1.

Modeling trend Linear trend:

AA037007/EQC7006 • Trend is slow, long-run evolution in the variables that • The trend increases or decreases like a straight
Time Series Analysis we want to model and forecast.䎻࣯ᱟᡁԜᜣ㾱ᔪ⁑઼亴⍻Ⲵਈ䟿 line.
ѝ㕃ធⲴǃ䮯ᵏⲴ╄ਈ

• Here we focus on model of deterministic trend, in • It is a simple linear function of time.

Lecture 3a which the trend evolves in a perfectly predictable
way.൘䘉䟼ˈᡁԜу⌘Ҿ⺞ᇊᙗ䎻࣯⁑රˈަѝ䎻࣯ԕа⿽ᆼ‫ޘ‬ਟ亴⍻Ⲵᯩᔿ╄
Tt E 0  E1TIMEt
ਈDŽ

Modeling Trend䎻࣯ᔪ⁑ • TIMEt is constructed artificially and called a time

• Two main patterns of time trend are:
trend or time dummy. d/DᬵᱟӪᐕᶴᔪⲴˈ㻛〠Ѫᰦ䰤䎻࣯˄ਈ䟿˅
(deterministic/regression method) (i) linear ᡆᰦ䰤㲊ᤏਈ䟿DŽ

(ii) nonlinear • For a sample of size T, TIMEt =1,2,3,…,T-1,T.

Example 1a: modeling trend. Example 1b: appropriate to be used on…

• If we put it differently, TIMEt = t, then:
40.00
Figure 5.1 Labor Force Participation Rate for Females
Tt E 0  E1t 65

20.00
Trend1t 10  0.25t 60
where t =1,2,3,….., T-1, T.
55
0.00

50
• ȕ0 is the regression intercept, it is the value of
Rate

-20.00 45
the trend at time t=0.
40
• ȕ1 is the regression slope, it is positive if the
-40.00
trend is increasing and negative if the trend is 35
Trend 2t 50  0.8t
decreasing. 30
-60.00 1950 1960 1970 1980 1990 2000
• The larger the absolute value of ȕ1, the steeper 1 11 21 31 41 51 61 71 81 91

Trend1 Trend2 Time

the trend’s slope. ɴІⲴ㔍ሩ٬䎺བྷˈ䎻࣯Ⲵᯌ⦷䎺䲑ጝDŽ
↻ᐞ੸⧠ࠪа⿽⁑ᔿ˄‫↓ݸ‬ਾ䍏ˈ޽↓޽䍏˅ˈ䘉㺘᰾ㆰঅⲴ㓯ᙗ䎻࣯ਟ㜭⋑ᴹᆼ‫ࡠ᥹ᦅޘ‬ᮠᦞⲴᡰᴹ㌫㔏
ᙗਈॆ˄ֻྲˈ໎䮯⦷Ⲵਈॆ˅DŽྲ᷌↻ᐞᱟᆼ‫ޘ‬䲿ᵪⲴˈ䛓Ѹ⁑රਟ㜭ቡ∄䖳ਸ䘲DŽ
Example 1c: appropriate to be used on…
Figure 5.4 Linear Trend: Labor Force Participation Rate for Females Figure 5.5 Linear Trend: Labor Force Participation Rate for Males

Figure 5.3 Labor Force Participation Rate for Males 90

60
90.0 85
50 87.5
80
85.0
40 75
4 2
82.5
2 30 70
80.0 1

Rate
0
77.5
0
-2
75.0
-4 -1
72.5
-6 -2
1950 1960 1970 1980 1990 2000 70.0 1950 1960 1970 1980 1990 2000
1950 1960 1970 1980 1990 2000
•ਣ‫ח‬㓥䖤 (30-60): Residual Actual Fitted Residual Actual Fitted
Time
• Actual (⛩⣦㲊㓯): ᇎ䱵Ⲵྣᙗࣣࣘ࣋৲о⦷ᮠᦞ˄਼ка亥Ⲵ㓯˅DŽ
• Fitted (䮯㲊㓯): ⭘㓯ᙗഎᖂ⁑රᤏਸࠪᶕⲴ䎻࣯㓯DŽ䘉ᱟаᶑㅄⴤⲴкॷⴤ㓯ˈԓ㺘Ҷ⁑රՠ䇑ࠪ
Ⲵ䮯ᵏᒣ൷໎䮯䎻࣯DŽ
•ᐖ‫ח‬㓥䖤 (-6 to 4):
• Residual (ᇎ㓯): ↻ᐞDŽ↻ᐞᱟᇎ䱵٬ (Actual) ߿৫ᤏਸ٬ (Fitted) ⲴᐞDŽᆳ㺘⽪⁑රᵚ㜭䀓䟺Ⲵ䜘࠶DŽ

Example 2b: modeling trend.

Nonlinear trend: Example 2a: modeling trend.
ȕ1 < 0 and ȕ2 < 0 : the trend is monotonically
ȕ1 > 0 and ȕ2 > 0 : the trend is monotonically, but decreasing. ɴІфϬф ɴЇфϬ˖䎻࣯ᱟঅ䈳䙂߿ⲴDŽ
Sometimes trend appears nonlinear, or curve (not nonlinearly, increasing. ɴІхϬф ɴЇхϬ˖䎻࣯ᱟঅ䈳Ⲵˈն䶎㓯ᙗ䙂໎DŽ 1000
a straight line). For example, when a variable 4000
500
increases at an increasing or decreasing rate.ᴹᰦ䎻࣯ 0
Trend t 10  0.4t  0.4t 2
3500
㺘⧠Ѫ䶎㓯ᙗˈᡆᴢ㓯˄нᱟⴤ㓯˅DŽֻྲˈᖃањਈ䟿ԕ䙂໎ᡆ䙂߿Ⲵ䙏⦷໎࣐ᰦ
-500
3000

2 -1000
2500 Trend t 10  0.3t  0.3t
(a) Quadratic trend. -1500
2000
¾ Quadratic trend models can potentially capture -2000

nonlinearities in a series.Ҽ⅑䎻࣯⁑රਟԕ▌൘ൠᦅ᥹ᒿࡇѝⲴ 1500

-2500
䶎㓯ᙗ⢩ᖱDŽ
1000 -3000
¾ It is defined by: -3500
500

-4000
Tt E 0  E1t  E 2t 2 0 1 11 21 31 41 51 61 71 81 91
1 11 21 31 41 51 61 71 81 91
Trend
Trend
 Example 2c: modeling trend. Example 2d: modeling trend.
ȕ1 < 0 and ȕ2 > 0 : the trend has U shape. ȕ1 > 0 and ȕ2 < 0 : the trend has an inverted U
shape.䎻࣯੸‫ ق‬U ᖒDŽ • Keep in mind that quadratic trends are used to
800

500 provide local approximations; one rarely has a

700

300
600
“U-shaped” trend. Instead, all of the data may
Trend t 10  25t  0.3t 2 lie on one or the other side of the U.䈧䇠տˈҼ⅑䎻
100 500
࣯䙊ᑨ⭘Ҿᨀ‫׋‬ተ䜘䘁լ˗ањᒿࡇᖸቁՊᆼᮤൠ੸⧠ĀUᖒā䎻࣯DŽ⴨৽ˈ
400 ᡰᴹᮠᦞਟ㜭ਚսҾUᖒⲴа‫ח‬ᡆਖа‫ח‬DŽ
-100

300
-300
200
Trend t 10  30t  0.3t 2
-500
100

-700 0
1 11 21 31 41 51 61 71 81 91 1 11 21 31 41 51 61 71 81 91

Trend Trend

•ᤏਸⲴҼ⅑䎻࣯㓯˄䮯㲊㓯˅൘ਾᵏ˄1985ᒤਾ˅䖳ྭൠᦅ᥹ҶӔ᱃䟿࣐䙏໎䮯Ⲵ⁑ᔿDŽ
•❦㘼ˈ൘ᰙᵏ˄1960-1980˅ˈᤏਸ㓯ᱟањл䱽ਾ޽㕃ធкॷⲴUᖒᓅ䜘ˈ㘼ᇎ䱵ᮠᦞ൘䛓⇥ᰦ䰤ᱟ⴨ሩᒣっᡆ
Example 2e: appropriate to be used on… 㕃ធ໎䮯ⲴDŽ䘉ሬ㠤൘ᰙᵏᤏਸ᭸᷌н֣DŽ Other types of nonlinear trend are sometimes
•↻ᐞമᱮ⽪ˈ൘UᖒⲴᓅ䜘४ฏ˄㓖1970-1980˅ˈ↻ᐞѪ䍏ф䖳བྷˈ㺘᰾⁑ර儈ՠҶᰙᵏⲴл䱽઼վՠҶᰙᵏⲴ
ᒣっDŽ㘼൘ਾᵏ࣐䙏໎䮯䱦⇥ˈ↻ᐞ㲭❦ᴹ⌒ࣘˈնᤏਸ㓯䐏кҶᙫփ䎻࣯DŽ appropriate. This includes:
•䘉ᨀ⽪ˈ㲭❦Ҽ⅑䎻࣯∄㓯ᙗ䎻࣯ᴤ㜭ᦅ᥹ਾᵏⲴ࣐䙏໎䮯ˈնᆳሩᮤњᒿࡇⲴᤏਸ˄⢩࡛ᱟᰙᵏ˅ਟ㜭ӽᴹተ
Figure 5.6 Volume on the New York Stock Exchange 䲀DŽањᆼᮤⲴUᖒਟ㜭нᱟሩ䘉њᮠᦞⲴᴰ֣᧿䘠ˈᮠᦞᴤ‫ۿ‬ᱟUᖒⲴਣॺ䗩˄࣐䙏кॷ䜘࠶˅DŽ

Figure 5.8 Quadratic Trend: Volume on the New York Stock Exchange (b) Exponential trend.
40,000
40,000
¾ Suitable if trend is characterized by constant
growth at rate ȕ1.ྲ᷌䎻࣯Ⲵ⢩⛩ᱟԕᚂᇊ໎䮯⦷ ɴІ໎䮯ˈࡉ䘲⭘DŽ
30,000 30,000

20,000 20,000
¾ It is defined by: 1
20,000

Volume
Tt E 0e E t
10,000
10,000
10,000 0 ¾ Also known as log-linear trend:
0
0 ln(Tt ) ln(E 0 e E1t ) ln(E 0 )  ln(e E1t )
60 65 70 75 80 85 90 95 00 05
-10,000
Time 60 65 70 75 80 85 90 95 00 05
ln(Tt ) ln(E 0 )  E1t
Residual Actual Fitted
Example 3a: modeling trend. Example 3b: modeling trend. Example 3c: modeling trend.
ȕ0 > 0 and ȕ1 > 0; ȕ0 > 0 and ȕ1 < 0; ȕ0 < 0 and ȕ1 > 0;
6
40 0

35 -5
5

30
Trend t 5e 0.02t Trend t 5e 0.02t Trend t 5e 0.02t
-10
4
25 -15

3
20 -20

15 -25
2

10 -30
1
5
-35

0
0
1 11 21 31 41 51 61 71 81 91 -40
1 11 21 31 41 51 61 71 81 91
1 11 21 31 41 51 61 71 81 91
Trend
Trend
Trend

↻ᐞമ˄ᇎ㓯˅ᱮ⽪ˈ㲭❦ӽᴹ⌒ࣘˈն㌫㔏ᙗⲴ‫ٿ‬ᐞլѾ∄Ҽ⅑䎻࣯⁑ර㾱ሿаӋDŽֻྲˈᰙᵏ⋑ᴹࠪ⧠Ҽ⅑
䎻࣯⁑රѝ䛓⿽Uᖒᓅ䜘Ⲵᤏਸ䰞仈DŽн䗷ˈ൘1987ᒤ㛑ᐲፙⴈ˄"唁㢢ᱏᵏа"˅䱴䘁ˈ↻ᐞࠪ⧠Ҷ䖳བྷⲴ⌒ࣘˈ
Example 3e: appropriate to be used on… 䘉㺘᰾⁑රᵚ㜭ᦅ᥹䘉⿽ケਁһԦDŽ
Example 3d: modeling trend. Figure 5.12 Exponential Trend: Volume on the New York Stock Exchange
First approach:
ȕ0 < 0 and ȕ1 < 0; Reconsider the volume on the New York Stock Exchange (Figure 40,000
5.6-Slide 16).
0
Figure 5.6 Volume on the New York Stock Exchange 30,000
20,000
-1 15,000 20,000
40,000
10,000 10,000
-2
30,000 5,000
Trend t 5e 0.02t 0
-3 0
20,000

Volume
-5,000
-4
-10,000
10,000 60 65 70 75 80 85 90 95 00 05
-5

Residual Actual Fitted

-6 0
1 11 21 31 41 51 61 71 81 91 60 65 70 75 80 85 90 95 00 05
It seems that the exponential trend is much better than the
Trend
quadratic trend (Slide 17) in modeling the volume on the
Time
New York Stock Exchange.ⴻ䎧ᶕ൘Ѫ㓭㓖䇱ࡨӔ᱃ᡰӔ᱃䟿ᔪ⁑ᯩ䶒ˈᤷᮠ䎻
࣯䘌ՈҾҼ⅑䎻࣯
 ↻ᐞമ˄ᇎ㓯˅ത㔅0㓯кл⌒ࣘˈ㲭❦нᱟᆼ‫ޘ‬䲿ᵪ˄ӽ❦ਟԕⴻࠪаӋ⁑ᔿˈֻྲ80ᒤԓᵛࡠ
Example 3f: appropriate to be used on… 90ᒤԓࡍⲴ⌒ࣘ䖳བྷ˅ˈն⴨∄Ҿⴤ᧕ሩ৏࿻ᮠᦞᤏਸ㓯ᙗᡆҼ⅑䎻࣯ˈ䘉䟼Ⲵ↻ᐞ⁑ᔿ㾱ྭаӋDŽ
Second approach: Figure 5.11 Linear Trend: Log Volume on the New York Stock Exchange 2. Estimating trend models
The following plot shows the logarithm of volume on the New 12
York Stock Exchange. • We fit our various trend models to data on a time
Figure 5.9 Log Volume on the New York Stock Exchange 10 series y using least-squares regression. That
11 ᡁԜ֯⭘ᴰሿҼ҈എᖂ⌅ሶ਴⿽
8
is, we use a computer to find 䎻࣯⁑රᤏਸࡠᰦ䰤ᒿࡇ y Ⲵᮠ
T 2 ᦞкDŽҏቡᱟ䈤ˈᡁԜ֯⭘䇑㇇
10
ᵪᶕ᢮ࡠᴰሿҼ҈⌅ቡᱟ㾱᢮ࡠ
2.0 6 T arg min ¦ yt  Tt (T ) , а㓴৲ᮠ șƄˈ֯ᗇᡰᴹ㿲⍻⛩Ⲵᇎ
9 T 䱵٬о⁑රᤏਸ٬ѻ䰤Ⲵᐞᔲ
1.5 t 1
4 ˄↻ᐞ˅Ⲵᒣᯩ઼ᴰሿDŽ
8 1.0
where ș denotes the set of parameters to be
0.5
7

Log Volume
0.0
estimated.
6
-0.5 • A linear trend, for example, has Tt(ș)=ȕ0+ȕ1t,
5 -1.0
60 65 70 75 80 85 90 95 00 05
and ș=(ȕ0,ȕ1), in which case the computer finds
4 T 2 ሩҾ㓯ᙗ䎻࣯ˈᡁԜ㾱᢮Ⲵ
60 65 70 75 80 85 90 95 00 05 ቡᱟ㜭֯ (yᴽ - (ȕ΋ +
Residual Actual Fitted ( E 0 , E1 ) arg min ¦ yt  E 0  E1t ȕΌt)) Ⲵᒣᯩ઼ᴰሿ
Time E 0 , E1 Ⲵ ȕ΋ ઼ ȕΌ Ⲵՠ䇑
t 1 ٬ ȕƄ΋ ઼ ȕƄΌDŽ

3. Selecting Forecasting Models Using

• Similarly, 4. Forecasting Trend
¾ Quadratic trend:
the Akaike and Schwarz Criteria. • Suppose we’re presently at time T, and we want
T 2 • Two very important selection criteria are the to use a trend model to forecast the h-step-
( E 0 , E1 , E 2 ) arg min ¦ yt  E 0  E1t  E 2t 2 Akaike information criterion (AIC) and the Schwarz ahead value of a series y. Therefore, point
E 0 , E1 , E 2 t 1 information criterion (SIC): forecasts; ‫ٷ‬䇮ᡁԜᖃࡽ༴Ҿᰦ䰤 TˈᒦфᡁԜᜣ֯⭘ањ䎻࣯⁑රᶕ亴⍻ᒿ
ࡇ y ᵚᶕ h ↕Ⲵ٬DŽഐ↔ˈ⛩亴⍻Ѫ˖
T T
¾ Exponential trend: 2 2
§ 2k ·¦e t §k· ¦e t ¾ Linear trend model: yt E 0  E1t  H t
2 ¨ ¸ ¨ ¸
T ¹ t 1 t 1
E1t AIC exp © T , and SIC T © T ¹ , Forecast equation: yT  h|T E 0  E1 T  h
( E 0 , E1 ) arg min ¦ yt  E 0 e T T
E 0 , E1 t 1
or where k = number of parameters to be estimated, * under assumption that H ~ N (0, V 2 ).
T 2 t Ht
T = sample size, et = regression residuals.
( E 0 , E1 ) arg min ¦ ln yt  ln E 0  E1t
E 0 , E1 t 1
=> Select model that gives the smallest values of 䘉ᱟањ‫ޣ‬Ҿ⁑ර䈟ᐞ亩 İᴽ Ⲵᑨ㿱‫ٷ‬䇮DŽ
•İᴽ ~ N(0, ı²_İ): 㺘⽪䈟ᐞ亩 İᴽ ᴽӾ൷٬Ѫ0ˈᯩᐞѪ ı²_İ Ⲵ↓ᘱ࠶ᐳDŽᒦф䙊ᑨ䘈‫ٷ‬
AIC and SIC. 䇮䈟ᐞ亩ᱟ⤜・਼࠶ᐳⲴ˄iid˅DŽ䘉њ‫ٷ‬䇮ሩҾᶴᔪ亴⍻४䰤ᱟ䟽㾱ⲴDŽ
 • Similarly,
¾ Quadratic trend: 5. Application: Forecasting Retail Sales
200,000
•ՠ䇑Ⲵ䎻࣯ᯩ〻Ѫ RTRRᴽ = -16391.25 +
2 349.7731 * tDŽ 150,000
yT  h|T E 0  E1 T  h  E 2 T  h • Linear trend regression. •㌫ᮠ C ઼ T 䜭儈ᓖᱮ㪇ˈR² ҏᖸ儈 (㓖0.90)ˈ㺘᰾⁑ර
൘ṧᵜ޵ᤏਸҶབྷ䜘࠶ਈᔲDŽ 100,000
•❦㘼ˈDurbin-Watson㔏䇑䟿ᶱվ (0.004682)ˈ䘉㺘᰾⁑
¾ Exponential trend: ර↻ᐞᆈ൘䶎ᑨѕ䟽Ⲵᒿࡇ⴨‫ޣ‬ᙗ䰞仈ˈ䈤᰾ㆰঅⲴ㓯 50,000
ᙗ䎻࣯⁑රਟ㜭н‫࠶ݵ‬ᡆн↓⺞DŽ
•䇠ᖅлAIC (22.19) ઼ SIC (22.20) Ⲵ٬ˈ⭘Ҿਾ㔝⁑ර∄ 40,000 0
š š 䖳DŽ
1 T h
-50,000
yT  h|T E 0e E or, ln y ln E 0  E1 T  h
T  h|T 20,000

0
• 95% forecast intervals;
-20,000
yT  h|T r 1.96V h
-40,000
1955 1960 1965 1970 1975 1980 1985 1990
where V h is an estimate of the standard deviation of
Residual Actual Fitted
the h-step forecast distribution.ᱟ h ↕ੁࡽ亴⍻䈟ᐞⲴḷ߶ᐞⲴՠ •↻ᐞ˄㬍㢢˅੸⧠ࠪ䶎ᑨ␵ᲠⲴ⁑ᔿ˖ᰙᵏѪ↓ˈѝᵏѪ䍏˄ᇎ䱵٬վҾ䎻࣯㓯˅ˈਾᵏ৸ਈ
䇑٬DŽ Ѫ↓ф䎺ᶕ䎺བྷ˄ᇎ䱵٬䘌儈Ҿ䎻࣯㓯˅DŽ䘉⿽ĀᗞㅁāᡆĀUᖒāⲴ↻ᐞ⁑ᔿᱟ㓯ᙗ䎻࣯⁑ර
H t ~ N (0, V H2 ). н䘲ਸ᧿䘠䶎㓯ᙗᮠᦞⲴިර㺘⧠DŽ
t •䘉ока亥Durbin-Watson㔏䇑䟿ᶱվᡰᤷ⽪Ⲵѕ䟽↓㠚⴨‫ޣ‬ᱟа㠤ⲴDŽᖃ↻ᐞ੸⧠䘉⿽㌫㔏ᙗ
* under assumption that ⁑ᔿᰦˈᆳԜ䙊ᑨᱟ㠚⴨‫Ⲵޣ‬DŽ

• Quadratic trend regression. • Exponential trend regression.

200,000

160,000

120,000
12,000
80,000
8,000
40,000
4,000
0
0

-4,000

-8,000
1955 1960 1965 1970 1975 1980 1985 1990

Residual Actual Fitted

•ՠ䇑Ⲵ䎻࣯ᯩ〻Ѫ RTRRᴽ = 18708.70 - 98.31130*t + 0.955404*t²DŽ •ՠ䇑Ⲵ䎻࣯ᯩ〻Ѫ RTRRᴽ = 12769.08 * e^(0.005783*t)DŽ

•ᡰᴹ㌫ᮠ䜭儈ᓖᱮ㪇DŽȕƄΌ < 0 ф ȕƄ΍ > 0 㺘᰾䘉ᱟањUᖒ䎻࣯DŽ •նᱟˈ↻ᐞӽ❦нᱟᆼ‫ޘ‬䲿ᵪⲴDŽᆳԜլѾӽ❦ᆈ൘аӋ⁑ •৲ᮠ䜭儈ᓖᱮ㪇DŽ
•R² བྷབྷᨀ儈 (0.997)ˈഎᖂḷ߶䈟བྷབྷ䱽վˈAIC (18.67) ઼ SIC (18.69) ҏ䘌վҾ㓯ᙗ䎻࣯⁑රˈ䘉Ӌ䜭㺘᰾Ҽ⅑䎻࣯
ᔿˈֻྲ൘1970ᒤԓѝᵏ઼1980ᒤԓࡍᵏˈ↻ᐞ‫ੁٮ‬ҾѪ䍏ˈ •R² (0.989) 儈Ҿ㓯ᙗ䎻࣯նվҾҼ⅑䎻࣯DŽAIC (19.94) ઼ SIC (19.96) ҏӻҾє㘵ѻ䰤ˈնᴤ᧕䘁Ҽ⅑䎻࣯DŽ
⁑ර∄㓯ᙗ䎻࣯⁑රᤏਸᗇᴤྭDŽ
㘼൘1970ᒤԓᵛᵏ઼1980ᒤԓᵛᵏˈ↻ᐞ‫ੁٮ‬ҾѪ↓DŽ䘉⿽⁑
ᔿ䀓䟺ҶѪӰѸD-W㔏䇑䟿ӽ❦ᖸվDŽ •Durbin-Watson㔏䇑䟿 (0.042) ӽ❦ᖸվˈ㺘᰾↻ᐞѝӽᴹѕ䟽Ⲵᒿࡇ⴨‫ޣ‬ᙗDŽ
•❦㘼ˈDurbin-Watson㔏䇑䟿 (0.151) ӽ❦ᖸվˈ㺘᰾↻ᐞѝӽᴹѕ䟽Ⲵᒿࡇ⴨‫ޣ‬ᙗDŽ •ӾAIC઼SICᶕⴻˈҼ⅑䎻࣯ (AIC=18.67, SIC=18.69) լѾՈҾᤷᮠ䎻࣯ (AIC=19.94, SIC=19.96)DŽ
 • Log-linear trend regression.
200,000 12.5
12.0
150,000 11.5
11.0
100,000
15,000 10.5

10,000 10.0
50,000 .3
9.5
5,000
.2 9.0
0
0
.1
-5,000
.0
-10,000

-15,000 -.1
1955 1960 1965 1970 1975 1980 1985 1990
-.2
Residual Actual Fitted 1955 1960 1965 1970 1975 1980 1985 1990

Residual Actual Fitted

•↻ᐞ˄㬍㢢˅Ⲵ⌒ࣘ㤳ത∄㓯ᙗ䎻࣯ሿˈ •ՠ䇑Ⲵሩᮠ㓯ᙗ䎻࣯ᯩ〻Ѫ LOG(RTRRᴽ) = 9.389975 + 0.005931*tDŽ
նլѾ∄Ҽ⅑䎻࣯Ⲵ↻ᐞ⮕བྷаӋˈᒦфӽ •㌫ᮠ䜭儈ᓖᱮ㪇DŽ
䙊䗷മᖒኅ⽪Ҷሩᮠ㓯ᙗ⁑රሩ䴦୞仍ᮠᦞⲴᤏਸᛵߥDŽ㲭❦ሩᮠ䴦୞仍
❦ᱮ⽪ࠪаӋ⁑ᔿ˄ֻྲˈ70ᒤԓѝᵏ઼80 •ሩᮠቪᓖкⲴR² ᖸ儈 (0.987)DŽ
ᒤԓࡍⲴ䍏↻ᐞˈ70ᒤԓᵛ઼80ᒤԓᵛⲴ↓ •Durbin-Watson㔏䇑䟿 (0.0199) ӽ❦ᖸվˈ㺘᰾ሩᮠ䖜ᦒਾⲴ↻ᐞѝӽᴹѕ䟽Ⲵᒿࡇ⴨‫ޣ‬ᙗDŽ ᮠᦞⴻ䎧ᶕᖸ㓯ᙗˈᒦф㓯ᙗ䎻࣯㓯ᤏਸᗇᖸྭˈն↻ᐞѝӽ❦ᆈ൘㌫㔏
↻ᐞ˅DŽ •䘉䟼ⲴAIC (-1.93) ઼ SIC (-1.91) н㜭ⴤ᧕оѻࡽ⁑රⲴAIC/SIC∄䖳ˈഐѪഐਈ䟿н਼DŽ ᙗ⁑ᔿˈ㺘᰾⁑රᵚ㜭ᦅ᥹ᡰᴹᰦ䰤‫׍‬䎆ᙗDŽ

Evaluating different types of models. Quadratic trend model, actual values and forecasts. Linear trend model, actual values and forecasts.
Figure 5.22 Retail Sales: History, 1990.01 - 1993.12 Linear Trend Forecast and Realization, 1994.01 - 1994.12
Figure 5.20 Retail Sales: History, 1990.01 - 1993.12 Quadratic Trend Forecast and Realization, 1994.01 - 1994.12
200,000
• Log-linear trend regression is excluded from 200,000

comparison since the dependent variable in logs, not 180,000

190,000
in levels. ሩᮠ㓯ᙗ䎻࣯എᖂ㻛ᧂ䲔൘∄䖳ѻཆˈഐѪަഐਈ䟿ᱟሩᮠ٬ˈ㘼нᱟ৏࿻٬DŽ
160,000
Linear Trend Quadratic Exponential 180,000

Trend Trend 140,000

170,000
AIC 22.19 18.67 19.94
120,000
SIC 22.20 18.69 19.96 160,000
History, Forecast and Realization

History, Forecast and Realization

100,000
90:01 91:01 92:01 93:01 94:01
• Quadratic trend model produces the smallest AIC and 150,000
90:01 91:01 92:01 93:01 94:01 Time
SIC. We select this model for forecasting purpose.Ҽ⅑䎻
࣯⁑රӗ⭏ҶᴰሿⲴAIC઼SICDŽᡁԜ䘹ᤙ䘉њ⁑ර⭘Ҿ亴⍻ⴞⲴDŽ Time If we use linear trend model for forecasting, the actual values are
Actual values fluctuate within the forecast intervals.ᇎ䱵٬൘亴 outside the forecast intervals. It proves that the quadratic trend
⍻४䰤޵⌒ࣘDŽ model is a better choice for forecasting.ྲ᷌ᡁԜ֯⭘㓯ᙗ䎻࣯⁑ර䘋㹼亴
⍻ˈᇎ䱵٬Պ㩭൘亴⍻४䰤ѻཆDŽ䘉䇱᰾ҶҼ⅑䎻࣯⁑රᱟᴤྭⲴ亴⍻䘹ᤙDŽ
 1. The Nature and Sources of • Seasonality arises from links of technologies,
Seasonality preferences, and institutions to the calendar.ᆓ㢲ᙗ
ⓀҾᢰᵟǃ‫ࡦ઼ྭٿ‬ᓖоᰕশⲴ㚄㌫DŽ
AA037007/EQC7006 • The weather is a trivial but very important
• A seasonal pattern is one that repeats itself every
Time Series Analysis year.ᆓ㢲ᙗ⁑ᔿᱟᤷ⇿ᒤ䜭Պ䟽༽㠚䓛Ⲵ⁑ᔿDŽ seasonal series, as it’s always hotter in the
summer than in the winter. Any technology that
• The annual repetition can be exact, in which case involves the weather, such as production of
Lecture 3b we speak of deterministic seasonality, or agricultural commodities, is likely to be seasonal
approximate, in which case we speak of as well.ཙ≄ᱟањн䀰㠚᰾ն䶎ᑨ䟽㾱Ⲵᆓ㢲ᙗᒿࡇˈഐѪ༿ཙᙫᱟ∄ߜཙ✝DŽ
ԫօ⎹৺ཙ≄Ⲵᢰᵟˈֻྲߌӗ૱Ⲵ⭏ӗˈҏᖸਟ㜭ᱟᆓ㢲ᙗⲴDŽ
stochastic seasonality.⇿ᒤⲴ䟽༽ਟԕᱟ㋮⺞Ⲵˈ䘉⿽ᛵߥлᡁԜ〠
Modeling Seasonality ѻѪ⺞ᇊᙗᆓ㢲ᙗ˗ᡆ㘵ਟԕᱟ䘁լⲴˈ䘉⿽ᛵߥлᡁԜ〠ѻѪ䲿ᵪᙗᆓ㢲ᙗDŽ • Preferences may also be linked to the calendar.
(deterministic/regression method) • In this topic, we focus on deterministic For example, gasoline sales in Figure 6.1.‫ྭٿ‬ҏਟ㜭оᰕ
শ⴨‫ޣ‬
seasonality.

Figure 6.1 Gasoline Sales

14,000 • Finally, social institutions that are linked to the
13,000 calendar, such as holidays, are responsible for
• The figure shows monthly U.S. current-dollar
12,000
seasonal variation in a variety of series. Purchase
gasoline sales, Jan 1980 - Jan 1992.
of retail goods skyrocket, for example, every
11,000
• It explains that people want to do more vacation Christmas season in some western countries.ᴰਾˈ
10,000 оᰕশ⴨‫⽮Ⲵޣ‬Պࡦᓖˈֻྲ㢲‫ٷ‬ᰕˈᱟ਴⿽ᒿࡇᆓ㢲ᙗਈॆⲴ৏ഐDŽֻྲˈ൘аӋ
travel in the summer, which tends to increase 㾯ᯩഭᇦˈ⇿ࡠ൓䈎㢲ᵏ䰤ˈ䴦୞୶૱Ⲵ䍝Ҡ䟿䜭Պ⥋໎DŽ
9,000

Gasoline Sales
both price and quantity of summertime gasoline • For example, monthly U.S. current-dollar liquor
8,000 sales, both of which feed into higher current- sales, Jan 1980 - Jan 1992, which are very high in
7,000 dollar sales.ᆳ䀓䟺ҶӪԜᐼᵋ൘༿ᆓ䘋㹼ᴤཊⲴ‫ٷ‬ᵏ᯵㹼ˈ䘉ᖰᖰՊ໎࣐༿ November and December.
ᆓ⊭⋩䬰୞ⲴԧṬ઼ᮠ䟿ˈ䘉є㘵䜭Պሬ㠤ᴤ儈Ⲵ⧠٬㖾‫ݳ‬䬰୞仍DŽ
6,000
80 81 82 83 84 85 86 87 88 89 90 91
䈕മ␵Რൠᱮ⽪Ҷ⇿ᒤ䟽༽Ⲵᆓ㢲ᙗ⁑ᔿ˄༿ᆓ Time
儈ˈߜᆓվ˅ˈ਼ᰦҏਟ㜭ᆈ൘䮯ᵏ䎻࣯DŽ
 Figure 6.2 Liquor Sales Figure 6.3 Durable Goods Sales

2,800 70,000

2,600 • Another example, sales of durable goods, which 60,000

2,400 fall in December, as holiday purchase tend to be
2,200 nondurables.ਖањֻᆀᱟ㙀⭘૱Ⲵ䬰୞仍ˈᆳ൘12ᴸԭл䱽ˈഐѪ‫ٷ‬ᰕ䍝 50,000
Ҡᖰᖰᱟ䶎㙀⭘૱DŽ
2,000
• This is given in Figure 6.3, in which show monthly 40,000

Liquor Sales
1,800
U.S. current-dollar durable goods sales, Jan 1980

Durable Goods Sales

1,600 30,000
- Jan 1992.
1,400
20,000
1,200
80 81 82 83 84 85 86 87 88 89 90 91
80 81 82 83 84 85 86 87 88 89 90 91
Time
Time

• Let’s construct seasonal dummy variables, which

2. Modeling Seasonality indicate which season we’re in. If, for example, there
• One way to deal with seasonality in a series is simply are four seasons, we create 䇙ᡁԜᶴᔪᆓ㢲ᙗ㲊ᤏਈ䟿ˈᆳԜᤷ⽪ᡁ
to remove it and then to model and forecast the Ԝᖃࡽ༴Ҿଚњᆓ㢲DŽֻྲˈྲ᷌ᴹഋњᆓ㢲˄ণs=4ˈྲᆓᓖᮠᦞ˅ˈᡁԜࡋᔪ
• A key technique for modeling seasonality is
seasonally adjusted time series. This strategy is regression on seasonal dummies.ᔪ⁑ᆓ㢲ᙗⲴањ‫ޣ‬䭞ᢰᵟᱟሩ D1 = (1,0,0,0,1,0,0,0,1,0,0,0,…);
appropriate in certain situations, such as when ᆓ㢲ᙗ㲊ᤏਈ䟿䘋㹼എᖂDŽ
D2 = (0,1,0,0,0,1,0,0,0,1,0,0,…);
interest centers explicitly on forecasting non-seasonal • Let s be the number of seasons in a year. Normally
D3 = (0,0,1,0,0,0,1,0,0,0,1,0,…);
fluctuations.༴⨶ᒿࡇѝᆓ㢲ᙗⲴа⿽ᯩ⌅ᱟㆰঅൠሶަ〫䲔ˈ❦ਾሩᆓ㢲ᙗ䈳ᮤਾ we’d think of four seasons in a year; but that
Ⲵᰦ䰤ᒿࡇ䘋㹼ᔪ⁑઼亴⍻DŽ䘉⿽ㆆ⮕൘ḀӋᛵߥлᱟਸ䘲Ⲵˈֻྲᖃ‫⛩⌘ޣ‬᰾⺞൘Ҿ亴⍻ notation is too restrictive for our purpose.‫ٷ‬䇮 s ᱟаᒤѝⲴ D4 = (0,0,0,1,0,0,0,1,0,0,0,1,…);
䶎ᆓ㢲ᙗ⌒ࣘᰦDŽ
ᆓ㢲ᮠDŽ䙊ᑨᡁԜՊ䇔Ѫаᒤᴹഋњᆓ㢲˗ն䘉⿽ḷ䇠ᯩᔿሩᡁԜⲴⴞⲴᶕ䈤䗷Ҿተ䲀DŽ
• Seasonal adjustment is often inappropriate in • Instead, think of s as the number of observations on
business forecasting situations, precisely because • D1 indicates whether we in the first quarter (it’s 1 in
a series in each year. Thus, s=4 if we have quarterly the first quarter and 0 otherwise), D2 indicates
interest typically centers on forecasting all the data, s=12 if we have monthly data, s=52 if we have
variation in a series, not just the non-seasonal part.ᆓ whether we’re in the second quarter (it’s 1 in the
㢲ᙗ䈳ᮤ൘୶ъ亴⍻ᛵຳѝ䙊ᑨнਸ䘲ˈ↓ᱟഐѪ‫⛩⌘ޣ‬䙊ᑨ䳶ѝ൘亴⍻ᒿࡇѝⲴᡰᴹਈࣘˈ
weekly data, and so forth.⴨৽ˈሶ s 㿶Ѫ⇿ᒤᒿࡇⲴ㿲⍻⅑ᮠDŽഐ↔ˈ second quarter and 0 otherwise), and so on. Іᤷ⽪ᡁԜᱟ
㘼нӵӵᱟ䶎ᆓ㢲ᙗ䜘࠶DŽ ྲ᷌ᡁԜᴹᆓᓖᮠᦞˈs=4˗ྲ᷌ᡁԜᴹᴸᓖᮠᦞˈs=12˗ྲ᷌ᡁԜᴹઘᓖᮠᦞˈs=52˗ԕ
↔㊫᧘DŽ ੖༴Ҿㅜаᆓᓖ˄൘ㅜаᆓᓖѪ1ˈ੖ࡉѪ0˅ˈЇᤷ⽪ᡁԜᱟ੖༴ҾㅜҼᆓᓖ˄൘ㅜҼᆓ
ᓖѪ1ˈ੖ࡉѪ0˅ˈ‫↔׍‬㊫᧘DŽ
 䘉ᱟᶴᔪᆓ㢲ᙗ㲊ᤏਈ䟿⁑රⲴਖа⿽˄ҏᱟᴤᑨ㿱Ⲵ˅ᯩ⌅ˈԕ䚯‫ݽ‬Ā㲊ᤏਈ䟿䲧䱡ā˄ᆼ‫ޘ‬ཊ䟽‫ޡ‬㓯ᙗ˅DŽ
• ᯩ⌅: 䘹ᤙањᆓ㢲֌ѪĀส߶ᆓ㢲ā˄reference season˅ˈнѪᆳ䇮㖞㲊ᤏਈ䟿DŽ❦ਾѪަ։Ⲵ s-
1 њᆓ㢲਴䇮㖞ањ㲊ᤏਈ䟿DŽ⁑රѝवਜ਼ањᙫⲴᡚ䐍亩˄ᑨᮠ亩˅DŽ
• The pure dummy model is•yyᴽ: ᰦ䰤ᒿࡇ൘ᰦ䰤⛩ t Ⲵ㿲⍻٬DŽ • Instead of including a full set of s seasonal dummies, • Trend may be included as well. Trend could be in
•D૾ᴽ: ㅜ i њᆓ㢲ᙗ㲊ᤏਈ䟿൘ᰦ䰤⛩ t Ⲵ٬˄0ᡆ1˅DŽ
௦ •Ȗ૾ (gamma-i): ㅜ i њᆓ㢲Ⲵᆓ㢲ᙗഐᆀᡆᆓ㢲ᙗ᭸ᓄDŽ we can include any (s-1) seasonal dummies and an
ᆳԓ㺘Ҷㅜ i њᆓ㢲Ⲵᒣ൷≤ᒣ˄ྲ᷌⁑ර⋑ᴹᡚ
form of linear, quadratic, exponential and etc. As
䐍˅DŽ intercept. Then the constant term is the intercept for
‫ݕ‬௧ = ෍ ߛ௜ ‫ܦ‬௜௧ + ߝ௧ example, model with linear trend is represented
௜ୀଵ the omitted season, and the coefficients on the as:䎻࣯ҏਟԕ㻛वਜ਼䘋ᶕDŽ䎻࣯ਟԕᱟ㓯ᙗǃҼ⅑ǃᤷᮠㅹᖒᔿDŽֻྲˈवਜ਼㓯ᙗ
seasonal dummies give the seasonal increase or 䎻࣯Ⲵ⁑ර㺘⽪Ѫ˖
• Effectively, we’re just regressing on an intercept, but ௦
decrease relative to the omitted season.䲔Ҷवਜ਼аᮤ྇ s њ ‫ݕ‬௧ = ߚଵ ‫ ݐ‬+ σ௜ୀଵ ߛ௜ ‫ܦ‬௜௧ + ߝ௧
we allow for a different intercept in each season, ߛ௜ᇱ s ᆓ㢲ᙗ㲊ᤏਈ䟿ཆˈᡁԜҏਟԕवਜ਼ԫ᜿ (s-1) њᆓ㢲ᙗ㲊ᤏਈ䟿઼ањᡚ䐍亩DŽ䛓Ѹˈᑨ
which called the seasonal factors. ᇎ䱵кˈᡁԜਚᱟ൘ሩањᡚ ᮠ亩ቡᱟ㻛ⴱ⮕Ⲵ䛓њᆓ㢲Ⲵᡚ䐍ˈ㘼ᆓ㢲ᙗ㲊ᤏਈ䟿Ⲵ㌫ᮠࡉ㔉ࠪҶ⴨ሩҾ㻛ⴱ⮕ᆓ㢲Ⲵ
ᆓ㢲ᙗ໎࣐ᡆ߿ቁ䟿DŽ
䐍䘋㹼എᖂˈնᡁԜ‫ݱ‬䇨⇿њᆓ㢲ᴹн਼Ⲵᡚ䐍ˈ䘉Ӌ ɶࣩ㻛〠Ѫᆓ㢲ᙗഐᆀDŽ
• However, we should not include s seasonal dummies • Here we want our model to account for trend, if
• In the absence of seasonality, the ߛ௜ᇱ s are all the it’s present, but we want to expand the model so
and an intercept. Inclusion of an intercept and a full
same, so we can drop all the seasonal dummies and that we can account for seasonality as well.൘䘉䟼ˈ
set of a seasonal dummies produces perfect ྲ᷌ᆈ൘䎻࣯ˈᡁԜᐼᵋᡁԜⲴ⁑ර㜭䀓䟺䎻࣯ˈնᡁԜҏᜣᢙኅ⁑රˈ֯ަҏ㜭䀓
instead simply include an intercept in the usual way.൘ 䟺ᆓ㢲ᙗDŽ
⋑ᴹᆓ㢲ᙗⲴᛵߥлˈᡰᴹⲴ ɶࣩ䜭⴨਼ˈᡰԕᡁԜਟԕ৫ᦹᡰᴹⲴᆓ㢲ᙗ㲊ᤏਈ䟿ˈ multicollinearity.❦㘼ˈᡁԜнᓄ䈕਼ᰦवਜ਼ s њᆓ㢲ᙗ㲊ᤏਈ䟿઼ањᡚ䐍亩DŽ
㘼ਚᱟ‫ۿ‬䙊ᑨ䛓ṧवਜ਼ањᡚ䐍亩DŽ वਜ਼ањᡚ䐍亩઼аᮤ྇ᆓ㢲ᙗ㲊ᤏਈ䟿Պሬ㠤ᆼ‫ޘ‬ཊ䟽‫ޡ‬㓯ᙗDŽ

• The behaviour of many series, such as sales, • Trading variation refers to the fact that different
• The idea of seasonality may be extended to allow for shipments, inventories, and so on depends in
more general calendar effects. Standard seasonality months contains different numbers of trading days
part on timing of such holiday. Thus, we may or business days, which is an important
is just one type of calendar effect. want to keep track of them in our forecasting
ᆓ㢲ᙗⲴᾲᘥਟԕᢙኅࡠवਜ਼ᴤᒯ⌋Ⲵᰕশ᭸ᓄDŽḷ߶ᆓ㢲ᙗਚᱟᰕশ᭸ᓄⲴа⿽DŽ consideration when modeling and forecasting
models.䇨ཊᒿࡇⲴ㹼Ѫˈྲ䬰୞仍ǃਁ䍗䟿ǃᓃᆈㅹˈ䜘࠶ਆߣҾ䘉Ӌ㢲‫ٷ‬ᰕ
• Two additional calendar effects are holiday variation Ⲵᰦ⛩DŽഐ↔ˈᡁԜਟ㜭ᐼᵋ൘ᡁԜⲴ亴⍻⁑රѝ䘭䑚ᆳԜDŽ certain series.Ӕ᱃ᰕਈࣘᤷⲴᱟн਼ᴸԭवਜ਼н਼ᮠ䟿ⲴӔ᱃ᰕᡆᐕ֌ᰕˈ䘉
൘ሩḀӋᒿࡇ䘋㹼ᔪ⁑઼亴⍻ᰦᱟањ䟽㾱Ⲵ㘳㲁ഐ㍐DŽ
and trading-day variation. ਖཆє⿽ᰕশ᭸ᓄᱟ‫ٷ‬ᰕਈ઼ࣘӔ᱃ᰕਈࣘDŽ • Holiday effects may be handled with dummy
• Holiday variation refers to the fact that some variables. In a monthly model, for example, in • For example, in a monthly forecasting model of
holiday’s dates change over time. Ester as an addition to a full set of seasonal dummies, we volume traded on the London Stock Exchange, in
example, arrives at approximately the same time might include an “Ester dummy”, which is 1 if the addition to a full set of seasonal dummies, we
each year, but the exact dates differ.‫ٷ‬ᰕਈࣘᤷⲴᱟаӋ㢲ᰕⲴᰕ month contains Ester and 0 otherwise. ‫ٷ‬ᰕ᭸ᓄਟԕ⭘ might include a trading-day variable, whose value
ᵏՊ䲿ᰦ䰤ਈॆDŽֻྲ༽⍫㢲ˈᆳ⇿ᒤབྷ㠤൘਼аᰦ䰤ࡠᶕˈնާփᰕᵏн਼DŽ 㲊ᤏਈ䟿ᶕ༴⨶DŽֻྲˈ൘ањᴸᓖ⁑රѝˈ䲔Ҷаᮤ྇ᆓ㢲ᙗ㲊ᤏਈ䟿ཆˈᡁԜਟ
㜭䘈Պवਜ਼ањĀ༽⍫㢲㲊ᤏਈ䟿āˈྲ᷌Ḁњᴸԭवਜ਼༽⍫㢲ˈࡉ䈕ਈ䟿Ѫ1ˈ੖ࡉ each month is the number of trading days that
Ѫ0DŽ
month.ֻྲˈ൘ањՖᮖ䇱ࡨӔ᱃ᡰӔ᱃䟿Ⲵᴸᓖ亴⍻⁑රѝˈ䲔Ҷаᮤ྇ᆓ㢲ᙗ
㲊ᤏਈ䟿ཆˈᡁԜਟ㜭䘈Պवਜ਼ањӔ᱃ᰕਈ䟿ˈަ⇿њᴸⲴ٬ᱟ䈕ᴸⲴӔ᱃ᰕᮠ䟿DŽ
• Inclusion of holiday and trading-day variation gives • Point forecast:⛩亴⍻нवਜ਼ᵚᶕⲴ䈟ᐞ亩˄‫ٷ‬ᇊަᵏᵋѪ0˅DŽ
the complete modelवਜ਼‫ٷ‬ᰕ઼Ӕ᱃ᰕਈࣘ㔉ࠪҶᆼᮤⲴ⁑ර 3. Forecasting Seasonal Series
s v1

• The full model is yT  h|T E1 (T  h)  ¦ J i Di ,T  h  ¦ G iHD HDVi ,T  h

i 1 i 1
s v1 v2 s v1 v2
HD TD HD TD v2
yt E1t  ¦ J i Dit  ¦ G i i
HDVit  ¦ G TDVit  H t yt E1t  ¦ J i Dit  ¦ G i i
HDVit  ¦ G TDVit  H t
i 1 i 1 i 1 i 1 i 1 i 1  ¦ G iTDTDVi ,T  h
i 1
where the HDVs are the relevant holiday variables,
and the TDVs are the relevant trading-day variables. so that at time T+h [where T = sample size], • 95% forecast intervals:
In most applications, v2=1 will be adequate.ަѝ HDV ᱟ⴨‫ޣ‬ s v1 yT  h|T r 1.96V h
HD
Ⲵ‫ٷ‬ᰕਈ䟿ˈTDV ᱟ⴨‫Ⲵޣ‬Ӕ᱃ᰕਈ䟿DŽ൘བྷཊᮠᓄ⭘ѝˈvЇ=1 ቡ䏣ཏҶDŽ䙊ᑨᡁԜਚ⭘ањ yT  h E1 (T  h)  ¦ J i Di ,T  h  ¦ G i HDVi ,T  h
ਈ䟿ᶕԓ㺘Ӕ᱃ᰕᮠDŽ i 1 i 1 • where V h is an estimate of the standard
• This is a standard regression equation and can be v2 deviation of the h-step forecast distribution. h ↕亴
 ¦ G iTDTDVi ,T  h  H T  h ⍻䈟ᐞⲴḷ߶ᐞՠ䇑
estimated by ordinary least squares.䘉ᱟањḷ߶Ⲵഎᖂᯩ
〻ˈਟԕ⭘Პ䙊ᴰሿҼ҈⌅ՠ䇑DŽ i 1
H t ~ N (0, V H2 ).
t
* under assumption that

•ᆓ㢲ᙗ⁑ᔿ䶎ᑨ␵Რ˖⇿ᒤ᱕༿ᆓ˄བྷ㓖3-5ᴸ䗮ࡠጠ٬˅ᔰᐕ䟿儈ˈ❦ਾ䙀⑀л䱽ˈࡠߜᆓ˄12ᴸ-2ᴸ˅䗮ࡠ䉧
ᓅˈ❦ਾ൘ᯠⲴаᒤ޽⅑എॷDŽ
4. Application: Forecasting Housing •䘉њ⁑ᔿ⇿ᒤ䜭൘䟽༽DŽ
•൘1990-1991ᒤᴹањվ䉧ˈѻਾ1992-1994ᒤᮤփ≤ᒣᴹᡰഎॷˈ䘉ਟ㜭৽᱐Ҷ⸝ᵏ䎻࣯ᡆઘᵏDŽ
Figure 6.4 Housing Starts, 1946.01-1994.11
Starts • Here we zoom in on the Jan 1990 – Nov 1994 to
250
• Figure 6.4 shows monthly data on U.S. housing inspect for seasonal pattern.
starts between Jan 1946 – Nov 1994, which are 200 Figure 6.5 Housing Starts, 1990.01-1994.11
160
seasonal because it’s usually preferable to start
150
houses in the spring, so that they’re completed 140

Starts
before winter arrives.മ6.4 ᱮ⽪Ҷ1946ᒤ1ᴸ㠣1994ᒤ11ᴸⲴ㖾ഭᴸᓖտ 120
ᡯᔰᐕ䟿ᮠᦞˈ䘉Ӌᮠᦞᱟᆓ㢲ᙗⲴˈഐѪ䙊ᑨᴰྭ൘᱕ᆓᔰ࿻ᔪᡯˈԕ‫ׯ‬൘ߜᆓࡠᶕѻ 100
ࡽᆼᐕDŽ
100
Starts

• We’ll use 50
80
¾ Jan 1946 – Dec 1993 for estimation,
0
60
¾ Jan 1994 – Nov 1994 for out-of-sample forecasting. 50 55 60 65 70 75 80 85 90

മѝ␵Რਟ㿱ᕪ⛸Ⲵᆓ㢲ᙗ⌒ࣘˈ਼ᰦ䮯ᵏ Time 40
䎻࣯н᰾ᱮᡆ䖳Ѫ༽ᵲˈਟ㜭䘈ᆈ൘୶ъઘ 90:01 91:01 92:01 93:01 94:01
ᵏDŽ Time
 Table 6.1; Regression results: Seasonal dummy variable model, • Notice that it has very low Durbin-Watson statistic
• The figures reveal that there is no trend, thus it is housing starts. 䘉㺘᰾⁑ර↻ᐞᆈ൘䶎
(see appendix).⌘᜿ࡠᆳⲴDurbin-Watson㔏䇑䟿䶎ᑨվˈ䘉
adequate to model the series using the pure •ՠ䇑ࠪҶ⇿њᴸⲴᒣ൷տᡯᔰᐕ䟿≤ᒣ ᑨѕ䟽Ⲵᒿࡇ⴨‫ޣ‬ᙗ䰞仈DŽ䘉᜿ણ⵰㓟ᆓ㢲ᙗ⁑රᱟн‫ˈⲴ࠶ݵ‬ᘭ⮕ҶᮠᦞѝⲴަԆ䟽㾱ࣘ
˄ᆓ㢲ᙗഐᆀ ȖƄ૾˅ˈֻྲ5ᴸᴰ儈ˈ1ǃ2ǃ ᘱ˄ྲ䎻࣯ǃઘᵏᡆ↻ᐞ㠚䓛Ⲵࣘᘱ˅DŽ
seasonal model that given by,മ㺘ᱮ⽪⋑ᴹ˄᰾ᱮⲴ䮯ᵏঅа˅ 12ᴸ䖳վDŽ
䎻࣯ˈഐ↔֯⭘ྲл㔉ࠪⲴ㓟ᆓ㢲ᙗ⁑රᶕᔪ⁑䈕ᒿࡇᱟ䏣ཏⲴˈ • The residual plot in Figure 6.6 makes clear the
s
䘉᜿ણ⵰ӵࠝᆓ㢲ᙗഐ㍐ቡ㜭䀓䟺տᡯ @EXPAND(@MONT strength and limitations of the model.
yt i it  Ht ᔰᐕ䟿ᮠᦞѝ㓖38%Ⲵ⌒ࣘDŽ䘉㺘᰾ᆓ
¦J D H)˖䘉ᱟEViewsⲴ
i 1 㢲ᙗ⺞ᇎᱟ䈕ᮠᦞⲴањ䶎ᑨ䟽㾱Ⲵ傡
ࣘഐ㍐DŽ ањ䶎ᑨᯩ‫࠭Ⲵׯ‬ • There is nothing in the model other than
ᮠˈᆳՊ㠚ࣘѪ֐
• Table 6.1 shows the estimation results. ࡋᔪа㌫ࡇԓ㺘ᴸ
deterministic seasonal pattern every year – it picks
ԭⲴ㲊ᤏਈ䟿DŽ up a lot of the variation in housing starts, but it
• The 12 seasonal dummies account for more than
a third of the variation in housing starts, as R2 = however, doesn’t pick up all of the variation as
0.38. 12њᆓ㢲ᙗ㲊ᤏਈ䟿䀓䟺Ҷտᡯᔰᐕ䟿ਈᔲⲴй࠶ѻаԕкˈഐѪ R² = 0.38DŽ evidenced by the serial correlation that’s apparent in
• At least some of the remaining variation is the residuals.⁑රѝ䲔Ҷ⇿ᒤ⺞ᇊⲴᆓ㢲ᙗ⁑ᔿཆˈ⋑ᴹަԆԫօь㾯——ᆳ⺞ᇎ
ᦅ᥹Ҷտᡯᔰᐕ䟿ѝⲴབྷ䟿ਈᔲˈնᱟˈ↻ᐞѝ᰾ᱮⲴᒿࡇ⴨‫ޣ‬ᙗ䇱᰾Ҷᆳᒦ⋑ᴹᦅ᥹ࡠᡰ
cyclical, which the model not designed to ᴹⲴਈᔲDŽ
•Ո⛩ (Strength): ᡀ࣏ᦅ᥹Ҷ⴨ᖃа䜘࠶⭡പᇊᆓ㢲ᙗ⁑ᔿᕅ䎧Ⲵਈࣘ˄R²=0.38˅DŽ
capture.㠣ቁ䜘࠶࢙։ⲴਈᔲᱟઘᵏᙗⲴˈ㘼䘉њ⁑රᒦ䶎䇮䇑⭘ᶕᦅ᥹ઘᵏᙗDŽ •ተ䲀ᙗ (Limitations): ⭡Ҿ⁑රਚवਜ਼ᆓ㢲ᙗ㲊ᤏਈ䟿ˈᆳᰐ⌅䀓䟺ᮠᦞѝަԆ㊫රⲴਈࣘ˄ྲ䎻࣯ǃ
୶ъઘᵏˈᡆᆓ㢲ᙗ⁑ᔿᵜ䓛䲿ᰦ䰤Ⲵਈॆ˅DŽ䘉Ӌᵚ㻛䀓䟺Ⲵਈࣘሬ㠤↻ᐞ੸⧠ࠪᒿࡇ⴨‫ޣ‬ᙗ
˄ণ↻ᐞнᱟ䲿ᵪⲴ˅DŽ

Figure 6.6 Housing Starts: Pure Seasonal Model Residual Plot

250

200 Figure 6.7 Estimated Seasonal Factors: Housing Starts

• The estimated seasonal factors are just the 12 160
150 estimated coefficients on the seasonal dummies 150
100 100 as given in Figure 6.7.ՠ䇑Ⲵᆓ㢲ᙗഐᆀቡᱟമ6.7ѝ㔉ࠪⲴᆓ㢲ᙗ㲊ᤏ 140
ਈ䟿Ⲵ12њՠ䇑㌫ᮠDŽ
50 50
130
0
• The seasonal effects are very low in January and
120
0 February, and then rise quickly and peak in May,
110
Seasonal Factors

-50 after which they decline, at first slowly and then

100
abruptly in November and December.ᆓ㢲ᙗ᭸ᓄ൘1ᴸ઼2
-100 ᴸ䶎ᑨվˈ❦ਾ䗵䙏кॷᒦ൘5ᴸ䗮ࡠ亦ጠˈѻਾл䱽ˈ䎧ࡍ㕃ធˈ❦ਾ൘11ᴸ઼12ᴸ 90
50 55 60 65 70 75 80 85 90 ᙕࢗл䱽DŽ
80
1 2 3 4 5 6 7 8 9 10 11 12
Residual Actual Fitted
•ᤏਸ٬˄⛩㓯˅⺞ᇎ⇿ᒤ䟽༽⵰⴨਼Ⲵ⁑ᔿˈ৽᱐Ҷ⁑රՠ䇑ࠪⲴᒣ൷ᆓ㢲ᙗDŽ Month
•↻ᐞ˄ᇎ㓯˅ത㔅0㓯⌒ࣘˈն㔍㔍䶎䲿ᵪDŽᆳᱮ⽪ࠪ䶎ᑨᕪ⛸Ⲵᤱ㔝ᙗ˄ྲ᷌ࡽаᵏ↻ᐞѪ↓ˈᵜᵏҏᖸਟ㜭Ѫ
↓˅઼㊫լ୶ъઘᵏⲴ䮯ᵏ⌒ࣘDŽֻྲˈ൘ḀӋᒤԭ˄ྲ70ᒤԓѝᵏǃ80ᒤԓࡍ˅ˈ↻ᐞᤱ㔝Ѫ䍏ˈ㺘᰾ᇎ䱵ᔰ
ᐕ䟿㌫㔏ᙗൠվҾᆓ㢲ᙗᒣ൷≤ᒣ˗㘼൘ਖаӋᒤԭˈ↻ᐞᤱ㔝Ѫ↓DŽ
•䘉⿽↻ᐞⲴ⁑ᔿ↓ᱟሬ㠤D-W㔏䇑䟿ᶱվⲴ৏ഐDŽᆳ␵ᾊൠ㺘᰾ˈ㓟ᆓ㢲ᙗ⁑ර䚇┿Ҷ䟽㾱Ⲵؑ᚟DŽ
• Figures 6.9 gives the history of housing starts through 1993,
together with the out-of-sample point and 95% interval
extrapolation forecasts for the first 11 months of 1994.മ6.9 㔉ࠪҶࡠ1993 Figure 6.9 Housing Starts: History, 1990.01-1993.12 Forecast and Realization, 1994.01-1994.11
ᒤⲴտᡯᔰᐕ䟿শਢᮠᦞˈԕ৺ሩ1994ᒤࡽ11њᴸⲴṧᵜཆ⛩亴⍻઼95%४䰤ཆ᧘亴⍻ˈ亴⍻ⴻ䎧ᶕਸ⨶ˈ
250
ഐѪ⁑රᱮ❦ᖸྭൠᦅ᥹Ҷᆓ㢲ᙗ⁑ᔿDŽ

• Forecasts look reasonable, as the model evidently done a good job

200
of capturing the seasonal pattern.
• The forecast intervals are quite wide, reflecting the fact that the 150
seasonal effects captured by the forecasting model are responsible
for only about a third of the variation in the variable being forecast. •S.E. (Optional for intervals): ൘亴⍻ሩ䈍Ṷѝˈᴹањਟԕປ߉ḷ߶䈟Ⲵൠᯩˈ⭘
100 Ҿ⭏ᡀ亴⍻४䰤DŽ֐൘䘉䟼ປ߉ҶseDŽ
亴⍻४䰤⴨ᖃᇭˈ৽᱐Ҷ亴⍻⁑රᦅ᥹ࡠⲴᆓ㢲ᙗ᭸ᓄӵ㜭䀓䟺㻛亴⍻ਈ䟿㓖й࠶ѻаⲴਈᔲ䘉аһᇎDŽ⁑
රⲴR²ਚᴹ0.38ᐖਣˈ᜿ણ⵰ᴹᖸབྷа䜘࠶˄㓖62%˅Ⲵਈᔲᱟ⁑රᰐ⌅䀓䟺ⲴDŽ䘉⿽ᵚ㻛䀓䟺Ⲵਈᔲሬ㠤 ⢩
•ӾMAPE (㓖4.97%) ઼ Theil U (㓖0.027) ᶕⴻˈ֐Ⲵ⁑ර൘1994ᒤ1ᴸ㠣11ᴸ䘉њ⢩
Ҷ亴⍻Ⲵн⺞ᇊᙗ䖳བྷˈӾ㘼֯ᗇ亴⍻४䰤䖳ᇭDŽ ᇊⲴṧᵜཆᵏ䰤㺘⧠ࠪҶ⴨ᖃྭⲴ亴⍻㋮ᓖDŽ
50 •ᴹᡚ䐍઼(s-1)њ㲊ᤏਈ䟿 (ᐖ‫⁑ח‬රˈণ֐ਾᶕ᭩ᡀⲴ hstarts s c

History, Forecast and Realization

• The forecast appears highly accurate as the realization and forecast @expand(@month,@dropfirst)) ᱟᴤᑨ㿱Ⲵ‫ˈ⌅ڊ‬ഐѪᆳ䚯‫ݽ‬ҶĀ㲊ᤏਈ䟿䲧䱡ā
˄ྲ਼᷌ᰦᴹᡚ䐍઼sњ㲊ᤏਈ䟿Պਁ⭏˅ˈᒦф㌫ᮠⲴᱮ㪇ᙗỰ傼˄Ự傼оส߶ᴸ
are quite close. In addition, the realization is everywhere well 0
ԭⲴᐞᔲ˅䙊ᑨᴤᴹ᜿ѹDŽ
within the 95% interval.亴⍻ⴻ䎧ᶕ儈ᓖ߶⺞ˈഐѪᇎ⧠٬઼亴⍻٬䶎ᑨ᧕䘁DŽ↔ཆˈ 90:01 91:01 92:01 93:01 94:01 аᴸԭ (ส߶ᴸԭ) Ⲵᒣ൷տᡯᔰᐕ䟿DŽ
•C (ᑨᮠ亩): 86.50417DŽ䘉њ٬ԓ㺘а
ᇎ⧠٬࿻㓸ᆼ‫ޘ‬൘95%४䰤޵DŽ ҼᴸԭⲴᒣ൷տᡯᔰᐕ䟿⴨
•@MONTH=2: 3.000000DŽ䘉њ٬ԓ㺘Ҽ ⴨ሩҾаᴸԭ儈ࠪ3.00њঅսDŽᡰԕҼᴸԭⲴ
Time ᒣ൷ᔰᐕ䟿ᱟ 86.50417 + 3.000000 = 89.50417DŽ
йᴸԭⲴᒣ൷տᡯᔰᐕ䟿⴨
•@MONTH=3: 36.37917DŽ䘉њ٬ԓ㺘й ⴨ሩҾаᴸԭ儈ࠪ36.37917њঅսDŽᡰԕйᴸ
ԭⲴᒣ൷ᔰᐕ䟿ᱟ 86.50417 + 36.37917 = 122.88334DŽ
•ԕ↔㊫᧘ˈަԆᴸԭ㲊ᤏਈ䟿Ⲵ㌫ᮠ䜭㺘⽪䈕ᴸԭ⴨ሩҾаᴸԭⲴᒣ൷ᐞᔲDŽ

Appendix : • The regression disturbance is serially correlated • DW takes values in the interval [0,4], and if all is
when M z 0. well, DW should be around 2.
• DW Ⲵਆ٬㤳ത൘ [0,4] ४䰤޵ˈྲ᷌а࠷↓ᑨ˄ণ⋑ᴹа䱦㠚⴨‫ˈ˅ޣ‬DW ᓄ䈕൘2䱴
• The Durbin-Watson statistic tests for correlation • Hypothesis: 䘁DŽ
over time, called serial correlation, in regression H0 :M 0
• If DW is substantially less than 2, there is
disturbances. Durbin-Watson 㔏䇑䟿⭘ҾỰ傼എᖂᢠࣘ亩ѝ䲿ᰦ䰤ᆈ൘Ⲵ⴨‫ޣ‬
ᙗˈ〠Ѫᒿࡇ⴨‫ޣ‬DŽ/Ự傼എᖂ⁑රⲴ↻ᐞ˄ᢠࣘ亩Ⲵՠ䇑˅ᱟ੖ᆈ൘㠚⴨‫˄ޣ‬ᒿࡇ⴨ H1 : M z 0 evidence of positive autocorrelation.
‫˅ޣ‬DŽ • ༷ᤙ‫ٷ‬䇮 (alternative hypothesis) ᱟ䈟ᐞ亩ᆈ൘а䱦㠚⴨‫ˈޣ‬DWỰ傼䙊ᑨᴤ‫↓⌘ޣ‬㠚⴨ • ྲ᷌ᆈ൘ᕪⲴ↓㠚⴨‫( ޣ‬ȡƄ ᧕䘁1)ˈࡉ DW Šʼn2(1-1) = 0DŽᡰԕDW٬᧕䘁0㺘⽪ᕪⲴ
‫( ޣ‬ij > 0)ˈഐѪ䘉൘㓿⍾ᰦ䰤ᒿࡇѝᴤᑨ㿱DŽ ↓㠚⴨‫ޣ‬DŽ↓㠚⴨‫ޣ‬᜿ણ⵰ྲ᷌ањ↻ᐞᱟ↓Ⲵˈлањ↻ᐞҏ‫ੁٮ‬Ҿᱟ↓Ⲵ˄৽ѻ
Ӗ❦˅DŽ䘉൘ѻࡽⲴ䴦୞仍઼տᡯᔰᐕ䟿⁑රⲴ↻ᐞമѝ䜭㿲ሏࡠҶDŽ
• The Durbin-Watson test works within the context of
• Durbin-Watson statistic
T
is • If DW is substantially greater than 2, there is
the regression model
2
yt E 0  E1 X 1,t    E k X k ,t  H t ¦ e e t t 1 evidence of negative autocorrelation.
t 2 • ྲ᷌ DW ᱮ㪇བྷҾ2ˈࡉᴹ䍏㠚⴨‫Ⲵޣ‬䇱ᦞDŽྲ᷌ᆈ൘ᕪⲴ䍏㠚⴨‫( ޣ‬ȡƄ ᧕䘁-1)ˈࡉ DW
DW T Šʼn2(1-(-1)) = 4DŽᡰԕDW٬᧕䘁4㺘⽪ᕪⲴ䍏㠚⴨‫ޣ‬DŽ䍏㠚⴨‫ޣ‬᜿ણ⵰ྲ᷌ањ
↻ᐞᱟ↓Ⲵˈлањ↻ᐞ‫ੁٮ‬Ҿᱟ䍏Ⲵ˄৽ѻӖ❦˅ˈ↻ᐞ੸⧠Ӕᴯ⁑ᔿDŽ
Ht MH t 1  vt , 2
¦e t
2 t 1
vt N 0, V