advanced time series and forecasting lecture 5 structural ...bhansen/crete/crete5.pdf ·...
TRANSCRIPT
![Page 1: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/1.jpg)
Advanced Time Series and ForecastingLecture 5
Structural Breaks
Bruce E. Hansen
Summer School in Economics and EconometricsUniversity of CreteJuly 23-27, 2012
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 1 / 99
![Page 2: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/2.jpg)
Organization
Detection of Breaks
Estimating Breaks
Forecasting after Breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 1 / 99
![Page 3: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/3.jpg)
Types of Breaks
Breaks in Mean
Breaks in Variance
Breaks in Relationships
Single Breaks
Multiple Breaks
Continuous Breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 2 / 99
![Page 4: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/4.jpg)
ExampleSimple AR(1) with mean and variance breaks
yt = ρyt−1 + µt + etet ∼ N(0, σ2t )
Eyt =µt1− ρ
var(yt ) =σ2t
1− ρ2
µt and/or σ2t may be constant or may have a break at some point inthe sampleSample size nQuestions: Can you guess:
I Is there a structural break?I If so, when?I Is the shift in the mean or variance? How large do you guess?
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 3 / 99
![Page 5: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/5.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 4 / 99
![Page 6: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/6.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 5 / 99
![Page 7: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/7.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 6 / 99
![Page 8: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/8.jpg)
Terminology
Sample Period: t = 1, ..., n
Breakdate: T1I Date of change
Breakdate fraction: τ1 = T1/nPre-Break Sample: t = 1, ...,T1
I T1 observations
Post-Break Sample: t = T1 + 1, ..., nI n− T1 observations
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 7 / 99
![Page 9: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/9.jpg)
Structural Break Model
Full structural break
yt = β′1xt + et , t ≤ T1yt = β′2xt + et , t > T1
oryt = β′1xt1 (t ≤ T1) + β′2xt1 (t > T1) + et
Partial structural break
yt = β′0zt + β′1xt1 (t ≤ T1) + β′2xt1 (t > T1) + et
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 8 / 99
![Page 10: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/10.jpg)
Variance Break Model
yt = β′xt + et ,
var (et ) = σ21, t ≤ T1var (et ) = σ22, t > T1
Breaks do not necessarily affect point forecasts
Affects forecast variances, intervals, fan charts, densities
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 9 / 99
![Page 11: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/11.jpg)
Detection of Breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 10 / 99
![Page 12: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/12.jpg)
Testing for a Break
Classic Test (Chow)
Assume T1 is known
Test H0 : β1 = β2Use classic linear hypothesis test (F, Wald, LM, LR)
Least-Squares
yt = β′0zt + β
′1xt1 (t ≤ T1) + β
′2xt1 (t > T1) + et
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 11 / 99
![Page 13: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/13.jpg)
Full Break Model
Y1 = X1β1 + e1Y2 = X1β2 + e2
β1 = (X ′1X1)−1(X ′1Y1)
β2 = (X ′2X2)−1(X ′2Y2)
e1 = Y1 − X1 β1
e2 = Y2 − X2 β2
SSE (T1) = e ′1 e1 + e′2 e2
σ2(T1) =1
n−m(e ′1 e1 + e
′2 e2)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 12 / 99
![Page 14: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/14.jpg)
F Test Statistic
F test
F (T1) =(SSE − SSE (T1)) /kSSE (T1)(n−m)
where k = dim(β1), m =all parameters,
SSE = e ′e
σ2 =1
n− k(e ′e)
e = Y − X β
(full sample estimate)I F test assumes homoskedasticity, better to use Wald test
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 13 / 99
![Page 15: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/15.jpg)
Wald Test Statistic
W (T1) = n(
β1 − β2
)′ (V1nT1+ V2
nn− T1
)−1 (β1 − β2
)where V1 and V2 are standard asymptotic variance estimators for β1 andβ2 (on the split samples:
V1 = Q−11 Ω1Q−11V2 = Q−12 Ω2Q−12
Q1 =1T1X ′1X1
Q2 =1
n− T1X ′2X2
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 14 / 99
![Page 16: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/16.jpg)
HAC variance optionsFor iid et
Ω1 = σ2Q1Ω2 = σ2Q2
For homoskedastic (within regiome
σ21 =1
T1 − k(e ′1 e1
)σ22 =
1n− T1 − k
(e ′2 e2
)For serially uncorrelated but possibly heteroskedastic
Ω1 =1
T1 − kT1
∑t=1xtx′t e
2t
Ω2 =1
n− T1 − kn
∑t=T1+1
xtx′t e2t
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 15 / 99
![Page 17: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/17.jpg)
For serially correlated (e.g. h > 1)
Ω1 =1
T1 − kT1
∑t=1xtx′t e
2t
+1
T1 − kh−1∑j=0
T1−j∑t=1
(xtx′t+j et et+j + xt+jx
′t et+j et
)Ω2 =
1n− T1 − k
n
∑t=T1+1
xtx′t e2t
+1
n− T1 − kh−1∑j=0
n−j∑
t=T1+1
(xtx′t+j et et+j + xt+jx
′t et+j et
)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 16 / 99
![Page 18: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/18.jpg)
Classic Theory
Under H0, if the number of observations pre- and post-break arelarge, then
F (T1)→dχ2kk
under homoskedasticity, and in general
W (T1)→d χ2k
We can reject H0 in favor of H1 if the test exceeds the critical valueI Thus “find a break” if the test rejects
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 17 / 99
![Page 19: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/19.jpg)
Modern Approach
Break dates are unknown
Sup tests (Andrews, 1993)
SupF = supT1F (T1)
SupW = supT1W (T1)
The sup is taken over all break dates T1 in the region [t1, t2] wheret1 >> 1 and t2 << n
I The region [t1, t2 ] are candidate breakdates. If the proposed break istoo near the beginning or end of sample, the estimates and tests will bemisleading
I Recommended rule t1 = [.15n], t2 = [.85n]
Numerically, calculate SupW using a loop
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 18 / 99
![Page 20: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/20.jpg)
Example
US GDP example presented yesterday
Quarterly data 1960:2011
k = 7
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 19 / 99
![Page 21: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/21.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 20 / 99
![Page 22: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/22.jpg)
Evidence for Structural Break?
SupW=27
Is this significant?
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 21 / 99
![Page 23: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/23.jpg)
Theorem (Andrews)
Under H0, if the regressors xt are strictly stationary, thenI SupF, SupW, etc, converge to non-standard asymptotic distributionswhich depend on
F k (the number of parameters tested for constancyF π1 = t1/nF π2 = t2/nF Only depend on π1 and π2 through λ = π2(1− π1)/(π1(1− π2))
Critical values in Andrews (2003, Econometrica, pp 395-397)
p-value approximation function in Hansen (1997 JBES, pp 60-67)
Critical values much larger than chi-square
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 22 / 99
![Page 24: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/24.jpg)
Evidence for Structural Break?
SupW=27
k = 7
1% asymptotic critical value = 26.72
Asymptotic p-value=0.008
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 23 / 99
![Page 25: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/25.jpg)
Non-Constancy in Marginal or Conditional?
The model is
yt = β′0zt + β′1xt1 (t ≤ T1) + β′2xt1 (t > T1) + et
The goal is to check for non-constancy in the conditional relationship(in the coeffi cients β) while being agnostic about the marginal (thedistribution of the regressors xt)Andrews assume that xt are strictly stationary, which excludesstructural change in the regressors
In Hansen (2000, JoE) I show that this assumption is bindingI If xt has a structural break in its mean or variance, the asymptoticdistribution of the SupW test changes
I This can distort inference (a large test may be due to instability in xt ,not regression instability)
There is a simple solution: Fixed Regressor BootstrapI Requires h = 1 (no serial correlation)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 24 / 99
![Page 26: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/26.jpg)
Fixed Regressor Bootstrap
Similar to a bootstrap, a method to simulate the asymptotic nulldistribution
Fix (zt , xt , et ), t = 1, ..., nLet y ∗t be iid N(0, e
2t ), t = 1, ..., n
Estimate the regression
y ∗t = β∗′0 zt + β
∗′1 xt1 (t ≤ T1) + β
∗′2 xt1 (t > T1) + e
∗t
Form the Wald, SupW statistics on this simulated data
W ∗(T1) = n(
β∗1(T1)− β
∗2(T1)
)′ (V ∗1 (T1)
nT1+ V ∗2 (T1)
nn− T1
)−1(
β∗1(T1)− β
∗2(T1)
)SupW∗ = sup
T1W ∗(T1)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 25 / 99
![Page 27: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/27.jpg)
Repeat this B ≥ 1000 times.Let SupW∗b denote the b’th value
Fixed Regressor bootstrap p-value
p =1B
N
∑b=1
1 (SupW∗b ≥ SupW∗)
Fixed Regressor bootstrap critical values are quantiles of empiricaldistribution of SupW∗bImportant restriction: Requires serially uncorrelated errors (h = 1)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 26 / 99
![Page 28: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/28.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 27 / 99
![Page 29: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/29.jpg)
Evidence for Structural Break?
SupW=27
Asymptotic p-value=0.008
Fixed regressor bootstrap p-value=0.106
Bootstrap eliminates significance!
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 28 / 99
![Page 30: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/30.jpg)
Recommendation
In small samples, the SupW test highly over-rejects
The Fixed Regressor Bootstrap (h = 1) greatly reduces this problem
Furthermore, it makes the test robust to structural change in themarginal distribution
For h > 1, tests not well investigated
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 29 / 99
![Page 31: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/31.jpg)
Testing for Breaks in the Variance
yt = β′xt + et ,
var (et ) = σ21, t ≤ T1var (et ) = σ22, t > T1
Since var (et ) = Ee2t , this is the same as a test for a break in aregression of e2t on a constant
Estimate constant-parameter model
yt = β′xt + et
Obtain squared residuals e2tApply Andrews SupW test to a regression of e2t on a constant
k = 1 critical values
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 30 / 99
![Page 32: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/32.jpg)
GDP Example: Break in Variance?
Apply test to squared OLS residuals
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 31 / 99
![Page 33: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/33.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 32 / 99
![Page 34: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/34.jpg)
Break in Variance?
SupW=15.96
k = 1
1% asymptotic critical value =12.16
Asymptotic p-value=0.002
Fixed regressor bootstrap p-value=0.000
Strong rejection of constancy in varianceI Great moderation
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 33 / 99
![Page 35: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/35.jpg)
End-of-Sample Breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 34 / 99
![Page 36: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/36.jpg)
End-of-Sample Breaks
The SupW tests are powerful against structural changes which occurin the interior of the sample
T1 ∈ [.15n, .85n]Have low power against breaks at the end of the sample
Yet for forecasting, this is a critical period
Classic Chow test allows for breaks at end of sampleI But requires finite sample normality
New end-of-sample instability test by Andrews (Econometrica, 2003)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 35 / 99
![Page 37: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/37.jpg)
End-of-Sample Test
Write model as
Y1 = X1β1 + e1Y2 = X1β2t + e2
where Y1 is n× 1, Y2 is m× 1 and X has k regressors
m is known but small
Test is for non-constancy in β2t
Let β be full sample (n+m) LS estimate, e = Y − X β full-sampleresiduals
Partition e = (e1, e2)
Test depends on Σ = E (e2e ′2)First take case Σ = Imσ2
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 36 / 99
![Page 38: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/38.jpg)
Ifm ≥ dS = e ′2X2
(X ′2X2
)−1 X ′2 e2If m < d
P = e ′2X2X′2 e2
If m is large we could use a chi-square approximation
But when m is small we cannot
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 37 / 99
![Page 39: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/39.jpg)
Andrews suggests a subsampling-type p-value
p =1
n−m+ 1n−m+1
∑j=1
1 (S ≤ Sj )
Sj = e ′jXj(X ′j Xj
)−1 X ′j ejXj = xt : t = j , ..., j +m− 1Yj = yt : t = j , ..., j +m− 1ej = Yj − Xj β(j)
and β(j) is least-squares using all observations except fort = j , ..., j + [m/2]Similar for P test
You can reject end-of-sample stability if p is small (less than 0.05)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 38 / 99
![Page 40: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/40.jpg)
Weighted Tests
Andrews suggested improved power by exploiting correlation in e2
S = e ′2Σ−1X2
(X ′2Σ
−1X2)−1
X ′2Σ−1 e2
where
Σ =1
n+ 1
n+1
∑j=1
(Yj − Xj β
) (Yj − Xj β
)′The subsample calculations are the same as before except that
Sj = e ′j Σ−1Xj
(X ′j Σ
−1Xj)−1
X ′j Σ−1 ej
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 39 / 99
![Page 41: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/41.jpg)
Example: End-of-Sample Instability in GDP Forecasting?
m = 12 (last 3 years)
S statistics (p-values)
Unweighted Weightedh = 1 .20 .21h = 2 .08 .36h = 3 .02 .29h = 4 .18 .27h = 5 .95 .94h = 6 .91 .83h = 7 .86 .70h = 8 .78 .86
Evidence does not suggest end-of-sample instability
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 40 / 99
![Page 42: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/42.jpg)
Breakdate Estimation
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 41 / 99
![Page 43: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/43.jpg)
Breakdate Estimation
The model is a regression
yt = β′0zt + β′1xt1 (t ≤ T1) + β′2xt1 (t > T1) + et
Thus a natural estimator is least squares
The SSE function is
S(β,T1) =1n
n
∑t=1
(yt − β′0zt − β′1xt1 (t ≤ T1)− β′2xt1 (t > T1)
)2(β, T1) = argmin S(β,T1)
The function is quadratic in β, nonlinear in T1I Convenient solution is concentration
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 42 / 99
![Page 44: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/44.jpg)
Least-Squares Algorithm
(β, T1) = argminβ,T1
S(β,T1)
= argminT1
minβS(β,T1)
= argminT1
S(T1)
where
S(T1) = minβS(β,T1)
=1n
n
∑t=1et (T1)2
and et (T1) are the OLS residuals from
yt = β′0zt + β
′1xt1 (t ≤ T1) + β
′2xt1 (t > T1) + et (T1)
with T1 fixed.Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 43 / 99
![Page 45: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/45.jpg)
Least-Squares Estimator
T1 = argminT1
S(T1)
For each T1 ∈ [t1, t2], estimate structural break regression, calculateresiduals and SSE S(T1)
Find T1 which minimizes S(T1)
Even if n is large, this is typically a quick calculation.
Plots of S(T1) against T1 are useful
The sharper the “peak”, then better T1 is identified
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 44 / 99
![Page 46: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/46.jpg)
Example: Breakdate Estimation in GDP
Plot SSE as function of breakdate
Break Date Estimate is lowest point of graph
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 45 / 99
![Page 47: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/47.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 46 / 99
![Page 48: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/48.jpg)
Break Date Estimate
T1 (minimum of SSE (T1) = 1980:4 (82nd observation)
Minimum not well defined
Consistent with weak break or no break
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 47 / 99
![Page 49: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/49.jpg)
Example: Breakdate Estimation for GDP Variance
Plot SSE as function of breakdate
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 48 / 99
![Page 50: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/50.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 49 / 99
![Page 51: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/51.jpg)
Break Date Estimate for Variacne
T1 (minimum of SSE (T1) = 1983:4 (93rd observation)
Well defined minimum
Sharp V shape
Consistent with strong single break
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 50 / 99
![Page 52: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/52.jpg)
Distribution Theory and Confidence Intervals
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 51 / 99
![Page 53: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/53.jpg)
Distribution of Break-Date Estimator
Bai (Review of Economics and Statistics, 1997)
DefineI Q = E (xtx′t )I Ω = E
(xtx′te2t
)I δ = β2 − β1
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 52 / 99
![Page 54: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/54.jpg)
Theorem
[If δ→ 0, and the distribution of (xt , et ) does not change at T1, then](δ′Qδ
)2δ′Ωδ
(T1 − T1
)→d ζ = argmax
s
[W (s)− |s |
2
]where W (s) is a double-sided Brownian motion. The distribution of ζ forx ≥ 0 is
G (x) = 1+
√x2π
exp(−x8
)− x + 5
2Φ(−√x2
)+3ex
2Φ(−3√x2
)and G (x) = 1− G (−x).If the errors are iid, then Ω1 = Q1σ21 and
δ′Q1δσ21
(T1 − T1
)→d ζ
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 53 / 99
![Page 55: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/55.jpg)
Critical Values (Bai Method)
Critical values for ζ can be solved by inverting G (x) :
Coverage c80% 4.790% 7.795% 11.0
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 54 / 99
![Page 56: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/56.jpg)
Confidence Intervals for Break Date (Bai Method)Point Estimate T1
Theorem: T1 ∼ T1 +δ′Ωδ(δ′Qδ
)2 ζ
Confidence interval is then
T1 ±δ′Ωδ(
δ′Q δ)2 c
where
δ = β2 − β1
Q =1n
n
∑t=1xtx′t
Ω =1
n− kn
∑t=1xtx′t e
2t +
1n− k
h−1∑j=0
T1−j∑t=1
(xtx′t+j et et+j + xt+jx
′t et+j et
)Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 55 / 99
![Page 57: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/57.jpg)
Confidence Intervals under Homoskedasticity
T1 ±nσ2(
β2 − β1
)′(X ′X )
(β2 − β1
)c
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 56 / 99
![Page 58: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/58.jpg)
Example
Break for GDP ForecastI Point Estimate: 1980:4I Bai 90% Interval: 1979:2 - 1982:2
Break for GDP VarianceI Point Estimate: 1983:3I Bai 90% Interval: 1983:2 - 1983:4I Very tight
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 57 / 99
![Page 59: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/59.jpg)
Confidence Intervals (Elliott-Mueller)
Elliott-Mueller (JoE, 2007) argue that Bai’s confidence intervalssystematically undercover when breaks are small to moderate
They recommend an alternative simple procedure
For each breakdate T1 for which the regression can be estimatedI Calculate the regression
yt = β′0zt + β
′1xt1 (t ≤ T1) + β
′2xt1 (t > T1) + et
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 58 / 99
![Page 60: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/60.jpg)
yt = β′0zt + β
′1xt1 (t ≤ T1) + β
′2xt1 (t > T1) + et
Ω1 =1
T1 − k
(T1
∑t=1xtx′t e
2t +
h−1∑j=0
T1−j∑t=1
(xtx′t+j et et+j + xt+jx
′t et+j et
))
Ω2 =1
n− T1 − k
(n
∑t=T1+1
xtx′t e2t +
h−1∑j=0
n−j∑
t=T1+1
(xtx′t+j et et+j + xt+jx
′t et+j et
))
Sj =j
∑t=1xt et
U(T1) =1T 21
T1
∑j=1S ′j Ω
−11 Sj +
1
(n− T1)2n
∑j=T1+1
S ′j Ω−12 Sj
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 59 / 99
![Page 61: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/61.jpg)
Theorem (Elliott-Muller)
U(T1)→d∫ 10 B(s)
′B(s)ds where B(s) is a 2k × 1 Brownian bridge,k = dim(xt )This is the Cramer-vonMises distribution
To form a confidence set for T1, find the set of T1 for which U(T1)are less than the critical value
The Elliott-Muller intervals can be much larger than Bai’sI Unclear if they are perhaps too large
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 60 / 99
![Page 62: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/62.jpg)
Critical Values for Confidence Intervals(Elliott-MullerMethod)
Coverage
99% 97.5% 95% 92.5% 90% 80%k = 1 1.07 0.90 0.75 0.67 0.61 0.47k = 2 1.60 1.39 1.24 1.14 1.07 0.88k = 3 2.12 1.89 1.68 1.58 1.49 1.28k = 4 2.59 2.33 2.11 1.99 1.89 1.66k = 5 3.05 2.76 2.54 2.40 2.29 2.03k = 6 3.51 3.18 2.96 2.81 2.69 2.41k = 7 3.90 3.60 3.34 3.19 3.08 2.77k = 8 4.30 4.01 3.75 3.58 3.46 3.14k = 9 4.73 4.40 4.14 3.96 3.83 3.50k = 10 5.13 4.79 4.52 4.36 4.22 3.86
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 61 / 99
![Page 63: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/63.jpg)
Examples: Breakdates for GDP Forecast and Variance
Plot U(T1) as function of T1Plot 90% critical value
90% confidence region is set of values where U(T1) is less thancritical value
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 62 / 99
![Page 64: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/64.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 63 / 99
![Page 65: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/65.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 64 / 99
![Page 66: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/66.jpg)
Example
Break for GDP ForecastI Point Estimate: 1980:4I Bai 90% Interval: 1979:2 - 1982:2I Elliott-Muller: 1981:2 - 2011:2
Break for GDP VarianceI Point Estimate: 1983:3I Bai 90% Interval: 1983:2 - 1983:4I Elliott-Muller: 1980:4 - 1993:4
Elliott-Muller intervals are much wider
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 65 / 99
![Page 67: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/67.jpg)
Slope Estimators
Estimate slopes from regression with estmate T1
yt = β′0zt + β
′1xt1 (t ≤ T1) + β
′2xt1 (t > T1) + et (T1)
In the case of full structural change, this is the same as estimation oneach sub-sample.
Asymptotic Theory:I The sub-sample slope estimates are consistent for the true slopesI If there is a structural break, their asymptotic distributions are“conventional”
F You can treat the structural break as if known
I Compute standard erros using conventional HAC formula
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 66 / 99
![Page 68: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/68.jpg)
Example: Variance Estimates
Pre 1983: σ21 = 14.8 (2.3)
Post 193: σ22 = 4.9 (1.0)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 67 / 99
![Page 69: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/69.jpg)
Multiple Breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 68 / 99
![Page 70: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/70.jpg)
Multiple Structural Breaks
Breaks T1 < T2
yt = β′0zt + β′1xt1 (t ≤ T1) + β′2xt1 (T1 < t ≤ T2)+β′3xt1 (t > T2) + et
Testing/estimation: Two approachesI Joint testing/estimationI Sequential
Major contributors: Jushan Bai, Pierre Perron
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 69 / 99
![Page 71: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/71.jpg)
Joint Methods
TestingI Test the null of constant parameters against the alternative of two(unknown) breaks
I Given T1,T2, construct Wald test for non-constancyI Take the largest test over T1 < T2I Asymptotic distribution a generalization of Andrews
EstimationI The sum-of-squared errors is a function of (T1,T2)I The LS estimates (T1, T2) jointly minimize the SSE
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 70 / 99
![Page 72: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/72.jpg)
Joint Methods - Computation
For 2 breaks, these tests/estimates require O(n2) regressionsI cumbersome but quite feasible
For 3 breaks, naive estimation requires O(n3) regressions,I not feasible
Bai-Perron developed effi cient computer code which solves theproblem of order O(n2) for arbitrary number of breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 71 / 99
![Page 73: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/73.jpg)
Sequential Method
If the truth is two breaks, but you estimate a one-break model, theSSE will (asymptotically) have local minima at both breakdates
Thus the LS breakdate estimator will consistently estimate one of thetwo breaks, e.g. T1 for T1Given an estimated break, you can split the sample and test forbreaks in each subsample
I You can then find T2 for T2
Refinement estimator:I Split the entire sample at T2I Now re-estimate the first break T1I The refined estimators are asymptotically effi cient
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 72 / 99
![Page 74: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/74.jpg)
Forecasting Focuses on Final Breakdate
If you only want to find the last break
First test for structural change on the full sample
If it rejects, split the sample
Test for structural change on the second half
If it rejects, split again...
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 73 / 99
![Page 75: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/75.jpg)
Forecasting After Breaks
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 74 / 99
![Page 76: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/76.jpg)
Forecasting After Breaks
There is no good theory about how to forecast in the presence ofbreaks
There is a multitude of comflicting recommendations
One important contribution:I Pesaran and Timmermann (JoE, 2007)
They show that in a regression with a single break, the optimalwindow for estimation includes all of the observations after the break,and some of the observations before the breakBy including more observations you decrease variance at the cost ofsome bias
They provide empirical rules for selecting sample sizes
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 75 / 99
![Page 77: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/77.jpg)
Recommentation
The simulations in Persaran-Timmermann suggest that there littlegain for the complicated procedures
The simple rule —Split the sample at the estimated break — seems towork as well as anything else
My recommendationI Test for structural breaks using the Andrews or Bai/Perron testsI If there is evidence of a break, estimate its date using Bai’sleast-squares estimator
I Calculate a confidence interval to assess accuracy (calculate both Baiand Elliott-Muller for robustness)
I Split the sample at the break, use the post-break period for estimationI Use economic judgment to enhance statistical findings
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 76 / 99
![Page 78: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/78.jpg)
Examples Revisited
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 77 / 99
![Page 79: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/79.jpg)
Examples from Beginning of Class
Simple AR(1) with mean and variance breaks
yt = ρyt−1 + µt (1− ρ) + etet ∼ N(0, σ2t (1− ρ2))
µt and/or σ2t may be constant or may have a break at some point inthe sample
Sample size n
Questions: Can you guess the timing and type of structural break?
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 78 / 99
![Page 80: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/80.jpg)
Model A
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 79 / 99
![Page 81: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/81.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 80 / 99
![Page 82: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/82.jpg)
Results - Regression
SupW = 0.01 (fixed regressor bootstrap p-value)
Breakdate Estimate = 62I Bai Interval = [55, 69]
Estimates
yt = 0.03(.60)
+ 0.69(.67)
yt−1 + et , t ≤ 62
yt = 0.69(.99)
+ 0.59(.53)
yt−1 + et , t > 62
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 81 / 99
![Page 83: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/83.jpg)
Results - Variance
SupW for Variance = 0.57 (fixed regressor bootstrap p-value)
Breakdate Estimate = 78I Bai Interval = [9, 100]
Estimates
σ2 = 0.37(.60)
, t ≤ 78
σ2 = 0.19(.24)
, t > 78
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 82 / 99
![Page 84: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/84.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 83 / 99
![Page 85: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/85.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 84 / 99
![Page 86: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/86.jpg)
DGP (Model A)
T1 = 60
µ1 = 0.2
µ2 = 0.4
σ21 = σ22 = 0.36
ρ = 0.8
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 85 / 99
![Page 87: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/87.jpg)
Model B
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 86 / 99
![Page 88: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/88.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 87 / 99
![Page 89: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/89.jpg)
Results - Regression
SupW = 0.07 (fixed regressor bootstrap p-value)
Breakdate Estimate = 37I Bai Interval = [20, 54]
Estimates
yt = 0.53(1.12)
+ 0.82(.40)
yt−1 + et , t ≤ 37
yt = 0.10(.70)
+ 0.85(.40)
yt−1 + et , t > 37
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 88 / 99
![Page 90: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/90.jpg)
Results - Variance
SupW for Variance = 0.06 (fixed regressor bootstrap p-value)
Breakdate Estimate = 15I Bai Interval = [0, 69]
Estimates
σ2 = 0.17(.17)
, t ≤ 15
σ2 = 0.40(.55)
, t > 15
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 89 / 99
![Page 91: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/91.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 90 / 99
![Page 92: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/92.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 91 / 99
![Page 93: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/93.jpg)
DGP (Model B)
T1 = 40
µ1 = 0.5
µ2 = 0.2
σ21 = σ22 = 0.36
ρ = 0.8
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 92 / 99
![Page 94: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/94.jpg)
Model C
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 93 / 99
![Page 95: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/95.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 94 / 99
![Page 96: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/96.jpg)
Results
SupW = 0.11 (fixed regressor bootstrap p-value)
Regression Breakdate Estimate = 84I Bai Interval = [78, 90]
Estimates
yt = 0.27(1.02)
+ 0.73(.70)
yt−1 + et , t ≤ 37
yt = 1.53(2.44)
+ −0.11(1.47)
yt−1 + et , t > 37
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 95 / 99
![Page 97: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/97.jpg)
Results - VarianceSupW for Variance = 0.13 (fixed regressor bootstrap p-value)Breakdate Estimate = 69
I Bai Interval = [65, 73]
Estimates
σ2 = 0.43(.51)
, t ≤ 69
σ2 = 1.77(3.06)
, t > 69
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 96 / 99
![Page 98: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/98.jpg)
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 97 / 99
![Page 99: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/99.jpg)
DGP (Model C)
T1 = 70
µ1 = µ2 = 0.2
σ21 = 0.36
σ22 = 1.44
ρ = 0.8
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 98 / 99
![Page 100: Advanced Time Series and Forecasting Lecture 5 Structural ...bhansen/crete/crete5.pdf · distribution of the SupW test changes I This can distort inference (a large test may be due](https://reader030.vdocument.in/reader030/viewer/2022040902/5e72e38d6738183817122c45/html5/thumbnails/100.jpg)
Assignment
Take your favorite model
Estimate the model allowing for one-time structure change in themean
Test the model for one-time structural change in the mean
If appropriate, revise your forecasts
Bruce Hansen (University of Wisconsin) Structural Breaks July 23-27, 2012 99 / 99