FORECASTING

INDEX NUMBERS
 INTRODUCTION TO FORECASTING
• After having studied the historical pattern of a time series, if there is
reason to believe that the most important features of the variable do not
change in the future, we can project the revealed pattern into the future
in order to develop forecasts.

• If a time series exhibits no (or hardly any) trend, cyclical and seasonal
variations, exponential smoothing can provide a useful forecast for one
period ahead:
Ft 1  S t

Ex 1:
We have applied exponential smoothing with w = 0.2 and w = 0.7 on quarterly
Australian unemployed persons (in thousands).
Since this time series does have some seasonal variations, exponential
smoothing cannot be expected to forecast unemployment reasonably well.
Nevertheless, just for illustration, let us forecast unemployment for the first
quarter of 1999.
2
 unemployed S (w=0.7)
1998 1 2461.4 2402.8
2 2210.9 2268.5
3 2221.3 2235.5 This is the smoothed value for the fourth
4 2102.6 2142.5 quarter of 1998, and thus the forecast for
1999 1 na the first quarter of 1999.

• If a time series exhibits a long-term (linear) trend and seasonal

variations, we can use regression analysis to develop forecasts in two
different ways.
1) We can forecast using the estimated trend and seasonal indices
as:
Ft  Tt  S t  ( ˆ 0  ˆ1t )  I t
2) Alternatively, we can forecast using the estimated multiple
regression model with a time variable and seasonal dummy
variables. This second approach is beyond the scope of your
syllabus.

3
Ex 2: Forecast retail turnover for households goods for the first quarter of
2001 applying the first approach can be implemented as follows.

Obtain the trend estimate from part a and the March seasonal
index from part b so that
t = 76, I76 = IMar = 0.930 and yˆ  1589.189  36.604t

F76  yˆ76  (1589.2  36.6  76)  0.930  4064.8

• We have predicted retail turnover for households goods for the first
quarter of 2001. Suppose we had another forecast value of 4203.4 for
the same data and the same time period using a different forecasting
model. How would we decide which forecast is more accurate?

4
 In retrospect it is easy to answer this question, we just have to calculate
the forecast error for each forecasting model.

The difference between the

actual and forecast values, i.e. et  yt  Ft

(Ex 2)
f) Suppose that one quarter passed since we predicted retail turnover for
households goods for the first quarter of 2001. The actual value of retail
turnover for households goods in this quarter was 4277.1 $m. Compare the
two forecasts from part e. Which model proved to be more accurate?
The forecast errors for the first quarter of 2001 (t = 76) are the following.
Model 1 : e76  y76  F76  4277.1  4064.8  212.3

Model 2 : e76  y76  F76  4277.1  4203.4  73.7

Both forecast errors are positive, i.e. both models overestimate retail
turnover for the first quarter of 2001, but the forecast from Model 2 is more
accurate.
5
 However, this does not imply by any means that Model 2 would produce
more accurate forecast for all time periods than Model 1.

• How can we decide which forecasting model is the most accurate in a

given situation?
Forecast the variable of interest for a number of of time periods using
alternative models and evaluate some measure(s) of forecast accuracy
for each of these models.
Among a number of possible criteria that can be used for this purpose
the two most commonly used are

1 n
Mean absolute deviation: MAD   yt  Ft
n t 1

1 n
Sum of squares of forecast error: SSFE    yt  Ft 
2

n t 1

Note: SSFE is the better measure if relatively large errors are to be penalised.

6
Ex 3:
Two forecasting models were used to predict the future values of a time series.
They are shown next, together with the actual values. For each model,
calculate MAD and SSFE to determine which was more accurate.
Ft et | et | et2
yt
Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Model 1 Model 2
6.0 7.5 6.3 -1.5 -0.3 1.5 0.3 2.25 0.09
6.6 6.3 6.7 0.3 -0.1 0.3 0.1 0.09 0.01
7.3 5.4 7.1 1.9 0.2 1.9 0.2 3.61 0.04
9.4 8.2 7.5 1.2 1.9 1.2 1.9 1.44 3.61

Total: 4.9 2.5 7.39 3.75

6.0 – 7.5 | -1.5 | (-1.5) 2

Model 1 : MAD = 4.9/4=1.225 and SSFE = 7.39/4=1.8475

Model 2 : MAD = 2.5/4=0.625 and SSFE = 3.75/4=0.9375

According to both criteria Model 2 is the more accurate.

7
 INDEX NUMBERS

• Index number: a descriptive measure of the relative change in

a time series.

– Data are collected at regular intervals

(e.g. annually, yearly, quarterly, monthly),
– at the same time (e.g. end of month, close of trade),
– for a single entity (e.g. country, family),
– using a fixed definition of the variable
(dollar amounts are usually satisfactory, the physical units
may not have a precise fixed meaning), and
– using a fixed method of data collection
(it should not vary from period to period).

Calendar year Financial year

8
An index number is a ratio (often in percentage form) of one value to another.

Why to use index numbers?

a) They are an important means of observing the ‘general movement’
in a time series
(e.g. we can easily check the percentage increase or decrease
since some fixed point in time).
b) They may also assist in comparing two or more time series
(e.g. we can use index numbers to average the changes in different
consumer prices).

How to construct index numbers?

– Calculations are conceptually simple and, providing we are
dealing with a single series, they are easy.
– Use of a spreadsheet is probably the best means of
calculation.
– dX will do it automatically for you if you are dealing with a
single time-series.
(Highlight the data then use Series/Index.)
9
 There are two types of index numbers.

Simple: Aggregate:
for a single time series for several time series

• Simple index number: it shows how the current value of a given

variable compares to its base period value.

– Say the time series is Yt (denoting the value of Y in period t);

– Choose a base period (t = 0);
– Index number for the base period: I0 = 100.0 (or 1.00);
– For other periods in general:

Index for period t Yt Period t value

I t ,0  100 (%)
with base period 0 Y0
Base period value

10
Note: 1) The absolute change in Y from the base period to period t is Yt – Y0 .

2) The percentage change in Y from the base period to period t is:

Yt  Y0
 100 (%)  I t,0  100 (%)
Y0
Ex 4:
Index 2254
Year Data I 0, 0   100
(%) 2254
Base period,
t=0 1995 2254 100.0
2400
t =1 1996 2400 106.5 I 1, 0   100
2254
t =2 1997 2585 114.7
2585
I 2,0   100
2254
It indicates that the time series
(Y) has increased by 6.5 % An increase of 14.7%
between 1995 and 1996. between 1995 and 1997.

11
 3) The percentage change in Y from period t1 to period t2 is:
Yt2  Yt1 I t2 ,0  I t1 ,0
100 (%)  100 (%)
Yt1 I t1 ,0

(Ex 4) The percentage change in Y from 1996 to 1997 is:

114.7  106.5
100 (%)  7.7%
106.5

4) The difference between index numbers in periods t1 and t2 is:

Yt2  Yt1 It compares the absolute
I t2 ,0  I t1 ,0  100 (%) change in Y from t1 to t2
Y0 to its base period value.

(Ex 4) For 1996 and 1997: 114.7 106.5  8.2

indicating that between 1996 and 1997 the index number changed
by 8.2 points.

12
 5) Changing the base period.
Suppose e.g. that we intend to compare two sets of index numbers
that have different base periods.
If t = 0 is the original base period and t = n is the new base period,
then
Yt I t ,0 Old index for
New index I t ,n  100(%)  100(%) period t
number for Yn I n,0 Old index for
period t period 0

(Ex 4) Let the new base period be 1996.

Year Data Old New

Index Index
100.0
1995 2254 100.0 93.9 I 0,1   100
106.5
1996 2400 106.5 100.0
114.7
1997 2585 114.7 107.7 I 2,1   100
106.5
13
 6) Splicing (joining) related series of index numbers.
Suppose that we have two sets of index numbers that have different
base periods, e.g. an ‘old’ series with base t = 0 and a ‘new’ series
with base t = n.
If in period n they overlap, the spliced series for t = n, n+1, …, is:

I nold
I tspliced
,n  I t ,n 
new ,0
(%)
100

Ex 5:
Year Old New Spliced They are the
same.
1996 129.9 n.a. 129.9
Overlap
1997 134.5 n.a. 134.5
139.8
1998 139.8 100.0 139.8 104.2 
100
1999 n.a. 104.2 145.7
139.8
2000 n.a. 106.7 149.2 106.7 
100
n.a.: not available 14
 • Aggregate index number: it shows the relative change for an
aggregate variable.

E.g. An aggregate price index shows how the current price for a group of items
compares to the base period price for the same group of items.

Unweighted aggregate price index: Weighted aggregate price index:

n n

P i ,t w P
i ,t i , t
I t ,0  i 1
n
100 (%) I t ,0  i 1
n
100 (%)
P
i 1
i ,0 w
i 1
P
i ,0 i ,0

Number Price of item i Weight Weight

of items. in period t. assigned to item assigned to item
i in period 0. i in period t.
Price of item i
in period 0.
15
• The unweighted aggregate price index is not really useful because
– it depends on the units individual items are measured at;
– it treats all items in the group equally important.

• How to choose the weights for the weighted aggregate price index?
– The weights should represent the importance of individual
items.
The importance of an item can be measured by the
volume of transaction, e.g. Q : quantity purchased,
in physical units.
Pi,t Qi,t : value of the transaction for item i in period t.

– If the base and current period quantities are not the same, the
index number will reflect the overall change that have occurred
in the aggregate value, due to changes in prices or/and
quantities.
In order to obtain a price index, and not an index of
expenditure, the quantities have to be fixed.

16
 Which set of quantities should be used?

Base period quantities (Qi,0) Current period quantities (Qi,t)

Laspeyres index: Paasche index:

n n

P Q i ,t i ,0 P Q i ,t i ,t
I tL,0  i 1
n
100 (%) I tP, 0  i 1
n
100 (%)
P
i 1
i ,0 Qi ,0 P
i 1
i ,0 Qi ,t

Total expenditure Total expenditure

Total base period needed to buy the needed to buy the Total current
expenditure. base period current period period
quantities on current quantities on base expenditure.
period prices. period prices.

17
Ex 6:
The following table shows the average prices ($) and monthly quantities
purchased for two products, A and B, and for two years, 1990 and 2000:

Product P1990 P2000 Q1990 Q2000

A 3.45 4.35 4567 4950

B 15.50 21.20 2444 3421

a) Calculate the Laspeyres index for 2000, using 1990 as the base.
4.35  4567  21.20  2444
,1990  100  133.6%
L
I 2000
3.45  4567  15.50  2444
Purchasing the 1990 quantities would cost 33.6% more in 2000
than in 1990.

18
b) Calculate the Paasche index for 2000, using 1990 as the base.

4.35  4950  21.20  3421

,1990  100  134.1%
P
I 2000
3.45  4950  15.50  3421
Purchasing the 2000 quantities would cost 34.1% higher in 2000
than in 1990.

Note:
– The Laspeyres index (133.6) is lower than the Paasche index (134.1).
This is usually the case (see next page).
– Apparently, with two products and two years it is easy to calculate
these index numbers. However, with many more products and many more
time periods these calculations become extremely tedious.
– Use a spreadsheet such as Excel.

19
 Which index number is better?

Properties of the Laspeyres index:

+ Since in each period it uses the base period quantities, it requires us
to observe only the prices for each period.
– The use of fixed base period quantities is unrealistic in the long run
when the consumption patterns change in response to price
changes.

Properties of the Paasche index:

+ Since in each period it uses the current period quantities, it always
reflects the actual consumption pattern, i.e. it is ‘up-to-date’.
– It requires us to observe both the prices and the quantities for each
period.

Moreover, in a period of rising prices and changing consumption patterns, the

Laspeyres index tends to overestimate, while the Paasche index tends to
underestimate the overall price rise.
The ‘truth’ is somewhere in between.
20