Chapter 8 Forecasting

Chapter 8 Revision

1. Describe a typical decomposition of a time series. Which components would you describe
as systematic, and how would you describe the others?

2. When might you forecast one period ahead, and when might you forecast five periods
ahead, say? Give an example of each.

3. Deduce the weights in the simple exponential smoothing forecasting model, and show that
the weights are decreasing when 0 <  < 1.

4. The monthly demand for a product for 2007 is given in the table. Find the trend in the
monthly demand using exponential smoothing with a smoothing constant of 0.8

Month in 2007 Demand

January 480
February 500
March 630
April 790
May 910
June 1050
July 1200
August 1480
September 1730
October 1990
November 2260

5. Which components would you expect to be important in modelling the following time
series?
1. Sales of sunspecs
2. The price of metals
3. Share price of Keppel Corporation.

6. The following forecasting equation is proposed for a set of sales data, Xt.
Xt = Xt – 1 + (1 – )Xt – 2 + εt
where εt is an error term;  = 0.6. Ten observations are given below:

t 1 2 3 4 5 6 7 8 9 10
Xt 60 52 58 47 49 54 51 46 52 59

(a) Evaluate the one period ahead forecasts provided by the equation.
(b) Calculate the mean square error.
(c) Find an expression for an N period ahead forecast.
Evaluate a forecast made at time t = 10, for N = 4 periods ahead.

MM 2023-24 Kumaresan 1
 Chapter 8 Forecasting

7. The monthly sales of a luxury item in 2007 are as follows:

Month in 2007 Sales

January 36
February 32
March 31
April 29
May 30
June 33
July 34
August 31
September 29
October 28
November 29
December 35

(a) Find the forecast demand for each month up to January 2008, starting with April 2007 by
using a three-month moving average.

(b) Find the forecast demand for January 2008 by using exponential smoothing with a
smoothing constant of 0.9.

(c) On the assumption that a forecasting method is assessed by its mean absolute deviation,
which is the better of the two methods in parts (a) and (b)?

8. The annual amount spent on advertising by a company between 2000 and 2008 is shown
in the following table. (The advertising figures are all in millions of dollars)

Year 00 01 02 03 04 05 06 07 08
Advertising 3.1 5.1 4.4 4.9 6.3 5.2 6.9 7.6 8.4

(a) Using a 3-point moving average, smooth out this time series and hence estimate
graphically the trend values in 2009 and 2010.

(b) Compare these estimates with those obtained using exponential smoothing using a
smoothing constant of 0.4

MM 2023-24 Kumaresan 2
 Chapter 8 Forecasting

9. The numbers of drunk drivers stopped each day over a four-week period are shown in the
following table:

Week 1 2 3 4
Day
Monday 11 7 7 6
Tuesday 6 8 3 7
Wednesday 17 19 12 16
Thursday 12 15 8 11
Friday 48 41 50 47
Saturday 42 36 40 35
Sunday 28 31 26 24

(a) Using 7-point moving averages, isolate the trend for this time series.
Using a graphical method or otherwise, estimate the trend during each day in week 5.

(b) Assuming a multiplicative model, estimate the ‘seasonal (i.e. daily) variations’ for the
data, and hence forecast the number of drunk drivers stopped on each day in week 5.

10. Two forecasting equations are proposed for a set of sales data, Xt.
(1) Xt = Xt – 1 + t
(2) Xt =  + t – 1 + t
where t is an error term;  is 1.1,  is 90, and  is 0.2. Ten observations are given:

t Xt
1 105.1
2 112.8
3 99.6
4 104.9
5 110.2
6 107.7
7 115.8
8 118.1
9 113.5
10 120.3

Evaluate the one-period-ahead forecasts provided by each equation.

Which forecasting equation gives the lowest root mean squared error?

MM 2023-24 Kumaresan 3
 Chapter 8 Forecasting

11. The table below shows the total export orders (in $m) for a company during 2004-07:
Total exports ($ millions)
Year Jan-Apr May-Aug Sep-Dec
2004 28 37 52
2005 34 43 61
2006 34 46 64
2007 40 58 79

(a) Using three-point moving averages, isolate the trend.

(b) Estimate the seasonal variations using a multiplicative model, and thus forecast the value
of exports for the company during the three periods in 2008.
(c) Repeat part (b) but using an additive model.
(d) If the exports for 2008 turn out to be Jan-Apr: 48; May-Aug: 65; Sep-Dec: 85, then
calculate the Root Mean Square Errors for your 2008 forecasts obtained from (b) and (c)
above.

12. Consider the following two forecasting models:

(1) yt = 1yt – 1 + t , (2) yt = 2t + t
(where yt is the observation at time t and t is an error term).
 1 n yt
If 2 is estimated by  2 =  where there are n observations of the time series, then
n t 1 t

(a) Carefully write down a similar equation for an estimate of 1.

(b) Estimate 1 and 2 for the following time series.
Period 1 2 3 4 5 6 7 8 9 10
Observation 50 70 95 90 130 160 145 175 230 300
(c) Which model gives the lowest mean absolute percentage error (for periods two to ten)?

13. Briefly discuss how one might choose between a multiplicative and additive seasonal
forecasting model.

14. Write down the general formula for each of the following.
(a) An autoregressive model, AR(3), of order 3.
(b) A moving average model, MA(3), of order 3.
(c) A mixed autoregressive moving average model, ARIMA(3, 3).

15. For each of the following three time series A, B and C (recorded from year 2002 to 2011),
decide whether a 2-year, 3-year or 4-year moving average is most appropriate.
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
A 12 9 14 10 16 12 19 13 21 15
B 60 55 54 58 51 45 43 48 41 35
C 30 38 35 33 40 37 34 43 36 34

MM 2023-24 Kumaresan 4
 Chapter 8 Forecasting

Solutions to Revision Problems

1. A typical decomposition of a time series is:

trend
seasonal variations
cyclic variations
random variation.
The first three components listed are systematic, while random variation is referred to as
noise.

Using a decomposition approach, each component is estimated, and then these estimates
are re-combined to form a forecast.

The trend of a time series is the systematic increase (or decrease) of the variable over a
‘long’ period of time. The trend line is a smooth line, indicating the path of the time series,
ignoring the other three components. When forecasting, the trend is usually taken to be a
straight line, which is extrapolated (extended) to estimate the trend in the future. The trend
could be found by using exponential smoothing or by taking a moving average of the data,
a process that is called ‘smoothing’.

Seasonal variations in a time series are periodic patterns that complete themselves within
a year, and are then repeated in subsequent years.

Seasonal variations might be due to holidays, the weather, or financial year ends, for
example. In the UK, with its changing climate, we might expect domestic fuel
consumption to peak in the winter, and to drop to a low in the summer, so the time series
for domestic fuel consumption in the UK has seasonal variation.

A multiplicative seasonal forecasting model (Xt = Tt  St  Ct  Rt) would be chosen if

it was thought that the seasonal fluctuations are proportional to the level of the time series,
while an additive seasonal forecasting model (Xt = Tt + St + Ct + Rt) would be chosen if
it was thought that the seasonal fluctuations are of a constant size.

Cyclical variation is the component of a time series that oscillates about the trend line,
with the length of a cycle being longer than one year. The oscillations do not arise from
seasonal influences.

After trend, cyclical and seasonal variations have been eliminated from a time series, there
is still an unpredictable factor called random variation that follows a random pattern.
Random variation in a time series might be caused by unusual events that cannot be
forecasted.

**************************************************************

MM 2023-24 Kumaresan 5
 Chapter 8 Forecasting

2. You might forecast one period ahead when there is a lot of random variation in the time
series, and no discernible trend. Example: when forecasting a stock market index.

You might forecast five periods ahead, say, when there does appear to be a discernible
trend. Example: when forecasting the sales of a well-established magazine.

**************************************************************
3. Simple exponential smoothing model: Ft = Xt – 1 + (1 – )Ft – 1 ……….. (*)
Suppose 0 <  < 1.
Ft = Xt – 1 + (1 – )[Xt – 2 + (1 – )Ft – 2] , replacing t by t – 1 in (*)
= Xt – 1 + (1 – )Xt – 2 + (1 – )2Ft – 2
= Xt – 1 + (1 – )Xt – 2 + (1 – )2[Xt – 3 + (1 – )Ft – 3] , replacing t by t – 2 in (*)
= Xt – 1 + (1 – )Xt – 2 + (1 – )2Xt – 3 + (1 – )3Ft – 3
= Xt – 1 + (1 – )Xt – 2 + (1 – )2Xt – 3 + (1 – )3[Xt – 4 + (1 – )Ft – 4]
= Xt – 1 + (1 – )Xt – 2 + (1 – )2Xt – 3 + (1 – )3Xt – 4 + (1 – )4Ft – 4
This process carries on until we reach F1.
0 <  < 1. So, the weights , (1 – ), (1 – )2, (1 – )3, … are decreasing since the
common ratio is 1 – . where 0 < 1 –  < 1.
**************************************************************
4. T1 = X1 = 480, Tt = Tt – 1 + 0.8(Xt – Tt – 1) for t  2.
For example, T2 = T1 + 0.8(X2 – T1) = 480 + 0.8(500 – 480) = 496,
T3 = T2 + 0.8(X3 – T2) = 496 + 0.8(630 – 496) = 603.2

Month in 2007 Month Demand Trend

( = 0.8)
t Xt Tt
January 1 480 480
February 2 500 496
March 3 630 603.2
April 4 790 752.64
May 5 910 878.53
June 6 1050 1015.71
July 7 1200 1163.14
August 8 1480 1416.63
September 9 1730 1667.33
October 10 1990 1925.47
November 11 2260 2193.09

**************************************************************

MM 2023-24 Kumaresan 6
 Chapter 8 Forecasting

5.
Time Series Important Components
Seasonal (expect increase in summer and
Sales of sunspecs
holidays), trend
The price of metals Cyclical, trend, random variation
Share price of Keppel Corporation Trend, random variation

**************************************************************
6. (a) Xt = Xt – 1 + (1 – )Xt – 2 + t, where t = Xt – Ft and  = 0.6
 Xt = 0.6Xt – 1 + 0.4Xt – 2 + Xt – Ft
 Ft = 0.6Xt – 1 + 0.4Xt – 2
A suitable forecast for the equation is

Ft = 0.6Xt – 1 + 0.4Xt – 2 for t  3

F1 = X1 = 60, F2 = X2 = 52

For example, F3 = 0.6X2 + 0.4X1 = (0.6  52) + (0.4  60) = 55.2

t Xt Ft (Xt – Ft)2
1 60
2 52
3 58 55.2 7.84
4 47 55.6 73.96
5 49 51.4 5.76
6 54 48.2 33.64
7 51 52 1
8 46 52.2 38.44
9 52 48 16
10 59 49.6 88.36
11 56.2
  265

  X t  Ft 2
(b) MSE =
n
265
MSE = = 33.125
8

MM 2023-24 Kumaresan 7
 Chapter 8 Forecasting

(c) Ft = 0.6Xt – 1 + 0.4Xt – 2

Replace X by F for the purpose of finding Ft + N. Then, Ft = 0.6Ft – 1 + 0.4Ft – 2
Ft – 0.6Ft – 1 – 0.4Ft – 2 = 0 ……………………………………….. ()

AE: m2 – 0.6m – 0.4 = 0  (m – 1)(m + 0.4) = 0  m = 1 or m = –0.4

CF: Ft = A(1)t + B(–0.4)t = A + B(–0.4)t
The RHS of () is zero, so a particular solution is Ft = 0
The general solution of () is Ft = A + B(–0.4)t
We want: Ft + N = A + B(–0.4)t + N

Put N = 0: Ft = A + B(–0.4)t ……………………………………………… (1)

t–1 (0.4)t 5
Put N = –1: Ft – 1 = A + B(–0.4) =A+B = A – B(–0.4)t …… (2)
 0.4 2
7 2( Ft  Ft  1 )
(1) – (2): Ft – Ft – 1 = B(–0.4)t  B =
2 7(0.4)t

2( Ft  Ft  1 )
From (1), A = Ft –
7
2( Ft  Ft  1 ) 2( Ft  Ft  1 )
Hence, the solution we want is Ft + N = Ft – + (–0.4)N
7 7
Reverting back to X, the N-period-ahead forecast is:
2( X t  X t  1 ) 2( X t  X t  1 ) 2( X t  X t  1 )
Ft + N = Xt – + (–0.4)N = Xt + [(–0.4)N – 1]
7 7 7
2( X 10  X 9 )
With t =10 and N = 4, we obtain F14 = X10 + [(–0.4)4 – 1]
7
= 59 + 2[(–0.4)4 – 1] = 57.0512
Checking:
Ft = 0.6Xt – 1 + 0.4Xt – 2
So, F12 = 0.6X11 + 0.4X10 = 0.6F11 + 0.4X10 = 0.6(56.2) + 0.4(59) = 57.32
So, F13 = 0.6X12 + 0.4X11 = 0.6F12 + 0.4F11 = 0.6(57.32) + 0.4(56.2) = 56.872
So, F14 = 0.6X13 + 0.4X12 = 0.6F13 + 0.4F12 = 0.6(56.872) + 0.4(57.32) = 57.0512
****************************************************************

MM 2023-24 Kumaresan 8
 Chapter 8 Forecasting

7. (a) (b)

Month Actual 3-month Exponential

in Demand moving average smoothing
2007 forecast Error = forecast Error =
t Xt Ft Xt – Ft Ft Xt – Ft
Jan 1 36 36
Feb 2 32 36 4
Mar 3 31 32.4 1.4
Apr 4 29 33 4 31.14 2.14
May 5 30 30.667 0.667 29.214 0.786
Jun 6 33 30 3 29.921 3.079
July 7 34 30.667 3.333 32.692 1.308
Aug 8 31 32.333 1.333 33.869 2.869
Sep 9 29 32.667 3.667 31.287 2.287
Oct 10 28 31.333 3.333 29.229 1.229
Nov 11 29 29.333 0.333 28.123 0.877
Dec 12 35 28.667 6.333 28.912 6.088
Jan 08 13 30.667 34.391

3-month moving average forecast = average of 3 most recent observations

X1  X 2  X 3 36  32  31
For example, F4 = = = 33
3 3

For exponential smoothing,

Predicted sales: F1 = X1 = 36, Ft = Ft – 1 + 0.9(Xt – 1 – Ft – 1) for t  2.
For example, F2 = F1 + 0.9(X1 – F1) = 36 + 0.9(36 – 36) = 36,
F3 = F2 + 0.9(X2 – F2) = 36 + 0.9(32 – 36) = 32.4

Forecast sales for January 2008  34.391

 X t  Ft 26
(c) For 3-month moving average, MAD = =  2.89
n 9

20 .66
For exponential smoothing, MAD   2.30
9

The MAD is lower for the exponential smoothing forecast, so exponential smoothing with
 = 0.9 is the better of the two forecasting methods for the given time series.

**************************************************************

MM 2023-24 Kumaresan 9
 Chapter 8 Forecasting

X t 1  X t  X t 1
8. (a) The 3-point moving average trend is given by Tt =
3
X1  X 2  X 3 3.1  5.1  4.4
For example, T2 = = = 4.2 (trend for 2001)
3 3
(a) (b)
Advertising 3-year moving Exp Smoothing
Year average trend Forecast ( = 0.4)
t Xt Tt Ft
2000 1 3.1 3.1
2001 2 5.1 4.2 3.1
2002 3 4.4 4.8 3.9
2003 4 4.9 5.2 4.1
2004 5 6.3 5.46667 4.42
2005 6 5.2 6.13333 5.172
2006 7 6.9 6.56667 5.1832
2007 8 7.6 7.63333 5.86992
2008 9 8.4 6.561952
2009 10 7.2971712

Extrapolating a best-fit line (by inspection), I estimate the 3-point trend to be 8.5 in 2009,
and 9.0 in 2010.

(b) F1 = X1 = 3.1, Ft = Ft – 1 + 0.4(Xt – 1 – Ft – 1) for t  2.

For example, F2 = F1 + 0.4(X1 – F1) = 3.1 + 0.4(3.1 – 3.1) = 3.1,
F3 = F2 + 0.4(X2 – F2) = 3.1 + 0.4(5.1 – 3.1) = 3.9
Forecast (Ft), using exponential smoothing with  = 0.4, is in the 5th column of table.
F11 = F10 + 0.4(X10 – F10) = F10 + 0.4(F10 – F10) = F10  7.30. So, F11  F10  7.30

9. (a) For a 7-point moving average, the trend is

MM 2023-24 Kumaresan 10
 Chapter 8 Forecasting

X t  3  X t  2  X t 1  X t  X t 1  X t  2  X t  3
Tt =
7
Example:
X1  X 2  X 3  X 4  X 5  X 6  X 7
T4 =
7
11  6  17  12  48  42  28
=  23.42857
7

Week Day # drunks Trend

t Xt Tt
Mon 1 11
Tue 2 6
Wed 3 17
1 Thu 4 12 23.42857
Fri 5 48 22.85714
Sat 6 42 23.14286
Sun 7 28 23.42857
Mon 8 7 23.85714
Tue 9 8 22.85714
Wed 10 19 22
2 Thu 11 15 22.42857
Fri 12 41 22.42857
Sat 13 36 21.71429
Sun 14 31 20.71429
Mon 15 7 19.71429
Tue 16 3 21
Wed 17 12 21.57143
3 Thu 18 8 20.85714
Fri 19 50 20.71429
Sat 20 40 21.28571
Sun 21 26 21.85714
Mon 22 6 22.28571
Tue 23 7 21.85714
Wed 24 16 21.14286
4 Thu 25 11 20.85714
Fri 26 47
Sat 27 35
Sun 28 24

20 .85714  23.42857
Trend is = –0.1224 per season, referring to table.
25  4
The latest trend figure we have is 20.857 on Thursday in week 4. 29 – 25 = 4

MM 2023-24 Kumaresan 11
 Chapter 8 Forecasting

Season Estimated trend

Monday 20.857 + 4(–0.1224)  20.37
Tuesday 20.37 – 0.1224  20.25
Wednesday 20.25 – 0.1224  20.13
Thursday 20.13 – 0.1224  20.01
Friday 20.01 – 0.1224  19.89
Saturday 19.89 – 0.1224  19.77
Sunday 19.77 – 0.1224  19.65

Xt X 12
(b) Multiplicative model: Seasonal ratio = . Example: S4 = 4 =  0.512
Tt T4 23.42857

Seasonal ratios
Mon Tue Wed Thu Fri Sat Sun
Week 1 0.512 2.1 1.815 1.195
Week 2 0.293 0.35 0.864 0.669 1.828 1.658 1.497
Week 3 0.355 0.143 0.556 0.384 2.414 1.879 1.190
Week 4 0.269 0.320 0.757 0.527
Average
A B C D E F G
Seasonal
0.306 0.271 0.726 0.523 2.114 1.784 1.294
Ratio
A B C D E F G
Seasonal       
H H H H H H H
variation
0.305 0.270 0.724 0.522 2.109 1.779 1.291

A B  C  D E  F G 7 .018
H= 
7 7

Forecast for nth season = Seasonal variation  Estimated trend Multiplicative model

Season Forecast
Monday 0.305  20.37  6.2
Tuesday 0.270  20.25  5.5
Wednesday 0.724  20.13  14.6
Thursday 0.522  20.01  10.4
Friday 2.109  19.89  41.9
Saturday 1.779  19.77  35.2
Sunday 1.291  19.65  25.4

**************************************************************

MM 2023-24 Kumaresan 12
 Chapter 8 Forecasting

10. (1) Xt = Xt – 1 + t , where  = 1.1 and t = Xt – Ft

So, Xt = 1.1Xt – 1 + Xt – Ft
So, Ft = 1.1Xt – 1
The forecasting equation for (1) is Ft = 1.1Xt – 1 for t  2

For example, F2 = 1.1X1 = 1.1  105.1 = 115.61

(2) Xt =  + t – 1 + t , where  = 90,  = 0.2 and t = Xt – Ft

So, Xt = 90 + 0.2(Xt – 1 – Ft – 1) + Xt – Ft
So, Ft = 90 + 0.2(Xt – 1 – Ft – 1)
The forecasting equation for (2) is Ft = 90 + 0.2(Xt – 1 – Ft – 1) for t  2
F1 = X1 = 105.1

For example, F2 = 90 + 0.2(X1 – F1) = 90 + 0 = 90

F3 = 90 + 0.2(X2 – F2) = 90 + 0.2(112.8 – 90) = 94.56

Forecast (1) Forecast (2)

t Xt Ft (Xt – Ft)2 Ft (Xt – Ft)2
1 105.1 105.1
2 112.8 115.61 7.8961 90 519.84
3 99.6 124.08 599.2704 94.56 25.4016
4 104.9 109.56 21.7156 91.008 192.9877
5 110.2 115.39 26.9361 92.7784 303.5121
6 107.7 121.22 182.7904 93.4843 202.0856
7 115.8 118.47 7.1289 92.8431 527.0176
8 118.1 127.38 86.1184 94.5914 552.6556
9 113.5 129.91 269.2881 94.7017 353.3751
10 120.3 124.85 20.7025 93.7597 704.3899
11 132.33 95.3081
  1221.847   3381.265

  X t  Ft  2
RMSE =
9
1
For forecast (1), RMSE  1221.847  11.65
3

1
For forecast (2), RMSE  3381.265  19.38
3

Hence, forecasting equation (1) gives the lower RMSE.

MM 2023-24 Kumaresan 13
 Chapter 8 Forecasting

X t 1  X t  X t 1
11. (a) The 3-point moving average trend is given by Tt =
3
X1  X 2  X 3 28  37  52
For example, T2 = = = 39
3 3

Season Period Time Series Trend

t Xt Tt
1 (2004) 1 28
2 (2004) 2 37 39
3 (2004) 3 52 41
1 (2005) 4 34 43
2 (2005) 5 43 46
3 (2005) 6 61 46
1 (2006) 7 34 47
2 (2006) 8 46 48
3 (2006) 9 64 50
1 (2007) 10 40 54
2 (2007) 11 58 59
3 (2007) 12 79

Xt X 37
(b) Multiplicative model: Seasonal ratio = . Example: S2 = 2 =  0.94872
Tt T2 39

Seasonal ratios
Jan-Apr May-Aug Sep-Dec
2004 0.94872 1.26829
2005 0.79070 0.93478 1.32609
2006 0.72340 0.95833 1.28
2007 0.74074 0.98305
Average seasonal ratio A  0.7516 B  0.9562 C  1.2915
A B C
Seasonal variation  0.7518  0.9564  1.2918
H H H

A BC 2.9993
H= 
3 3
59  39
Trend is = 2.2222 per season, from table in (a).
11  2
The latest trend figure we have is 59 in the 2nd season in 2007. 13 – 11 = 2

Estimated trend in 1st season, 2008 is 59 + 2(2.2222)  63.4444

Estimated trend in 2nd season, 2008 is 63.4444 + 2.2222  65.6667
Estimated trend in 3rd season, 2008 is 65.6667 + 2.2222  67.8889

MM 2023-24 Kumaresan 14
 Chapter 8 Forecasting

Forecast for nth season = Seasonal variation  Estimated trend Multiplicative model

Forecast for 1st season, 2008 is 0.7518  63.4444  47.70

Forecast for 2nd season, 2008 is 0.9564  65.6667  62.80
Forecast for 3rd season, 2008 is 1.2918  67.8889  87.70

(c)
Seasonal difference Xt – Tt
Jan-Apr May-Aug Sep-Dec
2004 –2 11
2005 –9 –3 15
2006 –13 –2 14
2007 –14 –1
Average seasonal difference A = –12 B = –2 C  13.3333
A–H B–H C–H
Seasonal variation
 –11.7778  –1.7778  13.5556

A BC  0.6667
Additive model: H = 
3 3

Forecast for nth season = Seasonal variation + Estimated trend … Additive Model

Forecast for 1st season, 2008 is –11.7778 + 63.4444  51.67

Forecast for 2nd season, 2008 is –1.7778 + 65.6667  63.89
Forecast for 3rd season, 2008 is 13.5556 + 67.8889  81.44

  X t  Ft  2
(d) RMSE =
3

2 2 2
(48  47.70)  (65  62.80)  (85  87.70)
For multiplicative model, RMSE =  2.02
3

2 2 2
(48  51.67)  (65  63.89)  (85  81.44)
For additive model, RMSE =  3.02
3

**************************************************************

MM 2023-24 Kumaresan 15
 Chapter 8 Forecasting

12. (a) (2) yt = 2t + t 2 is unknown, so we assume t = 0

 1 n yt
 yt  2t  2 
yt
t

and we are told that 2 
n t 1 t

Hence, an estimate for m is the average of all available values for m.

(1) yt = 1yt – 1 + t 1 is unknown, so we assume t = 0

 1 n yt
 yt  1yt – 1  1 
yt
yt  1
and so 1   (since y0 does not exist)
n  1 t  2 yt  1
(b)
Time t 1 2 3 4 5 6 7 8 9 10
Observation yt 50 70 95 90 130 160 145 175 230 300

 1 10 yt 1  70 95 90 130 160 145 175 230 300 

1   =         
9 t  2 yt  1 9  50 70 95 90 130 160 145 175 230 
  1.2346

 1 10 yt 1  50 70 95 90 130 160 145 300 

2  
10 t  1 t
=  
10  1 2
  
3 4 5

6

7
 ... 
10 
  28.9978

(c) Let Ft be the forecast for period t

Model 1: yt = 1yt – 1 + t  yt  1.2346yt – 1 + yt – Ft  Ft = 1.2346yt – 1
Model 2: yt = 2t + t  yt  28.9978t + yt – Ft  Ft = 28.9978t
Period Model 1 Model 2
t yt Ft % error Ft % error
1 50 28.9978
2 70 61.73 11.814 57.9956 17.149
3 95 86.422 9.029 86.9934 8.428
4 90 117.287 30.319 115.9912 28.879
5 130 111.114 14.528 144.989 11.53
6 160 160.498 0.311 173.9868 8.742
7 145 197.536 36.232 202.9846 39.989
8 175 179.017 2.295 231.9824 32.561
9 230 216.055 6.063 260.9802 13.470
10 300 283.958 5.347 289.978 3.341
Mean  12.9 Mean  18.2

61.73  70
Model 1: Absolute percentage error for period 2 is  100%  11.814%
70
The mean absolute percentage errors (for periods two to ten) are
1 1
(115.94)  12.9% for Model 1 and (164.09)  18.2% for Model 2,
9 9
Hence, model 1 gives the lowest mean absolute percentage error (for periods two to ten).

MM 2023-24 Kumaresan 16
 Chapter 8 Forecasting

13. A multiplicative seasonal forecasting model (Xt = Tt  St  Ct  Rt) would be chosen if

it was thought that the seasonal fluctuations are proportional to the level of the time series.

Tt , St are trend and seasonal components at time t,

Ct , Rt are cycle and random components at time t.
If we ignore Ct and Rt, then we obtain Xt = Tt  St

Xt
This re-arranges to give the seasonal ratio St = t
Tt
We hope that the sum of the average seasonal ratios for the n seasons will be close to n.

An additive seasonal forecasting model (Xt = Tt + St + Ct + Rt) would be chosen if it was

thought that the seasonal fluctuations are of a constant size.

If we ignore Ct and Rt, then we obtain Xt = Tt + St

This re-arranges to give the seasonal difference St = Xt – Tt t
We hope that the sum of the average seasonal differences for the n seasons will be close
to 0.
**************************************************************
14. (a) AR(3): Yt =  0 +  1Yt – 1 +  2Yt – 2 +  3Yt – 3 + t

(b) MA(3): Yt =  + t –  1t – 1 –  2t – 2 –  3t – 3

(c) ARIMA(3, 3): Yt =  0 +  1Yt – 1 +  2Yt – 2 +  3Yt – 3 + t –  1t – 1 –  2t – 2 –  3t – 3

**************************************************************
15.
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
A 12 9 14 10 16 12 19 13 21 15
change –3 +5 –4 +6 –4 +7 –6 +7 –6
B 60 55 54 58 51 45 43 48 41 35
change –5 –1 +4 –7 –6 –2 +5 –7 –6
C 30 38 35 33 40 37 34 43 36 34
change +8 –3 –2 +7 –3 –3 +9 –7 –2

Series Recommended m for an m-year moving average

2-year moving average, since there is a pattern that repeats itself every 2
A
years. The repeating sign pattern is of the form (– +).
4-year moving average, since there is a pattern that repeats itself every 4
B
years. The repeating sign pattern is of the form (– – + –).
3-year moving average, since there is a pattern that repeats itself every 3
C
years. The repeating sign pattern is of the form (+ – –).

MM 2023-24 Kumaresan 17