_t = v(B)a_t + e_t

then, multiply both sides by atka_{t-k} and take expectations

E(y^tatk)=v(B)E(atatk)+E(etatk)E(\hat y_t a_{t-k}) = v(B) E(a_t a_{t-k}) + E(e_t a_{t-k})
cov(y^t,atk)=vkσa2cov(\hat y_t, a_{t-k}) = v_k\sigma_a^2
vk=cov(y^t,atk)/σa2=corr(y^t,atk)se(y^t)se(at)corr(y^t,atk)v_k = cov(\hat y_t, a_{t-k}) / \sigma_a^2 = corr(\hat y_t, a_{t-k})\frac{se(\hat y_t)}{se(a_t)}\propto corr(\hat y_t, a_{t-k})

Therefore, we can test the statistical significance of vkv_k by examining the statistical significance of corr(y^t,atk)corr(\hat y_t, a_{t-k})

state the steps of using Box-Tiao transformation to estimate vjv_j

The steps of the estimation procedures

  1. Run the OLS regression on yt=j=1svjxtj+ety_t = \sum_{j=1}^s v_j x_{t-j} + e_t to collect the residuals {e^t}\{\hat e_t\}
  2. Identify an ARMA model for e^t\hat e_t
  3. Apply Box-Tiao transformation to filter yt,xty_t, x_t
  4. Run regression on the transformed equation
  5. check the correlation of regression residuals

Find the l-ahead optimal forecast of yt+l,y^t(l)y_{t+l}, \hat y_t(l) with at,eta_t,e_t.

Since v(B)=0kviBiv(B)=\sum_0^k v_i B^i has finite terms, it can be transformed v(B)=δ(B)/w(B)v(B)=\delta(B)/w(B), then

yt=δ(B)θ(B)w(B)ϕx(B)at+θn(B)ϕn(B)ety_t = \frac{\delta(B) \theta(B)}{w(B)\phi_x(B)}a_t + \frac{\theta_n(B)}{\phi_n(B)}e_t
yt=u(B)at+ψ(B)ety_t = u(B)a_t + \psi(B)e_t

all of them are finite order with max K of polynomials in B. Therefore,

yt+l=0kuiat+li+ψiet+liy_{t+l} = \sum_0^k u_i a_{t+l-i} + \psi_i e_{t+l-i}
y^t(l)=0kui+lati+ψi+leti\hat y_t(l) = \sum_0^k u_{i+l}^* a_{t-i} + \psi_{i+l}^* e_{t-i}

Derive the MSE

Consider ytly^t(l)y_{t-l} - \hat y_t(l), which equals to
=0l1uiat+li+ψiet+li(i)=\sum_0^{l-1}u_i a_{t+l-i} + \psi_i e_{t+l - i} (i) after time lag tt
0k(ui+lui+l)ati(ii)-\sum_0^k (u_{i+l}^* - u_{i+l})a_{t-i} (ii)
(ψi+lψi+l)eti(iii)-(\psi_{i+l}^* - \psi_{i+l})e_{t-i} (iii) up to time tt

E(yly^t(l))2=E(i)2+E(ii)2+E(iii)2E(y_l - \hat y_t(l))^2 = E(i)^2 + E(ii)^2 + E(iii)^2
=0l1ui2σa2+ψi2σe2=\sum_0^{l-1} u_i^2 \sigma_a^2 + \psi_i^2 \sigma_e^2
+0kσa2(ui+lui+l)+ \sum_0^k \sigma_a^2(u_{i+l^*}- u_{i+l})
+0kσe2(ψi+lψi+l)+ \sum_0^k \sigma_e^2(\psi_{i+l^*}- \psi_{i+l})

minimized when u=u,ψ=ψu^* = u, \psi^* = \psi

VAR

Consider a VAR(p) model of 2-d variables, i.e. yt=[yi,t,y2,t]Ty_t = [y_{i,t}, y_{2,t}]^T yt=1pAiyti+aty_t = \sum_1^p A_i y_{t-i} + a_t, and each Ai=[ϕi,11ϕi,12ϕi,21ϕi,22]A_i = \begin{bmatrix}\phi_{i,11}&\phi_{i,12}\\\phi_{i,21}&\phi_{i,22}\end{bmatrix}

state how to check the stationarity

All the roots of det(IkA1B...ApBp)=0\det(I_k - A_1B-...-A_pB^p) = 0 must lie outside of the unit circle, or the companion form ξt=Aξt1+vt\xi_t = A\xi_{t-1}+v_t must have the moduli of the eigenvalues of AA being <1.

Describe the methods to select the order for Equation (1)

  • Selection by information criteria, for example, BIC, AIC, DIC, HQ, SC, FPE
  • Using LRT for VAR(p) vs. VAR(p-1)

State how to test Granger causality for that X1tX_{1t} Granger causes X2tX_{2t} but not the other way around. Basd on the same condition, express X2tX_{2t} as the TFN model of X1tX_{1t}

X2,t=1pϕi,21X1,ti+ϕi,22X2,ti+a2,tX_{2,t} = \sum_1^p \phi_{i,21}X_{1,t-i} + \phi_{i,22}X_{2,t-i} + a_{2,t}
X2,t1pϕi,22X2,ti=1pϕi,21X1,ti+a2,tX_{2,t} - \sum_1^p \phi_{i,22}X_{2,t-i} = \sum_1^p \phi_{i,21}X_{1,t-i} + a_{2,t}
Φ22(B)X2,t=Φ21(B)X1,t+a2,t\Phi_{22}(B)X_{2,t} = \Phi_{21}(B)X_{1,t}+a_{2,t}
X2,t=Φ21(B)Φ22(B)X1,t+a2,tΦ22(B)X_{2,t} = \frac{\Phi_{21}(B)}{\Phi_{22}(B)}X_{1,t} + \frac{a_{2,t}}{\Phi_{22}(B)}

Let v(B)=Φ21(B)Φ22(B),Nt=a2,tΦ22(B)v(B) = \frac{\Phi_{21}(B)}{\Phi_{22}(B)}, N_t = \frac{a_{2,t}}{\Phi_{22}(B)}, the TFN model is

X2,t=v(B)X1,t+NtX_{2,t}=v(B)X_{1,t} + N_t

To see whether X1,tX_{1,t} Granger causes X2,tX_{2,t}, if not, then all ϕi,21=0\phi_{i,21} = 0

Describe how to test Granger causality using univariate approach. Suppose (Φ1(B)X1,t=Θ1(B)a1,t,Φ2(B)X2,t=Θ2(B)a2,t\Phi_1(B)X_{1,t}=\Theta_1(B)a_{1,t}, \Phi_2(B)X_{2,t}=\Theta_2(B)a_{2,t}\)

Using the Portmanteau test,

ρa1a2(k)=E(a1,t,a2,t+k)σ12σ22\rho_{a_1a_2}(k) = \frac{E(a_{1,t}, a_{2,t+k})}{\sqrt{\sigma_1^2\sigma_2^2}}

If k<0.ρ(k)=0\forall k < 0. \rho(k)=0, then X2X_2 does not Granger cause X1X_1.

Suppose X1,t,X2,tX_{1,t}, X_{2,t} not weakly stationary. How do you model the join dynamics of {X1,t,X2,t}\{X_{1,t}, X_{2,t}\} using co-integration.

  • differencing the two time series individually until each are stationary
  • use VAR(p) to fit the two stationary process and test the Granger causality
  • If (X1,t,X2,t)(X_{1,t}, X_{2,t}) are cointegrated, use Error correction model to include lagged disequilibrium terms as explanatory variables.

Discuss the reasons why we have to choose different models based on the condition of integration

If cointegration exists, if we use VAR(p) model directly fitted to the differenced stationary processes, the model will be misspecified.

Discuss the Engle-Granger approach for modeling cointegrated X1tX_{1t} and X2tX_{2t}

  1. test whether Xt,YtX_t,Y_t are I(1) using unit root test
  2. if both I(1), regress one against the other using least squares
  3. run a unit root test on regression residuals. If residuals are stationary, these two series are cointegrated.
  4. Where the regression line indicate the long-run equilibrium relationship between two variables, the disequilirium term is simply the regression residuals.
  5. Finally, consider the ECM
ΔXt=c1+ρ1(Yt1a^Xt1)+βx,1ΔXt1+...+βy,1ΔYt1+...+ϵx,t\Delta X_t = c_1 + \rho_1(Y_{t-1} - \hat a X_{t-1}) + \beta_{x,1}\Delta X_{t-1} + ... + \beta_{y,1}\Delta Y_{t-1}+...+\epsilon_{x,t}
ΔYt=c2+ρ2(Yt1a^Xt1)+γx,1ΔXt1+...+γy,1ΔYt1+...+ϵy,t\Delta Y_t = c_2 + \rho_2(Y_{t-1} - \hat a X_{t-1}) + \gamma_{x,1}\Delta X_{t-1} + ... + \gamma_{y,1}\Delta Y_{t-1}+...+\epsilon_{y,t}

Discuss the implication of Granger representation theorem

We have and should use ECM to model non-stationary time series

Bootstrap

Given an AR(2) model

yt=μ+ϕyt1+ϕ2yt2+aty_t = \mu + \phi y_{t-1} + \phi_2 y_{t-2} + a_t

(unconditional) parametric bootstrap

E(Yt)=μ+E(ϕ1Yt1)+E(ϕ2Yt2)+E(at)E(Y_t) = \mu + E(\phi_1 Y_{t-1}) + E(\phi_2 Y_{t-2}) + E(a_t)
E(Y)=μ+ϕ1E(Y)+ϕ2E(Y)E(Y)=\mu + \phi_1 E(Y) + \phi_2 E(Y)
E(Y)=μ1ϕ1ϕ2E(Y)=\frac{\mu}{1-\phi_1-\phi_2}

var(Y)=0+ϕ12var(Y)+ϕ22var(Y)+E(a)2var(Y) = 0 + \phi_1^2 var(Y) + \phi_2^2 var(Y) + E(a)^2
var(Y)=σ21ϕ12ϕ22var(Y) = \frac{\sigma^2}{1-\phi_1^2 - \phi_2^2}

  1. The unconditional distribution of YtN(μ1ϕ1ϕ2,σ21ϕ12ϕ22)Y_t\sim N(\frac{\mu}{1-\phi_1-\phi_2}, \frac{\sigma^2}{1-\phi_1^2 - \phi_2^2})
  2. simulate y0,y1y_0, y_1 by drawing a random number from YY.
  3. recursively simulate y2y_2 by the model, recursively

pros easy to compute
cons don't know whether it is correlated

GARCH

ARCH process

ARCH(1) The first order of autoregressive conditional heteroskedastic process is etN(0,σt2)e_t \sim N(0, \sigma_t^2) where σt2=a0+a1et12\sigma_t^2 = a_0 + a_1 e_{t-1}^2.
On defining vt=et2σt2v_t = e_t^2 - \sigma_t^2, the model can also be written as

et2=a0+a1et12+vte_t^2 = a_0 + a_1 e_{t-1}^2 + v_t

Since E(vtxt1,xt2,...)=0E(v_t \mid x_{t-1}, x_{t-2},...)=0, the model corresponds directly to an AR(1) model for the squared error et2e_t^2.

ARCH(q) Then, the ARCH(q) process is defined as

with

\[a_0 \geq 0, a_i > 0. \sum_{i=1}^q a_i < 1\]

or

\[e_t^2 = a_0 + \sum_1^q a_i e_{t-i}^2 + v_t\]
\[e_t^2 = a_0 + a(B)e_{t-1}^2 + v_t\]

GARCH(p,q)

\[e_t^2 = a_0 + \sum_1^q \beta_i \sigma_{t-i}^2 + \sum_1^q a_i e_{t-i}^2\]
\[\sigma_t^2 = a_0 + a(B)e_{t-1}^2 + \beta(B)\sigma_{t-1}^2\]

with \(a_0 \geq 0, a(B)\) and \(\beta(B)\) have no common roots and that the roots of \(1-\beta(B)\) all less than unit root.