2.6 Time Series
Question 1
For a certain time series, you have produced a correlogram with an autocorrelation function that includes twenty four monthly observations; m = degrees of freedom = 24 m = \text{degrees\;of\;freedom} = 24 m=degreesoffreedom=24. Your calculated Box-Pierce Q-statistic is 19.50 19.50 19.50 and your calculated Ljung-Box Q-statistic is 27.90 27.90 27.90. You want to determine if the series is white noise. Which is your best conclusion(given CHISQ.INV ( 0.95 , 24 ) = 36.41 \text{CHISQ.INV}(0.95, 24) = 36.41 CHISQ.INV(0.95,24)=36.41)?
A. With
95.0
%
95.0\%
95.0% confidence, you accept the series as white noise (more accurately, you fail to reject the null)
B. With
95.0
%
95.0\%
95.0% confidence, you accept the series as partial white noise (due to Box-Pierce) but reject the null (due to Ljung-Box)
C. With
95.0
%
95.0\%
95.0% confidence, you reject both null hypotheses and conclude the series is not white noise
D. With
95.0
%
95.0\%
95.0% confidence, you reject both null hypotheses but conclude the series is white noise because the sum of the statistics is greater than the critical value
Answer: A
Testing autocorrelation in the residuals is a standard specification check applied after fitting a model.
Null hypothesis: All its autocorrelations are jointly 0. (i.e.,
H
0
:
ρ
1
=
ρ
2
=
⋯
=
ρ
h
H_0 : \rho_1 =\rho_2=\cdots=\rho_h
H0:ρ1=ρ2=⋯=ρh).
Alternative hypothesis: At least one is non-zero. (i.e.,
H
1
:
ρ
j
≠
0
H_1 : \rho_j\neq 0
H1:ρj=0 for some
j
j
j).
The Box-Pierce test statistic is scaled by the sample size
T
T
T,
Q
B
P
=
T
∑
i
=
1
h
ρ
i
^
2
∼
χ
h
2
Q_{BP}=T\sum^h_{i=1}\hat{\rho_i}^2\sim \chi_h^2
QBP=T∑i=1hρi^2∼χh2
The Liung-Pierce test statistic works better in smaller samples,
Q
L
P
=
T
∑
i
=
1
h
(
T
+
2
T
−
i
)
ρ
i
^
2
∼
χ
h
2
Q_{LP}=T\sum^h_{i=1}(\frac{T+2}{T-i})\hat{\rho_i}^2\sim \chi_h^2
QLP=T∑i=1h(T−iT+2)ρi^2∼χh2
Values of the test statistic larger than the critical value indicate that the autocorrelations are not zero.
Question 2
All of the following traits characterize the covariance stationary of a time series process, except:
A. Stability of the mean.
B. Stability of the covariance structure.
C. A non-constant variance in the time series.
D. Stability of the autocorrelation
Answer: C
A time series is covariance stationary if its first two moments satisfy three key properties:
- The mean is constant and does not change over time.(i.e., E [ Y t ] = μ E[Y_t]=\mu E[Yt]=μ for all t t t)
- The variance is finite and does not change over time.(i.e., V [ Y t ] = γ 0 < ∞ V[Y_t]=\gamma_0<\infty V[Yt]=γ0<∞)
- The autocovariance is finite, does not change over time, and only depends on the distance between observations( h h h).(i.e., C o v [ Y t , Y t − h ] = γ h Cov[Y_t, Y_{t-h}]=\gamma_h Cov[Yt,Yt−h]=γh for all t t t)
Question 3
The following are statements about a moving average( MA \text{MA} MA) representation, and an autoregressive( AR \text{AR} AR) process. Which one describes the main difference between MA \text{MA} MA representation and AR \text{AR} AR process.
A. Moving average (
MA
\text{MA}
MA) representation shows an evidence of autocorrelation cutoff.
B. The autoregressive (
AR
\text{AR}
AR) process will never be covariance stationary.
C. The autoregressive (
AR
\text{AR}
AR) process shows evidence of autocorrelation cutoff.
D. An unadjusted moving average (
MA
\text{MA}
MA) process shows a clear evidence of a gradual autocorrelation decay.
Answer: A
Autocorrelation function(ACF):
γ
h
γ
0
\frac{\gamma_h}{\gamma_0}
γ0γh, refers to the degree of correlation and interdependency between data points in a time series.
Autoregressive(
AR
\text{AR}
AR) model is defined as the current value of a series is linearly related to its past values, plus an additive stochastic shock.
A first order
AR
\text{AR}
AR can be denoted
AR
(
1
)
\text{AR}(1)
AR(1),
Y
t
=
δ
+
ϕ
Y
t
−
1
+
ϵ
t
Y_t=\delta+\phi Y_{t-1}+\epsilon_t
Yt=δ+ϕYt−1+ϵt, where
ϵ
t
∼
W
N
(
0
,
σ
2
)
\epsilon_t \sim WN(0,\sigma^2)
ϵt∼WN(0,σ2)
- AR ( 1 ) \text{AR}(1) AR(1) is covariance stationary when ∣ ϕ ∣ < 1 |\phi|<1 ∣ϕ∣<1.
- Expectation: E [ Y t ] = E [ Y t − h ] = μ = δ 1 − ϕ E[Y_t]=E[Y_{t-h}]=\mu=\frac{\delta}{1-\phi} E[Yt]=E[Yt−h]=μ=1−ϕδ
- Variance: V [ Y t ] = V [ Y t − h ] = γ 0 = σ 2 1 − ϕ 2 V[Y_t]=V[Y_{t-h}]=\gamma_0=\frac{\sigma^2}{1-\phi^2} V[Yt]=V[Yt−h]=γ0=1−ϕ2σ2
- Autocovariance: C o v [ Y t , Y t h ] = γ ( h ) = ϕ ∣ h ∣ γ 0 Cov[Y_t, Y_{t h}]=\gamma(h)=\phi^{|h|}\gamma_0 Cov[Yt,Yth]=γ(h)=ϕ∣h∣γ0
- Autocorrelation(ACF): ρ ( h ) = ϕ ∣ h ∣ \rho(h)=\phi^{|h|} ρ(h)=ϕ∣h∣, the ACF geometrically decays to zero as h h h increases, it also oscillates between negative and positive if − 1 < ϕ < 0 -1<\phi<0 −1<ϕ<0
- Partial autocorrelation(PACF): α ( h ) = ϕ ∣ h ∣ , h ∈ { 0 , ± 1 } \alpha(h)=\phi^{|h|},\;h\in \lbrace0,\pm1\rbrace α(h)=ϕ∣h∣,h∈{0,±1}, α ( h ) = 0 , h ≥ 2 \alpha(h)=0,\;h\geq2 α(h)=0,h≥2, the PACF is non-zero only for the first lag.
The moving average (
MA
\text{MA}
MA) model is defined as the observed value of
Y
t
Y_t
Yt depends on both the contemporaneous shock
ϵ
t
\epsilon_t
ϵt and the previous shock.
A first order
MA
\text{MA}
MA can be denoted
MA
(
1
)
\text{MA}(1)
MA(1),
Y
t
=
μ
+
θ
ϵ
t
−
1
+
ϵ
t
Y_t=\mu+\theta \epsilon_{t-1}+\epsilon_t
Yt=μ+θϵt−1+ϵt, where
ϵ
t
∼
W
N
(
0
,
σ
2
)
\epsilon_t \sim WN(0,\sigma^2)
ϵt∼WN(0,σ2)
- MA \text{MA} MA is always covariance stationary.
- Expectation: E [ Y t ] = E [ Y t − h ] = μ E[Y_t]=E[Y_{t-h}]=\mu E[Yt]=E[Yt−h]=μ
- Variance: V [ Y t ] = V [ Y t − h ] = γ 0 = ( 1 + θ 2 ) σ 2 V[Y_t]=V[Y_{t-h}]=\gamma_0=(1+\theta^2)\sigma^2 V[Yt]=V[Yt−h]=γ0=(1+θ2)σ2
- Autocovariance: C o v [ Y t , Y t h ] = γ ( h ) = 0 Cov[Y_t, Y_{t h}]=\gamma(h)=0 Cov[Yt,Yth]=γ(h)=0, for h > 1 h>1 h>1
- Autocorrelation: ρ ( h ) = 1 \rho(h)=1 ρ(h)=1, when h = 0 h=0 h=0, ρ ( h ) = 0 \rho(h)=0 ρ(h)=0, when h ≥ 2 h\geq2 h≥2, ρ ( h ) = θ 1 + θ 2 \rho(h)=\frac{\theta}{1+\theta^2} ρ(h)=1+θ2θ, when h = 1 h=1 h=1
- Partial autocorrelation: The PACF has non-zero values at all lags and decays towards zero.
Question 4
PE2018Q21 / PE2019Q21 / PE2020Q21 / PE2021Q21 / PE2022PSQ14 / PE2022Q21
A risk manager at a major global bank is conducting a time series analysis of equity returns. The manager wants to know whether the time series is covariance stationary. Which of the following statements describes one of the requirements for a time series to be covariance stationary?
A. The distribution of a time series should have a kurtosis value near
3
3
3, ensuring no fat tails will distort stationarity.
B. The distribution of a time series should have a skewness value near
0
0
0, so that its mean will fall in the center of the distribution.
C. The autocovariance of a covariance stationary time series depends only on displacement term,
τ
\tau
τ, not on time.
D. When the autocovariance function is asymmetric with respect to displacement,
τ
\tau
τ, forward looking stationarity can be achieved.
Answer: C
Learning Objective: Describe the requirements for a series to be covariance stationary.
C is correct. One requirement for a series to be covariance stationary is that its covariance structure be stable over time. If the covariance structure is stable, then the autocovariances depend only on displacement, τ \tau τ, not on time, t t t. Also, covariance stationarity does not place restrictions on other aspects of the distributions or the series, such as kurtosis and skewness.
A and B are incorrect. Covariance stationarity does not place restrictions on other aspects of the distributions or the series, such as kurtosis and skewness.
D is incorrect. Covariance stationarity does not depend on the symmetry of the autocovariance function.
Question 5
PE2020Q70 / PE2021Q70 / PE2022Q70
A market risk manager would like to analyze and forecast a security performance and has obtained the historical time series for that security. The manager consults a colleague from the quantitative analytic team who provides the following Partial Autocorrelation Function (
PACF
\text{PACF}
PACF) plot:
Based on the plot above, which of the following is the best regression approach for the security?
A.
AR
(
1
)
\text{AR}(1)
AR(1)
B.
MA
(
1
)
\text{MA}(1)
MA(1)
C.
AR
(
2
)
\text{AR}(2)
AR(2)
D.
MA
(
2
)
\text{MA}(2)
MA(2)
Answer: C
Learning Objective: Define and describe the properties of autoregressive (
AR
\text{AR}
AR) processes.
The PACF \text{PACF} PACF cuts off after the second lag. This behavior indicates an AR ( 2 ) \text{AR}(2) AR(2) process.
ACF \text{ACF} ACF | PACF \text{PACF} PACF | |
---|---|---|
MA ( 1 ) \text{MA}(1) MA(1) Model | Cutoff | Decay |
AR( 1 ) \text{AR(}1) AR(1) Model | Decay | Cutoff |
Question 6
In regard to white noise, each of the following statements is true except which is false?
A. If a process is zero-mean white noise, then is must be Gaussian white noise.
B. If a process is Gaussian (aka, normal) white noise, then it must be (zero-mean) white noise.
C. If a process is Gaussian (aka, normal) white noise, then it must be independent white noise.
D. If a process is stationary, has zero mean, has constant variance and it serially uncorrelated, then the process is white noise.
Answer: A
White noise is denoted
ϵ
t
∼
WN
(
0
,
σ
2
)
\epsilon_t\sim\text{WN}(0,\sigma^2)
ϵt∼WN(0,σ2)
- Mean zero (i.e., E [ ϵ t ] = 0 E[\epsilon_t]=0 E[ϵt]=0)
- Constant and finite variance (i.e., V [ ϵ t ] = σ 2 < ∞ V[\epsilon_t]=\sigma^2<\infin V[ϵt]=σ2<∞)
- No autocorrelation or autocovariance (i.e., Cov [ ϵ t , ϵ t − h ] = 0 \text{Cov}[\epsilon_t,\epsilon_{t-h}]=0 Cov[ϵt,ϵt−h]=0 for all h ≠ 0 h\neq0 h=0)
Even though a white noise process is serially uncorrelated, it may not be serially independent or normally distributed.
If
ϵ
t
\epsilon_t
ϵt is serially independent, then we say
ϵ
t
\epsilon_t
ϵt is independent white noise.
If
ϵ
t
\epsilon_t
ϵt is serially uncorrelated and normally distributed, then we say
ϵ
t
\epsilon_t
ϵt is normal or Gaussian white noise.