smoothed stationary bootstrap bandwidth selection for density...
TRANSCRIPT
![Page 1: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/1.jpg)
Smoothed stationary bootstrap bandwidth selection fordensity estimation with dependent data
Ricardo CaoMODES group, University of A Coruna (Spain)
Joint work with Ines Barbeito Cal
Galician Seminar of Nonparametric Statistical Inference,June 8, 2016
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 1 / 81
![Page 2: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/2.jpg)
Index
1 Introduction and BackgroundAimsSmoothed Bootstrap iid caseBootstrap methods for dependent data
2 Already existing smoothing parameter selectors3 Bootstrap bandwidth selector under independence4 Smooth Stationary Bootstrap under dependence5 Smooth Moving Blocks Bootstrap under dependence6 Simulations7 Real data application8 Conclusions9 References
10 Contact info
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 2 / 81
![Page 3: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/3.jpg)
Motivation & Background Aims
Setup and aims
General dependent data, {Xt}t∈Z: stationary, α-mixing, φ-mixing, . . .Nonparametric Parzen-Rosenblatt kernel density estimation
fh(x) = 1nh
n∑i=1
K
(x−Xi
h
)= 1n
n∑i=1
Kh(x−Xi)
Smooth bootstrap methodsBandwidth (h) selection
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 3 / 81
![Page 4: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/4.jpg)
Motivation & Background Smoothed Bootstrap iid case
Smoothed bootstrap for independent data
Consider some statistic of interest: R(→X,F
)Smoothed bootstrap algorithm
1 Using the sample (X1, . . . , Xn) and the bandwidth h > 0, compute fh2 Draw bootstrap resamples
→X∗= (X∗1 , . . . , X∗n) from fh
3 Obtain the bootstrap version of the statistic: R∗ = R
( →X∗, Fh
)4 Repeat Steps 1-3, B times to obtain R∗(1), . . . , R∗(B)
5 Use the values R∗(1), . . . , R∗(B) to approximate the samplingdistribution of R.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 4 / 81
![Page 5: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/5.jpg)
Motivation & Background Smoothed Bootstrap iid case
How to draw from fh?
Considering two independent random variables: Y ∼ Fn and U withdensity K, it is easy to prove that Y + hU has density fhDrawing resamples from fh
1 Draw naive bootstrap resamples→
XNAIV E∗= (XNAIV E∗1 , . . . , XNAIV E∗
n ) from Fn
2 Draw a sample→U= (U1, . . . , Un) from the density K
3 Obtain the smoothed bootstrap resample→X∗= (X∗1 , . . . , X∗n), where
X∗i = XNAIV E∗i + hUi
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 5 / 81
![Page 6: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/6.jpg)
Motivation & Background Dependent data: bootstrap methods
Moving Blocks Bootstrap (MBB)
MBB algorithmKunsch (1989), Liu and Singh (1992)
1 Fix the block lenght, b ∈ N, and define k = min`∈N ` ≥ nb
2 Define:Bi,b = (Xi, Xi+1, . . . , Xi+b−1)
3 Draw ξ1, ξ2, . . . , ξk with uniform discrete distribution on{B1, B2, . . . , Bq}, with q = n− b+ 1
4 Define→X∗ as the vector formed by the first n components of
(ξ1,1, ξ1,2, . . . , ξ1,b, ξ2,1, ξ2,2 . . . , ξ2,b, . . . , ξk,1, ξk,2, . . . , ξk,b)
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 6 / 81
![Page 7: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/7.jpg)
Motivation & Background Dependent data: bootstrap methods
Stationary Bootstrap (SB)
SB algorithmPolitis and Romano (1994a)
1 Draw X∗1 from Fn
2 Once obtained X∗i = Xj , for some j ∈ {1, 2, . . . , n− 1}, i < n, defineX∗i+1 as follows:
X∗i+1 = Xj+1 (if j = n, Xj+1 = X1), with probability 1− pX∗i+1 is drawn from Fn with probability p
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 7 / 81
![Page 8: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/8.jpg)
Motivation & Background Dependent data: bootstrap methods
SubsamplingSubsampling algorithm (for dependent data)Politis and Romano (1994b)
1 Consider a dependent data sample (X1, . . . , Xn) with marginaldistribution F and θ = θ(F )
2 An estimator Tn = Tn(X1, . . . , Xn) of θ = θ(F ) is considered and
Jn(u, F ) = P (τn(Tn − θ) ≤ u)
3 Fix some b ∈ N such that b < n and define:
Sn,i = Tb(Bi,b), i = 1, 2, . . . , N, where N = n− b+ 1.
4 Use:
Ln(x) = 1N
N∑i=1
111{τb(Sn,i−Tn)≤x}
to approximate the sampling distribution of τn(Tn − θ):R. Cao SSB bandwidth selection GSNSI, June 8, 2016 8 / 81
![Page 9: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/9.jpg)
Already existing bandwidth selectors
Plug-in method under dependence (PI)Hall, Lahiri and Truong (1995)
Minimizing in h the asymptotic MISE:
AMISE(h) = 1nhR(K) + 1
4h4µ2
2R(f ′′)− h6 124µ2µ4R(f ′′′)
+ 1n
(2n−1∑i=1
(1− i
n
)∫gi(x, x)dx−R(f)
).
results in hAMISE =(J1n
)1/5+ J2
(J1n
)3/5, with
gi(x1, x2) = fi(x1, x2)− f(x1)f(x2), fi the density of (Xj , Xi+j),
J1 = R(K)µ2
2R(f ′′)and J2 = µ4R(f ′′′)
20µ2R(f ′′) .
Now hPI =(J1n
)1/5
+ J2
(J1n
)3/5
, with J1 and J2 some estimators
of J1 and J2.R. Cao SSB bandwidth selection GSNSI, June 8, 2016 9 / 81
![Page 10: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/10.jpg)
Already existing bandwidth selectors
Plug-in method under dependence (PI)
Replace R(f ′′) by I2 and R(f ′′′) by I3, where:
Ik = 2θ1k − θ2k, k = 2, 3,
θ1k = 2(n(n− 1)h2k+1
1
)−1 ∑∑1≤i<j≤n
K(2k)1
(Xi −Xj
h1
),
θ2k = 2(n(n− 1)h2(k+1)
1
)−1 ∑∑1≤i<j≤n
∫K
(k)1
(x−Xi
h1
)K
(k)1
(x−Xj
h1
)dx.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 10 / 81
![Page 11: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/11.jpg)
Already existing bandwidth selectors
Leave-(2l + 1)-out cross validation (CVl)Hart and Vieu (1990)
DefineCVl(h) =
∫f2(x)dx− 2
n
n∑j=1
f jl (Xj),
wheref jl (x) = 1
nl
n∑i:|j−i|>l
1hK
(x−Xi
h
).
Choose nl such that:
nnl = #{(i, j) : |i− j| > l}.
The CVl bandwidth selector is
hCVl= arg min
hCVl(h).
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 11 / 81
![Page 12: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/12.jpg)
Already existing bandwidth selectors
Penalized cross validation (PCV)
Estevez, Quintela and Vieu (2002) proposed it for hazard rate estimationThe PCV bandwidth selecctor is
hPCV = hCVl+ λ.
λ is chosen empirically as follows:
λn =(0.8e7.9ρ−1
)n−3/10hCVl
100 ,
where ρ is the estimated autocorrelation
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 12 / 81
![Page 13: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/13.jpg)
Already existing bandwidth selectors
Modified cross validation under dependence (SMCV )
Stute (1992) proposed it for independent dataDefine
SMCV (h) =1nh
∫K2(t)dt
+1
n(n− 1)h
∑i6=j
[1h
∫K
(x−Xi
h
)K
(x−Xj
h
)dx
]
−1
nnlh
n∑j=1
n∑i:|j−i|>l
[K
(Xi −Xj
h
)− dK′′
(Xi −Xj
h
)].
The SMCV bandwidth selector is
hSMCV = arg minhSMCV (h)
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 13 / 81
![Page 14: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/14.jpg)
Bootstrap bandwidth selector in the iid case
Exact MISE expression for the iid case
MISE (h) = E[∫ (
fh (x)− f (x))2dx
]= B (h) + V (h) ,
where
B (h) =∫ [
E(fh (x)
)− f (x)
]2dx, e
V (h) =∫V ar
(fh (x)
)dx
Exact expression for MISE(h):
B (h) =∫
(Kh ∗ f (x)− f (x))2 dx, and
V (h) = n−1h−1R (K)− n−1∫
(Kh ∗ f (x))2 dx.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 14 / 81
![Page 15: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/15.jpg)
Bootstrap bandwidth selector in the iid case
Smoothed bootstrap for the iid caseSmooth bootstrap algorithm for bandwidth selection Cao (1993)
1 Starting from (X1, . . . , Xn) (iid), and using a pilot bandwidth, g,compute fg
2 Draw bootstrap resamples (X∗1 , . . . , X∗n) from fg3 For every h > 0, obtain
f∗h(x) = 1nh
n∑i=1
K
(x−X∗ih
)4 Construct the bootstrap version of MISE:
MISE∗(h) =∫
E∗[(f∗h(x)− fg(x)
)2]dx
5 Obtain the bootstrap selector:
h∗MISE = arg minh>0
MISE∗(h).
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 15 / 81
![Page 16: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/16.jpg)
Bootstrap bandwidth selector in the iid case
Smoothed bootstrap for the iid case
Closed expression for the bootstrap MISEAn exact expression for MISE∗(h) can be found:
MISE∗(h) = 1n2
n∑i,j=1
[(Kh ∗Kg −Kg) ∗ (Kh ∗Kg −Kg)] (Xi −Xj)
+R(K)nh
− 1n3
n∑i,j=1
[(Kh ∗Kg) ∗ (Kg ∗Kg)] (Xi −Xj) ,
where ∗ denotes the convolution operator: f ∗ g(x) =∫∞−∞ f(x− y)g(y)dy.
Consequently, there is no need to draw bootstrap resamples by MonteCarlo to approximate MISE∗(h).
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 16 / 81
![Page 17: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/17.jpg)
Smooth Stationary Bootstrap (SSB)
Exact MISE expression under dependence and stationarity
Exact expression for MISE(h):
MISE (h) = B (h) + V (h) , where
B (h) =∫
(Kh ∗ f (x)− f (x))2 dx, and
V (h) = n−1h−1R (K)−∫
(Kh ∗ f (x))2 dx
+ 2n−2n−1∑`=1
(n− `)∫ ∫
Kh (x− y) f (y) (Kh ∗ f` (•|y)) (x) dx dy,
where f` (•|y) is the conditional density function of Xt+` given Xt = y.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 17 / 81
![Page 18: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/18.jpg)
Smooth Stationary Bootstrap (SSB)
Smooth Stationary BootstrapSSB resampling plan Barbeito and Cao (2016)
1 Draw X∗(SB)1 from Fn.
2 Draw U∗1 with density K and independently of X∗(SB)1 and define
X∗1 = X∗(SB)1 + gU∗1
3 Assume we have drawn X∗1 , . . . , X∗i and consider the indexj/X
∗(SB)i = Xj . Define I∗i+1, such that
P∗(I∗i+1 = 1
)= 1− p,
P∗(I∗i+1 = 0
)= p.
Assign X∗(SB)i+1 |I∗i+1=1 = X(jmodn)+1 and draw X
∗(SB)i+1 |I∗i+1=0 from
the empirical distribution function4 Define X∗i+1 = X
∗(SB)i+1 + gU∗i+1 (where U∗i+1 has density K). Go to
the previous step if i+ 1 < n.R. Cao SSB bandwidth selection GSNSI, June 8, 2016 18 / 81
![Page 19: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/19.jpg)
Smooth Stationary Bootstrap (SSB)
MISE closed expression for SSBAn explicit expression for MISE∗(h) can be obtained:
MISE∗(h) = n−1h−1R (K)
+[n− 1n3 −2
1− p− (1− p)n
pn3 + 2(n− 1) (1− p)n+1 − n (1− p)n + 1− p
p2n4
]·
n∑i,j=1
[(Kh ∗Kg) ∗ (Kh ∗Kg)] (Xi −Xj)
−2n−2n∑
i,j=1
(Kh ∗Kg ∗Kg) (Xi −Xj)
+n−2n∑
i,j=1
(Kg ∗Kg) (Xi −Xj) +2n−3n−1∑`=1
(n− `) (1− p)`
·n∑
k=1
[(Kh ∗Kg) ∗ (Kh ∗Kg)](Xk −Xd(k+`−1)modne+1
).
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 19 / 81
![Page 20: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/20.jpg)
Smooth Moving Blocks Bootstrap (SMBB)
Smooth Moving Blocks Bootstrap
SMBB resampling plan1 Fix the block lenght, b ∈ N, and define k = min`∈N ` ≥ n
b
2 Define:Bi,b = (Xi, Xi+1, . . . , Xi+b−1)
3 Draw ξ1, ξ2, . . . , ξk with uniform discrete distribution on{B1, B2, . . . , Bq}, with q = n− b+ 1
4 Define X∗(MBB)1 , . . . , X
∗(MBB)n as the first n components of
(ξ1,1, ξ1,2, . . . , ξ1,b, ξ2,1, ξ2,2 . . . , ξ2,b, . . . , ξk,1, ξk,2, . . . , ξk,b)
5 Define X∗i = X∗(MBB)i + gU∗i , where U∗i has been drawn with density
K and independently from X∗(MBB)i , for all i = 1, 2, . . . , n
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 20 / 81
![Page 21: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/21.jpg)
Smooth Moving Blocks Bootstrap (SMBB)
MISE closed expression for SMBBAn explicit expression for MISE∗(h) can be obtained, considering n anentire multiple of b.
If b = n,
MISE∗(h) =R(K)nh
+1n2
n∑i=1
n∑j=1
ψ(Xi −Xj)
−2n2
n∑i=1
n∑j=1
[(Kh ∗Kg) ∗Kg ] (Xi −Xj)
+1n2
n∑i=1
n∑j=1
[Kg ∗Kg ] (Xi −Xj)
+ψ(0)n
,
where ψ(Xi −Xj) = [(Kh ∗Kg) ∗ (Kh ∗Kg)] (Xi −Xj).
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 21 / 81
![Page 22: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/22.jpg)
Smooth Moving Blocks Bootstrap (SMBB)
MISE closed expression for SMBBIf b < n,
MISE∗(h) =R(K)nh
+n∑
i=1
ai
n∑j=1
aj · ψ(Xi −Xj)
−2n
n∑i=1
ai
n∑j=1
[(Kh ∗Kg) ∗Kg ] (Xi −Xj)
+1n2
n∑i=1
n∑j=1
[Kg ∗Kg ] (Xi −Xj)
−b− 1
n(n− b+ 1)2
n−b+1∑i=b−1
n−b+2∑j=b
ψ(Xi −Xj)
−1
nb · (n− b+ 1)2
[b−1∑i=1
b−1∑j=1
(min{i, j})ψ(Xi −Xj)
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 22 / 81
![Page 23: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/23.jpg)
Smooth Moving Blocks Bootstrap (SMBB)
MISE closed expression for SMBB
+b−1∑i=1
i
n−b+1∑j=b
ψ(Xi −Xj) +b−1∑i=1
n∑j=n−b+2
(min{(n− b+ i− j + 1), i})ψ(Xi −Xj)
+n−b+1∑
i=b
b−1∑j=1
j · ψ(Xi −Xj) +n∑
i=n−b+2
(min{(n− i+ 1), b})n−b+1∑
j=b
ψ(Xi −Xj)
+n−b+1∑
i=b
n∑j=n−b+2
(min{(n− j + 1), b}) · ψ(Xi −Xj)
+n∑
i=n−b+2
b−1∑j=1
(min{(n− b+ j − i+ 1), j})ψ(Xi −Xj) + b
n−b+1∑i=b
n−b+1∑j=b
ψ(Xi −Xj)
+n∑
i=n−b+2
n∑j=n−b+2
(n+ 1−max{i, j})ψ(Xi −Xj)
]
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 23 / 81
![Page 24: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/24.jpg)
Smooth Moving Blocks Bootstrap (SMBB)
MISE closed expression for SMBB
+2
nb(n− b+ 1)
b−1∑s=1
n−s∑j=1
(min{j, b− s} −max{1, j + b− n}+ 1)ψ(Xj+s −Xj)
−2
nb(n− b+ 1)2
b∑k,`=1k<`
[b−2∑i=k
b−1∑j=`
ψ(Xi −Xj) +n−b+k∑
i=n−b+2
n−b+`∑j=n−b+3
ψ(Xi −Xj)
+b−2∑i=k
n−b+`∑j=n−b+3
ψ(Xi −Xj) +n−b+k∑
i=n−b+2
b−1∑j=`
ψ(Xi −Xj)
]
+b−1∑k=1
(b− k)b−2∑i=k
n−b+2∑j=b
ψ(Xi −Xj) +b∑
`=2
(`− 1)n−b+1∑i=b−1
b−1∑j=`
ψ(Xi −Xj)
+b∑
`=2
(`− 1)n−b+1∑i=b−1
n−b+`∑j=n−b+3
ψ(Xi −Xj) +b−1∑k=1
(b− k)n−b+k∑
i=n−b+2
n−b+2∑j=b
ψ(Xi −Xj)
],
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 24 / 81
![Page 25: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/25.jpg)
Smooth Moving Blocks Bootstrap (SMBB)
MISE closed expression for SMBB
considering aj such that:
aj =
j
b(n− b+ 1) , if j = 1, . . . , b− 1
1n− b+ 1 , if j = b, . . . , n− b+ 1
n− j + 1b(n− b+ 1) , if j = n− b+ 2, . . . , n
.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 25 / 81
![Page 26: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/26.jpg)
Simulations
Simulated models
Six time series models have been considered
Model 1:Xt = −0.9Xt−1 − 0.2Xt−2 + at,
where the atd= N(0, 1) are independent. Thus Xt
d= N(0, 0.42)Model 2:
Xt = at − 0.9at−1 + 0.2at−2,
where atd= N(0, 1) are independent. Thus Xt
d= N(0, 1.85).
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 26 / 81
![Page 27: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/27.jpg)
Simulations
Simulated models
Model 3:Xt = φXt−1 + (1− φ2)1/2at,
with atd= N(0, 1), φ = 0,±0.3,±0.6,±0.9. Thus Xt
d= N(0, 1).Model 4:
Xt = φXt−1 + at,
where the distribution of at is given by P(It = 1) = φ,P(It = 2) = 1− φ, with at|It=1
d= 0 (constant), at|It=2d= exp(1),
and φ = 0, 0.3, 0.6, 0.9. We have Xtd= exp(1)
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 27 / 81
![Page 28: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/28.jpg)
Simulations
Simulated models
Model 5:Xt = φXt−1 + at,
where the distribution of at is P(It = 1) = φ2, P(It = 2) = 1− φ2,with at|It=1
d= 0 (constant), at|It=2d= Dexp(1), and
φ = 0,±0.3,±0.6,±0.9. Thus Xtd= Dexp(1).
Model 6:
Xt ={X
(1)t with probability 1/2
X(2)t with probability 1/2
,
where X(j)t = (−1)j+1 + 0.5X(j)
t−1 + a(j)t with j = 1, 2, ∀t ∈ Z,
a(j)t
d= N(0, 0.6) independent and Xtd= 1
2N(2, 0.8) + 12N(−2, 0.8)
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 28 / 81
![Page 29: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/29.jpg)
Simulations
Performance measures
The following results will be shown for the six models considered in thesimulations
log(
h
hMISE
)
log(
MISE(h)MISE(hMISE)
),
where h = hCVl, hSMCV , hPCV , hPI , h
∗SSB, h
∗SMBB.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 29 / 81
![Page 30: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/30.jpg)
Simulations
Approximating the optimal bandwidth
Consider some criterion function Ψ(h) (e.g. MISE∗(h) under SSB orSMBB; CVl(h) for Hart and Vieu’s CV, Stute’s MCV or Estevez, Quintelaand Vieu PCV).
1 Consider a set of five equispaced bandwidths, H1 between 0.01 and 102 Obtain hOPT1 = arg minh∈H1 Ψ(h)3 Consider ha the previous value of hOPT1 within H1 and hb the
following value to hOPT1 within H1
4 Construct a new set, H2, of equispaced bandwidths between ha andhb
5 Repeat Steps 2-4 10 times6 The approximated optimal bandwidth is the value obtained in the
10th repetition
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 30 / 81
![Page 31: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/31.jpg)
Simulations
Technical aspects
l = 5 for CVlhSMCV is considered as the smallest h for which SMCV (h) attains alocal minimum, not its global onePilot bandwidth for PI: h1 = 1Pilot bandwidth for h∗SSB and h∗SMBB as in the iid case: some normalreference estimator of
g0 =( ∫
K ′′(t)2dt
ndK∫f (3)(x)2dx
)1/7
p = 0.05 for SSBb = 20 for SMBBFor every model, 1000 random samples of size n = 100 were drawn
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 31 / 81
![Page 32: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/32.jpg)
Simulations
log(h/hMISE). Model 1
CV_l SMCV PCV SSB SMBB PI
−2
02
46
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 32 / 81
![Page 33: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/33.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 1
CV_l SMCV PCV SSB SMBB PI
01
23
4
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 33 / 81
![Page 34: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/34.jpg)
Simulations
log(h/hMISE). Model 2
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
23
45
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 34 / 81
![Page 35: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/35.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 2
CV_l SMCV PCV SSB SMBB PI
01
23
4
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 35 / 81
![Page 36: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/36.jpg)
Simulations
log(h/hMISE). Model 3, φ = −0.9
CV_l SMCV PCV SSB SMBB PI
−2
02
4
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 36 / 81
![Page 37: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/37.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = −0.9
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
3.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 37 / 81
![Page 38: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/38.jpg)
Simulations
log(h/hMISE). Model 3, φ = −0.6
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 38 / 81
![Page 39: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/39.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = −0.6
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 39 / 81
![Page 40: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/40.jpg)
Simulations
log(h/hMISE). Model 3, φ = −0.3
CV_l SMCV PCV SSB SMBB PI
−3.
0−
2.5
−2.
0−
1.5
−1.
0−
0.5
0.0
0.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 40 / 81
![Page 41: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/41.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = −0.3
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
3.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 41 / 81
![Page 42: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/42.jpg)
Simulations
log(h/hMISE). Model 3, φ = 0
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
2
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 42 / 81
![Page 43: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/43.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = 0
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 43 / 81
![Page 44: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/44.jpg)
Simulations
log(h/hMISE). Model 3, φ = 0.3
CV_l SMCV PCV SSB SMBB PI
−3.
0−
2.5
−2.
0−
1.5
−1.
0−
0.5
0.0
0.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 44 / 81
![Page 45: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/45.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = 0.3
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 45 / 81
![Page 46: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/46.jpg)
Simulations
log(h/hMISE). Model 3, φ = 0.6
CV_l SMCV PCV SSB SMBB PI
−2.
0−
1.5
−1.
0−
0.5
0.0
0.5
1.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 46 / 81
![Page 47: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/47.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = 0.6
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 47 / 81
![Page 48: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/48.jpg)
Simulations
log(h/hMISE). Model 3, φ = 0.9
CV_l SMCV PCV SSB SMBB PI
−3
−2
−1
01
2
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 48 / 81
![Page 49: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/49.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 3, φ = 0.9
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 49 / 81
![Page 50: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/50.jpg)
Simulations
log(h/hMISE). Model 4, φ = 0
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
2
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 50 / 81
![Page 51: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/51.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 4, φ = 0
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 51 / 81
![Page 52: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/52.jpg)
Simulations
log(h/hMISE). Model 4, φ = 0.3
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
23
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 52 / 81
![Page 53: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/53.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 4, φ = 0.3
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 53 / 81
![Page 54: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/54.jpg)
Simulations
log(h/hMISE). Model 4, φ = 0.6
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
23
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 54 / 81
![Page 55: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/55.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 4, φ = 0.6
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 55 / 81
![Page 56: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/56.jpg)
Simulations
log(h/hMISE). Model 4, φ = 0.9
CV_l SMCV PCV SSB SMBB PI
−3
−2
−1
01
23
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 56 / 81
![Page 57: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/57.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 4, φ = 0.9
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 57 / 81
![Page 58: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/58.jpg)
Simulations
log(h/hMISE). Model 5, φ = −0.9
CV_l SMCV PCV SSB SMBB PI
−4
−2
02
46
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 58 / 81
![Page 59: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/59.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = −0.9
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 59 / 81
![Page 60: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/60.jpg)
Simulations
log(h/hMISE). Model 5, φ = −0.6
CV_l SMCV PCV SSB SMBB PI
−3
−2
−1
01
23
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 60 / 81
![Page 61: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/61.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = −0.6
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
3.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 61 / 81
![Page 62: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/62.jpg)
Simulations
log(h/hMISE). Model 5, φ = −0.3
CV_l SMCV PCV SSB SMBB PI
−4
−2
02
46
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 62 / 81
![Page 63: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/63.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = −0.3
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 63 / 81
![Page 64: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/64.jpg)
Simulations
log(h/hMISE). Model 5, φ = 0
CV_l SMCV PCV SSB SMBB PI
−2
−1
01
2
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 64 / 81
![Page 65: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/65.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = 0
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
3.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 65 / 81
![Page 66: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/66.jpg)
Simulations
log(h/hMISE). Model 5, φ = 0.3
CV_l SMCV PCV SSB SMBB PI
−2
02
4
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 66 / 81
![Page 67: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/67.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = 0.3
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 67 / 81
![Page 68: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/68.jpg)
Simulations
log(h/hMISE). Model 5, φ = 0.6
CV_l SMCV PCV SSB SMBB PI
−2
02
4
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 68 / 81
![Page 69: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/69.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = 0.6
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
2.5
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 69 / 81
![Page 70: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/70.jpg)
Simulations
log(h/hMISE). Model 5, φ = 0.9
CV_l SMCV PCV SSB SMBB PI
−4
−2
02
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 70 / 81
![Page 71: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/71.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 5, φ = 0.9
CV_l SMCV PCV SSB SMBB PI
0.0
0.5
1.0
1.5
2.0
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 71 / 81
![Page 72: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/72.jpg)
Simulations
log(h/hMISE). Model 6
CV_l SMCV PCV SSB SMBB PI
−4
−3
−2
−1
01
23
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 72 / 81
![Page 73: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/73.jpg)
Simulations
log(MISE(h)/MISE(hMISE)). Model 6
CV_l SMCV PCV SSB SMBB PI
0.0
0.2
0.4
0.6
0.8
1.0
1.2
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 73 / 81
![Page 74: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/74.jpg)
Real data application
Real data application: Data sets considered
1 lynx data set: Number of Canadian lynxes trapped (114 observations).
Time
serie
1820 1840 1860 1880 1900 19200
1000
2000
3000
4000
5000
6000
7000
Time
serie
1820 1840 1860 1880 1900 1920
45
67
89
(1− φ1B − φ2B2)Yt = c+ (1 + θ1B + θ2B
2 + θ3B3)(1 + Θ1B
12)at.
2 sunspot.year data set: Yearly number of sunspots from 1700 to 1988(289 observations).
Time
serie
1700 1750 1800 1850 1900 1950
050
100
150
Time
serie
1700 1750 1800 1850 1900 1950
12
34
5
(1− φ1B − φ2B2 − φ2B3 − φ4B
4)(1−B)(1−B12)Yt =
c+ (1 + θ1B + θ2B2 + θ3B
3 + θ4B4) · (1 +B12Θ1)at.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 74 / 81
![Page 75: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/75.jpg)
Real data application
Real data application: lynx data set
4 6 8 10
0.00
0.05
0.10
0.15
0.20
0.25
0.30
Den
sity
h_SSBh_SMBBh_CV_lh_PCVh_SMCVh_PI
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 75 / 81
![Page 76: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/76.jpg)
Real data application
Real data application: sunspot.year data set
0 1 2 3 4 5 6
0.0
0.1
0.2
0.3
0.4
Den
sity
h_SSBh_SMBBh_CV_lh_PCVh_SMCVh_PI
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 76 / 81
![Page 77: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/77.jpg)
Real data application
Real data application: Bandwidth parameters
h∗SSB h∗SMBB hCVlhPCV hSMCV hPI
0.4345 0.4246 0.3173 0.6194 0.2585 0.4152
Table: Bandwidth parameters for lynx data set.
h∗SSB h∗SMBB hCVlhPCV hSMCV hPI
0.3173 0.3295 0.3002 0.5065 0.196 0.3392
Table: Bandwidth parameters for sunspot.year data set.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 77 / 81
![Page 78: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/78.jpg)
Conclusions
Main conclusions
New SSB and SMBB bootstrap resampling plans under dependence.Closed expressions for MISE∗ under SSB and SMBB. Monte Carlo isnot needed.Bandwidth selection for the KDE with dependent data:
Plug-inLeave-(2l + 1)-out cross validationPenalized cross validationModified cross validationSmooth Stationary BootstrapSmooth Moving Blocks Bootstrap
Good empirical behaviour of hPI , but sometimes it producesextremely large banwdidthsh∗SSB and h∗SMBB display the overall best performance.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 78 / 81
![Page 79: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/79.jpg)
Conclusions
References
Barbeito, I. and Cao, R. (2016). Smoothed stationary bootstrapbandwidth selection for density estimation with dependent data.Unpublished manuscript.
Cao, R. (1993). Bootstrapping the mean integrated squared error. J.Mult. Anal. 45, 137-160.Efron, B. and Tibishirani, R. (1986). An Introduction to theBootstrap. Chapman and Hall.
Estevez-Perez, G., Quintela-del-Rıo, A. and Vieu, P. (2002).Convergence rate for cross-validatory bandwidth in kernel hazardestimation from dependent samples. J. Statist. Planning andInference, 104, 1-30.Hall, P., Lahiri, S.N. and Truong, Y.K.(1995). On bandwidth choicefor density estimation with dependent data. Ann. Statist., 23, 6,2241-2263.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 79 / 81
![Page 80: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/80.jpg)
Conclusions
References (continued)
Hart, J. and Vieu, P. (1990). Data-driven bandwidth choice for densityestimation based on dependent data. Ann. Statist., 18, 873-890.
Kunsch, H.R. (1989). The jackknife and the bootstrap for generalstationary observations. Ann. Statist., 17, 1217-1241.
Liu, R.Y., and Singh, K. (1992). Moving blocks jackknife andbootstrap capture weak dependence, in Exploring the Limits of theBootstrap, ed. by R. LePage and L. Billiard. New York: Wiley.
Politis, D.N. and Romano, J.R. (1994a). The stationary bootstrap. J.Amer. Statist. Assoc., 89, 1303-1313.Politis, D.N. and Romano, J.R. (1994b). Large Sample ConfidenceRegions Based on Subsamples under Minimal Assumptions. Ann.Statist., 22, 2031-2050 .Stute, W. (1992). Modified cross-validation in density estimation. J.Statist. Planning and Inference, 3, 293-305.
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 80 / 81
![Page 81: Smoothed stationary bootstrap bandwidth selection for density …eio.usc.es/pub/gsnsi/descargas/Cao_GSNSI.pdf · 2016-06-08 · Smoothed stationary bootstrap bandwidth selection for](https://reader034.vdocument.in/reader034/viewer/2022042206/5ea928953e0c463d555cf17a/html5/thumbnails/81.jpg)
Conclusions
Contact info
Thank you for your attention!
You can contact me at [email protected]
R. Cao SSB bandwidth selection GSNSI, June 8, 2016 81 / 81