Regression Example 1
Initial Plots
Logarithmic Fit
Indicators vs Harmonic Seasonal Terms
Forecasting
Exercises
Source Coverage 1
General review of ideas from Cowpertwait and Metcalfe, chapter 5
Initial Plots 2
Consider time series {xt } with following plot:
5.0
150
4.5
100
4.0
x
3.5
50
3.0
2.5
0 20 40 60 80 100 120 0 20 40 60 80 100 120
t t
Seasonal effect size increases with trend, {xt } > 0, consider yt = log[xt ]
Seasonal effect size seems stable for {yt }
Trend looks linear in t
Analyzing yt = log[xt ] 3
Fit initial model yt = α1 t + β1+(t−1)mod 12 + zt
No intercept term b/c using distinct β for each season.
Looking at ACF for residuals for yt : Series residuals([Link])
1.0
0.6
ACF
0.2
-0.2
0 5 10 15 20
Lag
Only r0 is significantly different than 0 ⇒ residuals look like white noise
Remark: Don’t need PACF plot b/c ACF shows it is white noise.
Choice of Seasonal Model 4
Have used Model 1: (Indicators for each month)
yt = α1 t + β1+(t−1)mod 12 + zt
Have 14 parameters: α1 , 12 βi values, σ 2 = Var[zt ]
Consider Model 2: (Harmonic seasonal terms)
5 h
X t t i
yt = α0 + α1 t + si sin 2πi + ci cos 2πi + c6 cos(πt) + zt
12 12
i=1
Separated i = 6 term b/c sin(πt) = 0 for all integer t.
Still have 14 parameters: α0 , α1 , σ 2 , 5 si values, 6 ci values
These are equivalent, both have AIC 53.06, same α1 , etc.
Harmonic Selection 5
Parameters for Model 2 are:
Parameter Estimate SE t-value p-value
α0 3.14 0.052 59.93 0.000
α1 0.01 0.001 13.07 0.000
c1 0.24 0.037 6.62 0.000
s1 0.66 0.037 18.05 0.000
c2 0.01 0.037 0.26 0.795
s2 −0.02 0.037 −0.55 0.585
c3 0.07 0.037 1.82 0.072
s3 0.09 0.037 2.53 0.012
c4 −0.01 0.037 −0.17 0.865
s4 −0.03 0.037 −0.80 0.426
c5 −0.04 0.037 −1.01 0.313
s5 0.07 0.037 2.01 0.047
c6 −0.01 0.026 −0.49 0.624
This suggests we can drop c2 , s2 , c3 , c4 , s4 , c5 , c6 terms.
Some people require keeping both harmonics if either is significant,
i.e. would keep c3 , c5 b/c keep s3 , s5 . Text mentions this, does not require it.
Updated Model 6
Refitting without those terms:
t
1t 3t 5t
yt = α0 + α1 + c1 cos 2π + s1 sin 2π + s3 sin 2π + s5 sin 2π + zt
12 12 12 12
Updated parameters are:
Parameter Estimate SE t-value p-value
α0 3.14 0.052 59.93 0.000
α1 0.01 0.001 13.07 0.000
c1 0.24 0.037 6.66 0.000
s1 0.66 0.037 18.17 0.000
s3 0.09 0.037 2.55 0.012
s5 0.07 0.036 2.02 0.045
Parameters essentially unchanged b/c harmonics are independent
# dof increased, slightly decreases SE, increases t-value, decreases p-value.
Effects small here, even smaller with larger data sets.
AIC decreases from 53.06 to 45.19
Reduction in loglikelihood more than compensated by decrease in parameters.
Forecasting 7
Full model with indicator functions outputs
1X
α
b1 = 0.00985 βb6 = 2.8619 exp[b
zt ] = 1.0368
n
To predict for t = 126, (126 mod 12) ≡ 6 and get
b1 · 126 + βb6
yb126 = α
= 0.00985 · 126 + 2.8619 = 4.103
xb126 = exp(4.103) · (1.0368) = 62.75
For comparison
Var[b
zt ] = 0.07275
exp[Var[b
zt ]/2] = 1.0370
so if zt were normal, would forecast 62.76
Exercise 1 8
Based on the following plot, would you use an additive or multiplicative model for xt ?
What model for trend would you consider?
50
40
x
30
20
0 50 100 150
Exercise 1 8
Based on the following plot, would you use an additive or multiplicative model for xt ?
What model for trend would you consider?
50
40
x
30
20
0 50 100 150
Seasonal effect seems not to depend on trend value, use additive model.
Clear curve, try mt = α0 + α1 t + α2 t 2
Exercise 2 9
Series [Link]
The PACF from the residuals of the previous model are:
0.0 0.2 0.4
Partial ACF
-0.4
5 10 15 20
Lag
What residual correlation structure seems appropriate?
Exercise 2 9
Series [Link]
The PACF from the residuals of the previous model are:
0.0 0.2 0.4
Partial ACF
-0.4
5 10 15 20
Lag
What residual correlation structure seems appropriate?
Have PACF significant at lag 2, roughly 0 afterwards.
Use AR(2) residuals.