Lesson 6: Time Series 1

Let’s start doing some time series analysis.

To create a TS object in R, it is often useful to use the xts function in the xts package. By running the
following line, you can have more info about xts.
?xts

## No documentation for 'xts' in specified packages and libraries:

## you could try '??xts'
Assume we want to create a white noise TS, based on a N(0,1). Recall: it is not necessary for a white noise
to be normally distributed! This time series contains monthly observations starting from January 1, 2018
until May 1, 2020.
library(xts)

## Loading required package: zoo

##
## Attaching package: 'zoo'
## The following objects are masked from 'package:base':
##
## as.Date, as.Date.numeric
set.seed(10) # so we all get the same results
x=rnorm(29,0,1)
dates <- seq(as.Date("2018-01-01"),length=29,by="months")
white_noise=xts(x,order.by =dates,frequency=12)
white_noise

## [,1]
## 2018-01-01 0.01874617
## 2018-02-01 -0.18425254
## 2018-03-01 -1.37133055
## 2018-04-01 -0.59916772
## 2018-05-01 0.29454513
## 2018-06-01 0.38979430
## 2018-07-01 -1.20807618
## 2018-08-01 -0.36367602
## 2018-09-01 -1.62667268
## 2018-10-01 -0.25647839
## 2018-11-01 1.10177950
## 2018-12-01 0.75578151
## 2019-01-01 -0.23823356
## 2019-02-01 0.98744470
## 2019-03-01 0.74139013
## 2019-04-01 0.08934727
## 2019-05-01 -0.95494386
## 2019-06-01 -0.19515038
## 2019-07-01 0.92552126

1
## 2019-08-01 0.48297852
## 2019-09-01 -0.59631064
## 2019-10-01 -2.18528684
## 2019-11-01 -0.67486594
## 2019-12-01 -2.11906119
## 2020-01-01 -1.26519802
## 2020-02-01 -0.37366156
## 2020-03-01 -0.68755543
## 2020-04-01 -0.87215883
## 2020-05-01 -0.10176101
plot(white_noise)

white_noise 2018−01−01 / 2020−05−01

1.0 1.0

0.5 0.5

0.0 0.0

−0.5 −0.5

−1.0 −1.0

−1.5 −1.5

−2.0 −2.0

Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr
2018 2018 2018 2018 2019 2019 2019 2019 2020 2020
Let’s us now import some financial data from quantmod. We can use Google data in the last 2 years. When
you import from quantmod, the object you create is already a xts time series.
library(quantmod)
getSymbols("GOOGL",from="2017-05-01",to="2020-05-11",src="yahoo")

## [1] "GOOGL"
stock=Ad(`GOOGL`) # Stock Values
plot(stock)

2
 stock 2017−05−01 / 2020−05−08

1500 1500

1400 1400

1300 1300

1200 1200

1100 1100

1000 1000

May 01 Nov 01 May 01 Nov 01 May 01 Nov 01 Apr 30

2017 2017 2018 2018 2019 2019 2020 I
guess you can appreciate immediately the different behavior of the white noise and Google stock. Which one
has a trend?
Let’s have a look at their ACF.
acf(white_noise)

3
 Series white_noise
0.8
0.4
ACF

0.0
−0.4

0 2 4 6 8 10 12 14

Lag As ex-
pected, the acf sees no significant autocorrelation in a white noise. Apart from the first bar, which we can
ignore, all the others are within the confidence lines (blue dashed ones). Please be always careful about the
first bar in the ACF, that it is always 1 being the autocorrelation of xt with xt , i.e. ρ(0).
What about Google?
acf(stock,100)

4
 Series stock
1.0
0.8
0.6
ACF

0.4
0.2
0.0

0 20 40 60 80 100

Lag
Given the long and strong decay of a yet-to-be-cleaned TS, an ACF like the one above suggests the presence
of a strong trend in the data. We need to remove it before proceeding with the analysis of the type of time
series we are dealing with.
We can use two methods: regression and differences.
# Regression
t=1:length(stock)
fit=lm(stock~t)
summary(fit)

##
## Call:
## lm(formula = stock ~ t)
##
## Residuals:
## Min 1Q Median 3Q Max
## -245.225 -52.886 -6.225 47.112 236.196
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 960.82285 5.73806 167.45 <2e-16 ***
## t 0.46438 0.01303 35.64 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 79.12 on 760 degrees of freedom
## Multiple R-squared: 0.6256, Adjusted R-squared: 0.6252
## F-statistic: 1270 on 1 and 760 DF, p-value: < 2.2e-16

5
clean_stock=as.xts(fit$residuals)
plot(clean_stock)

clean_stock 2017−05−01 / 2020−05−08

200 200

100 100

0 0

−100 −100

−200 −200

May 01 Nov 01 May 01 Nov 01 May 01 Nov 01 Apr 30

2017 2017 2018 2018 2019 2019 2020
We may also want to visualize the trend in the original plot. This is a useful characteristic of regression: we
know the trend, we do not just remove it.
trend=fit$coefficients[1]+t*fit$coefficients[2]
trend=as.xts(trend,order.by =time(stock))
plot(stock)

6
 stock 2017−05−01 / 2020−05−08

1500 1500

1400 1400

1300 1300

1200 1200

1100 1100

1000 1000

May 01 Nov 01 May 01 Nov 01 May 01 Nov 01 Apr 30

2017 2017 2018 2018 2019 2019 2020
lines(trend,col=2)

stock 2017−05−01 / 2020−05−08

1500 1500

1400 1400

1300 1300

1200 1200

1100 1100

1000 1000

May 01 Nov 01 May 01 Nov 01 May 01 Nov 01 Apr 30

2017 2017 2018 2018 2019 2019 2020
We can also detrend by differencing, i.e.

7
diff_stock=diff.xts(stock,na.action=na.pass)
plot(diff_stock)

diff_stock 2017−05−01 / 2020−05−08

100 100

50 50

0 0

−50 −50

−100 −100

May 01 Nov 01 May 01 Nov 01 May 01 Nov 01 Apr 30

2017 2017 2018 2018 2019 2019 2020
Why do we get plots that are so different?
Let’s compare the ACFs
acf(clean_stock,na.action=na.omit)

8
 Series clean_stock
1.0
0.8
0.6
ACF

0.4
0.2
0.0

0 5 10 15 20 25

Lag
acf(diff_stock,na.action = na.pass)

Series diff_stock
1.0
0.8
0.6
ACF

0.4
0.2
−0.2

0 5 10 15 20 25

Lag
Regression only removes the linear (deterministic) trend, while differencing also accounts for a the variable
trends and (partially) for the seasonal effect.

9
Both methods are unable to deal with cluster volatility, however.
And the apparently nice plot we get via differencing can be misleading. . . Why?
In any case, both new time series are just the starting point of the analysis. The best is yet to come!

10