W1

House TSEx NDef BP LRMTS FDLM HACSE OJP
MPO1A: Advanced Quantitative Methods

Introduction to Time Series
Distributed lag models
Paul Kattuman
Judge Business School
January 23, 2012
Paul Kattuman MPO1A: Session 1 January 23, 2012
House TSEx NDef BP LRMTS FDLM HACSE OJP Time table Assessment Applied Time Series
Course organisation
Timetable
Lectures
First 7 weeks: Mondays 14 : 00 to 16 : 00
Last lecture: Thursday 15 March 14 : 00 to 16 : 00
Lab Sessions
2 Feb 14 : 00 to 16 : 00
16 Feb 14 : 00 to 16 : 00
2 Mar 13 : 00 to 15 : 00
16 Mar 14 : 00 to 16 : 00
Emails:
Paul Kattuman : [email protected]
Michael Freeman : [email protected]
Module assessment
Workbooks and Contests
Book 1 Book 2 Book 3 Book 4 Contests
Handed out Feb 20 Feb 20 Mar 19 Mar 19 Mon: 30/1
Submit on Mar 12 Mar 12 May 16 May 16 Sun: 5/2
Weight 20 % 20 % 20 % 30 % 10 %
Applied Time Series Analysis
Objective, Use, Focus
Time series models useful for dierent types of research

problems
Uses:
Forecasting - with no causal interpretation eg: what will

employment level / rate of growth be next month?
Causal inference - dynamic causal eects, eg: if the 10 year

interest rate declines by 10 basis points, what eect on
employment growth in 3 months? in 12 months? [quantitative
easing]
Modelling risk (e.g., in nancial markets) - volatility in a

returns series, changing volatility, clustering of volatility
Applied course real world focus
Software: R
House TSEx NDef BP LRMTS FDLM HACSE OJP TSD Ex1 Ex2 Ex3 Ex4 Ex5 Ex6 Ex7 Ex8 Ex9 SPX
Time series data
Data collected on the same observational unit at consecutive,

equally spaced intervals over time
Y
t
= Random variable Y in period t
Collection of random variables {Y

t
}, t = 1, 2, . . ., indexed by
time of occurrence
Observed time series data set: {y

1
, . . . , y
T
} : T dated
observations on the variable Y
One realisation of the stochastic process generating the

collection of random variables {Y
t
}
A time series model is a mechanism that generates the series

of interest
... in terms of the moments of the joint distribution of the set

of random variables {Y
t
}
Employment UK
Employment US
Unemployment rate: UK
Unemployment rate: US
Ination rate: China
Ination rate: India
T-bills rate: US
10 year interest rates
Corporate bond rates
S&P Broad Market Index
House TSEx NDef BP LRMTS FDLM HACSE OJP Iss LDG L D ACV ACR TSP ACREx St
Issues in Time series analysis
Time lags
Correlation over time: serial correlation, or autocorrelation
Stationarity of the series
Correct estimation of Standard Errors when errors are serially

correlated
Forecasting models: Concerns are dierent from those in

estimation of causal eects
R
2
is important in forecasting
Model using historical data must be valid and useful for the
(near) future
No concern about interpreting coecients - omitted variable

bias is not a problem
Transforming time series random variables: lags, rst
dierences, logarithms, growth rates
Lags, rst dierences, logarithms, and growth rates
The st lag of time series Y

t
: Y
t1
; j
th
lag of Y
t
: Y
tj
The rst dierence of Y

t
: Y
t
= Y
t
Y
t1
The rst dierence of the logarithm of Y

t
:
ln(Y
t
) = ln(Y
t
) ln(Y
t1
)
The percentage change of a Y

t
between t 1 and t:
approximately 100ln(Y
t
); approximation is more accurate
when the percentage change is small
Lag operator
The lag operator denoted L:

LY
t
= Y
t1
L can be manipulated as if it were a constant. e.g.:

Y
t2
= LY
t1
= L(LY
t
) = L
2
Y
t
More generally:
L
p
Y
t
= Y
tp
L
0
1
Applying L to a constant leaves the constant unaected, e.g.:

L =
Dierence operator
Dierence operator denoted :

Y
t
= Y
t
Y
t1
Applying to a constant yields 0:

= 0
Lag and dierence operator:

= 1 L
Operators can be treated as scalars

Y
t
Y
t1
= (Y
t
Y
t1
) = (Y
t
) =
2
Y
t
Autocovariance
The covariance of a series with own lagged values:

Autocovariance
The rst order Autocovariance of Y

t
: Cov(Y
t
, Y
t1
)
The j
th
order Autocovariance of Y
t
:
j
= Cov(Y
t
, Y
tj
) = E(Y
t
E(Y
t
))(Y
tj
E(Y
tj
))
The j
th
order sample Autocovariance of Y
t
:

j
=

Cov(Y
t
, Y
tj
) =
1
T
T
t=j +1
(Y
t
Y
j +1,T
)(Y
tj
Y
1,Tj
)
Where Y
j +1,T
is the sample average of Y
t
computed over
observations t = j + 1, . . . , T.
The summation is over t = j + 1 to T why?

Autocorrelation
The correlation of a series with own lagged values:

Autocorrelation or Serial correlation.
The rst order Autocorrelation of Y

t
: Corr (Y
t
, Y
t1
) =
j
The j
th
order Autocorrelation of Y
t
:
j
= Corr (Y
t
, Y
tj
) =
Cov(Y
t
, Y
tj
)
_
Var (Y
t
)Var (Y
tj
)
These population Autocorrelations describe the population

joint distribution of (Y
t
, Y
tj
)
The j
th
order sample Autocorrelation of Y
t
(Y
t
stationary):

j
=

Corr (Y
t
, Y
tj
) =
Cov(Y
t
, Y
tj
)
Var (Y
t
)
A Time series property
One fundamental (visual) characteristic that distinguishes

dierent times series: their degrees of smoothness
Smoothness is induced by correlation between adjacent

random variables
To what extent does the value of the series at t, Y

t
, depends
on the past values Y
t1
, Y
t2
, , and/or past values of
shocks, disturbances, u
t1
, u
t2
,
A time series model expresses the mechanism that (we think)

generates the observed time series with this characteristic
property, probabilistically
Autocorrelation example : Unemployment rate UK
Highly serially correlated:

1
=??
Last months unemployment rate contains most information

about this months unemployment rate
Multiyear swings
Stationarity
Stationarity : history (the path to date) is irrelevant to the

future
A time series is stationary if the probability distribution of the

random variable Y
t
does not change over time
Stationarity: the joint distribution of (Y

t
, Y
t+1
, . . . , Y
t+T
)
does not depend on t
Note that stationarity does not rule out serial correlation; but
Corr (Y
t+k
, Y
t+s
) must not depend on t for any k, s
Stationarity is a key requirement for time series regressions
It is a testable property - next lecture
For now we assume that the time series variables we use in

any regression are stationary
House TSEx NDef BP LRMTS FDLM HACSE OJP TSD WN WN2 WN3 MA1 MA2 AR1 AR2 RW RWD
Time series data
Univariate time series data: one realisation of a stochastic

process
White Noise
The simplest kind of time series: a collection of uncorrelated

random variables, u
t
, with mean 0 and nite variance
2
White Noise
The plot: a collection of Standard Normal random variables,

in the order in which observations are drawn
The series is not smooth enough to serve as a plausible model

for any observed time series
If the stochastic behaviour of time series could be modelled as

white noise, classical statistical methods would suce.
With no serial correlation, and constant mean and variance,

random sample properties obtain
White Noise
A process {Y
t
} is White Noise if:
1. E[Y
t
] = constant mean
2. Var [Y
t
] =
2
< constant, nite variance
3. Cov[Y
t
, Y
ts
] = 0, s = 0 No serial correlation
If E[Y
t
] = 0 zero mean White Noise
Denote the zero mean white noise by u

t
Moving Average processes
One way of introducing serial correlation (and more

smoothness) into time series models:
Replace the White Noise series {u

t
} by its moving average
For example, replace u

t
by an average of its current value and
its immediate two past neighbours,
Y
t
= (1/3)(u
t
+ u
t1
+ u
t2
)
More generally: the current value Y

t
depends on some
weighted sum over q past shocks
Y
t
= + u
t
+
1
u
t1
+
2
u
t2
+ . . . +
q
u
tq
Moving Average
Smoother series as the larger shocks are averaged out
Plausible model (visually) for some observed time series

variables?
Autoregressive processes
Consider the White Noise series u

t
as input and calculate
output as:
Y
t
=
1
Y
t1
+ u
t
successively for t = 1, 2,
More generally: the current value Y

t
is a function of the past
value of the series; hence the name AutoRegression
Y
t
= c +
1
Y
t1
+
2
Y
t2
+ . . . +
p
Y
tp
+ u
t
Need the initial condition (Y

0
) to generate Y
1
Autoregressive process
The Autoregressive process and its generalizations can serve

to model many observed series. For example, GDP growth.
Random Walk
A special case of an Autoregressive process:

Y
t
= Y
t1
+ u
t
for t = 1, 2, , with initial condition Y

0
= 0.
The value of the time series at time t is the value of the series
at time t 1 plus a completely random movement determined
by u
t
N(0, 1).
Random walk
Random Walk with drift
A model for time series with a trend is the Random Walk with
drift:
Y
t
= c + Y
t1
+ u
t
t = 1, 2, , Y
0
= 0
The constant c is the drift; if c = 0, Random Walk
Rewrite the model as a cumulative sum of White Noise

variates:
Y
t
= ct +
t
i =0
u
t
replacing Y
t1
by c + Y
t2
+ u
t1
and so on
Note the time trend

The problem we have to solve: Univariate case
Estimate the best model given one realisation of the time

series (Moving average? Autoregressive? ARIMA? ).
Key terms and concepts so far
Time series
Lag operator
Dierence operator
Autocovariance
Autocorrelation
Stationarity
White noise
Moving average
Autoregression
Random Walk
Random Walk with drift

References
Robert H. Shumway and David S. Stoer (2006), Time Series

Analysis and Its Applications. With R Examples. Springer,
second edition.
Stock, J. and M. Watson (2006), Introduction to

Econometrics, Addison-Wesley. Second Edition. Ch. 14.
House TSEx NDef BP LRMTS FDLM HACSE OJP TSWT LRM DCEDLM DCE
Time Series with Trends
In Time Series analysis, will need to alter classical linear

regression model assumptions to take into account that we no
longer have a random sample of observations
Need to recognise the temporal ordering of observations in the

realization of the stochastic process
Time series are often not stationary often have trends of

some kind or the other, or trend breaks
Two time series trending together may both be driven by

some other unobserved factor, and may not have a causal
relationship between them
How do we handle trending variables - next lecture
For now assume that we are working with stationary time

series
Time Series vs. Cross Sectional
The linear model : as dominant in Time Series context as in

classical (cross-section) analysis
Time series Y
t
, expressed as a (response) to a linear
combination of other input time series
E.g., a simple model that is linear in parameters:

Y
t
= +
0
X
t
+
1
X
t1
+ +
k
X
tk
+ u
t
Can estimate the coecients ,

0
,
1
, ,
k
by least
squares, under various assumptions
Dynamic Causal Eects: Distributed Lag model
Agenda
Dynamic Causal Eects
Example: Orange Juice price process
Estimation of Dynamic Causal Eects with Exogenous

Regressors: The Distributed Lag Model
HAC Standard Errors
Dynamic Causal Eects (Cumulative multipliers) in the

Orange Juice Price process
Dynamic Causal Eect
Simple Static models using time series relate one time series
variable to other time series variables
with the eect assumed to operate within the period.
In many cases the eect of a causal variable is not contained

within the period
Dynamic eects:
-eect of a stimulus / austerity package on output over time

this year, next year
-eect of a change in the Bank of England rate (or the Fed

Funds rate) on ination, this month, in 6 months, and 1 year
-eect of a freeze in Florida on the price of orange juice

concentrate in 1 month, 2 months, 3 months,...
Dynamic causal eects are the eects on Y over time, of a

unit change in X
The size and nature of the eect can vary over time
House TSEx NDef BP LRMTS FDLM HACSE OJP OJP D OJSR Ex Ex Ex DLE DLE DL1 DL2 DM
Orange Juice Price
Data
Monthly, Jan. 1950-Dec.2000(T=612)
Price=price of frozen OJ
%ChgP=percentage change in price at an annual rate, so

%ChgP
t
= 100ln(Price
t
)
FDD=number of freezing degree-days during the month,

recorded in Orlando FL
Example: If November has 2 days with lows < 32F, one at 30

and at 25, then FDD
Nov
= 2 + 7 = 9
Data
OJ prices: Static regression
%ChgP
t
= 0.4 + 0.47FDD
t
(0.22) (0.13)
Statistically signicant positive relation
More freezing degree days leads to contemporaneous price

increase
Standard errors are heteroskedasticity and autocorrelation

consistent (HAC SE)
What is the eect of FDD over time?

Causal eect: outcome of an ideal randomized controlled
experiment
Eect of fertilizer on tomato yield: 3 year experiment:

Fertilize some plots, not others (random assignment)
Third year harvest measured: Y

t
One lot fertilised in year 1: X

t2
= 1, Eect on Y
t
of unit
change in X
t
2:
2
Another lot fertilised in year 2: X

t1
= 1, Eect on Y
t
of unit
change in X
t
1:
1
A third lot fertilised in year 3: X

t
= 1, contemporaneous eect
of X
t
on Y
t
:
0
Fourth lot left unfertilised (Omitted dummy)
Causal eect: di in yields fertilised and unfertilised
Dynamic causal eect: sequence of causal eects: of X

t
on
Y
t
, Y
t+1
, . . . sequence of coecients:
0
,
1
,
2
, . . .
In time series, we generally cannot conduct this ideal randomized
controlled experiment
Only one OJ market...
Cannot randomly assign FDD to dierent replicates of the OJ

market
Cannot measure the average (across subjects) outcome at

dierent times only one subject
So cannot estimate the causal eect at dierent times using

the dierences between groups
An alternative thought experiment
Randomly give the same subject dierent treatments (FDD

t
)
at dierent times (as they occur)
Measure the outcome variable (%ChgP

t
)
The population of subjects consists of the same subject (OJ

market) but at dierent dates
If the dierent subjects are drawn from the same

distribution that is, if Y
t
, X
t
are stationary then the
dynamic causal eect can be estimated by an OLS regression
of Y
t
on lagged values of X
t
This estimator (regression of Y

t
on X
t
and lags of X
t
) is the
nite distributed lag estimator
Eects of X on Y over time
Distributed Lags in Eect
Modelling Approach: Distributed lag model
A static model relates contemporaneous variables:

Y
t
=
0
+
1
X
t
+ u
t
A nite distributed lag model of order q will relate Y

t
to up to
q lags of X
t
Y
t
= +
0
X
t
+
1
X
t1
+
2
X
t2
+ +
q
X
tq
+ u
t
The coecient
0
is the short run or impact multiplier: the
contemporaneous eect of X
t
on Y
t
, holding past X
t
constant
Modelling Approach: Distributed lag model
Y
t
= +
0
X
t
+
1
X
t1
+
2
X
t2
+ +
q
X
tq
+ u
t

1
is the 1-period dynamic multiplier: the eect of change in
X
t1
on Y
t
, holding X
t
, and X
t2
constant

2
is the 2-period dynamic multiplier: the eect of change in
X
t2
on Y
t
, holding X
t
, X
t1
, and X
t3
, X
t
4
constant
To be able to estimate such a regression, the X

s
must
exogenous (otherwise cannot use OLS use VAR)
Dynamic Multipliers
Each
j
is known as the multiplier at lag j
The sum of partials

0
+
1
,
0
+
1
+
2
, etc. are called the
intermediate or cumulative multipliers;
The sum of all the partials, that is,

0
+
1
+ +
q
is the
long run multiplier it reects the long-run change in Y after
a permanent change in X
How far back should we go in specifying the model? What lag

length? Let the data decide
Problem: the lag length can be very long (too many

variables!)
Solution Autoregressive Distributed Lag model - next lecture

House TSEx NDef BP LRMTS FDLM HACSE OJP Why M1 M2 M3 F1 F2 R
HAC Standard Errors
When u
t
is autocorrelated, conventional OLS Std. Errors
(heteroskedasticity-robust or not) are wrong
Tests of hypothesis about dynamic causal eects will be wrong
Errors in forecast condence intervals
We need Std. Errors that are robust to autocorrelation as well

as heteroskedasticity
Heteroskedasticity-and Autocorrelation-consistent(HAC)
Standard Errors
HAC Standard Errors
Simple example, with single regressor X

t
:
Y
t
=
0
+
1
X
t
+ u
t
The OLS estimator: (Stock and Watson, App. 4.3)
1
=
1
+
1
T
T
t=1
(X
t

X)u
t
1
T
T
t=1
(X
t

X)
2
Denoting (X
t

X)u
t
by v
t
, in large samples
1
=
1
+
1
T
T
t=1
v
t
2
X
HAC Standard Errors, ctd.
So in large samples, the variance of the estimator:

Var (
1
) = Var
_
1
T
T
t=1
v
t
_
/(
2
X
)
2
=
1
T
2
T
t=1
T
s=1
Cov (v
t
, v
s
) /(
2
X
)
2
In i.i.d. cross sectional data, Cov(v

i
, v
j
) = 0 for i = j
Var (
1
) =
1
T
2
T
t=1
Var (v
t
) /(
2
X
)
2
=

2
v
T(
2
X
)
2
The usual i.i.d. result
HAC Standard Errors, ctd.
The need
But in time series data, Cov(v

t
, v
s
) = 0 in general
Let T = 2. Consider the numerator in the expression for

Var (
1
)
Var
_
1
T
T
t=1
v
t
_
= Var [(v
1
+ v
2
)/2]
= [Var (v
1
) + Var (v
2
) + 2Cov(v
1
, v
2
)] /4
=
2
v
/2 +
2
v
/2 (where
1
= Corr (v
1
, v
2
))
=
2
v
f
2
/2, where f
2
= (1 +
1
)
In i.i.d. data,
1
= 0 so f
2
/2 = 1, yielding the usual formula
for Var (
1
)
In time series data, if

1
= 0 then Var (
1
) is not given by the
usual formula
HAC SE: Expression for Var(estimator), general T
Var
_
1
T
T
t=1
v
T
_
=

2
v
T
f
T
So
Var (
1
) =
_
1
T
2
v
(
2
X
)
2
_
f
T
Where
f
T
= 1 + 2
T1
i =1
_
T j
T
_
If we knew the factor f

T
= 1 + 2
T1
i =1
_
Tj
T
_
j
, we could
make the adjustment
HAC SE: Expression for Var(estimator), general T
Var (
1
) =
_
1
T
2
v
(
2
X
)
2
_
f
T
, where f
T
= 1 + 2
T1
i =1
_
Tj
T
_
The most commonly used estimator of f

T
is:
f
T
= 1 + 2
m1
i =1
_
m j
m
_

j

j
is an estimator of
j
The Newey-West HAC SE estimator
m is called the truncation parameter
Why not just set m = T?
How should you choose m?
Rule of thumb, m = 0.75T

1/3
Recap: When do you need to use HAC S.e.s?
use HAC SEs When u

t
is serially correlated: if u
t
is serially
uncorrelated, then OLS SEs are ne
In AR and ADL models, the errors are serially uncorrelated if

you have included enough lags of Y
If you include enough lags of Y, then the error term cannot be

predicted using past Y, or equivalently by past u. So u is
serially uncorrelated
House TSEx NDef BP LRMTS FDLM HACSE OJP OJP DM1 DM2 CM1 CM2 Rev
Analysis of the OJ Price Data
What is the dynamic causal eect (what are the dynamic
multipliers) of a unit increase in FDD on OJ prices?
%ChgP
t
=
0
+
0
FDD
t
+ +
q
FDDt q + u
t
What q to use?
Let the data determine (are the coecients signicant?)
Use HAC standard errors
What m(Newey-West truncation parameter) to use?

m = 0.75 612
1/3
= 6.4
= 7 (observations=612)
The Dynamic Eect
Estimates
Multipliers: Dynamic eect of an FDD on the price of
Orange Juice
Cumulative multipliers
Estimates
Cumulative multipliers
Estimates
What did we do?
Notation, Denitions
Basic Time Series Processes
Finite Distributed Lag Models
HAC Standard Errors
Estimating dynamic and cumulative multipliers and their

standard errors

W1

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

W1

Uploaded by

Copyright:

Available Formats

House TSEx NDef BP LRMTS FDLM HACSE OJP

MPO1A: Advanced Quantitative Methods

Time series models useful for dierent types of research

Forecasting - with no causal interpretation eg: what will

Causal inference - dynamic causal eects, eg: if the 10 year

Modelling risk (e.g., in nancial markets) - volatility in a

Applied course real world focus

Data collected on the same observational unit at consecutive,

Collection of random variables {Y

Observed time series data set: {y

One realisation of the stochastic process generating the

A time series model is a mechanism that generates the series

... in terms of the moments of the joint distribution of the set

Correlation over time: serial correlation, or autocorrelation

Stationarity of the series

Correct estimation of Standard Errors when errors are serially

Forecasting models: Concerns are dierent from those in

No concern about interpreting coecients - omitted variable

The st lag of time series Y

The rst dierence of Y

The rst dierence of the logarithm of Y

The percentage change of a Y

The lag operator denoted L:

L can be manipulated as if it were a constant. e.g.:

Applying L to a constant leaves the constant unaected, e.g.:

Dierence operator denoted :

Applying to a constant yields 0:

Lag and dierence operator:

Operators can be treated as scalars

The covariance of a series with own lagged values:

The rst order Autocovariance of Y

The summation is over t = j + 1 to T why?

The correlation of a series with own lagged values:

The rst order Autocorrelation of Y

These population Autocorrelations describe the population

One fundamental (visual) characteristic that distinguishes

Smoothness is induced by correlation between adjacent

To what extent does the value of the series at t, Y

A time series model expresses the mechanism that (we think)

Highly serially correlated:

Last months unemployment rate contains most information

Stationarity : history (the path to date) is irrelevant to the

A time series is stationary if the probability distribution of the

Stationarity: the joint distribution of (Y

Stationarity is a key requirement for time series regressions

It is a testable property - next lecture

For now we assume that the time series variables we use in

Univariate time series data: one realisation of a stochastic

The simplest kind of time series: a collection of uncorrelated

The plot: a collection of Standard Normal random variables,

The series is not smooth enough to serve as a plausible model

If the stochastic behaviour of time series could be modelled as

With no serial correlation, and constant mean and variance,

Denote the zero mean white noise by u

One way of introducing serial correlation (and more

Replace the White Noise series {u

For example, replace u

More generally: the current value Y

Smoother series as the larger shocks are averaged out

Plausible model (visually) for some observed time series

Consider the White Noise series u

More generally: the current value Y

Need the initial condition (Y