Statistics 5350/7110

Forecasting

Lecture 21
Building SARIMA Models

Professor Robert Stine

 Preliminaries
• Questions?

• Assignments
• Assignment 5 posted
• Due next Thursd y

• Quick review
• Building SARIMA models
• Adding expl n tory f ctors
• C lend r v ri bles Lecture_20.Rmd & Lecture_21.Rmd (tod y)

2
a
a
a
a
a
a
a
a
 Today’s Topics Text, Ch pter 5

• Building SARIMA models

• Identifying model
• Incorpor ting exogenous f ctors

• Ex mples
• Housing construction
• S les in furniture stores (d t from Assignment)

3
a
a
a
a
a
a
a
a
New Housing Construction
 Housing Time Series
• New housing construction
• Thous nds of units, priv tely owned
• Cle rly se son l with non-st tion ry trends

FRED HOUSTNSA 5
a
a
a
a
a
a
a
 Housing Time Series
• New housing construction
• Thous nds of units, priv tely owned
• Zoom in on recent trend to ex mine se son lity
• Is se son l p ttern st ble?

FRED HOUSTNSA 6
a
a
a
a
a
a
a
a
a
 Housing Starts
• Se son l st bility
• Overl y devi tions from nnu l me n

2000-2009 2010-2019

7
a
a
a
a
a
a
a
a
 Housing Starts
• Se son l st bility
• Overl y devi tions from nnu l me n

2010-2019 2020-2023

8
a
a
a
a
a
a
a
a
 Housing Starts
• Recognize n ture of ch nges
• Di erent se son lity
• Coll pse of construction during 2008-2009

• Focus on period since housing bubble burst…

• Use d t of 10 ye rs, 2010-2019
• Objective: Predict 2 ye rs out

• Revised focus reve ls…

• Softening p ttern prior to Covid
• M gnitude of se son l v ri tion prompts question:

Should we be modeling on log sc le?

9
a
ff
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
 Housing: To Log or Not To Log?
• Subst ntive thinking
• Will eventu lly di erence d t for st tion rity
• Di erences of logs nicely interpret ble s percent ge ch nge

• D t n lysis
• Ex mine rel tionship of me n to v ri nce
• Would like these to be “sep r ble”

Conclude?

10
ff
a
a
a
a
a
a
a
a
ff
a
a
a
a
a
a
a
a
a
a
a
 Housing: Differences
• Sequence plots
• Log counts
• Di erenced log counts

• P tterns worth noticing in the di erences?

• P rticul rly with reg rd to se son lity

11
a
a
ff
a
a
a
a
ff
 Housing: Model Identi ication
• ACF nd PACF
• Di erenced log housing st rts

• We ker p tterns
• M gnitudes of the utocorrel tions re
sm ller th n in other ex mples

• Model
• Initi l choices?

• Anything else
• Wh t else should we be thinking bout
before ARIMA modeling?

Build non-stoch stic fe tures such s length of month, d tes v il ble for construction.
12
a
ff
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
 Housing: SARIMA Modeling
• Initi l model
• p=0, d=1, q=1, P=1, D=0, Q=0
• Includes monthly c lend r fe tures (d ys in month, weekd ys in month)

13
a
a
a
a
a
 Housing: SARIMA Modeling
• Revised model Incre se P gr du lly nd
• p=0,d=1,q=1, P=4, D=0, Q=0, S=12 check it e ch time

• Continue to include monthly c lend r fe tures

14
a
f
a
a
a
a
a
a
a
 Housing: SARIMA Modeling
• Switch to se son l moving ver ge
• Lots of AR suggests m ybe better with MA
• p=0,d=1,q=1, P=1, D=0, Q=1, S=12
• Includes monthly c lend r fe tures… still not signi ic nt
• Nice it, but coe icients re “speci l”

15
f
ff
a
a
a
a
a
a
a
a
a
a
 Housing: SARIMA Modeling
• Fin l model
• p=0, d=1, q=1, P=4, D=0, Q=0, S=12
• Removes monthly c lend r fe tures (one t time, never signi ic nt)
• Simil r error v ri nce to other models, simpler form
• Residu ls ≈ Norm l, residu l ACF mostly sm ll

16
a
a
a
a
a
a
a
a
a
a
a
a
a
 Housing: SARIMA Forecasts
• Forec st two ye rs out
• Predicts growth in subsequent ye rs
• Obviously miss Covid drop (which w sn’t so big for housing)

17
a
a
a
a
 Housing: SARIMA Forecasts
• Forec st two ye rs out
• Add two ye rs of ctu l d t
• Anticip te recovery in 2021 fter Covid… Luck?

How do you think

regression model
with trend would f re
in this ex mple?

18
a
a
a
a
a
a
a
a
a
a
a
a
Regression Model
 Regression Trend Models
• Wh t would you use?

20
a
 Regression Trend Model
• Qu dr tic regression model
• AR(1) for residu ls
• M ny se son l coe icients re simil r (April-July)
• Cl ims better it th n SARIMA (0.0054 vs 0.0064)

21
a
a
a
a
a
a
f
a
a
ff
a
a
 Housing: Regression Forecasts
• Forec st two ye rs out
• Comp re to ctu l d t
• Forec sts re too low… Qu dr tic trend doesn’t continue

22
a
a
a
a
a
a
a
a
a
a
a
 Furniture Sales

Are housing st rts le ding indic tor of subsequent s les?

a
a
a
a
 Furniture Sales and Housing Starts
• Le ding indic tor?
• Does the volume of housing st rts nticip te l ter s les t furniture stores?
• How long does it t ke to build nd then sell home?

24
a
a
a
a
a
a
a
a
a
 Cross-Correlations
• Cross-correl tion, levels
• Modest correl tion ≈ 0.20 using st rts nd s les
• Pr ctic l le ding indic tor le ds by bout 14 months (Is 2 months re son ble?)

25
a
a
a
a
a
a
a
a
a
a
a
 Cross-Correlations
• Cross-correl tion, di erences
• Stronger (though still we k) correl tion ≈ 0.30 using di erences of st rts nd s les
• Pr ctic l le ding indic tor le ds by bout 14 months (Is 2 months re son ble?)

26
a
a
a
a
a
a
ff
a
a
a
 Cross-Correlations
• Cross-correl tion, di erences
• Stronger (though still we k) correl tion ≈ 0.30 using di erences of st rts nd s les
• Pr ctic l le ding indic tor le ds by bout 14 months (Is 2 months re son ble?)
• Zoom in on correl tion t 14 months
• Wh t’s going on?

27
a
a
a
a
a
a
a
a
ff
a
a
a
a
 What’s next?
• Di erent ppro ch to underst nding time series … Periodicity
• Di erent w y to ppro ch st tion ry processes
• Novel di gnostic method
• More directed to science th n for modern economics

28
ff
ff
a
a
a
a
a
a
a
a
a
a