Time Series Analysis by State Space Methods

RK
May 5, 2014

Abstract
The purpose of this document is to summarize Part I of the book Time Series Analysis by State Space
Methods , written by James Durbin and Siem Jan Koopman.

Time Series Analysis by State Space Methods

Contents
1 Introduction

2 Local Level Model

2.1

Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2

Forecast errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.3
2.4

State Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Disturbance smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8
9

2.5

Simulation Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

2.6

Missing observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12

2.7

Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

2.8

Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

15

2.9

Parameter estimation

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.10 Steady state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.11 Diagnostic checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16

3 Linear Gaussian state space models

19

3.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

19

3.2

Structural time series models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

19

3.3

ARMA models and ARIMA models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

19

3.4

Exponential smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

3.5

State space versus Box-Jenkins approaches

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

3.6

Regression with time-varying coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

3.7

Regression with ARMA error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

4 Filtering, smoothing and forecasting

25

4.1

Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

4.2

State smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

4.3

Disturbance smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

26

4.4

Covariance matrices of smoothed observations . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

4.5

Weight functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

4.6

Simulation Smoothing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

4.7

Simulation of State vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

28

4.8

Missing observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

36

4.9

Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

36

5 Initialisation of filter and smoother

37

6 Further Computational Aspects

38

6.1

Square root filter and smoother . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

38

6.2

Univariate treatment of multivariate series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

6.3

The algorithms of SsfPack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

Time Series Analysis by State Space Methods

7 Maximum Likelihood estimation

40

7.1

Likelihood function, Hessian and Score function . . . . . . . . . . . . . . . . . . . . . . . . . . .

40

7.2

EM Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

8 Bayesian Analysis

44

8.1

Posterior analysis of state vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

44

8.2

MCMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

44

9 Illustrations of the use of the linear Gaussian model

45

9.1

Structural time series models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

45

9.2

Box-Jenkins analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

52

Time Series Analysis by State Space Methods

Preface
The distinguishing feature of state space time series models is that observations are regarded as made up
of distinct components such as trend, seasonal, regression elements and disturbance terms, each of which is
modeled separately. These models for the components are put together to form a single model called a state
space model which provides the basis for analysis. The book is primarily aimed at applied statisticians and
econometricians. Not much of math background is needed to go through the book. State space time series
analysis began with the path breaking paper of Kalman and early developments of the subject took place in
the field of engineering. The term state space comes form engineering. Statisticians and econometricians
tend to stick to the same terminology.
Part I of the book deals with linear Gaussian models and Part II deals with the extensions to non linear,
non Gaussian world. Extensions like exponentially distributed observations , allowing non-linearities in the
model , allowing heavy-tailed densities to deal with outliers in the observations and structural shifts in the
state are dealt in the latter part of the book. The treatment given in the book is a simulation based approach, as excepting for a few models, there are no closed form solutions available. Instead of following
the MCMC treatment, the book shows the use of importance sampling and antithetic variables in analyzing
state space models. The book provides analysis from the classical treatment as well as the Bayesian treatment.
In this document I will summarize Part I of the book that take up 175 out of 240 odd pages in the book. Part
II of the book is all about nonlinear non Gaussian models and the use of importance sampling to estimate the
filtering and smoothing estimates. In contrast to Part II of the book, the math in the first part of the book is
definitely manageable. In fact one can go through the entire Part I of the book by knowing just one Lemma.In
any case, let me summarize the contents of the book and also provide R code to replicate some of the visuals
and numbers from the book.

Introduction

The developments of the system under study are guided by unobserved series of state vectors 1 , 2 , . . . with
which are associated a series of observations y1 , y2 , . . . , yn , the relation between the s and yt s are specified
by the state space model. The purpose of the state space analysis is to infer the relevant properties of the s
from a knowledge of observations y1 , . . . , yn . The introductory chapter of the book gives the road map to the
entire book by explaining a bit about each of the 14 chapters in the book.

Local Level Model

This chapter is like a trailer to the first part of the book. Instead of bombarding the reader with the math
right from the start, the authors take a more tempered approach and present the simplest of all state space
models, Local Level Model. This model takes the following form

yt =t + t ,

t N (0, 2 )

t+1 = t + t ,

t N (0, 2 ),

1 N (a1 , P1 )

Time Series Analysis by State Space Methods

Even though the form looks very simple, it is the basis for exponentially weighted moving average method,
one of the most popular methods for smoothing a time series. Given the above model, one can write down the
distribution in a multivariate form and do whatever stats one needs to do. However the routine computations
become very cumbersome. This is where Kalman Filter comes in. By filtering and smoothing the series, KF
offers an efficient way of obtaining results that arise from Multivariate theory. The chapter gives extensive
visuals for each of the concepts that are discussed in the context of local level model. I guess this is the
chapter with the maximum visuals out of all the fourteen chapters. That also means that this book is filled
with math and one has to patiently work through the details. The authors are not in a mood to dumb down
the math stuff and one has to slog through the book to understand the concepts. No easy path here. At least
thats what I felt while going over the book. Through the use of local level model, the authors touch upon the
following principles of state space analysis
Filtering
Smoothing
Disturbance Smoothing
Simulation Smoothing
How to deal with missing observations ?
How to forecast the state vector n steps ahead ?
How to initialize the recursive system ?
How to estimate the meta parameters of the local level model ?
Diagnostics to check model validity

All the visuals in this chapter can be replicated via KFAS package in R. My goto package for state space modeling
till now has been dlm. However for this book, I have found that KFAS closely mirrors all the algorithms from
the book.
The first question that comes up for anyone just beginning to understand local level model is that,Why do we
need a recursive updating of various estimates? Isnt the local level model after all a multivariate realization?.
y N (1a1 , ), with = 110 P1 +
where the (i, j)th element of the n n matrix is given by

(i 1)2 , i < j

ij = 2 + (i 1)2 , i = j

(j 1) 2 , i > j

Given the above equations, one can compute away all the conditional means, variances and covariances.
However there is problem here. As soon as n increases, the computations become very cumbersome. Thats
why you need a recursive framework so that you dont carry much luggage as you chug along. By using filtering
and smoothing, the estimation can be improved considerably.

Time Series Analysis by State Space Methods

2.1

Filtering

This section derives the filtering recursive equations for the local level model. They are
t = yt at
Ft = Pt + 2
Kt = Pt /Ft
at+1 = at + Kt t
Pt+1 = Pt (1 Kt ) + 2
where at = E(t |Yt1 ) and Pt = V ar(t |Yt1 )
The above relations constitute the celebrated Kalman Filter for this problem.
I will use the samedata set that the authors use in the book and replicate the figures via R packages.
data(Nile)
model

<- SSModel( Nile ~ SSMtrend(1, Q = list(matrix(NA))) , H = matrix(NA))

model

<- fitSSM( inits = c(log(var(Nile)) , log(var(Nile) ) ) ,

model = model, method

out

='BFGS' )$model

<- KFS(model, filtering='state', smoothing=c('state','disturbance'),simplify=F)

par(mfrow=c(2,2))
conf.bands <- cbind(out$a, as.vector(out$a) +
(sqrt(cbind(out$P))* qnorm(0.95))%*%
temp

<- cbind(Nile, conf.bands[-1,])

cols

<- c("grey","blue","red","red")

lwd

<- c(1,2,1,1)

lty

(t(c(-1,1))))

<- c(1,1,2,2)

plot.ts(temp, plot.type="single", col = cols,lwd = lwd,lty=lty,xlab="",ylab="",main="i")

legend("topright",
legend = c("Observation Data","Filtered state","Confidence intervals"),
col = c("grey","blue","red"), lty = c(1,1,2),
lwd = c(1,2,1),bty="n", cex=0.9 )
temp

<- ts( c(out$P)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="ii")

temp

<- ts( c(out$v[-1]),start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="iii")

abline(h=0,col="grey")
temp

<- ts( c(out$F)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="", main = "iv")

Time Series Analysis by State Space Methods

6000

600

10000

1000

Observation Data
Filtered state
Confidence intervals

16000

ii

1400

1880

1900

1920

1940

1960

1880

1900

1940

1960

1940

1960

iv

400

22000

28000

200

iii

1920

1880

1900

1920

1940

1960

1880

1900

1920

Figure 2.1: Nile data and output of Kalman filter : (i) filtered state and confidence intervals, (ii) filtered state variance,
(iii) prediction error , (iv) prediction variance

The most obvious feature of the above illustrations is that Ft and Pt converge rapidly to constant values which
confirms that the local level model has a steady state solution.

2.2

Forecast errors

This section shows that the relation between innovation terms in the model, i.e t = yt at are observations
= C(y 1a1 )
where C is a lower triangular matrix.
N (0, CC 0 ),

N (0, F )

The transformation of a symmetric positive matrix in to a diagonal matrix F using a lower triangular matrix
C is known as Cholesky decomposition. The Kalman filter can therefore be regarded as essentially a Cholesky
decomposition of the variance matrix implied by the local level model.
One can also restate the state space model in its innovation form

t = xt + t ,

t N (0, 2 )

xt+1 = Lt xt + t Kt t
Lt = 1 Kt

Time Series Analysis by State Space Methods

2.3

State Smoothing

Smoothing essentially comprises estimation of 1 , . . . , n given the entire sample Yn . The basic behind these
computations is that the forecast errors 1 , . . . , n are mutually independent and linear transformation of
y1 , . . . , yn , and t , . . . , n are independent of y1 , . . . , yt1 with zero means. Using the standard multivariate
theory, the smoothed state estimate and variance are derived. The following are the state smoothing recursions
and state variance smoothing recursions:

t = at + Pt rt1
t
rt1 =
+ Lt rt
Ft
t+1
t+2
n
rt =
+ Lt+1
+ . . . + Lt+1 + . . . + Ln1
Ft+1
Ft+2
Fn
rn = 0, for t = n, n 1, . . . , 1
Nt1 = Ft1 + L2t Nt
Vt = Pt Pt2 Nt1
1
1
1
Nt =
+ L2t+1
+ . . . + L2t+1 . . . L2n1
Ft+1
Ft+2
Fn
Nn = 0

par(mfrow=c(2,2))
conf.bands <- cbind(out$alphahat, as.vector(out$alphahat) +
(sqrt(cbind(out$V))* qnorm(0.95))%*%
temp

<- cbind(Nile, conf.bands)

cols

<- c("grey","blue","red","red")

lwd

<- c(1,2,1,1)

lty

(t(c(-1,1))))

<- c(1,1,2,2)

plot.ts(temp, plot.type="single", col = cols,lwd = lwd,

lty=lty,xlab="",ylab="",main="i")
legend("topright",
legend = c("Observation Data","Smoothed state","Confidence intervals"),
col = c("grey","blue","red"), lty = c(1,1,2),
lwd = c(1,2,1),bty="n", cex=0.7 )
temp

<- ts( c(out$V),start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="ii")

temp

<- ts( c(out$r)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="iii")

abline(h=0,col="grey")
temp

<- ts( c(out$N)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main = "iv")

Time Series Analysis by State Space Methods

2500

600

1000

Observation Data
Smoothed state
Confidence intervals

3500

ii

1400

1880

1900

1920

1940

1960

1880

1900

1940

1960

1940

1960

iv

0.03

0.01

0.01

0e+00 4e05 8e05

iii

1920

1880

1900

1920

1940

1960

1880

1900

1920

Figure 2.2: Nile data and output of state smoothing recursion: (i) smoothed state and confidence intervals, (ii) smoothed
state variance, (iii) smoothing cumulant , (iv) smoothing variance cumulant

2.4

Disturbance smoothing

When I first learnt about disturbance smoothing, I did not understand the point in computing smoothed
observation disturbance E(t |y) and E(t |y). Later I realized that they are immensely useful in making MLE
computations quicker and in simulating state vectors. The section does not derive these equations and instead
refers to the later chapters for the complete derivation. It merely states the results for the local level model.
The following are the recursive relations for disturbance smoothing

t = 2 ut
ut = Ft1 t Kt rt
Var(t |y) = 2 4 Dt
Dt = Ft1 + Kt2 Nt
t = 2 rt
ut = Ft1 t Kt rt
Var(t |y) = 2 4 Nt

par(mfrow=c(2,2))
temp

<- ts( c(out$epshat),start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="i")

abline(h=0,col="grey")
temp

<- ts( c(out$V_eps),start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="ii")

Time Series Analysis by State Space Methods

temp

<- ts( c(out$etahat),start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="iii")

abline(h=0,col="grey")
temp

<- ts( c(out$V_eta),start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main="iv")

ii

300

2500

100

3500

100

1880

1900

1920

1940

1960

1880

1900

1940

1960

1940

1960

iv

1250

40

1350

20

1450

iii

1920

1880

1900

1920

1940

1960

1880

1900

1920

Figure 2.3: Outout of disturbance smoothing recursion : (i) observation error, (ii) observation error variance , (iii) state
error , (iv) state error variance

2.5

Simulation Smoothing

The authors have come up with a better simulation smoothing algorithm than the one mentioned in this book
after the book went for publishing. It is available here. KFAS implements this simulation smoother.
set.seed(1)
n

<- length(Nile)

theta.uncond <- numeric(n)

theta.uncond[1] <- Nile[1]
for(i in 2:n){
theta.uncond[i] <- theta.uncond[i-1] + rnorm(1,0,sqrt(out$model$Q))
}
theta.cond <- simulateSSM(model, type = c("states"))
eps.sim

<- simulateSSM(model, type = c("epsilon"))

eta.sim

<- simulateSSM(model, type = c("eta"))

10

Time Series Analysis by State Space Methods

par(mfrow=c(2,2))
ylim = c(400,max(c(out$alphahathat,theta.cond,theta.uncond))+50)
plot(1:100, Nile, type="l", col = "grey", ylim = ylim, main = "i")
points(1:100, out$alphahat , type="l", col = "green", lwd = 2)
points(1:100, c(theta.uncond), type="l", col = "red", lwd = 2)
leg <-c("actual data","smoothed estimate","unconditional sample")
legend("topleft",leg,col=c("grey","green","red"),lwd=c(1,2,2),cex=0.7,bty="n")

plot(1:100, Nile, type="l", col = "grey", ylim = ylim, main = "ii")

points(1:100, out$alphahat , type="l", col = "green", lwd = 2)
points(1:100, c(theta.cond), type="l", col = "blue", lwd = 2)
leg <-c("actual data","smoothed estimate","conditional sample")
legend("topleft",leg,col=c("grey","green","blue"),lwd=c(1,2,2),cex=0.7,bty="n")
temp

<- ts( c(out$epshat),start = 1871)

ylim

<- range(out$epshat,eps.sim)

plot(temp, type="l", col = "blue",ylim=c(-400,300),

, main = "iii")
points(1871:1970,eps.sim,pch = 19,col="sienna",cex=0.5)
temp

<- ts( c(out$etahat),start = 1871)

ylim

<- range(out$etahat,eta.sim)

plot(temp, type="l", col = "blue",ylim=c(-400,300),

, main = "iv")
points(1871:1970,eta.sim,pch = 19,col="sienna", cex=0.5)

11

Time Series Analysis by State Space Methods

ii

40

60

80

100

20

40

60

1:100

1:100

iii

iv

80

100

temp

400

400

100 100

100 100

300

20

300

temp

1200
400

400

800

Nile

1200

actual data
smoothed estimate
conditional sample

800

Nile

actual data
smoothed estimate
unconditional sample

1880

1900

1920

1940

1960

1880

1900

Time

1920

1940

1960

Time

Figure 2.4: Simulation : (i) smoothed state and unconditional sample, (ii) smoothed state and conditional sample ,
(iii) smoothed observation error and conditional sample, (iv) smoothed state error and conditional sample

2.6

Missing observations

Missing observations are easily handled in the state space setting where all the recursive equations hold good
with the respective innovation term and Kalman gain equated to 0 for the missing observations.
Nile.miss
<- Nile
Nile.miss[21:40] <- NA
Nile.miss[61:80] <- NA
model.miss <- SSModel( Nile.miss ~ SSMtrend(1, Q = c(model$Q)) , H = c(model$H))
out.miss

<- KFS(model.miss, filtering='state',

smoothing=c('state','signal','disturbance'), simplify=F)

par(mfrow=c(2,2))
temp

<- cbind(Nile.miss, c(out.miss$a)[-1])

cols

<- c("grey","blue")

plot.ts(temp, plot.type="single", col = cols,xlab="",ylab="",

main="i", lwd=c(1,2,2))
legend("topright",
legend = c("Observation Data","Filtered"),
col = c("grey","blue"),
lwd = c(1,2),bty="n", cex=0.8 )

12

Time Series Analysis by State Space Methods

temp

<- ts( c(out.miss$P)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main ="ii")

temp

<- cbind(Nile.miss, c(out.miss$alphahat))

cols
<- c("grey","blue")
plot.ts(temp, plot.type="single", col = cols,xlab="",ylab="",
main="iii", lwd=c(1,2,2))
legend("topright",
legend = c("Observation Data","Smoothed"),
col = c("grey","blue"),
lwd = c(1,2),bty="n", cex=0.8 )

temp

<- ts( c(out.miss$V)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main ="iv")

600

5000 15000

1000

Observation Data
Filtered

30000

ii

1400

1880

1900

1920

1940

1960

1880

1900

1940

1960

1940

1960

600

4000

1000

Observation Data
Smoothed

8000

iv

1400

iii

1920

1880

1900

1920

1940

1960

1880

1900

1920

Figure 2.5: Filtered and smoothing output when observations are missing (i) filtered state , (ii) filtered state variance
, (iii) smoothed state (iv) smoothed state variance

2.7

Forecasting

Forecasting in State space setting is done via simply assuming that the forecast period observations are
missing and hence the Kalman gain and innovation for the period are assumed to 0. The chapter illustrates
by extending the Nile dataset and placing NA for the observations

13

Time Series Analysis by State Space Methods

data(Nile)
Nile2

<- c(Nile,rep(NA,30))

model

<- SSModel( Nile2 ~ SSMtrend(1, Q = c(model$Q)) , H = c(model$H))

out

<- KFS(model, filtering='state',

smoothing=c('state','signal','disturbance'),simplify=F)

par(mfrow=c(2,2))
conf.bands <- cbind(out$a, as.vector(out$a) +
(sqrt(cbind(out$P))* qnorm(0.95))%*%
temp

<- cbind(Nile2, conf.bands[-1,])

cols

<- c("grey","blue","red","red")

lwd
lty

(t(c(-1,1))))

<- c(1,2,1,1)
<- c(1,1,2,2)

plot.ts(temp, plot.type="single", col = cols,lwd = lwd,lty=lty,xlab="",ylab="",main="i")

legend("topright",
legend = c("Observation Data","Forecasts","Confidence intervals"),
col = c("grey","blue","red"), lty = c(1,1,2),
lwd = c(1,2,1),bty="n", cex=0.8 )

temp

<- ts( c(out$P)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",main ="ii")

temp

<- ts( c(out$alphahat)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",

ylim=c(700,1250),main = "iii")
temp

<- ts( c(out$F)[-1],start = 1871)

plot(temp, col = "blue",lwd = 2,

xlab="",ylab="",main = "iv")

14

Time Series Analysis by State Space Methods

ii

400

10000

800

30000

1200

Observation Data
Forecasts
Confidence intervals

50000

20

40

60

80

100

120

1880

1900

1920

1960

1980

2000

1960

1980

2000

iv

700

20000

900

40000

1100

60000

iii

1940

1880

1900

1920

1940

1960

1980

2000

1880

1900

1920

1940

Figure 2.6: Nike data and output of forecasting (i) state forecast and confidence intervals, (ii) state variance , (iii)
observation forecast (iv) observation forecast variance

2.8

Initialization

This section explains about initialization of KF. When one does not know anything about the prior density of
the initial state, it is reasoanable to represent 1 as having diffuse prior density.
1 = y1 a1 ,

F1 = P1 + 2

By substituting the above values in to the equations for a2 and P2 , and letting P1 , we get a2 = y1 , P2 =
2 + 2 , and we can then proceed normally with KF for t = 2, . . . , n. Based on the diffuse prior, the state
and smoothing equations also have to be changed. They are not affected for t = 2, . . . , n but for the smoothed
state for initial state and smoothed conditional variance of initial state change. So do smoothed mean and
variance of the disturbances.

1 = y1 + 2 r1
V1 = 2 4 N1
Use of diffuse prior might find resistance from some people as they may regard the assumption of infinite
variance as unnatural. An alternative approach is to assume 1 is an unknown constant to be estimated
from data. Using this approach one initializes a1 = y1 and P1 = 2 . It follows that we obtain the same
initialization of KF by representing 1 as a random variable with infinite variance as by assuming that it is
fixed and unknown.

15

Time Series Analysis by State Space Methods

2.9

Parameter estimation

Parameters in the state space models are often called hyperparameters to distinguish them from elements
of state vector which can plausibly be thought of as random parameters.The loglikelihood function of the
observation data is written using the conditional probability of innovation terms and a straightforward MLE
estimation can be done. However a better way to estimate parameters is by using EM algorithm that is dealt
in Chapter 7. To do MLE using KFAS, one can use fitSSM function.
model

<- SSModel( Nile ~ SSMtrend(1, Q = list(matrix(NA))) , H = matrix(NA))

model

<- fitSSM( inits = c(log(var(Nile)) , log(var(Nile) ) ) ,

model = model, method

='BFGS' )$model

c(model$H)
## [1] 15099
c(model$Q)
## [1] 1469
The section also shows a way to re-parametrise the model to obtain a concentrated diffuse likelihood function.
This is maximized to estimate the parameters.

2.10

Steady state

KF converging to a steady state is important thing to know as it makes the recursive computations even faster.
The condition for steady state is given by the condition
x > 0, where x = h +

p
h2 + 4 h/2,

h = 2 /2

This holds for all non-trivial local level models.

2.11

Diagnostic checking

The assumptions underlying the local level model are that the disturbances t and t are normally distributed
and serially independent with constant variances. On these assumptions the standardized one-step forecast
errors are also normally distributed and serially independent with unit variance. KFAS has functions to retrieve
the standardized residuals and one step ahead forecast errors
out

<- KFS(model, filtering='state', smoothing=c('state','signal','disturbance'),

simplify=F)

stdres

<- rstandard(out)

par(mfrow=c(2,2))
temp

<- ts(stdres,start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="", main ="i")

abline(h=0, col = "grey")

16

Time Series Analysis by State Space Methods

hist(stdres, prob=T, col = "grey", main = "ii", xlab ="",ylab="")

lines(density(stdres[-1]), col = "blue",lwd = 2)
qqnorm(c(stdres), col="blue",main = "iii")
qqline(c(stdres),col="red")
acf(stdres[-1], ,main="iv")

ii

0.0

0.2

1 0

0.4

1880

1900

1920

1940

1960

iv

ACF

0.2

0.2

0.6

2
1
1 0
3

Sample Quantiles

1.0

iii

Theoretical Quantiles

10

15

Lag

Figure 2.7: Diagnostic plots for standardised prediction errors : (i) standardised residual, (ii) histogram plus estimated
density, (iii) ordered residuals (iv) correlogram

<- KFS(model, filtering='state', smoothing=c('state','signal','disturbance'),

out

simplify=F)
temp1

<-

out$epshat/sqrt(c(out$V_eps))

temp2

<-

out$etahat/sqrt(c(out$V_eta))

par(mfrow=c(2,2))
temp

<- ts(temp1,start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="",

main="i")
abline(h=0, col = "grey")
hist(temp, prob=T, col = "grey", ,ylim=c(0,0.3),
main="ii",xlab="",ylab="")
lines(density(temp), col = "blue",lwd = 2)

17

Time Series Analysis by State Space Methods

temp

<- ts(temp2,start = 1871)

plot(temp, col = "blue",lwd = 2,xlab="",ylab="" ,main ="iii")

abline(h=0, col = "grey")
hist(temp, prob=T, col = "grey", main = "iv",
ylim=c(0,1.2),xlab="",ylab="")
lines(density(temp), col = "blue",lwd = 2)

ii

0.00

0.10

0.20

0.30

1880

1900

1920

1940

1960

iv

0.0

1.0

0.4

0.0

0.8

1.2

iii

1880

1900

1920

1940

1960

1.5

1.0

0.5

0.0

0.5

1.0

Figure 2.8: Auxilary Residual plots : (i) observation residual , (ii) histogram and estimated density for observation
residual, (iii) state residual , (iv) histogram and estimated density for state residual

By sprinkling the chapter with adequate visuals, the authors provide the reader a nice trailer of all the principles
that are relevant for state space modeling.

18

Time Series Analysis by State Space Methods

3
3.1

Linear Gaussian state space models

Introduction

The general linear Gaussian state space model can be written in a variety of ways. The version used in the
book is

y t = Z t t + t ,
t+1 = Tt t + Rt ,

t N (0, Ht )
t N (0, Qt )

where yt is a p 1vector of observations and t is an unobserved m 1 vector called the state vector. The
first equations is called the observation equation and the second equation is called the state equation. The
initial state vector 1 N (a1 , P1 ). The matrices Zt , Tt , Rt , Ht , Qt are initially assumed to be known and
the error terms t and t are assumed to be serially independent and independent of each other at all time
points. Matrices Zt and Tt1 are permitted to depend on y1 , y2 , . . . yt1 . The first equation has the structure
of a linear regression model where the coefficient vector t varies over time. The second equation represents
a first order vector autoregressive process, the Markovian nature of which accounts for many of the elegant
properties of the state space model.

3.2

Structural time series models

A structural time series model is one in which the trend, seasonal and error terms in the local level, plus
other relevant components are modelled explicitly. This is in sharp contrast to the philosophy underlying
the philosophy underlying Box-Jenkins ARIMA models where trend and seasonal are removed by differencing
prior to the detailed analysis. This section shows the state space model version of trend , trend + seasonality
, regression models, seemingly unrelated time series equations model.

3.3

ARMA models and ARIMA models

ARIMA models eliminate the trend and seasonality by differencing and then try to fit a stationary ARMA
process to the differenced series. These models can be readily put in state space framework as follows :
Consider the ARMA(p,q) process defined as
Yt =

r
X
j=1

j Ytj +

r
X

j tj + t ,

r = max(p, q + 1), j = 0 for p > q, j = 0 for j > q,

j=1

19

Time Series Analysis by State Space Methods

Zt = [1 0 . . .

.
Tt = ..

r1
r

0]
1

...

0
..
.

1
..
.

...
..
.

...

...

1
0

Rt = [1 1 . . . r1 ]

The basic idea of the above representation is to shift the ARMA dependence from Yt to the state vector
Yt = 1,t .
The DLM representation of an ARIMA(p,d,q) model with d > 0 is : For a general d, set Yt d Yt .
dj Yt = Yt +

j
X

di Yt1 ,

j = 1, . . . , d

i=1

Yt1

Yt1

.
..

d1

Yt1

t =
Yt

2 Yt1
+ . . . r Ytr+1
+ 1 t + . . . + r1 tr+2

3 Y + . . . r Y

t1
tr+2 + 2 t + . . . + r1 tr+3

..

r Yt1
+ r1 t

Zt = [1 1 . . . 1 0

1
1

0
1

. . . . . .

0 ...

Tt = 0 . . .

0 ...

..
..
.
.

. . . . . .

. . . 0]

...

...

...

...

...

...

...

...

...

...

...

0
..
.

2
..
.

0
..
.

1
..
.

...

r1

...

...

. . . 0

. . . 0

..

0 1
0 0
0

Rt = [0 . . . 0 1 1 . . . r1 ]0
ARMA estimation using dlm package

20

Time Series Analysis by State Space Methods

set.seed(1)
y

<- arima.sim(n = 100, list(ar = c(0.8897), ma = c(-0.2279)),sd = sqrt(0.1796))

arima(y,c(1,0,1),include.mean=T)
## Series: y
## ARIMA(1,0,1) with non-zero mean
##
## Coefficients:
##

ar1

ma1

intercept

##

0.932

-0.294

-0.015

## s.e.

0.041

0.099

0.360

##
## sigma^2 estimated as 0.15:
## AIC=103.4

AICc=103.8

log likelihood=-47.7

BIC=113.8

fn <- function(params){
model <- dlmModARMA(ar = params[1],
ma = params[2],
sigma2 = exp(params[3]),
dV = 1e-7)
}
fit <- dlmMLE(y,rep(0,3),fn)
#ARMA parameters
print(c(fit$par[1:2]))
## [1]

0.9324 -0.2966

# error variance
exp(fit$par[3])
## [1] 0.1499
ARMA estimation using KFAS package
It took me sometime to figure out the way to write my own function that can be used in MLE.
model

<- SSModel(y ~ SSMarima(ar = c(0.5), ma = c(0.5), d = 0, Q = 1e-7),H=1e-7)

fn

<-function(pars,model,...){

model$T[2,2,1]

<- pars[1]

model$R[3,,1]

<- pars[2]

model$Q[,,1]

<- exp(pars[3])

model
}
model

<- fitSSM(inits=c(0.5,0.5,0.5),
model=model,updatefn = fn,method='BFGS')$model

#ar component

21

Time Series Analysis by State Space Methods

model$T[2,2,1]
## [1] 0.9735
#ma component
model$R[3,,1]
##

arima2

## -0.3161
#Q
c(model$Q)
## [1] 0.1533

3.4

Exponential smoothing

EWMA started appearing in 1950s and became very popular method as it required little storage of past data
and could be easily implemented on a computer. For a one step ahead forecasting of yt+1 given a univariate
time series yt , yt1 , . . ., it has the form
yt+1 = (1 )yt +
y
This can be shown to be an ARIMA model. EWMA forecasts produced by the recursion are minimum mean
square error forecasts in the sense that they minimize E(
yt+1 yt+1 )2 for the observations from a local level
model. The EWMA model was extended by Holt(1957) and Winters(1960) to a series containing trend and
seasonal. This extension is also a type of ARIMA model. It was shown later that the forecasts produced by
these Holt-Winters recursions are minimum mean square error forecasts for the local linear trend model. This
pattern continues with adding seasonal components too. The state space models and the Box-Jenkins ARIMA
models which appear to be very different conceptually but both give minimum mean square error forecasts
from EWMA recursions. The explanation is that when the time series has an underlying structure which is
sufficiently simple, then the appropriate state space and ARIMA models are essentially equivalent. It is when
we move towards more complex structures that the differences emerge.

3.5

State space versus Box-Jenkins approaches

I found this section brilliant as it brings out so many aspects that Box-Jenkins approach misses out. There
are a dozen are so, differences mentioned that show that state space models are superior to ARIMA models.
This is the nicest comparison between state space models and Box-Jenkins methods that I have come across
till date. I think this sort of comparison must be mentioned by all the faculty who teach undergrad classic
time series stuff, so that interested students can dig up the content on State space models and learn to model
at a more generic level.

22

Time Series Analysis by State Space Methods

State Space Models

Box-Jenkins approach

Wide class of problems can be encapsulated in

specific type of models only

a simple linear model

2

Due to Markovian nature of the model, the

non recursive nature

calculations needed for practical application of

the model could be set up in recursive form
in a way that is particularly convenient on a
computer
3

The key advantage of the state space approach

more a black box approach where the model

is that is based on structural analysis of the

adopted depends purely on the data without

problem. The different components that make

prior analysis of the structure of the system

up the series such as trend, seasonal , cy-

over time.

cle, and calendar variations, together with the

effects of explanatory variables and interventions
4

Multivariate observations can be handled by

not the case here

straightforward extensions
5

Explanatory variables can be easily incorpo-

not the case here

rated
6

Associated regression coefficients can be made

not the case here

to vary over time

7

Because of Markovian form, the calculations

not the case here

can be put in a recursive form. This enables

increasingly large models to be handled effectively without disproportionate increases in
the computational burden.
No dependency on differencing

Reliance on differencing before fitting an stationary series

No dependency on differencing

Artificial requirement that differenced series is

10

Easy to handle missing data

not the case here

11

No reliance on auto correlation

Main tool is the sample auto correlation func-

stationary

tion which is notoriously imprecise due to high

sampling variability
12

3.6

Highly transparent

not the case here

Regression with time-varying coefficient

If we want to the coefficient vector to vary over time, state space models are flexible enough to incorporate
this requirement. All one needs to do is that write a random walk equation for each of the coefficients.

3.7

Regression with ARMA error

Without any difficulty, one can easily set up the required matrices for a regression framework where the errors
follow ARMA. These matrices are given in the section.
23

Time Series Analysis by State Space Methods

The chapter ends with a section on state space models in continuous time.

24

Time Series Analysis by State Space Methods

Filtering, smoothing and forecasting

This chapter is the core chapter from the book that shows the derivations behind all the main estimates from
a state space model.
y t = Z t t + t ,

t N (0, Ht )

t+1 = Tt t + Rt t ,

t N (0, Ht )
1 N (a1 , P1 )

The key theme running behind all these derivations is recursion. All the recursive formula derivations flow
from the matrix form of conditional normal moments.

4.1

Filtering

The key term that makes computations easier is the innovation term denoted by t where t = yt E(yt |Yt1 ),
where Yt = {y1 , . . . , yt }. Some useful relations and the recursive formula for Kalman Filtering.
t = yt Zt at
Mt = Cov(t , t ) = Pt Zt0
Ft = Var(t ) = Zt Pt Zt0 + Ht
Kt = Tt Mt Ft = Tt Pt Zt0 Ft1
Lt = Tt Kt Zt
at+1 = Tt at + Kt t
Pt+1 = Tt Pt L0t + Rt Qt Rt0
When dealing with a time-invariant state space model in which system matrices are constant over time, the
Kalman recursion of Pt+1 converges to a constant matrix P which is the solution to the matrix equation
P = T P T 0 T P Z 0F

ZP T 0 + RQR0 ,

where F = ZP Z 0 + H

State space model can be written in terms of state estimation errorxt = t at and this form is called
innovation form of state space model.

t = Zt xt + t ,

t N (0, Ht )

xt+1 = Lt xt + Rt t Kt t
The advantage of writing an equation in terms of forecast errors is that these error are independent of each
other and hence make computations easier.

4.2

State smoothing

This involves estimation of t given the entire series y1 , . . . , yn and is denoted by

t . For deriving smoothing recursion innovation form is used as the estimation can be split in to two independent set of variables

25

Time Series Analysis by State Space Methods

{y1 , . . . , yt1 } and {t , t+1 , . . . n }

Smoothed state vector
rt1 = Zt0 Ft1 t + L0t rt

t = at + Pt rt1
rn = 0
1
0
0
0
rt = Zt+1
Ft+1
t+1 + L0t+1 Zt+2
Ft+2
t+2 + . . . + L0t+1 . . . L0n1 Zn0 Fn1 n

As one can see the smoothed estimate is a linear combination of filtered mean and one step ahead forecast
errors.
Smoothed state variance
Nt1 = Zt0 Ft1 Zt + L0t Nt Lt
Vt = Pt Pt Nt1 Pt
Nn = 0
1
0
0
0
Nt = Zt+1
Ft+1
t+1 + L0t+1 Zt+2
Ft+2
Zt+2 Lt+1 + . . . + L0t+1 . . . L0n1 Zn0 Fn1 Zn Ln1 . . . Lt+1

As one can see the smoothed estimate is a linear combination of filtered mean and one step ahead forecast
errors. The above filtering and smoothing recursions enable us to update our knowledge of the system each
time a new observations comes in. The key advantage of the recursion is that we do not have to invert a
(pt pt) matrix to fit the model each time the tth observation comes in for t = 1, . . . , n,We have to invert the
(pp) matrix Ft and p is generally much smaller than n. For the multivariate case, an additional improvement
can be made by introducing the elements of the new observation one at time rather than the whole observation.

4.3

Disturbance smoothing

This section derives the recursions for t = E(t |y) and t = E(t |y). The recursive equations are as follows:
t = Ht (Ft1 t Kt rt0 )
Var(t |y) = Ht Ht (Ft1 + Kt0 Nt Kt )Ht
t = Qt Rt0 rt
Var(t |y) = Qt Qt Rt0 Nt Rt Qt
rt1 = Zt0 Ft1 t + L0t rt
Nt1 = Zt0 Ft1 Zt + L0t Nt Lt
rn = 0
Nn = 0
These estimates have a variety of uses, particularly for parameter estimation and diagnostic checking. The
above equations are collectively referred to as disturbance smoothing recursion

26

Time Series Analysis by State Space Methods

4.4

Covariance matrices of smoothed observations

This section contains pretty laborious derivations of covariance matrices. pretty painful to go over. Thanks
to KFAS, you get all the output without having to code recursive equations.

4.5

Weight functions

The conditional mean vectors and variance matrices from the filtering , smoothing and disturbance smoothing
estimates are all weighted sums of y1 , . . . , yn . It is of interest to study these weights to gain a better understanding of the properties of the estimators. Models which produce weight patterns for trend component which
differ from what is regarded as appropriate should be investigated. In effect, the weights can be regarded as
what are known as kernel functions in the field of nonparametric regression.

4.6

Simulation Smoothing

The drawing of samples of state or disturbance vectors conditional on the observations held fixed is called
simulation smoothing. Such samples are useful for investigating the performance of techniques of analysis
proposed for the linear Gaussian model and for Bayesian analysis based on this model. The primary purpose
of simulation smoothing in the book is to serve as the basis for importance sampling techniques. The recursive
algorithms in this section have been replaced by a simpler version, written by the same authors after this book
was published. The paper is titled, A simple and efficient simulation smoother for state space time series
analysis.
I will take a brief detour here and summarize the new and efficient simulation smoother.
Simple and Efficient Smoother
For any model that you build, it is imperative that you are able to sample random realizations of the various
parameters of the model. In a structural model, be it a linear Gaussian or nonlinear state space model, an
additional requirement is that you have to sample the state and disturbance vector given the observations.
To state it precisely, the problem we are dealing is to draw samples from the conditional distributions of
 = (01 , 02 , . . . , 0n )0 , = (10 , 20 , . . . , n0 )0 and = (10 , 20 , . . . , n0 )0 given y = (y10 , y20 , . . . , yn0 )0 .
The past literature on this problem is :
Fruhwirth-Schnatter(1994), Data augmentation and dynamic linear models. : The method comprised

drawing samples of |y recursively by first sampling n |y, then sampling n1 |n , y, then n2 |n1 , n , y.
de Jong and Shephard(1995),The simulation smoother for time series models. : This paper was a signif-

icant advance as the authors first considered sampling the disturbances and subsequently sampling the
states. This is more efficient than sampling the states directly when the dimension of is smaller than
the dimension of
In this paper, the authors present a new simulation smoother which is simple and is computationally efficient
relative to that of de Jong and Shephard(1995).

27

Time Series Analysis by State Space Methods

The dataset used in the paper is Nile dataset. The State space model used is Local level model
yt = t + t ,

t N (0, 2 )

t+1 = t + t ,

t N (0, 2 )

The New Simulation Smoother

Algorithm 1
1. Draw a random vector w+ from density p(w) and use it to generate y + by means of the recursive
observation and system equation with the error terms related by w+ , where the recursion is initialized
by the draw 1+ N (a1 , P1 )
p(w) = N (0, ),

= diag(H1 , . . . , Hn , Q1 , . . . , Qn )

0
0
2. Compute w
= (
0 , 0 )0 = E(w|y) and w
+ = (+ , + )0 = E(w+ |y + ) by means of standard Kalman

filtering and disturbance smoothing using the following equations

t = Ht Ft1 t Ht Kt0 rt
t = Qt Rt0 rt
rt1 = Zt Ft1 t + L0t rt
3. Take w
=w
w
+ + w+

4.7

Simulation of State vector

Algorithm 2
1. Draw a random vector w+ from density p(w) and use it to generate + , y + by means of the recursive
observation and system equation with the error terms related by w+ , where the recursion is initialized
by the draw 1+ N (a1 , P1 )
p(w) = N (0, ),

= diag(H1 , . . . , Hn , Q1 , . . . , Qn )

2. Compute
== E(|y) and
+ = E(+ |y + ) by means of standard Kalman filtering and smoothing
equations
3. Take
=

+ + +

R excursion
Now lets use the above algorithms to generate samples from the state vector and disturbance vector. Let me
use the same Nile dataset.

28

Time Series Analysis by State Space Methods

Algorithm 1 - Implementation
# MLE for estimating observational and state evolution variance.
fn

<- function(params){
dlmModPoly(order= 1, dV= exp(params[1]) , dW = exp(params[2]))
}

<- c(Nile)

fit

<- dlmMLE(y, rep(0,2),fn)

mod

<- fn(fit$par)

(obs.error.var

<- V(mod))

##

[,1]

## [1,] 15100
(state.error.var <- W(mod))
##

[,1]

## [1,] 1468
filtered

<- dlmFilter(y,mod)

smoothed

<- dlmSmooth(filtered)

smoothed.state

<- dropFirst(smoothed$s)

w.hat

<- c(Nile- smoothed.state,c(diff(smoothed.state),0))

Step 1 :
set.seed(1)
n

<- length(Nile)

# Step 1
w.plus

<- c( rnorm(n,0,sqrt(obs.error.var)),rnorm(n,0,sqrt(state.error.var)))

alpha0

<- rnorm(1, mod$m0, sqrt(mod$C0))

y.temp

<- numeric(n)

alpha.temp

<- numeric(n)

#Step 2
for(i in 1:n){
if(i==1){
alpha.temp[i] <- alpha0 + w.plus[100+i]
}else{
alpha.temp[i] <- alpha.temp[i-1] + w.plus[100+i]
}
y.temp[i]

<- alpha.temp[i] + w.plus[i]

}
temp.smoothed

<- dlmSmooth(y,mod)

29

Time Series Analysis by State Space Methods

alpha.smoothed

<- dropFirst(temp.smoothed$s)

Step 2 :
w.hat.plus

<- c(Nile- alpha.smoothed,c(diff(alpha.smoothed),0))

Step 3 :
w.tilde

<- w.hat - w.hat.plus + w.plus

This generates one sample of the observation and state disturbance vector.
Lets simulate a few samples and overlay them on the actual observation and state disturbance vectors
set.seed(1)
n

<- length(Nile)

disturb.samp

<- matrix(data= 0,nrow = n*2, ncol = 25)

for(b in 1:25){
# Step 1
w.plus

<- c( rnorm(n,0,sqrt(obs.error.var)),rnorm(n,0,sqrt(state.error.var)))

alpha0

<- rnorm(1, mod$m0, sqrt(mod$C0))

y.temp

<- numeric(n)

alpha.temp

<- numeric(n)

#Step 2
for(i in 1:n){
if(i==1){
alpha.temp[i] <- alpha0 + w.plus[100+i]
}else{
alpha.temp[i] <- alpha.temp[i-1] + w.plus[100+i]
}
y.temp[i]

<- alpha.temp[i] + w.plus[i]

}
temp.smoothed

<- dlmSmooth(y.temp,mod)

alpha.smoothed

<- dropFirst(temp.smoothed$s)

w.hat.plus

<- c(y.temp- alpha.smoothed,c(diff(alpha.smoothed),0))

#Step 3
w.tilde

<- w.hat - w.hat.plus + w.plus

disturb.samp[,b]

<- w.tilde

temp

<- cbind(disturb.samp[101:200,],w.hat[101:200])

plot.ts(temp, plot.type="single",col =c(rep("grey",25),"blue"),

30

Time Series Analysis by State Space Methods

ylim = c(-150,200),lwd=c(rep(0.5,25),3), ylab = "",xlab="")

leg <-c("realized state error","simulated state error")

100 150 200

legend("topright",leg,col=c("blue","grey"),lwd=c(3,1),cex=0.7,bty="n")

150

50

50

realized state error

simulated state error

20

40

60

80

100

Figure 4.1: Comparison of Simulated and Realized State disturbance

31

Time Series Analysis by State Space Methods

temp

<- cbind(disturb.samp[1:100,],w.hat[1:100])

plot.ts(temp, plot.type="single",col =c(rep("grey",25),"blue"),

,lwd=c(rep(0.5,25),3), ylab = "",xlab="")
leg <-c("realized observation error","simulated observation error")

400

legend("topright",leg,col=c("blue","grey"),lwd=c(3,1),cex=0.7,bty="n")

400

200

200

realized observation error

simulated observation error

20

40

60

80

100

Figure 4.2: Comparison of Simulated and Realized Observation disturbance

32

Time Series Analysis by State Space Methods

Algorithm 2 - Implementation
set.seed(1)
n

<- length(Nile)

alpha.samp

<- matrix(data= 0,nrow = n, ncol = 25)

for(b in 1:25){
# Step 1
w.plus

<- c( rnorm(n,0,sqrt(obs.error.var)),rnorm(n,0,sqrt(state.error.var)))

alpha0

<- rnorm(1, mod$m0, sqrt(mod$C0))

y.temp

<- numeric(n)

alpha.temp

<- numeric(n)

#Step 2
for(i in 1:n){
if(i==1){
alpha.temp[i] <- alpha0 + w.plus[100+i]
}else{
alpha.temp[i] <- alpha.temp[i-1] + w.plus[100+i]
}
y.temp[i]

<- alpha.temp[i] + w.plus[i]

}
temp.smoothed

<- dlmSmooth(y.temp,mod)

alpha.smoothed

<- dropFirst(temp.smoothed$s)

alpha.hat.plus

<- alpha.smoothed

#Step 3
alpha.tilde

<- smoothed.state - alpha.hat.plus + alpha.temp

alpha.samp[,b]

<- alpha.tilde

33

Time Series Analysis by State Space Methods

temp

<- cbind(alpha.samp,smoothed.state)

plot.ts(temp, plot.type="single",col =c(rep("grey",25),"blue"),

,lwd=c(rep(0.5,25),3), ylab = "",xlab="")
leg <-c("realized state vector","simulated state vector")
legend("topright",leg,col=c("blue","grey"),lwd=c(3,1),cex=0.7,bty="n")

700

800

900

1000

1200

realized state vector

simulated state vector

20

40

60

80

100

Figure 4.3: Comparison of Simulated and Realized State vector

34

Time Series Analysis by State Space Methods

Simulation via KFAS package

KFAS package closely follows the implementations suggested by the authors. One can simulate the state
observations using simulateSSM() function. One can also turn on the antithetic sampling suggested in this
paper via the functions argument.
modelNile

<-SSModel(Nile~SSMtrend(1,Q=list(matrix(NA))),H=matrix(NA))

modelNile

<-fitSSM(inits=c(log(var(Nile)),log(var(Nile))),
model=modelNile,
method='BFGS',
control=list(REPORT=1,trace=0))$model
<- KFS(modelNile,filtering='state',smoothing='state')

out

state.sample <- simulateSSM(modelNile, type = c("states"))

temp

<- cbind(state.sample,smoothed.state)

plot.ts(temp, plot.type="single",col =c(rep("grey",1),"blue"),

,lwd=c(rep(1,1),2), ylab = "",xlab="")
leg <-c("realized observation error","simulated observation error")
legend("topright",leg,col=c("blue","grey"),lwd=c(3,1),cex=0.7,bty="n")

700

800

900

1000 1100 1200

realized observation error

simulated observation error

20

40

60

80

100

Figure 4.4: Comparison of Simulated and Realized State vector

35

Time Series Analysis by State Space Methods

Conclusion
The paper presents a simulation smoother for drawing samples from the conditional distribution of the disturbances given the observations. Subsequently the paper highlights the advantages of this simulation technique
over the previous methods.
derivation is simple
the method requires only the generation of simulated observations from the model together with the

Kalman Filter and standard smoothing algorithms

no inversion of matrices are needed beyond those in the standard KF
diffuse initialization of state vector is handled easily
this approach solves problems arising from the singularity of the conditional variance matrix W

4.8

Missing observations

One of the advantages of dealing with state space models is that the missing observations can be easily
integrated in to the framework. When the set of observations yt for t = , . . . , 1 is missing, the vectors t
and Kalman gain matrix Kt are set to 0. Thus the updates become

at+1 = Tt at
Pt+1 = Tt Pt Tt0 + Rt Qt Rt0 ,

t = , . . . , 1

rt1 = Tt0 rt
Nt1 = Tt0 Nt Tt
A similar treatment is done in the multivariate case When some but not all elements of the observation vector
yt are missing.Of course, the dimensionality of the observation vector varies over time, but this does not affect
the validity of the formula.

4.9

Forecasting

Forecasting in State space framework is very easy. All you have to is to extend the observation series to
whatever number of steps ahead you want to forecast and replace it with missing values. Why does this work
? In the case of missing observations, the filtered and smoothed recursions go through after replacing the
s associated with the missing observations as 0 and replacing Kalman gain matrices as 0. In the case of
forecasting, if you take the loss function as minimum mean squared error, then the conditional expectation
given data turns out to have the same form as recursions based on missing values. The authors in fact derive
the forecasts in the book and show them to be equal to those formed by assuming missing observations.
The last section of the chapter casts the state space model in a matrix format, which enables the reader to
get a quicker derivation of smoothing estimates, state disturbance vectors etc.

36

Time Series Analysis by State Space Methods

Initialisation of filter and smoother

In all the previous chapters the starting value , the starting value came from a known distribution, 1
N (a1 , P1 ). This chapter develops methods of starting up the series when at least some of the elements of a1 or
P1 are unknown. The methods fall under the name, initialisation. A general model for the initial state vector
1 is
1 = a + A + R0 0 ,

N (0, Q0 )

where the m 1 vector a is known, is a q 1 vector of unknown quantities, the m q matrix A, and the
m (m q) matrix R0 are selection matrices Im , they are defined so that when taken together, their columns
constitute a set of g columns of Im with g m and A0 R0 = 0. The matrix Q0 is assumed to be positive
definite and known. There are two cases to be considered here, first case is where is fixed and unknown, the
second case is where is random. In the second case, the authors assume a specific form
N (0, Iq )
The initial conditions are such such that a1 = E(1 ) and P1 = Var(1 ), where
P1 = P + P ,

where P = AA0 and P = R0 Q0 R00

A vector with distribution N (0, Iq ) as k is said to be diffuse. Initialisation of the Kalman filter
when some elements of 1 are diffuse is called diffuse initialisation of the filter. One approximate technique
is to replace by a large number and run the kalman filter. The authors use a different approach where they
expand Ft1 . The mean square error matrix Pt has the decomposition
Pt = P,t + P,t + O(1 )
where P,t , P,t do not depend on . The chapter also shows that P,t = 0, t > d where d is a positive
integer, which is small relative to n. The consequence is that the usual Kalman filter applies without change
for t = d + 1, . . . , n. The recursion for exact initial state smoothing, exact initial disturbance smoothing and
exact initial simulation smoothing are derived. I felt very painful to go over these derivations where there is
nothing new here exact mind numbing math. A major relief for me was going through an example of structural
time series, ARMA model ,non stationary ARIMA models and Regression model with ARMA errors, for whom
exact initial Kalman filter and smoother are given. I skipped the augmented Kalman filter part of the chapter
as it was kind of overwhelming for me after the heavy dose of initialisation process in the previous sections.
The documentation in KFAS mentions this process in precise words
KFS uses exact diffuse initialization where the unknown initial states are set to have a zero mean
and infinite variance, so
P1 = P + P ,

where P = AA0 and P = R0 Q0 R00

with going to infinity and P,1 being diagonal matrix with ones on diagonal elements corresponding to unknown initial states. Diffuse phase is continued until rank of P,1 becomes zero.
Rank of P decreases by 1, if F > tol > 0. Usually the number of diffuse time points equals the
number unknown elements of initial state vector, but missing observations or time varying Z can
affect this.
37

Time Series Analysis by State Space Methods

Further Computational Aspects

This chapter is very interesting as it talks about the implementation issues with the state space recursive
framework. However it begins with a method to incorporate regression effects within the Kalman filter which
I felt had nothing to do with the title of the chapter. One of the methods that I followed was the inclusion of
independent variables as state vectors and write trivial state vector equations for those coefficients. The only
reason that this has been mentioned I guess is to make a point that the enlargement of the state space model
will not cause extra computing because of the sparse nature of the system matrices.

6.1

Square root filter and smoother

Because of rounding errors and matrices being to singularity, the possibility arises that the calculated value
of Pt is negative definite, or close to this, giving rise to unacceptable rounding errors. Since Pt appears in
all the recursive equations, it can happen that the calculated value of Pt becomes negative definite when,
for example, erratic changes occur in the system matrices over time.The problem can be avoided by using a
transformed version of the Kalman filter called the square root filter. However the amount of computation
required is substantially larger than that required for the standard Kalman filter. First the usual recursive
equations used
Ft = Zt Pt Zt0 + Ht
Kt = Tt Pt Zt0 + Ft1
Pt+1 = Tt Pt Tt0 + Rt Qt Rt0 Kt0 Ft Kt0
Define the partitioned matrix Ut by
"
Zt Pt
Ut =
Tt Pt

Ht

0
t
Rt Q

where
Pt = Pt Pt0 ,
"
Ut Ut0

tH
t0 ,
Ht = H

tQ
0t
Qt = Q

Ft

Zt Pt Tt0

Tt Pt Zt0

Tr Pt Tt0 + Rt Qt Rt0

The matrix Ut is transformed in to a lower triangular matrix using the orthogonal matrix G , Ut G = Ut . The
lower triangular matrix can be written as
Ut

"

U1,t

U2,t

U3,t

By equating the Ut Ut0 , we get that Pt+1 = U3,t

U3,t
. Thus one finds that U3,t
= Pt+1 . The recursive equation

becomes
1

at+1 = Ta at + U2,t
U1,t
t

The above equation is numerically stable.Matrices U1,t

, U2,t
, U3,t
are used for deriving numerically smoothing

recursions. The square root formulation is not that useful for initialisation since it usually only requires a
limited number of d updates, the numerical problems are not substantial.

38

Time Series Analysis by State Space Methods

6.2

Univariate treatment of multivariate series

This section converts a multivariate KF in to a set of univariate KF equations, the main motivation being
computational efficiency. This approach avoids the inversion of Ft matrix and two matrix multiplications.
The results from the section show that there is considerable computational savings for higher dimensional
observation data.

6.3

The algorithms of SsfPack

This is a commerical software and the authors provide all the functions relevant to state space model analysis.
However we are living in a world of open source. My goto packages for state space models are DLM and KFAS.
In fact the latter package closely mirrors this book and all the analysis done in the book can be replicated
using KFAS. There are certain functions though that are not in the open source packages such as
SimSmoWgt
SsfLikConc
SsfLikSco

However one can write these functions from the KFAS output with ease.

39

Time Series Analysis by State Space Methods

Maximum Likelihood estimation

7.1

Likelihood function, Hessian and Score function

Nobody is going to hand down the parameters for a state space model for they need to be estimated for the
data. There are generally two ways one can estimate the meta parameters of the model, one via frequentist
a.k.a MLE approach, the second is via Bayesian approach. This chapter deals with the first approach of
computing the log likelihood function and then using Newton Raphson method to compute the MLE. The
method basically solves the equation
1 () =

log L(y|)
=0

using the Taylor series

1 () 1 () + 2 ()( )
where
for some trial value ,

()
= 1 ()|= ,

2 () = 2 ()|=

The revised equation for

= 2 ()1 1 ()
The gradient 1 () determines the direction of the step taken to the optimum and the Hessian modifies the
size of the step. It is possible to overstep the maximum in the direction determined by the vector

() = 2 ()1 1 ()
and therefore it is common to include a line search along the gradient vector within the optimisation process.
The first object needed for MLE is the likelihood function. There are three forms of likelihood functions
depending on the initial conditions.
1. Prediction error decomposition - When the initial conditions are known
2. Diffuse Loglikelihood - When some of the elements of 1 are diffuse
3. Likelihood when elements of the initial state vector are fixed but unknown
All the three forms are derived in this chapter.
The second object that one needs to get going on the Newton Raphson is the Hessian and the authors suggest
the use of BFGS method. This method approximates the Hessian matrix in such a way that the resulting
matrix remains negative definite.
The third object needed is the gradient vector or score vector. The chapter gives the score vector in all the
possible cases based on the initial distribution. The math is laid out in a crystal clear way. Whats the score
vector ?
1 () =

log L(y|)

For the first case, the initial vector has the distribution 1 N (a1 , P1 ) where a1 , P1 are known.
Let p(, y|) be the joint density of and y, p(|y, ) be the conditional density of given y, and let p(y|)

40

Time Series Analysis by State Space Methods

be the marginal density of y for given .

log p(y|) = log p(, y|) log p(|y, )
be the expectation with respect to density p(|y, ) and taking expectations on both sides
Let E

log p(y|) = E[log

p(, y|)]

as E[log
p(|y, )] = 0 and log p(y|) is independent of .
Thus the score vector becomes
n
log L(y|)
1 X
0
log |Ht |+log |Qt1 |+tr[(
t 0t +Var(t |y))Ht1 ]+tr[(
|= =
t1 t1
+Var(t1 |y))Q1
t1 ]

2 t=1

The score vector for the diffuse case is also derived

where t1 , t , Var(t |y), Var(t |y), are evaluated for = .
in this section.

7.2

EM Algorithm

EM algorithm crops in variety of places. In contexts where you see that the log likelihood function that is
difficult to maximize, it makes sense to introduce additional latent variables and then make it tractable. In
the context of state space models, EM algorithm for many state space models has a particularly neat form.

The EM consists of E step which involves the evaluation of the conditional expectation of E(log
p(, y|) and
the M step that involves maximizing the expectation with respect to the element of . From the previous
section the conditional expectation can be written down as
The log likelihood function
n

log p(, y|) =

1X
0
log |Ht | + log |Qt1 | + 0t Ht1 t + t1
Q1
t1 t1 +
2 t=1

E step - Evaluating the loglikelihood function with respect to density p(|y, )

1

E(log
p(, y|)) =
2

n
X

0
log |Ht | + log |Qt1 | + tr[(
t 0t + Var(t |y))Ht1 ] + tr[(
t1 t1
+ Var(t1 |y))Q1
t1 ]

t=1

M step - Maximize Expectation by taking the derivative

n

1 X
0
E(log p(, y)|) =
log |Ht |+log |Qt1 |+tr[(
t 0t +Var(t |y))Ht1 ]+tr[(
t1 t1
+Var(t1 |y))Q1
t1 ]

2 t=1

The values in the next iteration are

0

old
Htnew = n1 [(old
+ Var(old
t t
t |y)]
0

old old
old
Qnew
= n1 [(t1
t1 + Var(t1
|y)]
t

More specific information on the EM algorithm can be found in the paper, An approach to timeseries smooth-

41

Time Series Analysis by State Space Methods

ing and forecasting by Shumway and Stoffer. The only irritant in state space model literature is that every
one follows his/her own notation. As of now, I have not found a convenient way to quickly grasp various
notations. In anycase, thats the way it is, so better get used to it.
Here is the R code that implements EM. I will run the algo on a known dataset so that the estimates can be
compared. The dataset is Nile and the model used is local level model. EM algo will be used to estimate the
observation variance and state evolution variance.
new

<- c(var(Nile),var(Nile))

old

<- numeric(2)

<- length(Nile)

while(sum(abs(new - old)) > 1e-7){

old

<- new

modelNile

<- SSModel(Nile~SSMtrend(1,Q=old[1]),H=old[2])

out

<- KFS(modelNile,filtering='state',smoothing=c('state','disturbance'))

<- diag(n)

diag(W)

<- c(out$V_eta)

new[1]

<- sum(diag((out$etahat)%*% t(out$etahat) + W))/n

<- diag(n)

diag(V)

<- out$V_eps

new[2]

<- sum(diag((out$epshat)%*% t(out$epshat) + V))/n

}
#Parameters - Observation error variance and State error variance
params

<- data.frame( H= new[2], Q = new[1])

params
##

## 1 15099 1469
To compare the output of EM algo with the output from using fitSSM
modelNile

<-SSModel(Nile~SSMtrend(1,Q=list(matrix(NA))),H=matrix(NA))

modelNile

<-fitSSM(inits=c(log(var(Nile)),log(var(Nile))),
model=modelNile,
method='BFGS',
control=list(REPORT=1,trace=0))$model

c(modelNile$H) #15098.65
## [1] 15099
c(modelNile$Q) #1469.163
## [1] 1469

42

Time Series Analysis by State Space Methods

As one can see the EM algorithm concurs with the MLE values.
Given the MLE estimate, it can be shown that under reasonable assumptions about the stability of the model
over time, the distribution of for large n is approximately
N (, ),


=

2 log L
0

1

The section introduces the antithetic sampling to reduce the bias amongst various estimates, conditional on
, the metaparameters. The authors touch upon briefly the model selection criterion (based on AIC and BIC)
and diagnostics. The diagnostic tests involve standardized residuals and auxiliary residuals, the former are
used to check for normality, serial correlation etc, while the latter are used to identify structural breaks.

43

Time Series Analysis by State Space Methods

8
8.1

Bayesian Analysis
Posterior analysis of state vector

When the vector is not fixed and known, it is treated as a random vector with prior density p(). The
section introduces importance sampling as a technique to estimate
x = E[x()|y]
The basic idea of importance sampling is to generate random samples from a distribution that is as close as
possible to p(|y) and then take a weighted average of the samples to obtain posterior mean x. The book
gives a chapter length treatment to importance sampling in Part II of the book.

8.2

MCMC

This basic idea here is to start with some initial parameter vector 0 and generate a state sample (i) from
p(|y, 0 ). The second step involves sampling a new from p(|y, (i) ). This is repeated until convergence.
The first step of sampling given metaparameters can be accomplished via Simulation smoothing algorithm
from KFAS or dlmBSample. However the speed from the former package is much faster than the latter. The
algorithm used in the former is from the paper titled, A simple and efficient simulation smoother for state
space time series analysis. After generating samples from the posterior distribution of metaparameters, one
has to apply the usual methods like burning-in, thinning etc to estimate the parameters. The second step
in the above algorithm where one has to sample from p(|y, ) depends partly on the model for and is
usually only possible up to proportionality. To work around this problem, accept-reject methods have been
developed in the MCMC literature. A convenient way is to assume appropriate conjugate priors for hyper
parameters so that the posterior distributions have a closed form.

44

Time Series Analysis by State Space Methods

Illustrations of the use of the linear Gaussian model

This chapter gives a set of examples that illustrate various aspects of State space model analysis. I will try to
replicate the analysis using R.

9.1

Structural time series models

In this section, a level and seasonal model is used to fit the monthly numbers(logged) of drivers who were
killed or seriously injured in road accidents in cars in Great Britain in a certain time period.
data.1

<- log(read.table("data/UKdriversKSI.txt",skip=1))

colnames(data.1)

<- "logKSI"

data.1

<- ts(data.1, start = c(1969),frequency=12)

par(mfrow=c(1,1))
plot(data.1,

col = "darkgrey", xlab="",ylab = "log KSI",pch=3,cex=0.5,

cex.lab=0.8,cex.axis=0.7)

7.4
7.0

7.2

log KSI

7.6

7.8

abline(v=1969:2003, lty= "dotted",col="sienna")

1970

1975

1980

1985

Figure 9.1: Monthly numbers(logged) of drivers who were killed or seriously injured in road accidents in cars in Great
Britain in a certain time period.

Using dlm package

fn
mod
V(mod)

<- function(params){
<- dlmModPoly(order = 1 ) + dlmModSeas(frequency =12)
<- exp(params[1])

45

Time Series Analysis by State Space Methods

diag(W(mod))[1:12] <- exp(c(params[2],rep(params[3],11)))

return(mod)
}
fit

<- dlmMLE(data.1, c(var(data.1),rep(0,12)),fn)

mod
(obs.error.var

<- fn(fit$par)
<- V(mod))

##

[,1]

## [1,] 0.003371
(seas.var

<- (diag(W(mod))[2]))

## [1] 2.378e-06
(level.var

<- (diag(W(mod))[1]))

## [1] 0.0009313
res

<- data.frame(V = obs.error.var,

W = level.var,
Ws = seas.var)

res
##
V
W
Ws
## 1 0.003371 0.0009313 2.378e-06
Using KFAS package
model <-SSModel(log(drivers)~SSMtrend(1,Q=list(NA))+
SSMseasonal(period=12,sea.type='trigonometric',Q=NA),data=Seatbelts,H=NA)
ownupdatefn
model$H[]

<-

function(pars,model,...){

<- exp(pars[1])

diag(model$Q[,,1])<- exp(c(pars[2],rep(pars[3],11)))
model
}
fit

<-fitSSM(inits=
log(c(var(log(Seatbelts[,'drivers'])),0.001,0.0001)),
model=model,
updatefn=ownupdatefn,method='BFGS')

res

<- data.frame(V = fit$model$H[,,1],

W = diag(fit$model$Q[,,1])[1],
Ws = diag(fit$model$Q[,,1])[2])

res

46

Time Series Analysis by State Space Methods

##

Ws

## 1 0.003416 0.000936 5.004e-07

out

<- KFS(fit$model,filtering=c('state'), smoothing=c('state','disturbance','mean'))

level.sm

<- out$alphahat[,1]

seas.sm

<- rowSums(out$alphahat[,2:12])

epshat

<- out$epshat

47

Time Series Analysis by State Space Methods

par(mfrow=c(3,1))
temp

<- ts( cbind(data.1,level.sm) ,start = 1969,frequency =12)

plot.ts(temp,plot.type="single", ylim=c(7,8),
xlab="",ylab = "log KSI",
col = c("darkgrey","blue"),lwd=c(1,2),cex.lab=1.3,main = "i")
temp

<- ts( seas.sm,start = 1969,frequency = 12)

plot(temp,xlab="",ylab = "log KSI",

col = "blue",lwd=1,main="ii")

abline(h=0,col="grey")
temp

<- ts( epshat ,start = 1969,frequency

plot(temp,xlab="",ylab = "log KSI",

=12)

col = "blue",lwd=1,main="iii")

abline(h=0,col="grey")

7.6
7.4
7.0

7.2

log KSI

7.8

8.0

1970

1975

1980

1985

1980

1985

1980

1985

0.0
0.2

0.1

log KSI

0.1

0.2

0.3

ii

1970

1975

0.00
0.10

log KSI

0.05

0.10

iii

1970

1975

Figure 9.2: Estimated components : (i) level, (ii) seasonal, (iii) irregular

48

Time Series Analysis by State Space Methods

par(mfrow=c(1,1))
temp

<- ts( cbind(data.1,out$a[,1],out$alphahat[,1])[c(-1:-12),] ,

start = c(1970,2),frequency =12)

plot.ts(temp,plot.type="single", ylim=c(7,8),
xlab="",ylab = "log KSI",
col = c("darkgrey","blue","green"),lwd=c(1,2,2))
leg <-c("actual data","filtered level","smoothed level")

8.0

legend("topright",leg,col=c("grey","blue","red"),lwd=c(1,2,2),cex=0.7,bty="n")

7.0

7.2

7.4

log KSI

7.6

7.8

actual data
filtered level
smoothed level

1970

1975

1980

1985

Figure 9.3: Filtered and Smoothed level estimate

49

Time Series Analysis by State Space Methods

par(mfrow=c(3,1))
one.step.ahead <- residuals(out, "recursive")
temp

<- ts( one.step.ahead,start = 1969,frequency = 12)

plot(temp,xlab="",ylab = "log KSI",

col = "blue",lwd=1,main = "i")

abline(h=0,col="grey")
temp

<- residuals(out, type="response")

temp

<- ts( temp,start = 1969,frequency = 12)

plot(temp,xlab="",ylab = "log KSI",

col = "blue",lwd=1, main = "ii")

abline(h=0,col="grey")
temp

<- residuals(out, type="state")[,1]

temp

<- ts( temp,start = 1969,frequency = 12)

plot(temp,xlab="",ylab = "log KSI",

col = "blue",lwd=1, main = "iii")

abline(h=0,col="grey")

log KSI

0.3

0.2

0.1

0.0

0.1

1970

1975

1980

1985

1980

1985

1980

1985

0.00
0.10

log KSI

0.05

0.10

ii

1970

1975

0.00
0.02
0.04
0.06

log KSI

0.02

iii

1970

1975

Figure 9.4: (i)One step ahead residuals,(ii) auxiliary irregular residuals, (iii) auxiliar level residuals

50

Time Series Analysis by State Space Methods

The residuals show that there is a large negative value for Feb 1983. This suggests a need to incorporate an
intervention variable to measure the level shift. The analysis is redone using the intervention variable which
equals one from Feb 1983 and is zero prior to that.
n

<- dim(Seatbelts)[1]

model

<- SSModel(log(drivers)~SSMtrend(1,Q=list(NA))+
SSMseasonal(period=12,sea.type='trigonometric',Q=NA)+
log(PetrolPrice)+law,data=Seatbelts,H=NA)

ownupdatefn

<-function(pars,model,...){

model$H[]

<-exp(pars[1])

diag(model$Q[,,1]) <-exp(c(pars[2],rep(pars[3],11)))
model
}
<- fitSSM(inits=log(c(var(log(Seatbelts[,'drivers'])),0.001,0.0001)),

fit

model=model,updatefn=ownupdatefn,method='BFGS')
<- KFS(fit$model,smoothing=c('state','mean','disturbance'))

out

(out$alphahat[n,1])
## log(PetrolPrice)
##

-0.2914

(out$alphahat[n,2])
##

law

## -0.2377
par(mfrow=c(3,1))
level.sm

<- out$alphahat[,1]*log(as.data.frame(Seatbelts)$PetrolPrice) +

out$alphahat[,2]*(as.data.frame(Seatbelts)$law) + out$alphahat[,3]
seas.sm

<- rowSums(out$alphahat[,2:14])

epshat

<- out$epshat

temp

<- ts( cbind(data.1,level.sm) ,start = 1969,frequency =12)

plot.ts(temp,plot.type="single",
xlab="",ylab = "log KSI",
col = c("darkgrey","blue"),lwd=c(1,2),cex.lab=1.3,main = "i")
temp

<- ts( seas.sm,start = 1969,frequency = 12)

plot(temp,xlab="",ylab = "log KSI",

temp

col = "blue",lwd=1,main="ii")

<- ts( epshat ,start = 1969,frequency

plot(temp,xlab="",ylab = "log KSI",

=12)

col = "blue",lwd=1,main="iii")

abline(h=0,col="grey")

51

Time Series Analysis by State Space Methods

7.4
7.0

7.2

log KSI

7.6

7.8

1970

1975

1980

1985

1980

1985

1980

1985

6.6
6.5
6.2

6.3

6.4

log KSI

6.7

6.8

6.9

ii

1970

1975

0.00
0.10

log KSI

0.05

0.10

0.15

iii

1970

1975

Figure 9.5: Estimated components for model with intervention and regression effects : (i) level, (ii) seasonal, (iii)
irregular

9.2

Box-Jenkins analysis

data(internet)
internet.miss <- internet
internet.miss[c(6,16,26,36,46,56,66,72,73,74,75,76,86,90)] <- NA
d1

<- diff(internet)

d2

<- diff(internet.miss)

par(mfrow=c(2,1))
plot(diff(internet),xlab="",ylab = "",

col = "blue",lwd=1,main="i")

abline(h=0)
plot(diff(internet.miss),xlab="",ylab = "",

col = "blue",lwd=1,main="ii")

abline(h=0)

52

Time Series Analysis by State Space Methods

15

10

10

15

20

40

60

80

100

60

80

100

15

10

10

15

ii

20

40

Figure 9.6: (i) First difference of number of users logged on to Internet server each minute (ii) The same data with 14
observations deleted

ps

<- 0:5

qs

<- 0:5

results <- matrix(0, nrow = length(ps), ncol= length(qs))

for( p in ps){
for(q in qs ){
if(p==0 & q==0) next
if(q > 0 ){
ar

<- rep(-0.05,p)

ma

<- rep(0.001,q)

model

<- SSModel(d1 ~ SSMarima(ar = ar, ma = ma,

d = 0, Q = 1e-7),H=1e-7)

fn

<- function(pars,model,...){

model$T[2:(2+p-1),2,1]

<- (pars[1:p])

model$R[3:(3+q-1),,1]

<- pars[(p+1):(p+q)]

model$Q[,,1]

<- exp(pars[p+q+1])

model

53

Time Series Analysis by State Space Methods

}
model

<- fitSSM(inits=c(ar,ma,1e-7),model=model,updatefn = fn,

method='BFGS', control=list(REPORT=1,trace=0))$model

}else{
ar

<- rep(-0.05,p)

model

<- SSModel(d1 ~ SSMarima(ar = ar,

d = 0, Q = 1e-7),H=1e-7)

fn

<- function(pars,model,...){

model$T[2:(2+p-1),2,1]

<- (pars[1:p])

model$Q[,,1]

<- exp(pars[p+q+1])

model
}
model

<- fitSSM(inits=c(ar,1e-7),model=model,updatefn = fn,

method='BFGS', control=list(REPORT=1,trace=0))$model

}
aic.model <- -2*logLik(model) +2*(p+q)
results[p+1,q+1] <- aic.model
}
}
results <- results/length(d1)
rownames(results) <- paste("AR",0:5,sep="")
colnames(results) <- paste("MA",0:5,sep="")
results[results>100] <- 0

AR0
AR1
AR2
AR3
AR4
AR5

MA0
0.0000
5.3331
5.2947
5.1541
5.1741
5.1942

MA1
5.2210
5.2155
43.6969
5.1741
5.1943
5.2144

MA2
5.2115
5.2072
5.3608
0.0000
0.0000
5.2231

MA3
5.1785
5.1813
42.0075
5.9370
0.0000
5.2043

MA4
5.1980
9.8745
4.4523
6.8736
5.1484
5.0350

MA5
5.2182
3.5363
4.7491
10.2940
9.4143
5.0954

Table 1: AIC for different ARMA Models

54

Time Series Analysis by State Space Methods

ps

<- 0:5

qs

<- 0:5

results2 <- matrix(0, nrow = length(ps), ncol= length(qs))

for( p in ps){
for(q in qs ){
if(p==0 & q==0) next
if(q > 0 ){
ar

<- rep(-0.05,p)

ma

<- rep(0.001,q)

model

<- SSModel(d1 ~ SSMarima(ar = ar, ma = ma,

d = 0, Q = 1e-7),H=1e-7)

fn

<- function(pars,model,...){

model$T[2:(2+p-1),2,1]

<- (pars[1:p])

model$R[3:(3+q-1),,1]

<- pars[(p+1):(p+q)]

model$Q[,,1]

<- exp(pars[p+q+1])

model
}
model

<- fitSSM(inits=c(ar,ma,1e-7),model=model,updatefn = fn,

method='BFGS', control=list(REPORT=1,trace=0))$model

}else{
ar

<- rep(-0.05,p)

model

<- SSModel(d2 ~ SSMarima(ar = ar,

d = 0, Q = 1e-7),H=1e-7)

fn

<- function(pars,model,...){

model$T[2:(2+p-1),2,1]

<- (pars[1:p])

model$Q[,,1]

<- exp(pars[p+q+1])

model
}
model

<- fitSSM(inits=c(ar,1e-7),model=model,updatefn = fn,

method='BFGS', control=list(REPORT=1,trace=0))$model

}
aic.model <- -2*logLik(model) +2*(p+q)
results2[p+1,q+1] <- aic.model
}
}

55

Time Series Analysis by State Space Methods

results2 <- results2/length(d1)

rownames(results2) <- paste("AR",0:5,sep="")
colnames(results2) <- paste("MA",0:5,sep="")
results2[results2>100] <- 0

AR0
AR1
AR2
AR3
AR4
AR5

MA0
0.0000
4.1606
13.7568
14.4141
14.9953
10.6752

MA1
5.2210
5.2155
43.6969
5.1741
5.1943
5.2144

MA2
5.2115
5.2072
5.3608
0.0000
0.0000
5.2231

MA3
5.1785
5.1813
42.0075
5.9370
0.0000
5.2043

MA4
5.1980
9.8745
4.4523
6.8736
5.1484
5.0350

MA5
5.2182
3.5363
4.7491
10.2940
9.4143
5.0954

Table 2: AIC for different ARMA Models with missing observations

Choosing ARMA(1,1) model,

d1.miss

<- c(d1,rep(NA,5))

<- 1

<- 1

ar

<- rep(-0.05,p)

ma

<- rep(0.001,q)

model

<- SSModel(d1.miss ~ SSMarima(ar = ar, ma = ma,

d = 0, Q = 1e-7),H=1e-7)

fn

<- function(pars,model,...){

model$T[2:(2+p-1),2,1]

<- (pars[1:p])

model$R[3:(3+q-1),,1]

<- pars[(p+1):(p+q)]

model$Q[,,1]

<- exp(pars[p+q+1])

model
}
model

<- fitSSM(inits=c(ar,ma,1e-7),model=model,updatefn = fn,

method='BFGS', control=list(REPORT=1,trace=0))$model

## Warning:

Not a valid SSModel object.

Returning -Inf

out

<- KFS(model,filtering=c('state'), smoothing=c('state','disturbance','mean'))

d.smoothed

<-rowSums(as.matrix(out$alphahat))

par(mfrow=c(1,1))
plot.ts(d.smoothed, col="blue",xlab="",ylab="",main="",lwd =2)
points(d1.miss, pch = 19, col ="red",cex=0.7)

56

15

10

10

15

Time Series Analysis by State Space Methods

20

40

60

80

100

Figure 9.7: Internet series with in-sample one step forecasts and out of sample corecasts(points represent the actual
data)

57