【计量经济学】时间序列分析笔记 Models with Trend
时间序列复习 chapter 5
密码 2022年6月11日-6月13日
额外的参考资料:北大金融时间序列的备课笔记
文章目录
- 时间序列复习 chapter 5
- Chapter 5 Models with Trend
- Removing the Deterministic Trend : Detrending
- Removing the Stochastic Trend: Differencing
- ARIMA(p,d,q) processes
- Test for Unit Root
- Case 1 : 1: 1: both { y t } \left\{y_{t}\right\} {yt} and { z t } \left\{z_{t}\right\} {zt} are stationary
- Case 2 : spurious regression
- Case 3 : y t y_t yt and z t z_t zt are cointegrated
- Dickey-Fuller Tests
- Analysis of Case 1
- Augmented Dickey-Fuller Tests for Unit Roots
- Testing for Higher Orders of Integration
Chapter 5 Models with Trend
Deterministic Trend
y t = y 0 + δ t + X t y_t=y_0+\delta t+X_t yt=y0+δt+Xt
Stochastic Trend : Random Walk
Δ y t = ϵ t or y t = y t − 1 + ϵ t Hence y t = y 0 + ∑ i = 1 t ϵ i ⏟ stochastic trend \begin{gathered} \Delta y_{t}=\epsilon_{t} \quad \text { or } \quad y_{t}=y_{t-1}+\epsilon_{t} \\ \text { Hence } y_{t}=y_{0}+\underbrace{\sum_{i=1}^{t} \epsilon_{i}}_{\text {stochastic trend }} \end{gathered} Δyt=ϵt or yt=yt−1+ϵt Hence yt=y0+stochastic trend i=1∑tϵi
- { ϵ t } \left\{\epsilon_{t}\right\} {ϵt} is a white noise process.
- y t y_{t} yt is said to have a stochastic trend since each ϵ i \epsilon_{i} ϵi shock imparts a permanent, albeit random, change in the conditional mean of the series.
Stochastic Trend : Random Walk with Drift
Δ y t = δ + ϵ t or y t = y t − 1 + δ + ϵ t Hence y t = y 0 + δ t ⏟ deterministic trend + ∑ i = 1 t ϵ i ⏟ stochastic trend \begin{aligned} &\Delta y_{t}=\delta+\epsilon_{t} \text { or } y_{t}=y_{t-1}+\delta+\epsilon_{t} \\ &\text { Hence } y_{t}=y_{0}+\underbrace{\delta t}_{\text {deterministic trend }}+\underbrace{\sum_{i=1}^{t} \epsilon_{i}}_{\text {stochastic trend }} \end{aligned} Δyt=δ+ϵt or yt=yt−1+δ+ϵt Hence yt=y0+deterministic trend δt+stochastic trend i=1∑tϵi
- { ϵ t } \left\{\epsilon_{t}\right\} {ϵt} is a white noise process.
- obvious tendency to increase or decrease over time
Generalizations of the Stochastic Trend Model
Δ y t = δ + u t or y t = y t − 1 + δ + u t , \Delta y_{t}=\delta+u_{t} \quad \text { or } y_{t}=y_{t-1}+\delta+u_{t}, Δyt=δ+ut or yt=yt−1+δ+ut,
where u t u_{t} ut is a stationary zero mean ARMA ( p − 1 , q ) \operatorname{ARMA}(p-1, q) ARMA(p−1,q) process.
Hence, y t = y 0 + δ t ⏟ deterministic trend + ∑ i = 1 t u i ⏟ stochastic trend \text { Hence, } y_{t}=y_{0}+\underbrace{\delta t}_{\text {deterministic trend }}+\underbrace{\sum_{i=1}^{t} u_{i}}_{\text {stochastic trend }} Hence, yt=y0+deterministic trend δt+stochastic trend i=1∑tui
- Eq.(53) includes Eq.(49)( Deterministic Trend )as a special case δ = p − 1 = q = 0 \delta=p-1=q=0 δ=p−1=q=0.
- Eq.(53) includes Eq.(51) as a special case: p − 1 = q = 0 p-1=q=0 p−1=q=0.
- Allow Δ y t \Delta y_{t} Δyt to be serially correlated, e.g, logarithm of quarterly GDP (possibly δ > 0 \delta>0 δ>0 ); interest rates (possibly δ = 0 \delta=0 δ=0 ).
- Is X t = ∑ i = 1 t u i X_{t}=\sum_{i=1}^{t} u_{i} Xt=∑i=1tui an ARMA process?
Types of Model from the General Formulation y t = α + δ t + X t y_{t}=\alpha+\delta t+X_{t} yt=α+δt+Xt, where ϕ ( L ) X t = θ ( L ) ϵ t \phi(L) X_{t}=\theta(L) \epsilon_{t} ϕ(L)Xt=θ(L)ϵt
Eq(1): Deterministic Trend
Eq(4): Generalizations of the Stochastic Trend Model
Removing the Deterministic Trend : Detrending
y t = y 0 + δ t + X t y_t=y_0+\delta t+X_t yt=y0+δt+Xt
去趋势:在一个常数和时间上回归yt,并保存残差。
更一般地说,如果存在多项式趋势(例如二次时间趋势),则通过在确定性多项式时间趋势上回归yt并保存残差来实现去趋势。
差分法会丢失一些信息
Removing the Stochastic Trend: Differencing
- Unit root processes (also known as difference stationary processes):
y t = α + δ t + X t , ϕ ( L ) X t = θ ( L ) ϵ t y_{t}=\alpha+\delta t+X_{t}, \quad \phi(L) X_{t}=\theta(L) \epsilon_{t} yt=α+δt+Xt,ϕ(L)Xt=θ(L)ϵt
where ϕ ( L ) = ( 1 − L ) ψ ( L ) \phi(L)=(1-L) \psi(L) ϕ(L)=(1−L)ψ(L), and the roots of ψ ( z ) = 0 \psi(z)=0 ψ(z)=0 lie outside the unit circle. - We induce stationarity by differencing once :
ψ ( L ) Δ y t = ψ ( 1 ) δ + θ ( L ) ϵ t . \psi(L) \Delta y_{t}=\psi(1) \delta+\theta(L) \epsilon_{t} . ψ(L)Δyt=ψ(1)δ+θ(L)ϵt. - If we try to detrend a series which has a stochastic trend, then we will not remove the trend.
ARIMA(p,d,q) processes
- We can generalize this concept to consider the case where the series contains more than one “unit root” :
y t = α + δ t + X t , ( 1 − L ) d ψ ( L ) X t = θ ( L ) ϵ t , y_{t}=\alpha+\delta t+X_{t}, \quad(1-L)^{d} \psi(L) X_{t}=\theta(L) \epsilon_{t}, yt=α+δt+Xt,(1−L)dψ(L)Xt=θ(L)ϵt,
where the roots of ψ ( z ) = 0 \psi(z)=0 ψ(z)=0 lie outside the unit circle. - y t y_{t} yt must be differenced d d d times before it becomes stationary
ψ ( L ) Δ d y t = θ ( L ) ϵ t \psi(L) \Delta^{d} y_{t}=\theta(L) \epsilon_{t} ψ(L)Δdyt=θ(L)ϵt
Then y t y_{t} yt is said to be integrated of order d d d. We write y t ∼ I ( d ) y_{t} \sim I(d) yt∼I(d). So if y t ∼ I ( d ) y_{t} \sim I(d) yt∼I(d), then △ d y t ∼ I ( 0 ) \triangle^{d} y_{t} \sim I(0) △dyt∼I(0). - Taking d d d th differences of an ARIMA(p, d, q) process produces a stationary ARMA ( p , q ) (p, q) (p,q) process.
Test for Unit Root
We consider the following regression model : y t = a 0 + a 1 z t + e t y_{t}=a_{0}+a_{1} z_{t}+e_{t} yt=a0+a1zt+et.
y t , z t ∼ I ( 0 ) y_{t}, z_{t} \sim I(0) yt,zt∼I(0)
y t , z t , e t ∼ I ( 1 ) y_{t}, z_{t}, e_{t} \sim I(1) yt,zt,et∼I(1).
This is the case in which the regression is spurious. 残差不平稳
In this case, it is often recommended that the regression equation be estimated in first differences.
y t , z t ∼ I ( 1 ) y_{t}, z_{t} \sim I(1) yt,zt∼I(1) and e t ∼ I ( 0 ) e_{t} \sim I(0) et∼I(0). In this circumstance, y t y_{t} yt and z t z_{t} zt are cointegrated.
y t y_{t} yt and z t z_{t} zt are integrated of different orders. Meaningless regression!
Case 1 : 1: 1: both { y t } \left\{y_{t}\right\} {yt} and { z t } \left\{z_{t}\right\} {zt} are stationary
Assumptions of the classical model :
- both { y t } \left\{y_{t}\right\} {yt} and { z t } \left\{z_{t}\right\} {zt} sequences are stationary.
- the errors { e t } \left\{e_{t}\right\} {et} have a zero conditional mean and a finite variance.
- the errors { e t } \left\{e_{t}\right\} {et} may be serially correlated.
检验有两种方法
- 中心极限定理:利用之前的AR,MA阶数的test t ( a 0 ^ − 0 ) → N ( 0 , v a r ) \sqrt t(\hat{a_0}-0)\to N(0,var) t (a0^−0)→N(0,var)
- 可以使用t检验和F检验
- simulation 模拟:Monte Carlo
Case 2 : spurious regression
y t = y t − 1 + ϵ y t , ϵ y t ∼ i . i . d . N ( 0 , 1 ) z t = z t − 1 + ϵ z t , ϵ z t ∼ i.i.d. N ( 0 , 1 ) \begin{array}{ll} y_{t}=y_{t-1}+\epsilon_{y t}, & \epsilon_{y t} \sim \stackrel{i . i . d .}{ } \mathcal{N}(0,1) \\ z_{t}=z_{t-1}+\epsilon_{z t}, & \epsilon_{z t} \sim \text { i.i.d. } \mathcal{N}(0,1) \end{array} yt=yt−1+ϵyt,zt=zt−1+ϵzt,ϵyt∼i.i.d.N(0,1)ϵzt∼ i.i.d. N(0,1)
where ϵ y t \epsilon_{y t} ϵyt and ϵ z t \epsilon_{z t} ϵzt are independent, and T = 1000 \mathrm{T}=1000 T=1000.
- Regression result : y ^ t = 23.94 + 0.905 z t \widehat{y}_{t}=23.94+0.905 z_{t} y t=23.94+0.905zt with R 2 = 0.4069 R^{2}=0.4069 R2=0.4069. The residuals from the regression are nonstationary.
However,
since y t y_{t} yt and z t z_{t} zt are independent, the regression is spurious.
A spurious regression has a high R 2 R^{2} R2 and t-ratios that appear to be significant, but the results are of no economic meaning.
The regression output “looks good” because the OLS estimates are inconsistent and the customary tests of statistical inference do not hold.
Case 3 : y t y_t yt and z t z_t zt are cointegrated
只能用Monte Carlo ,不能用t检验和F检验,因为这个分布函数比较偏不能跟正态分布很好的拟合
两大原因
- spurious
- conditional t test and F test don’t apply
Dickey-Fuller Tests
{ ϵ t } \left\{\epsilon_{t}\right\} {ϵt} is i.i.d. with mean zero and variance σ 2 \sigma^{2} σ2. Denote ρ ~ = ρ − 1 \tilde{\rho}=\rho-1 ρ~=ρ−1.(判断是否存在单位根)
- Case 1 : no constant term in the regression
- y t = ρ y t − 1 + ϵ t y_{t}=\rho y_{t-1}+\epsilon_{t} yt=ρyt−1+ϵt.
- We test H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 against H 1 : ρ < 1 H_{1}: \rho<1 H1:ρ<1.
- Equivalently, Δ y t = ρ ~ y t − 1 + ϵ t \Delta y_{t}=\tilde{\rho} y_{t-1}+\epsilon_{t} Δyt=ρ~yt−1+ϵt.
- We test H 0 : ρ ~ = 0 H_{0}: \tilde{\rho}=0 H0:ρ~=0 against H 1 : ρ ~ < 0 H_{1}: \tilde{\rho}<0 H1:ρ~<0.
- Case 2 : constant term included in the regression
- y t = α + ρ y t − 1 + ϵ t y_{t}=\alpha+\rho y_{t-1}+\epsilon_{t} yt=α+ρyt−1+ϵt (or △ y t = α + ρ ~ y t − 1 + ϵ t \triangle y_{t}=\alpha+\tilde{\rho} y_{t-1}+\epsilon_{t} △yt=α+ρ~yt−1+ϵt ).
- We test H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 against H 1 : ρ < 1 H_{1}: \rho<1 H1:ρ<1 (or H 0 : ρ ~ = 0 H_{0}: \tilde{\rho}=0 H0:ρ~=0 v.s. H 1 : ρ ~ < 0 H_{1}: \tilde{\rho}<0 H1:ρ~<0 ).
- Case 3 : constant term and time trend included in the regression
- y t = α + β t + ρ y t − 1 + ϵ t y_{t}=\alpha+\beta t+\rho y_{t-1}+\epsilon_{t} yt=α+βt+ρyt−1+ϵt (or Δ y t = α + β t + ρ ~ y t − 1 + ϵ t ) \left.\Delta y_{t}=\alpha+\beta t+\tilde{\rho} y_{t-1}+\epsilon_{t}\right) Δyt=α+βt+ρ~yt−1+ϵt).
- We test H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 against H 1 : ρ < 1 H_{1}: \rho<1 H1:ρ<1 (or H 0 : ρ ~ = 0 H_{0}: \tilde{\rho}=0 H0:ρ~=0 v.s. H 1 : ρ ~ < 0 ) \left.H_{1}: \tilde{\rho}<0\right) H1:ρ~<0).
上述回归中的两个估计量都没有极限正态分布。case2,3 也不能使用通常的F检验
Analysis of Case 1
H 0 : y t = y t − 1 + ϵ t H_{0}: y_{t}=y_{t-1}+\epsilon_{t} H0:yt=yt−1+ϵt.
No constant term in the test regression : y t = ρ y t − 1 + ϵ t y_{t}=\rho y_{t-1}+\epsilon_{t} yt=ρyt−1+ϵt. Then the OLS estimator ρ T ^ = ∑ t = 1 T y t − 1 y t ∑ t = 1 T y t − 1 2 \hat{\rho_{T}}=\frac{\sum_{t=1}^{T} y_{t-1} y_{t}}{\sum_{t=1}^{T} y_{t-1}^{2}} ρT^=∑t=1Tyt−12∑t=1Tyt−1yt.
To get better sense of the asymptotic of ρ T ^ \hat{\rho_{T}} ρT^, we derive
ρ T ^ − 1 = ∑ t = 1 T y t − 1 ϵ t ∑ t = 1 T y t − 1 2 . \hat{\rho_{T}}-1=\frac{\sum_{t=1}^{T} y_{t-1} \epsilon_{t}}{\sum_{t=1}^{T} y_{t-1}^{2}} . ρT^−1=∑t=1Tyt−12∑t=1Tyt−1ϵt.
The denominator is O p ( T 2 ) O_{p}\left(T^{2}\right) Op(T2), while the numerator is O p ( T ) O_{p}(T) Op(T) ( F C L T ) (F C L T) (FCLT)So it makes sense to multiply this equation by T T T
T ( ρ T ^ − 1 ) = 1 T ∑ t = 1 T y t − 1 ϵ t 1 T 2 ∑ t = 1 T y t − 1 2 . T\left(\hat{\rho_{T}}-1\right)=\frac{\frac{1}{T} \sum_{t=1}^{T} y_{t-1} \epsilon_{t}}{\frac{1}{T^{2}} \sum_{t=1}^{T} y_{t-1}^{2}} . T(ρT^−1)=T21∑t=1Tyt−12T1∑t=1Tyt−1ϵt.By FCLT and continuous mapping theorem,
T ( ρ T ^ − 1 ) ⟶ D 1 2 ( W 2 ( 1 ) − 1 ) ∫ 0 1 W 2 ( r ) d r superconsistent. T\left(\hat{\rho_{T}}-1\right) \stackrel{\mathcal{D}}{\longrightarrow} \frac{\frac{1}{2}\left(W^{2}(1)-1\right)}{\int_{0}^{1} W^{2}(r) d r} \text { superconsistent. } T(ρT^−1)⟶D∫01W2(r)dr21(W2(1)−1) superconsistent.The probability that T ( ρ T ^ − 1 ) T\left(\hat{\rho_{T}}-1\right) T(ρT^−1) is negative approaches 0.68 0.68 0.68 as T T T becomes large. The limiting distribution of T ( ρ T ^ − 1 ) T\left(\hat{\rho_{T}}-1\right) T(ρT^−1) is skewed to the left.
OLS t t t test of H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 :
t T = ρ T ^ − 1 s T / ∑ t = 1 T y t − 1 2 D i c k e y F u l l e r t e s t t_{T}=\frac{\hat{\rho_{T}}-1}{s_{T} / \sqrt{\sum_{t=1}^{T} y_{t-1}^{2}}}\quad Dickey ~~Fuller~~ test tT=sT/∑t=1Tyt−12 ρT^−1Dickey Fuller test
where s T / ∑ t = 1 T y t − 1 2 s_{T} / \sqrt{\sum_{t=1}^{T} y_{t-1}^{2}} sT/∑t=1Tyt−12 is the usual OLS standard error for the estimated coefficient, and s T 2 = ∑ t = 1 T ( y t − ρ T ^ y t − 1 ) 2 / ( T − 1 ) s_{T}^{2}=\sum_{t=1}^{T}\left(y_{t}-\hat{\rho_{T}} y_{t-1}\right)^{2} /(T-1) sT2=∑t=1T(yt−ρT^yt−1)2/(T−1) is the OLS estimate of the residual variance.Again, we can use FCLT and continuous mapping theorem to show
t T ⟶ D 1 2 ( W 2 ( 1 ) − 1 ) [ ∫ 0 1 W 2 ( r ) d r ] 1 / 2 . t_{T} \stackrel{\mathcal{D}}{\longrightarrow} \frac{\frac{1}{2}\left(W^{2}(1)-1\right)}{\left[\int_{0}^{1} W^{2}(r) d r\right]^{1 / 2}} . tT⟶D[∫01W2(r)dr]1/221(W2(1)−1).
Augmented Dickey-Fuller Tests for Unit Roots
ADF检验的基本思想是通过在回归中包含高阶自回归项来控制序列相关性。
Case I: no constant term in the regression.
We can rewrite the model as
y t = ρ y t − 1 + η 1 Δ y t − 1 + ⋯ + η p − 1 Δ y t − p + 1 + ϵ t , or Δ y t = ρ ~ y t − 1 + η 1 Δ y t − 1 + ⋯ + η p − 1 Δ y t − p + 1 + ϵ t \begin{aligned} y_{t} &=\rho y_{t-1}+\eta_{1} \Delta y_{t-1}+\cdots+\eta_{p-1} \Delta y_{t-p+1}+\epsilon_{t}, \\ \text { or } \Delta y_{t} &=\tilde{\rho} y_{t-1}+\eta_{1} \Delta y_{t-1}+\cdots+\eta_{p-1} \Delta y_{t-p+1}+\epsilon_{t} \end{aligned} yt or Δyt=ρyt−1+η1Δyt−1+⋯+ηp−1Δyt−p+1+ϵt,=ρ~yt−1+η1Δyt−1+⋯+ηp−1Δyt−p+1+ϵt
where ρ ≡ ϕ 1 + ϕ 2 + ⋯ + ϕ p , η j ≡ − [ ϕ j + 1 + ϕ j + 2 + ⋯ + ϕ p ] \rho \equiv \phi_{1}+\phi_{2}+\cdots+\phi_{p}, \eta_{j} \equiv-\left[\phi_{j+1}+\phi_{j+2}+\cdots+\phi_{p}\right] ρ≡ϕ1+ϕ2+⋯+ϕp,ηj≡−[ϕj+1+ϕj+2+⋯+ϕp] for j = 1 , 2 , ⋯ , p − 1 j=1,2, \cdots, p-1 j=1,2,⋯,p−1, and ρ ~ ≡ ρ − 1 \tilde{\rho} \equiv \rho-1 ρ~≡ρ−1.
- It can be shown that testing unit roots is equivalent to testing ρ = 1 ( \rho=1( ρ=1( or ρ ~ = 0 ) \tilde{\rho}=0) ρ~=0).
- We test H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 against H 1 : ρ < 1 H_{1}: \rho<1 H1:ρ<1 (or H 0 : ρ ~ = 0 H_{0}: \tilde{\rho}=0 H0:ρ~=0 v.s. H 1 : ρ ~ < 0 ) \left.H_{1}: \tilde{\rho}<0\right) H1:ρ~<0)
Case II : constant term included in the regression
y t = α + ρ y t − 1 + η 1 Δ y t − 1 + ⋯ + η p − 1 △ y t − p + 1 + ϵ t y_{t}=\alpha+\rho y_{t-1}+\eta_{1} \Delta y_{t-1}+\cdots+\eta_{p-1} \triangle y_{t-p+1}+\epsilon_{t} yt=α+ρyt−1+η1Δyt−1+⋯+ηp−1△yt−p+1+ϵt (or Δ y t = α + ρ ~ y t − 1 + η 1 Δ y t − 1 + ⋯ + η p − 1 Δ y t − p + 1 + ϵ t \Delta y_{t}=\alpha+\tilde{\rho} y_{t-1}+\eta_{1} \Delta y_{t-1}+\cdots+\eta_{p-1} \Delta y_{t-p+1}+\epsilon_{t} Δyt=α+ρ~yt−1+η1Δyt−1+⋯+ηp−1Δyt−p+1+ϵt.
We test H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 against H 1 : ρ < 1 H_{1}: \rho<1 H1:ρ<1 (or H 0 : ρ ~ = 0 H_{0}: \tilde{\rho}=0 H0:ρ~=0 v.s. H 1 : ρ ~ < 0 ) \left.H_{1}: \tilde{\rho}<0\right) H1:ρ~<0)
Case III: constant term and time trend included in the regression
y t = α + β t + ρ y t − 1 + η 1 Δ y t − 1 + ⋯ + η p − 1 Δ y t − p + 1 + ϵ t y_{t}=\alpha+\beta t+\rho y_{t-1}+\eta_{1} \Delta y_{t-1}+\cdots+\eta_{p-1} \Delta y_{t-p+1}+\epsilon_{t} yt=α+βt+ρyt−1+η1Δyt−1+⋯+ηp−1Δyt−p+1+ϵt (or Δ y t = α + β t + ρ ~ y t − 1 + η 1 Δ y t − 1 + ⋯ + η p − 1 Δ y t − p + 1 + ϵ t \Delta y_{t}=\alpha+\beta t+\tilde{\rho} y_{t-1}+\eta_{1} \Delta y_{t-1}+\cdots+\eta_{p-1} \Delta y_{t-p+1}+\epsilon_{t} Δyt=α+βt+ρ~yt−1+η1Δyt−1+⋯+ηp−1Δyt−p+1+ϵt.
We test H 0 : ρ = 1 H_{0}: \rho=1 H0:ρ=1 against H 1 : ρ < 1 H_{1}: \rho<1 H1:ρ<1 (or H 0 : ρ ~ = 0 H_{0}: \tilde{\rho}=0 H0:ρ~=0 v.s. H 1 : ρ ~ < 0 ) \left.H_{1}: \tilde{\rho}<0\right) H1:ρ~<0)
以上t检验和F检验不需要为序列相关进行调整,直接用通常的t检验和f检验即可
Testing for Higher Orders of Integration
Other Approaches to Testing for Unit Roots
- PP test - Phillips, P.C.B. and Perron, P. (1988). “Testing for a Unit Root in Time Series Regression.”
- DF-GLS test - Elliott, G., Rothenberg, T.J. and Stock, J.H. (1996). “Efficient Tests for an Autoregressive Unit Root.”
- Hylleberg, S., Engle, R.F., Granger, C.W.J. and Yoo, B.S. ( 1990 ) (1990) (1990). “Seasonal Integration and Cointegration.”
- Panel unit root test-Im, K.S., Pesaran, M.H., and Shin, Y. (2003), “Testing for Unit Roots in Heterogeneous Panels.”
【计量经济学】时间序列分析笔记 Models with Trend相关推荐
- 时间序列分析笔记(待整理)
时间序列有三种基本模式: 平稳性 / 随机性(Stationarity):当数据没有明显的模式特征的话,我们认为它是平稳的,Y值在一个范围内随着时间上下浮动. 趋势性(Trend):当Y值在一段时间内 ...
- 数据分析学习总结笔记15:时间序列分析及Python实现
文章目录 1 引言 2 时间序列的特性 2.1 自相关 2.2 季节性 2.3 平稳性 3 时间序列建模 3.1 移动平均法 3.2 指数平滑法 3.3 双指数平滑法 3.4 三重指数平滑法 3.5 ...
- 《统计学》学习笔记之时间序列分析和预测
鄙人学习笔记 文章目录 时间序列分析和预测 时间序列及其分解 时间序列的描述性分析 时间序列预测的程序 确定时间序列成分 选择预测方法 预测方法的评估 平稳序列的预测 简单平均法 移动平均法 指数平滑 ...
- matlab 牛顿向后差分,Matlab在时间序列分析中的应用--笔记
<Matlab在时间序列分析中的应用--笔记>由会员分享,可在线阅读,更多相关<Matlab在时间序列分析中的应用--笔记(12页珍藏版)>请在人人文库网上搜索. 1.MATL ...
- 运用spss实现时间序列分析(清风视频笔记)
时间序列是某个指标数值长期变化的数值表现,所以时间序列变化背后必然蕴含着数值变化的规律性.时间序列的数值变化规律: 1.长期变动趋势(T secular trend) 2.季节变动规律(S seaso ...
- 《统计学》笔记:第13章 时间序列分析和预测
时间序列 times series 时间序列是同一现象在不同时间上的相继观察值排列而成的序列.经济数据大多数以时间序列的形式给出.根据时间的不同,时间序列中的事件可以是年份.季度.月份或其他任何时间形 ...
- 时间序列分析个人学习笔记2
时间序列分析方法的依据: 时间序列描述:n个随机变量在任意n个时间点t1,t2,-,tnt1,t2,-,tnt_1,t_2,\dots,t_n的集合 一个时间序列有联合分布函数,定义如下: Ft1,t ...
- 浅尝辄止_数学建模(笔记_时间序列分析及其SPSS实现)
本文多是广泛的概念和SPSS运用,没有具体的推导过程和深入的探究 文章目录 一.时间序列分析 1.具体步骤: 二.基本知识 1.时间序列数据 2.时间序列的基本概念 3.时间序列分解 4.叠加模型和乘 ...
- 王燕《应用时间序列分析》学习笔记2
平稳时间序列分析 一个序列经过预处理被识别为平稳非白噪声序列,就说明该序列是一个蕴含着相关信息的平稳序列.在统计上,我们通常是建立一个线性模型来拟合该序列的发展,借此提取该序列中的有用信息.ARMA( ...
最新文章
- 2021年春季学期-信号与系统-第三次作业参考答案-第十道题
- ITK:二维高斯混合模型期望最大化
- 【STM32】输入捕获程序
- 明晚直播丨一次特殊的 Oralce 硬解析性能问题的技术分享
- python对日期型数据排序_如何对日期执行数学运算并用Python对它们进行排序?
- sql server 登录与用户绑定
- Zabbix 结合 bat 脚本实现多个应用程序状态监控
- 关闭 kali 警报音
- screen linux卸载,Ubuntu常用软件安装(附截图软件、FTP、卸载命令)
- python正则匹配单词和字符
- Ubuntu安装Caffe过程和BUG以及解决方案(全网最全)
- 微信分享开发:准备工作[微信公众平台以及微信中控服务配置](一)
- C++ 读取wav文件中的PCM数据
- 网络知识详解之:网络攻击与安全防护
- 汉堡菜单html加logo,HTML+Sass实现HambergurMenu(汉堡包式菜单)
- 计算机的音乐设置方法,让电脑开机和关机音乐更个性的设置方法(图文)
- c# 时间格式化为英文_C#中如何将日期中的月份转化成英文
- 接口,类与接口的关系,接口与抽象类的区别
- 【小猫爪】AUTOSAR学习笔记12-功能安全之E2E模块
- QGraphicsView放大和缩小下鼠标位置使用鼠标滚轮