Chapter 15: Time Series Regression and Forecasting

>Time Series DAtA:What's Different:

c
examples
 >urble
correlated wit
self-

>
Introduction time
to series Data 2 Serial Correlation:

(1)

(2)

(3)
notation:

lags, first differences, and growth rates:

c
((Yt) m(Y=-Y-)
note:
bc I
InCYt-Yer) E
=
 , we multiply by 4 Annualize
to

c
Autocorrelation Serial Correlation):

1 Thisi sthe same

pop.

W
correlation mrible
used
1
Ing before.
Sample Auto correlations:
 months
L 14 ago
1 month

Ago

As
you
see that
Asj" CPUC As
j" his
cy,
less
predictive power

c
forecasting, stationality, and the mean squared Error:
forecast

* c important
stationality:

I have
adort this
about
a re

> This is
what we

Care About

,
we return to this later
in the chapter.
 think
c So
of stationality as
follows:

-
you
have series
a
of data:( 1 ) Y,, Y, Yy,..., Y time series
out-of-sample
to data.

· And Another series

of dataofthe same
length thatstarts when series (1) at
is
4,:(2) Y, s'Yats'...Y++ s

Then the distribution distribution

·
probability sampling both time
of series;which are
of the same
length;sampling
Ys;butthey start
at differentperiods the
frequency a same are same

· note:
they may
have differentmiles ofcoefficients, data, etc;but
they are the same in
probability distribution

sprobalistic sense)

differentpoints have the this

· ..time series
starting At in time same mean a mariance, is

the i.. but time data.

equimlentto is series

· Since each y, . . . , tehas same

sampling sprobabilitydistributions a
they have the same mean a

Variance
(Gly,) =...
E(y)
=
a
nrcy,) ...

varcy,

> This the

is same
concepty i.i.d;butfor time series data (dathAcross time

forecasts and forecast Errors:

a
dAtAup period I
to

Actual Realization ( W

forecasto utofsample prediction

That why
is

it called
is

M5ft

We Care About
C
this alotistime
Series
Regression
C
Autoregressions:

b Autocorrelation
> ARSI) C
ofSerial

) ARCP)

The
firstorder Autoregressive model ARCI):

Random error
noise
Example:ARCI) model for the
growth rate
of614:

> So Ho:B, O

H,: , 0 c
only malid if
stationary scries

>Estimated model

>
GPPGR2017:
Q3
CYT
w

Y +

15
 he ARCP) Model: multiple lags
using for forecasting:

So jointhypothesis that:Ho:, 0, p0
=

>
=
...,

more on the ARCI) model:

consider the model: Pot PY- 4=

I following ARCI
It
+

(1) CAses
ofB:

c (I) B 0 <
history does not matter

be: P+ random
c
4 U,intercept noise
+
 model
(2) O<, 1c
stationary (Series
stationary
is

(3)
i B,
1 C
Y Po +
Y. Y=C
+

history permanent is

& This known

is As A Random walk model

show the
· We will that
if P, =1> series
nonstationary
is

note:Inter discuss
~ . on we will cases (1) 9(2) in the
monstrationality section
ofthis chapter

(4) there a
is case where B, 1 a Another where
0, however in economics are do not
study those cases.

> .
they are me in Economics a we will focus more or (2) 2(3).

Implications of 0
B, 1 and that
P 0:

other than the looka tthe implications q:0

concepty stationarity;lets
so
a B 1

consider the model:

3 So
following Y B+ BY- 4
+
=

that
Yo0c initial condition
- .
suppose

Also that:U, =0.3, R2 0.2, 4,

·
suppose =
=
-0.1, 4, 0.4

thatEIU)
. so we are
justassuming 0 C
they cansel out
on
Average

that:B0 So the
· now
suppose &
B0.3, Series
stationary
is

0 0.310) 0.30.3
C...
,
+
+

0.510.03) 0.05
y2 0 +
0.2

Y3 0 + 0.510.05) 0.1 0.075

0 0.50.075) + 0.2 0.1625

Yu
-
+

So As see that here: start i A "shock occurs" (n,=0.3) & get y,

c you you y =
0.
you

c then
Y, obtained
is As
half of the firstshock(0.5(0.3) A
+
second shock (U.=-0.2)
 obtained
>
Y is As half
of
the
hay shock
ofthe first a
halfof the second shock 0.510.051) +

Athird shock
(m, 0.1)
=

3
Yy
obtained
is As half
of
the
hay ofthe hay shock
ofthe first
and half
of
the
hoty ofthe second shock and ye the
of
third

Shock (0.5) 0.075) A

+

fourth shock (Un 0.2)

C.: As
you
see that the firstshock's impact less
is a less as t"(first shockmatters less a less)

> Same in and, 3rd shock, etc.

goes

C... this shows that

"today"is as result
a
of the initial condition in
+

declining importance All

previous
shocks leader shocks

C So whatt his tells us isthat

for stationary series (models;history matters;but
recent
history matters more a earlier
history
matters less.

Implications of
0
B, 1 and that
no introduction
of
An
intercept:

I now;we will
study the same
implications above;buti sintro of intercept

&.
SAy B, I

1 0.5(0) 0.3 1.3

3.
Y,
+
+

Y2 1 + 0.511.3) 0.2 1.45

Y3 1 0.5(1.45)
+
0. I 1.625

1 + 0.5(1.625 +0.2-2.0125
Yu

the
3 So As
you
see that
implications are the same As in
p0 =

could think trend drift

y P.
> So As a or a
you

> So history continues matter

to here (At
decreasing
A
Rates;the same
way
when P. =0; buto n
topy thatwe have

trend.
A

and around thattrend.

declining
shocks happen
a
(Ata Rites

Handline

See >
 C note:
stationary models Are Actually P, I
C-1, 1 c however;i n this course are rule out
the
negative

its advanced classes

c
for more
they are common
not in economics.
-

Implications y B, 1
P 0.
+
=

< So here are assume

non-stationary
A model;since ,

0 (1)(0) 0.3 0.3

3.
Y,
+
+

(1) (0.3) 0.1

Y20 +
0.2

Y0 +
(1)(0.1) 0. I O

Yy0 (I) (0 10.2 0.2

-
+

> So this known

is As the radom walk models more on this in inter sections

c
but
as
you
see that
by
is
perfect history ofAll
A
previous
shocks
(they are
declining)
not

, . us My

Since 0 + 0.3 0.2 + 0.2

Y,
C 0.1

-
W
All
ofhistory, sumo previous shocks
initial
condition

> So
"today"is the
sumy All
previous
shocks.

3 So All
of history incorporated
is into
todity; isno
defining down
history.

Implications ofB. 1 i An introduction An

intercept:
of

I now;we will
study the same
implications above;buti sintro of intercept

0.S
SAy
.
&
 0.5 (l) (0) 0.3 0.8
3.
Y,
+ +

Y2 0.5 + (1) 10.8) 0.2 1. I

Y 0.5+ (1)) 1.1) 0.1 1.5

(1) (1.5 2.2

Yy 10.2
0.5 + -

the B. =0;buti tcomes trend

history of B
so matters same as in A
in 0.5
c
=

C. So
by J. +Gd=
+

410.5)
30
+

initial
history +
0.3 0.2 0.1 +0.2

trend -
condition
history y shocks

this
> So Also
is
non-stationary
A series

- .
i P 0
=
> Simple Andom walkmodel

random walkmodel a trend

·
if $,0 c in

Time additional model:

c series
regressions with
predictors and the Auto
Regressive
Distributed
Lag (ADL)

>error shock
Example y ADL:nterest
rates and the term spread:

spread btw c

2 ts
 Istimation the and
MSAE forecastIntervals:
> of

forecastErrow:

of y's up Tto
c vector

vector X,
c of up to T

unexplained
~error

&
tid arrow due being
to
wrong

The mean
squared forecast
er ror (msft)
of
ADL model:

itwool be the MSSE the B,

a so
Are true
if
pop. Bs

cie:((u + +
1)
 < So Above whatwe are
trying say
to that:MSSE ECUT" if I are estimated truthfully
0

*
e

& since > Ms+ E Elu+1)" ECIB Bol Y+ 23

((u,,, e(0)
+

E(u 1)
+

Also the
of
number
ofpredictors
are small relative sample
to size (most of
world
MSFE arise
from ECU++,
and ECCB Bol Y, + B B 2] 'would be
very
small

c iopposite (ECU+) E21B Bo)Y, + (BY B 23's this would be

large

The root mean

squared forecastError RMSAE):

c i.e. similar
standard
to
error
ofRegression
 construct
Using the RMSAE to n ternals:

normal distribution
c
from a 2 1.x
of

Example:forecastinterval:

fanchart

here darker
3 So forecasts are more
likely be
to in
regions.
 Selection
>
Lag Using Information
criteria:

models.
Lag selection in
Arcps

3 So we don't need to

About
worry Omiting
Ariables

making multiple t-testo r test

f
I so

like chapter
is 8.

too
large Cdfv b estimates" would be
large efficient) less precise estimate
i pis number St (not
> So
y ( gp, a

the
of B, C less variance
of Y large.
is

c
ip
too
is small (bins ofI,"be we have omitted lays thathave still
strong predictive power
(b)
they are still

strongly serially
correlated
4,

I so thati swhen BIC's SAIC's come in

handy.
 The
Bayes Information Criterion BIC:

so think random mariable

c. I p
as
pop,
· And BICCP) estimates " P

2 The Aknike
Information Criterion MIC:

3 As

AIC
you
uses
see that

different
a
Al

penalty
term

i.e:
c
monthly frequency is
quarterly.
Example:BICSP) in ARP) model:

C Since itminimizes BICSP)

c
there is a here
trade of

RV
that
predictors
regressors
We know As

But
BICCP) has chosen instead 46
·
p=2 y
=

That should be used.

· is
why Again is
Lag selection in
ADLp,r) models.

>note ACL model Looks like this:

general

⑱
 reads:
nonstationary
·
a

(A)

(B)

x)

(D)

(2)
What is Trend?:
a

trend line

>has some sort

of quadratic trend

c no observable trend.
Chard tell
to
< does not have trend
a
Deterministic 2 Stochastic Trends:

3 So the firstseries appeared be

to
polynomial in time.

>Random shock

I
serially
uncorrected

> As we will show Random

later; walk models are
special
a
type of Arcis model where
,
I
 here have differentminuses

"... they
So in time:
ifyou compare
& . a series

>
proof:
uncorrelated Y,)
Since UI serially
is
(EIKLY,, ...,
0

And that
· we know stationarity:VArsy
is
Varsyz...vAriYz-p) rcYss) etc

so random walk model is:

·
Ye *t

.
·
WrCY) WrCY, +mrCUL

mrcy) wrcy.
+

On

c since 6? F0 C Ary,) wrCYE

 > So
Assuming:1 0:
Y, Y k, u,
+
= =

Y =Y,
+

k, k, 4,
=
+

yy Y =
kz k, kz ks
+
=
+ +

Y= Yz
=

-
1
4
+

=
= k, +
kc ...
+

kz
+

C..
mrcY) wr K, t...+
Us to f(t)

time
c function of

c so what we saw have

again
that
is the
jointdistribution of (Y,s:...Yits) depends on 5 since
mrcy) for example;differs across time

dift I periods
 Random
for whether Y,has walk (thus
stationary)
tests A
a non or not

Stochastic trends and Unit

Autoregressive roots:
what caused trends:
problems are
by

Industrial production

Japanese

& Statistically Significant

C positive
URATE

JAP IP

C both are
very stratiscally significants they are
strongly correlated;buti n realitysince they are both trending mimbles,
they seem

correlated but
ingestthis correlation
sparios.
is
 How do detect
stochastic trends?
you

As discussed earlier, the random walk is special

a case the much model
P,
of
· to

& contains stochastic trend.

<if, non-stationary
1
CYseries is A
=

hus model;the trend

· the
in ARCI)
hypothesis that ARCI has stochastic
a is:

Ho:, I
non-stationary be Unit Root, Stochastic trend
c
of
1,: , I be stochastic trend
a
stationary y
no unit Root (no

This the model: ARCI model

>
hypothesis testi s uses
following iP +
Bi, *=)
+

so
rejectthe null hypothesis n on
stationary.
if y
a we series is
· now this test
ismost
easily implemented by estimating modified
a version the
of AR (II model;which obtain
is
by subtracting it

from both sides:

=El Pot PY 1*t1

IP+P, Y- -+*t

I Do Pr tl t
+

P E1, t
+ +
.

C... : 8 0
I non
stationary

-:8 0 >
Stationary

so B, I Rootexists
stationarity
I
. C Unit (non

so Root
stationary
unit exists
·

I B I no (non

distributed
c isnotnormally it not
is
normally distributed

..

Clinear)
le
deterministic
WriAb ~
time
trand

ne

↓
Loef
The
Dickey-fuller Test
in ARP):

Ho:8 D

H,:84

deterministic trand.
c
including
When should include time
a trend in the of test?
you

3 mene

S merman men

3 Hers are must include

time
A
trend
 M
-

Bo Bi

>I-sided

I fail to rejectHo: 80 => P, 1=

3 I(CPP) is non stationary (we saw thatthere

was trend
a
graphically
 How Address Raised trends:
mitigate problems by
to

i < is
stationary.

s nonstationarity:Breaks and model

stability:

a so this
non-stationary
is by , differ
Across time
 Tests Breaks change
for in
Regression (officients:

(1) Case
IBreak change)
1: is known:

Durm iAble

3 Jo is for intercept

5, is
for ,

U, for
is
5,

c8 0, 4 =
=
0
=
) nostructural Break

null Structural break

3
if
we
reject this
hypothesis occurs t
At
 understand
, this
diagram will
help to structural Breaks in this case:

O
T
I

D(I) 0 0 00 I , 1 I 1 I

So that structural breakoccurs 5

suppose we know a t=
at

So their introduced technological charge, pandemic that

some
governmentpolicy innovation,
or some
3 · is or or

At
-
causes a structural break.

their structural
change;then model
if
a so is no ou r ADLs1, 1 is:

> Y B+PY- , +4t

+

c their
if
is structural breaks;then before is the model will look. As
follows:

> Y B+PY- , +4t

+

however;after I, the model looka s

c
will
follows:

>
+

Vo +

V)Ye-,
+ +

8 + Un x
= 1
+

Ut
 C Since this is
jointdistribution.
a

break
c f
we
reject H.:8 0, 2= = 0
= ) structural occurs

evidence breaks.
ifthe fail reject
structural
of
s to (no

&
Y, Read the slides.
for case (2)