2 FRM9649

MAY/JUNE 2020

TIME SERIES ANALYSIS AND FORECASTING (FRM9649)

ASSIGNMENT 2

Due Date: 07 August 2020 by 23:59

Instruction(s):
• Answer ALL the questions.

• Write your answers legibly (preferably typed, but I won’t hold it against you if not typed, unless I
can’t read what you wrote).

• Submit by email to mkamga-pene@unam.na in response to the email you received about this
assignment.

• Your assignment must be submitted in one PDF file (you can have it handwritten and scanned,
or typed in Word and then exported to PDF, or if you work in Excel, copy your answers in Word
and then export to PDF. Bottom line, it must be one PDF file).

• The PDF file name should look like FRM9649-2020-S1-A2-YourStudentNumber.pdf (for

example, if your student number is 209912345, then the file name should be FRM9649-2020-S1-
A2-209912345.pdf ).

• Provide your name and student number in the PDF file (as you usually do on the cover page).

• Late submissions will not be considered and will be allocated a zero mark.

• You may loose marks for not following these instructions.

• GOOD LUCK!

QUESTION 1

Answer each of the following questions by TRUE or FALSE.

(1.1) In regression, a dummy variable is a quantitative variable which is assigned a qualitative value (2)
based on whether a condition is satisfied or not.
Solution: False. X2 A dummy variable is a qualitative variable which is assigned a quanti-
tative value based on whether a condition is satisfied or not.

(1.2) In forecasting, the seasonal index is an indicator of seasonality. (2)

Solution: False. X2 The seasonal index helps in deseasonalizing data (if there is seasonality).
(1.3) Moving average is a useful tool that helps dealing with seasonality in time-series. (2)
Solution: True. X2
(1.4) The F ratio is an indicator of good fit in regression analysis. (2)
Solution: False. X2 The F ratio is an indicator of the overall significance of regression. The
goodness of fit is indicated by the R2 .

(1.5) The Durbin-Watson test is used to assess the level of serial correlation of residuals in auto- (2)
regressive models.

[TURN OVER]
 3 FRM9649
MAY/JUNE 2020

Solution: False. X2 The Durbin-Watson test is used to assess the level of serial correlation
of residuals in linear regression not involving time-series, or in trend models for time-series.
Serial correlation is assessed in auto-regressive models via autocorrelation.
[10]

QUESTION 2

You are a risk analyst hired in January 2020 by the management team of a big restaurant. Your first
task is to forecast sales for every month of 2020. To that end, you are given monthly records of sales from
January 2017 to December 2019 as summarized in the following table:
Month 2017 2018 2019
January 242 263 282
February 235 238 255
March 232 247 265
April 178 193 205
May 184 193 210
June 140 149 160
July 145 157 166
August 152 161 174
September 110 122 126
October 130 130 148
November 152 167 173
December 206 230 235

(2.1) Provide a graphical representation for the given time-series. (2)

Solution: The required graph is given below (Figure 2.1):

X2
Figure 1: Restaurant Sales - [Question (2.1)]

(2.2) Conduct an analysis of seasonality of the given time-series. This includes calculating the (8)

[TURN OVER]
 4 FRM9649
MAY/JUNE 2020

seasonal indices, detect high and low seasonal sales months, and the trend if any, for sales for
similar months throughout the yeaars.
Solution: From the given data, the highest sales happen in January and the lowest sales in
September.XIn similar months throughout the years, the sales are steadily increasing.XThe
seasonal indices are computed using the ratio-to-moving-average method, and the results (de-
seasonalized sales and trend) are summarized in the following table, where all the calculation
was rounded off to two decimal places:

Month Sales SIX6 Deseasonalized Trend

2017 2018 2019 2017 2018 2019 2017 2018 2019
Jan 242 263 282 1.44 168.06 182.64 195.83 170.52 182.76 194.99
Feb 235 238 255 1.30 180.77 183.08 196.15 171.54 183.77 196.01
Mar 232 247 265 1.34 173.13 184.33 197.76 172.56 184.79 197.03
Apr 178 193 205 1.04 171.15 185.58 197.12 173.58 185.81 198.05
May 184 193 210 1.05 175.24 183.81 200.00 174.60 186.83 199.07
Jun 140 149 160 0.80 175.00 186.25 200.00 175.62 187.85 200.09
Jul 145 157 166 0.83 174.70 189.16 200.00 176.64 188.87 201.11
Aug 152 161 174 0.85 178.82 189.41 204.71 177.66 189.89 202.13
Sep 110 122 126 0.63 174.60 193.65 200.00 178.68 190.91 203.15
Oct 130 130 148 0.70 185.71 185.71 211.43 179.70 191.93 204.17
Nov 152 167 173 0.85 178.82 196.47 203.53 180.72 192.95 205.19
Dec 206 230 235 1.16 177.59 198.28 202.59 181.74 193.97 206.21

(2.3) Deseasonalize the time-series. Does there appear to be any trend in the deseasonalized time- (4)
series?
Solution: We deseasonalize the time-series by multiplying the sales values by the correspond-
ing seasonal indices. The results are in the Deseasonalized portion of the above table.X2 A
graphical representation of the deseasonalized time-series (Figure 2.3) given below suggests a
linear trend.X

X
Figure 2: Restaurant Sales (deseasonalized) - [Question (2.3)]

[TURN OVER]
 5 FRM9649
MAY/JUNE 2020

(2.4) Forecast sales for January through December of the year 2020. (6)
Solution: We find the corresponding estimates using linear trend information from above,
and apply to values of t from 37 to 48, then multiply by the corresponding seasonal indices.
The results (for 2020 forecast) are summarized in the following table:

Month t SI Trend Forecast

Jan 37 1.44 207.23 298.41
Feb 38 1.30 208.25 270.72
Mar 39 1.34 209.27 280.42
Apr 40 1.04 210.29 218.70
May 41 1.05 211.31 221.87
Jun 42 0.80 212.33 169.86 X6
Jul 43 0.83 213.35 177.08
Aug 44 0.85 214.37 182.21
Sep 45 0.63 215.39 135.69
Oct 46 0.70 216.40 151.48
Nov 47 0.85 217.42 184.81
Dec 48 1.16 218.44 253.40

[20]

QUESTION 3

You are a risk analyst hired in 2015 by the management team of a big sales company. One of your tasks is
to analyze and make predictions on the net worth of the company as time evolves, based on all company’s
activities. To that end, you are given yearly records of the company’s net worth from 1997 to 2014 as
summarized in the following table:

Year Worth Year Worth Year Worth

1997 4.40 2003 55.00 2009 106.00
1998 10.55 2004 68.50 2010 111.85
1999 23.20 2005 88.50 2011 122.20
2000 38.70 2006 89.80 2012 141.55
2001 44.00 2007 88.05 2013 148.75
2002 49.50 2008 94.15 2014 150.90

(3.1) Provide a graphical representation for the given time-series. (2)

Solution: The required graph is given below (Figure 3.1):

(3.2) Does the time-series seem to follow a trend? If yes, which type of trend? Justify your answer. (2)
Solution: Yes, the time-series seems to follow a linear trendXbecause the points appear to
be positioned along a straight line.X

(3.3) Is the time-series covariance-stationary? Justify your answer. (2)

Solution: No, the time-series is not covariance-stationaryXbecause from the trend, the data
values are steadily increasing.X

(3.4) Assess the level of heteroskedasticity. Is there a reason for concern? Justify your answer. (3)

[TURN OVER]
 6 FRM9649
MAY/JUNE 2020

X2
Figure 3: Company Net Worth - [Question (3.1)]

Solution: We run the Breusch-Pagan test. Regressing the squared residuals over time (the
independent variable),Xwe obtain the t-statistic χ = 18R2 = 0.18723,Xwhich suggests that
(check the chi-squared table where df = 18 − 2 = 16, our calculated χ is much smaller than
the smallest entry on the corresponding row) heteroskedasticity is not a reason for concern.X

(3.5) Assess the level of serial correlation. Is there a reason for concern? Justify your answer. (3)
Solution: We run the DW test. We find DW = 1.1951.XAccording to the DW table, for 18
observations, we have dl = 1.16 and du = 1.39.XSince dl < DW < du , we cannot confidently
conclude if positive serial correlation is a reason for concern or not.X

(3.6) Conduct an AR(1) process on a new time-series obtained after a carefully chosen transforma- (5)
tion, and assess the level of serial correlation via an auto-correlation test.
Solution: We transform the data as follows: Zt = Yt − Yt−1 .XThe regression equation for
the AR(1) model is Zt = b0 + b1 Zt−1 + εt . Performing simple linear regression (or ordinary
least squares method), we get b0 = 7.73892 and b1 = 0.11449, thus the corresponding equation
is Zt = 7.73892 + 0.11449Zt−1 + εt .XMoreover, autocorrelation test is summarized in the
following table:
Lag Autocorrelation t-statistic
1 0.08809 0.36321
2 −0.56547 −2.33149
3 −0.15538 −0.64065
4 0.10708 0.44149
5 0.00842 0.03470 X2
6 −0.04461 −0.18392
7 0.19689 0.81179
8 0.03342 0.13779
9 −0.15175 −0.62567
10 −0.11747 −0.48434
The initial time-series has 18 observations, but the Z variable has 17 observations. Thus our
AR(1) model has 17 − 2 = 15 degrees of freedom. According to the t−table, the lowest value

[TURN OVER]
 7 FRM9649
MAY/JUNE 2020

for the t−statistic for 15 degrees of freedom is 0.691. Hence, by inspection, lags 2 and 7 are
a reasons for concern about autocorrelation since the confidence level for autocorrelation is
quite high (about 98% for lag 2 and 60% for lag 7).X

(3.7) Provide an estimate of the value of the time-series for 2015. (3)
Solution: For 2015, we have t = 19. To find an estimate of Y19 , we need to find Z19 and for
that, we need to find Z18 . We have

Z18 = Y18 − Y17 = 150.90 − 148.75 = 2.15.X

Then
Z19 = b0 + b1 Z18 = 7.73892 + 0.11449 × 2.15 = 7.98507.X
Now,
Z19 = Y19 − Y18 =⇒ Y19 = Z19 + Y18 = 7.98507 + 150.90 = 158.88507.X
Thus an estimate for the year 2015 is 158.88507.
[20]

TOTAL MARKS: [50]

END