Solutions of Forecast: Methods and Applications (1998)

Download as pdf or txt
Download as pdf or txt
You are on page 1of 83

D/Solutions to exercises

Chapter 1: The forecasting perspective

1.1 Look for pragmatic applications in the real world. Note that there are no fixed
answers in this problem.

(a) Dow theory: There is an element of belief that past patterns will continue
into the future. So first, look for the patterns (support and resistance levels)
and then project them ahead for the market and individual stocks. This is a
quantitative time series method.
(b) Random walk theory: This is quantitative, and involves a time series rather
than an explanatory approach. However, the forecasts are very simple because
of the lack of any meaningful information. The best prediction of tomorrow’s
closing price is today’s closing price. In other words, if we look at first differences
of closing prices (i.e., today’s closing price minus yesterday’s closing price) there
will be no pattern to discover.
(c) Prices and earnings: Here instead of dealing with only one time series (i.e., the
stock price series) we look at the relation between stock price and earnings per
share to see if there is a relationship—maybe with a lag, maybe not. There-
fore this is an explanatory approach to forecasting and would typically involve
regression analysis.

1.2 Step 1: Problem definition This would involve understanding the nature of the indi-
vidual product lines to be forecast. For example, are they high-demand prod-
ucts or specialty biscuits produced for individual clients? It is also important
to learn who requires the forecasts and how they will be used. Are the forecasts
to be used in scheduling production, or in inventory management, or for bud-
getary planning? Will the forecasts be studied by senior management, or by
the production manager, or someone else? Have there been stock shortages so
that demand has gone unsatisfied in the recent past? If so, would it be better
to try to forecast demand rather than sales so that we can try to prevent this

76
Chapter 1: The forecasting perspective 77

happening again in the future? The forecaster will also need to learn whether
the company requires one-off forecasts or whether the company is planning on
introducing a new forecasting system. If the latter, are they intending it to
be managed by their own employees and, if so, what software facilities do they
have available and what forecasting expertise do they have in-house?
Step 2: Gathering information It will be necessary to collect historical data on each
of the product lines we wish to forecast. The company may be interested in
forecasting each of the product lines for individual selling points. If so, it is
important to check that there are sufficient data to allow reasonable forecasts
to be obtained. For each variable the company wishes to forecast, at least a
few years of data will be needed.
There may be other variables which impact the biscuit sales, such as economic
fluctuations, advertising campaigns, introduction of new product lines by a
competitor, advertising campaigns of competitors, production difficulties. This
information is best obtained by key personnel within the company. It will be
necessary to conduct a range of discussions with relevant people to try to build
an understanding of the market forces.
If there are any relevant explanatory variables, these will need to be collected.
Step 3: Preliminary (exploratory) analysis Each series of interest should be graphed
and its features studied. Try to identify consistent patterns such as trend
and seasonality. Check for outliers. Can they be explained? Do any of the
explanatory variables appear to be strongly related to biscuit sales?
Step 4: Choosing and fitting models A range of models will be fitted. These models
will be chosen on the basis of the analysis in Step 3.
Step 5: Using and evaluating a forecasting model Forecasts of each product line will
be made using the best forecasting model identified in Step 4. These forecasts
will be compared with expert in-house opinion and monitored over the period
for which forecasts have been made.
There will be work to be done in explaining how the forecasting models work
to company personnel. There may even be substantial resistance to the in-
troduction of a mathematical approach to forecasting. Some people may feel
threatened. A period of education will probably be necessary.
A review of the forecasting models should be planned.
78 Part D. Solutions to exercises

Chapter 2: Basic forecasting tools

2.1 (a) One simple answer: choose the mean temperature in June 1994 as the forecast
for June 1995. That is, 17.2 ◦ C.
(b) The time plot below shows clear seasonality with average temperature higher
in summer.

20
18
16
Celsius

14
12
10
8
6

1994 1994 1994 1994 1994 1994 1995 1995 1995


Jan Feb May Jul Sep Nov Jan Mar May

Month

Exercise 2.1(b): Time plot of average monthly temperature in Paris (January 1994–May
1995).

2.2 (a) Rapidly increasing trend, little or no seasonality.


(b) Seasonal pattern of period 24 (low when asleep); occasional peaks due to stren-
uous activity.
(c) Seasonal pattern of period 7 with peaks at weekends; possibly also peaks during
holiday periods such as Easter or Christmas.
(d) Strong seasonality with a weekly pattern (low on weekends) and a yearly pat-
tern. Peaks in either summer (air-conditioning) or winter (heating) or both
depending on climate. Probably increasing trend with variation increasing with
trend.

2.3 (a) Smooth series with several large jumps or direction changes; very large range
of values; logs help stabilize variance.
(b) Downward trend (or early level shift); cycles of about 15 days; outlier at day
8; no transformation necessary.
(c) Cycles of about 9–10 years; large range and little variation at low points indi-
cating transformation will help; logs help stabilize variance.
Chapter 2: Basic forecasting tools 79

(d) No clear trend; seasonality of period 12; high in July; no transformation neces-
sary.
(e) Initial trend; level shift end of 1982; seasonal period 4 (high in Q2 and Q3, low
in Q1); no transformation necessary.

2.4 1-B, 2-A, 3-D, 4-C. The easiest approach to this question is to first identify D.
Because it has a peak at lag 12, the time series must have a pattern of period 12.
Therefore it is likely to be monthly. The slow decay in plot D shows the series has
trend. The only series with both trend and seasonality of period 12 is Series 3. Next
consider plot C which has a peak at lag 10. Obviously this cannot reflect a seasonal
pattern since the only series remaining which is seasonal is series 2 and that has
period 12. Series 4 is strongly cyclic with period approximately 10 and series 1 has
no seasonal or strong cyclic patterns. Therefore C must correspond to series 4. Plot
A shows a peak at lag 12 indicating seasonality of period 12. Therefore, it must
correspond with series 2. That leaves plot B aligned with series 1.

2.5 (a)
X Y
Mean 52.99 43.70
Median 52.60 44.42
MAD 3.11 2.47
MSE 15.94 8.02
St.dev. 4.14 2.94
(b) Mean and median give a measure of center; MAD, MSE and St.dev. are mea-
sures of spread.
(c) r = −0.660. See plot on next page.
(d) It is inappropriate to compute autocorrelations since there is no time component
to these data. The data are from 14 different runners. (Autocorrelation would
be appropriate if they were data from the same runner at 14 different times.)

2.6 (a) See plot on following page.


(b) and (c)
Notation: Error 1 = (actual demand) − (method 1 forecast)
Error 2 = (actual demand) − (method 2 forecast)
80 Part D. Solutions to exercises

48
Y: maximal aerobic capacity

46
44
42
40

48 50 52 54 56 58 60

X: running times

Exercise 2.5(c): Plot of running times versus maximal aerobic capacity.


240
220
200
Demand

180

Actual
160

Forecast Method 1
Forecast Method 2
140

5 10 15 20

Month

Exercise 2.6(a): Time plots of data and forecasts.


Chapter 2: Basic forecasting tools 81

Period Actual Method 1 Error 1 Method 2 Error 2


1 139 157 −18 170 −31
2 137 145 −8 162 −25
3 174 140 34 157 17
4 142 162 −20 173 −31
5 141 149 −8 164 −23
6 162 144 18 158 4
7 180 156 24 166 14
8 164 172 −8 179 −15
9 171 167 4 177 −6
10 206 169 37 180 26
11 193 193 0 199 −6
12 207 193 14 202 5
13 218 202 16 211 7
14 229 213 16 221 8
15 225 223 2 232 −7
16 204 224 −20 235 −31
17 227 211 16 225 2
18 223 221 2 232 −9
19 242 222 20 233 9
20 239 235 4 243 −4

Analysis of errors ME 6.25 −4.80


(periods 1–20) MAE 14.45 14.29
MSE 307.25 294.00
MPE 2.55 −3.61
MAPE 7.87 8.24
Theil’s U 0.94 0.85

On MAE and MSE, Method 2 is better than Method 1. On MAPE, Method 1


is better than Method 2. Note that this is different from the conclusion drawn
in Section 4/2/3 where these two methods are compared. The difference is that
we have used a different time period over which to compare the results. Holt’s
method (Method 2) performs quite poorly at the start of the series. In Chapter
4, this period is excluded from the analysis of errors.

2.7 (a) Changes: −0.25, −0.26, 0.13, . . . , −0.09, −0.77. There are 78 observations in
the DOWJONES.DAT file. Therefore there are 77 changes.
(b) Average change: 0.1336. So the next 20 changes are each forecast to be 0.1336.
(c) The last value of the series is 121.23. So the next 20 are forecast to be:

X̂79 = 121.23 + 0.1336 = 121.36


X̂80 = 121.36 + 0.1336 = 121.50
X̂81 = 121.50 + 0.1336 = 121.63 etc.
82 Part D. Solutions to exercises

In general, X̂79+h = 121.23 + h(0.1336).


(d) See the plot below.
120
Dow Jones index

115
110

0 20 40 60 80 100

day

Exercise 2.7(d): Plot of Dow Jones index (DOWJONES.DAT)

1 Pn
(e) The average change is c = n−1 t=2 (Xt − Xt−1 ) and the forecasts are X̂n+h =
Xn + hc. Therefore,
n
1 X
X̂n+h = Xn + h (Xt − Xt−1 )
n − 1 t=2
h
= Xn + (Xn − X1 ).
n−1
This is a straight line with slope equal to (X n − X1 )/(n − 1). When h = 0,
X̂n+h = Xn and when h = −(n − 1), X̂n+h = X1 . Therefore, the line is drawn
between the first and last observations.

2.8 (a) See the plot on the next page. The variation when the production is low is
much less than the variation in the series when the production is high. This
indicates a transformation is required.
(b) See the plot on the next page.
(c) See the table on page 84.
Chapter 2: Basic forecasting tools 83


10 12

Forecast
Vehicles (thousands)

8
6
4
2
0

1950 1960 1970 1980 1990


Logarithms of vehicles

8
6
4

1950 1960 1970 1980 1990

Exercise 2.8 (a) and (b): Time plots of Japanese automobile production and the logarithms
of Japanese automobile production.
84 Part D. Solutions to exercises

Year Data Log Forecast Error Error2 |Error/Log|


1947 11 2.40
1948 20 3.00 2.40 0.598 0.357 0.1996
1949 29 3.37 3.00 0.372 0.138 0.1103
1950 32 3.47 3.37 0.098 0.010 0.0284
1951 38 3.64 3.47 0.172 0.030 0.0472
1952 39 3.66 3.64 0.026 0.001 0.0071
1953 50 3.91 3.66 0.249 0.062 0.0635
1954 70 4.25 3.91 0.337 0.113 0.0792
1955 69 4.23 4.25 −0.014 0.000 0.0034
1956 111 4.71 4.23 0.475 0.226 0.1009
1957 182 5.20 4.71 0.495 0.245 0.0950
1958 188 5.24 5.20 0.032 0.001 0.0062
1959 263 5.57 5.24 0.336 0.113 0.0602
1960 482 6.18 5.57 0.606 0.367 0.0981
1961 814 6.70 6.18 0.524 0.275 0.0782
1962 991 6.90 6.70 0.197 0.039 0.0285
1963 1284 7.16 6.90 0.259 0.067 0.0362
1964 1702 7.44 7.16 0.282 0.079 0.0379
1965 1876 7.54 7.44 0.097 0.009 0.0129
1966 2286 7.73 7.54 0.198 0.039 0.0256
1967 3146 8.05 7.73 0.319 0.102 0.0396
1968 4086 8.32 8.05 0.261 0.068 0.0314
1969 4675 8.45 8.32 0.135 0.018 0.0159
1970 5289 8.57 8.45 0.123 0.015 0.0144
1971 5811 8.67 8.57 0.094 0.009 0.0109
1972 6294 8.75 8.67 0.080 0.006 0.0091
1973 7083 8.87 8.75 0.118 0.014 0.0133
1974 6552 8.79 8.87 −0.078 0.006 0.0089
1975 6942 8.85 8.79 0.058 0.003 0.0065
1976 7842 8.97 8.85 0.122 0.015 0.0136
1977 8514 9.05 8.97 0.082 0.007 0.0091
1978 9269 9.13 9.05 0.085 0.007 0.0093
1979 9636 9.17 9.13 0.039 0.002 0.0042
1980 11043 9.31 9.17 0.136 0.019 0.0146
1981 11180 9.32 9.31 0.012 0.000 0.0013
1982 10732 9.28 9.32 −0.041 0.002 0.0044
1983 11112 9.32 9.28 0.035 0.001 0.0037
1984 11465 9.35 9.32 0.031 0.001 0.0033
1985 12271 9.41 9.35 0.068 0.005 0.0072
1986 12260 9.41 9.41 −0.001 0.000 0.0001
1987 12249 9.41 9.41 −0.001 0.000 0.0001
1988 12700 9.45 9.41 0.036 0.001 0.0038
1989 13026 9.47 9.45 0.025 0.001 0.0027
1990 9.47

Exercise 2.8 (c) and (d).


Chapter 2: Basic forecasting tools 85

(d) MSE=0.059 (average of column headed Error 2 )


MAPE=3.21% (average of values in last column multiplied by 100).
(e) See graph. Forecast is e9.47 = 13026.
(f ) There are a large number of possible methods. One method, which is discussed
in Chapter 5, is to consider only data after 1970 and use a straight line fitted
through the original data (i.e. without taking logarithms).
(g) The data for 1974 is lower than would be expected. If this information could be
included in the forecasts, the MSE and MAPE would both be smaller because
the forecast error in 1974 would be smaller.
86 Part D. Solutions to exercises

Chapter 3: Time series decomposition

3.1
Y 3-MA 5-MA 7-MA 3 × 3-MA 5 × 5-MA
42 55.50 70.33 81.50 62.92 81.14
69 70.33 81.50 91.60 73.50 88.71
100 94.67 91.60 99.83 93.56 96.65
115 115.67 111.40 107.57 113.22 111.10
132 129.33 128.40 126.00 129.11 125.92
141 142.33 142.60 141.86 142.33 141.60
154 155.33 155.60 156.71 155.33 156.80
171 168.33 170.00 172.86 169.56 172.32
180 185.00 187.40 189.29 185.78 189.80
204 204.00 206.00 210.71 205.11 210.96
228 226.33 230.00 236.86 228.56 236.72
247 255.33 261.40 268.29 257.78 262.54
291 291.67 298.80 283.00 295.56 289.27
337 339.67 316.50 298.80 331.78 304.09
391 364.00 339.67 316.50 351.83 318.32
400

3-MA
3x3 MA
5-MA
5x5 MA
7-MA
300
200
100

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Exercise 3.1: Smoothers fitted to the shipments data.


Chapter 3: Time series decomposition 87

The graph on the previous page shows the five smoothers. Because moving average
smoothers are “flat” at the ends, the best smoother in this case is the one with the
smallest number of terms, namely the 3-MA.

3.2 1
1
T̂t = 3 5 (Yt−3 + Yt−2 + Yt−1 + Yt + Yt+1 )
+ 51 (Yt−2 + Yt−1 + Yt + Yt+1 + Yt+2 )

+ 51 (Yt−1 + Yt + Yt+1 + Yt+2 + Yt+3 )
1 2 1 1 1 2 1
= 15 Yt−3 + 15 Yt−2 + 5 Yt−1 + 5 Yt + 5 Yt+1 + 15 Yt+2 + 15 Yt+3 .

3.3 (a) The 4 MA is designed to eliminate seasonal variation because each quarter
receives equal weight. The 2 MA is designed to center the estimated trend at
the data points. The combination 2 × 4 MA also gives equal weight to each
quarter.
(b) T̂t = 18 Yt−2 + 14 Yt−1 + 41 Yt + 14 Yt+1 + 81 Yt+2 .

3.4 (a) Use 2×4 MA to get trend. If the end-points are ignored, we obtain the following
results.
Data: Trend:
Y1 Y2 Y3 Y4 Y1 Y2 Y3 Y4
Q1 99 120 139 160 Q1 110.250 129.250 150.125
Q2 88 108 127 148 Q2 114.875 134.500 154.750
Q3 93 111 131 150 Q3 100.375 119.635 138.875
Q4 111 130 152 170 Q4 105.500 124.375 145.125

Data – trend:
Y1 Y2 Y3 Y4 Ave
Q1 9.750 9.750 9.875 9.792
Q2 –6.875 –7.500 –6.500 –7.042
Q3 –7.375 –8.625 –8.875 –8.292
Q4 5.500 5.625 6.875 6.000
(b) Hence, the seasonal indices are:
Ŝ1 = 9.8, Ŝ2 = −7.0, Ŝ3 = −8.3 and Ŝ4 = 6.0.
The seasonal component consists of replications of these indices.
(c) End points ignored. Other approaches are possible.

3.5 (a) See the top plot on the next page. There is clear trend which appears close to
linear, and strong seasonality with a peak in August–October and a trough in
January–March.
(b) Calculations are given at the bottom of the next page. The decomposition plot
is shown at the top of the next page.
88 Part D. Solutions to exercises

1600
1200
Plastic sales
data

800
trend-cycle

1200
1000
110
seasonal

70 80 90

104
remainder

100
96
2 3 4 5

Year J F M A M J J A S O N D
Data
1 742 697 776 898 1030 1107 1165 1216 1208 1131 971 783
2 741 700 774 932 1099 1223 1290 1349 1341 1296 1066 901
3 896 793 885 1055 1204 1326 1303 1436 1473 1453 1170 1023
4 951 861 938 1109 1274 1422 1486 1555 1604 1600 1403 1209
5 1030 1032 1126 1285 1468 1637 1611 1608 1528 1420 1119 1013
2×12 MA Trend
1 977.0 977.0 977.1 978.4 982.7 990.4
2 1000.5 1011.2 1022.3 1034.7 1045.5 1054.4 1065.8 1076.1 1084.6 1094.4 1103.9 1112.5
3 1117.4 1121.5 1130.7 1142.7 1153.6 1163.0 1170.4 1175.5 1180.5 1185.0 1190.2 1197.1
4 1208.7 1221.3 1231.7 1243.3 1259.1 1276.6 1287.6 1298.0 1313.0 1328.2 1343.6 1360.6
5 1374.8 1382.2 1381.2 1370.6 1351.2 1331.2
Ratios
1 119.2 124.5 123.6 115.6 98.8 79.1
2 74.1 69.2 75.7 90.1 105.1 116.0 121.0 125.4 123.6 118.4 96.6 81.0
3 80.2 70.7 78.3 92.3 104.4 114.0 111.3 122.2 124.8 122.6 98.3 85.5
4 78.7 70.5 76.2 89.2 101.2 111.4 115.4 119.8 122.2 120.5 104.4 88.9
5 74.9 74.7 81.5 93.8 108.6 123.0
Seasonal indices
Ave 77.0 71.3 77.9 91.3 104.8 116.1 116.8 122.9 123.6 119.3 99.5 83.6

Exercise 3.5(a) and (b): Multiplicative classical decomposition of plastic sales data.
Chapter 3: Time series decomposition 89

(c) The trend does appear almost linear except for a slight drop at the end. The
seasonal pattern is as expected. Note that it does not make much difference
whether these data are analyzed using a multiplicative decomposition or an
additive decomposition.
3.6 Period Trend Seasonal Forecast
t Tt St Ŷt = Tt St /100
61 1433.96 76.96 1103.6
62 1442.81 71.27 1028.3
63 1451.66 77.91 1131.0
64 1460.51 91.34 1334.0
65 1469.36 104.83 1540.3
66 1478.21 116.09 1716.1
67 1487.06 116.76 1736.3
68 1495.91 122.94 1839.1
69 1504.76 123.55 1859.1
70 1513.61 119.28 1805.4
71 1522.46 99.53 1515.3
72 1531.31 83.59 1280.0

3.7 (a) See the top of the figure on the previous page.
(b) The calculations are given below.
Year Q1 Q2 Q3 Q4
Data
1 362 385 432 341
2 382 409 498 387
3 473 513 582 474
4 544 582 681 557
5 628 707 773 592
6 627 725 854 661
4×2 MA
1 382.5 388.0
2 399.3 413.3 430.4 454.8
3 478.3 499.6 519.4 536.9
4 557.9 580.6 601.5 627.6
5 654.8 670.6 674.9 677.0
6 689.4 708.1
Ratios
1 112.9 87.9
2 95.7 99.0 115.7 85.1
3 98.9 102.7 112.1 88.3
4 97.5 100.2 113.2 88.7
5 95.9 105.4 114.5 87.4
6 91.0 102.4
Seasonal indices
Ave 95.8 101.9 113.7 87.5
90 Part D. Solutions to exercises

600
data

400

700
trend-cycle

600
500
400
seasonal

105
90 95
remainder

100
98
96
2 3 4 5 6

Exercise 3.7: Decomposition plot for exports from French company.

(c) Multiplicative decomposition seems appropriate here because the variance is


increasing with the level of the series. The most interesting feature of the
decomposition is that the trend has levelled off in the last year or so. Any
forecast method should take this change in the trend into account.

3.8 (a) The top plot shows the original data followed by trend-cycle, seasonal and
irregular components. The bottom plot shows the seasonal sub-series.
(b) The trend-cycle is almost linear and the small seasonal component is very small
compared to the trend-cycle. The seasonal pattern is difficult to see in time
plot of original data. Values are high in March, September and December and
low in January and August. For the last six years, the December peak and
March peak have been almost constant. Before that, the December peak was
growing and the March peak was dropping. There are several possible outliers
in 1991.
Chapter 3: Time series decomposition 91

(c) The recession is seen by several negative outliers in the irregular component.
This is also apparent in the data time plot. Note: the recession could be made
part of the trend-cycle component by reducing the span of the loess smoother.

3.9 (a) and (b) Calculations are given below. Note that the seasonal indices are
computed by averaging the de-trended values within each half-year.
Data 2×2 MA Detrended Seasonal Seasonal
Trend Data Component Adjusted Data
1.09 0.017 1.073
1.07 1.0825 -0.0125 -0.014 1.084
1.10 1.0825 0.0175 0.017 1.083
1.06 1.0750 -0.0150 -0.014 1.074
1.08 1.0625 0.0175 0.017 1.063
1.03 1.0450 -0.0150 -0.014 1.044
1.04 1.0300 0.0100 0.017 1.023
1.01 1.0225 -0.0125 -0.014 1.024
1.03 1.0075 0.0225 0.017 1.013
0.96 -0.014 0.974
(c) With more data, we could take moving averages of the detrended values for
each half-year rather than a simple average. This would result in a seasonal
component which changed over time.
92 Part D. Solutions to exercises

Chapter 4: Exponential smoothing methods

4.1
Period Data MA(3) SES(α = 0.7)
t Yt Ŷt Et Ŷt Et
1974 1 1 5.4
2 2 5.3 5.40 -0.10
3 3 5.3 5.33 -0.03
4 4 5.6 5.33 0.27 5.31 0.29
1975 1 5 6.9 5.40 1.50 5.51 1.39
2 6 7.2 5.93 1.27 6.48 0.72
3 7 7.2 6.57 0.63 6.99 0.21
4 8 7.10 7.14
Accuracy statistics from period 4 through 7
ME 0.92 0.65
MAE 0.92 0.65
MAPE 13.22 9.56
MSE 1.08 0.64
Theil’s U 1.40 1.14

Theil’s U statistic suggests that the naı̈ve (or last value) method is better than
either of these. If SES is used with an optimal value of α chosen, then α = 1 is
selected. This is equivalent to the naı̈ve method. Note different packages may give
slightly different results for SES depending on how they initialize the method. Some
packages will also allow α > 1.

4.2 (a) Forecasts for May 1992


Method MA(3) MA(5) MA(7) MA(9) MA(11)
Forecast 24.0 48.6 55.6 51.7 53.1
MSE 1484.3 1031.2 757.5 860.8 1313.8
(b) Forecasts for May 1992
Method α = 0.1 α = 0.3 α = 0.5 α = 0.7 α = 0.9
Forecast 45.5 41.9 33.1 29.1 28.7
MSE 1421.35 1211.80 1193.98 1225.40 1298.49
(c) Of these forecasting methods, the best MA(k) method has k = 7 and the best
SES method has α = 0.5. However, it should be noted that the MSE values for
the MA methods are taken over different periods. For example, the MSE for the
MA(7) method is computed only over 9 observations because it is not possible
to compute an MA(7) forecast for the first seven observations. So the MSE
Chapter 4: Exponential smoothing methods 93

values are not strictly comparable for the MA forecasts. It would be better to
use a holdout sample but there are too few data.
4.3 Optimizing α for SES over the period 3 through 10:

α 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
MAPE 65.60 53.46 44.43 37.60 32.32 28.16 24.82 22.08 19.80 17.86
MSE 79.34 47.24 29.95 20.10 14.17 10.41 7.91 6.17 4.92 4.00

The optimal value is α = 1.


With Holt’s method, any combination of α and β will give MAPE=0. This is so
because the differences between successive values of (4.13) are always going to be
zero with this errorless series. Using α = 1 for SES and α = 0.5 and β = 0.5 for
Holt’s method gives the following results.

Data SES Holt’s


Yt Ŷt Et Ŷt Et
2
4 2 2 4 0
6 4 2 6 0
8 6 2 8 0
10 8 2 10 0
12 10 2 12 0
14 12 2 14 0
16 14 2 16 0
18 16 2 18 0
20 18 2 20 0

(a) Clearly Holt’s method is better as it allows for the trend in the data.
(b) For SES, α = 1. Because of the trend, the forecasts will always lag behind the
actual values so that the forecast errors will always be at least 2. Choosing
α = 1 makes the forecast errors as small as possible for SES.
(c) See above.

4.4 (a) (b) and (c) See the table on the following page.
(d) There’s not much to choose between these methods. They are both bad! Look
at Theil’s U values for instance. The last value method over the same period
(13–28) gives MSE=6.0, MAPE=2.05 and Theil’s U=1.0.
94 Part D. Solutions to exercises

Period Data Forecast Errors Forecast Errors


t Yt MA(12) Et MA(6) Et
1 108
2 108
3 110
4 106
5 108
6 108
7 105 108.00 -3.00
8 100 107.50 -7.50
9 97 106.17 -9.17
10 95 104.00 -9.00
11 95 102.17 -7.17
12 92 100.00 -8.00
13 95 102.67 -7.67 97.33 -2.33
14 95 101.58 -6.58 95.67 -0.67
15 98 100.50 -2.50 94.83 3.17
16 97 99.50 -2.50 95.00 2.00
17 101 98.75 2.25 95.33 5.67
18 104 98.17 5.83 96.33 7.67
19 101 97.83 3.17 98.33 2.67
20 99 97.50 1.50 99.33 -0.33
21 95 97.42 -2.42 100.00 -5.00
22 95 97.25 -2.25 99.50 -4.50
23 96 97.25 -1.25 99.17 -3.17
24 96 97.33 -1.33 98.33 -2.33
25 97 97.67 -0.67 97.00 0.00
26 98 97.83 0.17 96.33 1.67
27 94 98.08 -4.08 96.17 -2.17
28 92 97.75 -5.75 96.00 -4.00
29 97.33 95.50
Accuracy criteria: periods 13–28
ME -1.51 -0.10
MAE 3.12 2.96
MSE 14.40 12.64
MAPE 3.23 3.03
Theil’s U 1.58 1.45

Calculations for Exercise 4.4


Chapter 4: Exponential smoothing methods 95

4.5 (a) (b) and (c)


Paperbacks Hardcovers
SES Holt SES Holt
Smoothing parameters α = 0.213 α = 0.335 α = 0.347 β = 0.437
β = 0.453 β = 0.157
Forecast Day 31 210.15 224.24 240.38 250.73
Forecast Day 32 210.15 231.79 240.38 254.63
Forecast Day 33 210.15 239.33 240.38 258.53
Forecast Day 34 210.15 246.88 240.38 262.43
MAE 29.6 33.9 27.3 28.6
MSE 1252.2 1701.7 1060.6 1273.0
MAPE 17.1 18.4 13.5 14.3
Theil’s U 0.68 0.92 0.81 0.92
For both series, SES forecasting is performing better than Holt’s method.
(d) SES forecasts are “flat” and Holt’s forecasts show a linear trend. Both series
show an upward linear trend and we would expect the forecasts to reflect that
trend. Perhaps an out-of-sample analysis would give a better indication of the
merits of the two methods.
(e) The autocorrelation functions of the forecast errors in each case are plotted on
the next page. In each case, there is no noticeable pattern. Only a few spikes
are just outside the critical bounds which is expected.

4.6 Here is a complete run for one set of values (β = 0.1 and α 1 = 0.1). Note that in this
program we have chosen to make the first three values of α be equal to the starting
value. This is not crucial, but it does make a difference.

t Yt Ft Et At Mt αt
1 200.0 200.00 0.00 0.00 0.00 0.100
2 135.0 200.00 -65.00 -6.50 6.50 0.100
3 195.0 193.50 1.50 -5.70 6.00 0.100
4 197.5 193.65 3.85 -4.74 5.79 0.950
5 310.0 197.31 112.69 7.00 16.48 0.820
6 175.0 289.74 -114.74 -5.18 26.30 0.425
7 155.0 241.00 -86.00 -13.26 32.27 0.197
8 130.0 224.08 -94.08 -21.34 38.45 0.411
9 220.0 185.43 34.57 -15.75 38.06 0.555
10 277.5 204.62 72.88 -6.89 41.55 0.414
11 235.0 234.77 0.23 -6.17 37.41 0.166
12 234.81 0.165
96 Part D. Solutions to exercises

0.4 SES paperbacks Holt paperbacks

0.4
0.2

0.2
ACF

ACF
0.0

0.0
-0.2

-0.2
-0.4

-0.4
2 4 6 8 10 12 14 2 4 6 8 10 12 14

Lag Lag

SES hardbacks Holt hardbacks


0.4

0.4
0.2

0.2
ACF

ACF
0.0

0.0
-0.2

-0.2
-0.4

-0.4

2 4 6 8 10 12 14 2 4 6 8 10 12 14

Lag Lag

Exercise 4.5 (e): Autocorrelation functions of forecast errors.

For other combinations of values for β and starting values for α, here is what the
final α value is:

β = 0.1 β = 0.3 β = 0.5 β = 0.7


α = 0.1 0.165 0.327 0.732 0.143
α = 0.2 0.058 0.454 0.783 0.133
α = 0.3 0.140 0.618 0.797 0.133

The time series is not very long and therefore the results are somewhat fickle. In
any event, it is clear that the β value and the starting values for α have a profound
effect on the final value of α.

4.7 Holt-Winters’ method is best because the data are seasonal. The variation increases
with the level, so we use Holt-Winters’ multiplicative method. The optimal smooth-
ing parameters (giving smallest MSE) are a = 0.479, b = 0.00 and c = 1.00. These
give the following forecasts (read left to right):
Chapter 4: Exponential smoothing methods 97

309.1 312.1 315.2 318.3 321.3 324.4


327.5 330.5 333.6 336.7 339.7 342.8
345.9 348.9 352.0 355.0 358.1 361.2
364.2 367.3 370.4 373.4 376.5 379.6

4.8 First choose any values for the three parameters. Here we have used α = β = γ = 0.1.
Different values will choose different initial values. Our program uses the method
described in the textbook and gave the following results:

ME MAE MSE MAPE r1 Theil’s U


-240.5 240.5 62469.9 37.5 0.70 2.6

Now compare with the optimal values: α = 0.917, β = 0.234 and γ = 0.000. Using
the same initialization, we obtain the results in Table 4-11, namely

ME MAE MSE MAPE r1 Theil’s U


-9.46 24.00 824.75 3.75 0.17 0.29
98 Part D. Solutions to exercises

Chapter 5: Simple regression

5.1 (a) −0.7, almost 1, 0.2


(b) False. The correlation is negative. So below-average values of one are associated
with above-average values of the other variable.
(c) Wages have been increasing over time due to inflation. At the same time,
population has been increasing and consequently, new houses need to be built.
So, because they are both increasing with time, they are positively correlated.
(d) There are many factors affecting unemployment and it is simplistic to draw a
causal connection with inflation on the basis of correlation. As in the previous
question, both vary with time and the correlation could be induced by their
time trends. Or they could both be related to some third variable such as
business confidence or government spending.
(e) The older people in the survey had much less opportunity for education than the
younger people. This negative correlation is caused by the increase in education
levels over time.
P P
5.2 (a) X̄ = 5, Ȳ = 25, (Xi − X̄)2 = 20, (Xi − X̄)(Yi − Ȳ ) = 78. So b = 78/20 = 3.9
and a = 25 − 3.9(5) = 5.5. Hence, the regression line is Ŷ = 5.5 + 3.9X.
P P p
(b) (Ŷi − Ȳ )2 = 304.20, (Yi − Ŷi )2 = 25.80, σe = 25.80/(5 − 2) = 2.933. So

304.20/(2 − 1)
F = = 35.4.
25.80/(5 − 2)

This has (2−1) = 1 df for the numerator and (5−2) = 3 df for the denominator.
From Table C in Appendix III, the P -value is slightly smaller than 0.010. (Using
a computer, it is 0.0095.) Standard errors:
q
s.e.(a) = (2.93) 15 + 25 20 = 3.53
q
1
s.e.(b) = (2.93) 20 = 0.656.

On 3 df, t∗ = 3.18 for a 95% confidence interval. Hence 95% intervals are

α: 5.500 ± 3.18(3.53) = [−5.7, 16.7]


β: 3.900 ± 3.18(0.656) = [1.8, 6.0]
Chapter 5: Simple regression 99

Output from Minitab for Exercise 5.2:


Regression Analysis

The regression equation is


Y = 5.50 + 3.90X

Predictor Coef StDev T P


Constant 5.500 3.531 1.56 0.217
X 3.9000 0.6557 5.95 0.010

S = 2.933 R-Sq = 92.2% R-Sq(adj) = 89.6%

Analysis of Variance

Source DF SS MS F P
Regression 1 304.20 304.20 35.37 0.010
Error 3 25.80 8.60
Total 4 330.00

(c) R2 = 0.922, rXY = rY Ŷ = 0.960.


(d) The line through the middle of the graph is the line of best fit. The 95%
prediction interval shown is the interval which would contain the Y value with
probability 0.95 if the X value was 17. The 80% prediction interval shown is
the interval which would contain the Y value with probability 0.80 if the X
value was 26. The dotted line at the boundary of the light shaded region gives
the ends of all the 95% prediction intervals. The dotted line at the boundary
of the dark shaded region gives the ends of all the 80% prediction intervals.

5.3 (a) See the plot on the next page and the Minitab output on page 101. The straight
line is Ŷ = 0.46 + 0.22X.
(b) See the plot on the next page. The residuals may show a slight curvature (Λ
shaped). However, the curvature is not strong and the fitted model appears
reasonable.
(c) R2 = 90.2%. Therefore, 90.2% of the variation in melanoma rates is explained
by the linear regression.
(d) From the Minitab output:
Prediction: 9.286. Prediction interval: (6.749, 11.823)
100 Part D. Solutions to exercises

10
8
melanoma

6
4
2

10 20 30 40

ozone

Exercise 5.3(a): Scatterplot of melanoma rate against ozone depletion.


1.5
1.0
Residuals

0.5
0.0
-0.5
-1.0

10 20 30 40

ozone

Exercise 5.3(b): Scatterplot of residuals from the linear regression.


Chapter 5: Simple regression 101

Output using Minitab for Exercise 5.3:

MTB > Regress ’Melanoma’ 1 ’Ozone’;


SUBC> predict 40.

The regression equation is


Melanoma = 0.460 + 0.221 Ozone

Predictor Coef StDev T P


Constant 0.4598 0.6258 0.73 0.481
Ozone 0.22065 0.02426 9.09 0.000

S = 0.9947 R-Sq = 90.2% R-Sq(adj) = 89.1%

Analysis of Variance

Source DF SS MS F P
Regression 1 81.822 81.822 82.70 0.000
Error 9 8.905 0.989
Total 10 90.727

Fit StDev Fit 95.0% CI 95.0% PI


9.286 0.517 ( 8.116, 10.456) ( 6.749, 11.823)

Note that it is the prediction interval (PI) we want here. Minitab also gives
the confidence interval (CI) for the line at this point, something we have not
covered in the book.
(e) This analysis has assumed that the susceptibility to melanoma among people
living in the various locations is constant. This is unlikely to be true due to
the diversity of racial mix and climate over the locations. Apart from ozone
depletion, melanoma will be affected by skin type, climate, culture (e.g. is
sun-baking encouraged?), diet, etc.

5.4 (a) See plot on the next page and computer output on page 103.
(b) Coefficients: a = 4.184, b = 0.9431. Only b is significant, showing the relation-
ship is significant. (We could refit the model without the intercept term.)
(c) If X = 80, Ŷ = 4.184 + 0.9431(80) = 79.63. Standard error of forecast is 1.88
(from computer output).
102 Part D. Solutions to exercises

90
80
70
Production rating

60
50
40
30

20 40 60 80

Manual dexterity

Exercise 5.4(a): Scatterplot of production rating against manual dexterity test scores.
100



80
Production rating

• •
• •


60

• •

• •

• • •
40

• •
• •
20

20 40 60 80 100

Manual dexterity

Exercise 5.4(e): 95% prediction intervals for production rating.


Chapter 5: Simple regression 103

Output using Minitab for Exercise 5.4

MTB > Regress ’Y’ 1 ’X’;


SUBC> Predict ’newX’.

The regression equation is


Y = 4.18 + 0.943 X

Predictor Coef StDev T P


Constant 4.184 3.476 1.20 0.244
X 0.94306 0.05961 15.82 0.000

S = 5.126 R-Sq = 93.3% R-Sq(adj) = 92.9%

Analysis of Variance

Source DF SS MS F P
Regression 1 6576.8 6576.8 250.29 0.000
Error 18 473.0 26.3
Total 19 7049.8

Fit StDev Fit 95.0% CI 95.0% PI


23.05 2.38 ( 18.04, 28.05) ( 11.17, 34.92)
41.91 1.46 ( 38.85, 44.97) ( 30.71, 53.10)
60.77 1.18 ( 58.28, 63.26) ( 49.71, 71.82)
79.63 1.88 ( 75.68, 83.58) ( 68.16, 91.10)

(d) For confidence and prediction intervals, use Table B with 18 df. 95% CI for β
is 0.94306 ± 2.10(0.05961) = [0.82, 1.07].
(e) See output. Again it is the prediction interval (PI) we want here, not the
confidence interval (CI). The prediction intervals are shown in the plot on the
previous page.

5.5 (a) See the plot on the following page. The straight line regression model is Ŷ =
20.2−0.145X where Y = electricity consumption and X = temperature. There
is a negative relationship because heating is used for lower temperatures, but
there is no need to use heating for the higher temperatures. The temperatures
are not sufficiently high to warrant the use of air conditioning. Hence, the
electricity consumption is higher when the temperature is lower.
104 Part D. Solutions to exercises

19
Electricity consumption (Mwh)

18
17
16

10 15 20 25 30

Temperature

Exercise 5.5(a): Electricity consumption (Mwh) plotted against temperature (degrees Cel-
sius).

Possible outlier
2
1
Residuals

0
-1

10 15 20 25 30

Temperature

Exercise 5.5(c): Residual plot for the straight line regression of electricity consumption
against temperature.
Chapter 5: Simple regression 105

(b) r = −0.791
(c) See the plot on the previous page. Apart from the possible outlier, the model
appears to be adequate. There are no highly influential observations.
(d) If X = 10, Ŷ = 20.2 − 0.145(10) = 18.75. If X = 35, Ŷ = 20.2 − 0.145(35) =
15.12. The first of these predictions seems reasonable. The second is unlikely.
Note that X = 35 is outside the range of the data making prediction danger-
ous. For temperatures above about 20 ◦ C, it is unlikely electricity consumption
would continue to fall because no heating would be used. Instead, at high
temperatures (such as X = 35◦ C), electricity consumption is likely to increase
again due to the use of air-conditioning.

5.6 (a) When H = 130 and W = 45, r = 0.553.


(b) When H = 40 and W = 150, r = −0.001.
(c) The following table shows the influence of outliers at various positions.
H W r
129 0 -0.393
128 22 0.032
122 44 0.527
112 64 0.773
99 83 0.846
83 99 0.810
65 112 0.627
44 122 0.151
22 128 -0.365
0 129 -0.624
The point about all this is that an outlier (and skewness in general) can seriously
affect the correlation coefficient. It is a good idea to look at the scatterplot
before computing any correlation.

5.7 (a) See the plot on the next page. The winning time has been decreasing with year.
There is an outlier in 1896.
(b) The fitted line is Ŷ = 196−0.0768X where X denotes the year of the Olympics.
Therefore the winning time has been decreasing an average 0.0768 seconds per
year.
(c) The residuals are plotted on the next page. The residuals show random scatter
about 0 with only one usual point (the outlier in 1896). But note that the
last five residuals are positive. This suggests that the straight line is “levelling
out”—the winning time is decreasing at a slower rate now than it was earlier.
106 Part D. Solutions to exercises

54
52
50
winning.time

48
46
44

1900 1920 1940 1960 1980 2000

year

Exercise 5.7(a): Scatterplot of winning times against year.


3
2
fit$resid

1
0
-1

1900 1920 1940 1960 1980 2000

year

Exercise 5.7(c): Residual plot for linear regression model of winning times.
Chapter 5: Simple regression 107

(d) The predicted winning time in the 2000 Olympics is

Ŷ = 196 − 0.0768(2000) = 42.50 seconds.

This would smash the world record. But given the previous five results (with
positive residuals), it would seem more likely that the actual winning time
would be higher. A prediction interval is

42.50 ± 2.0796(1.1762) = 42.50 ± 2.45 = [40.05, 44.95].

5.8 (a) There is strong seasonality with peaks in November and December and a trough
in January. The surfing festival shows as a smaller peak in March from 1988.
The variation in the series is increasing with the level and there is a strong
positive trend due to sales growth.
(b) Logarithms are necessary to stabilize the variance so it does not increase with
the level of the series.
(c) See the plot on the next page and the computer output on page 109. The fitted
line is Ŷ = −526.57 + 0.2706X where X is the year and Y is the logged annual
sales.
(d)
X = 1994 : Ŷ = −526.57 + 0.2706(1994) = 12.98
X = 1995 : Ŷ = −526.57 + 0.2706(1995) = 13.25
X = 1996 : Ŷ = −526.57 + 0.2706(1996) = 13.52

Prediction intervals (from computer output):

X = 1994 : [12.57, 13.40]


X = 1995 : [12.80, 13.71]
X = 1996 : [13.03, 14.02]

(e) We transform the forecasts and intervals with the exponential function:

Total annual sales for 1994 exp(12.98) = $434, 443


Total annual sales for 1995 exp(13.25) = $569, 439
Total annual sales for 1996 exp(13.52) = $746, 383

Prediction intervals:

X = 1994 : [e12.57 , e13.40 ] = [286673, 658385]


X = 1995 : [e12.80 , e13.71 ] = [361994, 895764]
X = 1996 : [e13.03 , e14.02 ] = [455060, 1224208]
108 Part D. Solutions to exercises

80000
60000
Sales

40000
20000
0

1988 1990 1992 1994

Exercise 5.8(a): Time plot of sales figures.


12.5
Log Total annual sales

12.0
11.5

1987 1988 1989 1990 1991 1992 1993

Exercise 5.8(c): Regression line fitted to the logged sales data.


Chapter 5: Simple regression 109

Output using Minitab for Exercise 5.8:


MTB > regress ’Log Sales’ 1 ’Year’;
SUBC> predict ’new years’;

The regression equation is


Log Sales = - 527 + 0.271 Year

Predictor Coef StDev T P


Constant -526.57 46.44 -11.34 0.000
Year 0.27059 0.02334 11.60 0.000

S = 0.1235 R-Sq = 96.4% R-Sq(adj) = 95.7%

Analysis of Variance

Source DF SS MS F P
Regression 1 2.0501 2.0501 134.45 0.000
Error 5 0.0762 0.0152
Total 6 2.1263

Fit StDev Fit 95.0% CI 95.0% PI


12.9818 0.1044 ( 12.7135, 13.2502) ( 12.5661, 13.3975)
13.2524 0.1257 ( 12.9293, 13.5755) ( 12.7994, 13.7054) X
13.5230 0.1476 ( 13.1435, 13.9025) ( 13.0282, 14.0178) X
X denotes a row with X values away from the center

These prediction intervals are very wide because we are only using annual totals
in making these predictions. A more accurate method would be to fit a model
to the monthly data allowing for the seasonal patterns. This is discussed in
Chapter 7.
(f ) One way would be to calculate the proportion of sales for each month compared
to the total sales for that year. Averaging these proportions will give a rough
guide as to how to split the annual totals into 12 monthly totals.
110 Part D. Solutions to exercises

14
12
Percentage mortality

10
8
6
4

0 20 40 60 80 100

Percentage Type A Birds

Exercise 5.9(a): Scatterplot of percentage mortality against percentage of Type A birds.

5.9 (a) The plot is shown above. The fitted line is

Ŷ = 4.38 + 0.0154X

where X = percentage of type A birds and Y = percentage mortality.


(b) From the computer output:
Predictor Coef StDev T P
Constant 4.3817 0.6848 6.40 0.000
% Type A 0.015432 0.007672 2.01 0.046

So the t-test is significant (since P < 0.05). A 95% confidence interval for the
slope is

0.01543 ± 1.976(0.007672) = 0.01543 ± 0.01516 = [0.003, 0.031].

This suggests that the Type A birds have a higher mortality than the Type B
birds, the opposite to what the farmers claim.
(c) For a farmer using all Type A birds, X = 100. So Ŷ = 4.38 + 0.0154(100) =
5.92%. For a farmer using all Type B birds, X = 0. So Ŷ = 4.38%. Prediction
intervals for these are [2.363, 9.487] and [0.587, 8.177] respectively.
(d) R2 = 2.6. So only 2.6% of the variation in mortality is due to bird type.
Chapter 5: Simple regression 111

140
Model 1
Model 2

120
100
consumption

80
60
40

40 60 80 100

price

Exercise 5.10(b): Scatterplot of gas consumption against price.

(e) This information suggests that heat may be a lurking variable. If Type A birds
are being used more in summer and the mortality is higher in summer, than the
increased mortality of Type A birds may be due to the summer rather than the
bird type. A proper randomized experiment would need to be done to properly
assess whether bird type is having an effect here.
5.10 (a) Cross sectional data. There is no time component.
(b) See the plot above.
(c) When the price is higher, the consumption may be lower due to the pressure of
increased cost. Therefore, we would expect b 1 < b2 < 0.
(d) Model 1: First take logarithms of Y i , then use simple linear regression to obtain
a = 5.10, b = −0.0153, σe2 = 0.0735.
Model 2: Split data into two groups. Fit each group separately using simple
linear regression to obtain
a1 = 221, b1 = −2.91 and a2 = 84.8, b2 = −0.447.
Using the equation given in the question, we obtain
σe2 = 2913.7/16 = 182.06.
The fitted lines are shown on the graph above.
112 Part D. Solutions to exercises

(e) Model 1: R2 = rY2 Ŷ = 0.721.


1
Model 2: R2 = rY2 Ŷ = 0.859. The second model is better with higher R 2 value.
2
The residual plots are given on the following page. Again, the second model is
much better showing random scatter about zero. The first model show pattern
in the residuals.
(f ) The graph on page 114 shows a local linear regression through the data. The
fitted curve resembles the fitted lines for model 2. This suggests that model 2
is a reasonable model for the data. However, our approach has also meant the
two lines do not join at X = 60. A better model would force them to join. This
means the parameters must be restricted which makes the estimation much
harder.
(g) and (h) Using model 2, forecasts are obtained by

220.9 − 2.906X when X ≤ 60
Ŷ =
84.8 − 0.447X when X > 60.

and standard errors are obtained from (5.19):


r
√ 1 (X − 63)2
s.e.(Ŷ ) = 182.06 1 + + .
20 10672.11

The 95% PI are obtained using Ŷ ± t∗ (s.e.) where t∗ = 2.12 (from Table B with
16 df). Hence, we obtain the following values.
X Ŷ s.e. [ 95% PI ]
40 104.67 14.15 [74.7 , 134.7]
60 46.55 13.83 [17.2 , 75.9]
80 49.03 14.00 [19.3 , 78.7]
100 40.09 14.65 [ 9.0 , 71.1]
120 31.15 15.70 [ -2.1 , 64.4]
For example, at a price of 80c, the gas consumption will lie between 19.3 and
78.7 for 95% of towns.
Chapter 5: Simple regression 113

40
20
Residuals model 1

0
-20

40 60 80 100

Price
20
10
Residuals model 2

0
-10
-20

40 60 80 100

Price

Exercise 5.10(e): Residual plots for the two models.


114 Part D. Solutions to exercises

140
120
100
Consumption

80
60
40

40 60 80 100

Price

Exercise 5.10(f ): Local linear regression through the gas consumption data. The fitted line
suggests that model 2 is more appropriate.
140
120
100
Consumption

80
60
40
20
0

40 60 80 100 120

Price

Exercise 5.10(h): 95% prediction intervals for gas consumption.


Chapter 6: Multiple regression 115

Chapter 6: Multiple regression

6.1 (a) df for numerator = k and for denominator = n − k − 1 where n = number of


observations and k = number of explanatory variables. Here, k = 16 so that
n − 16 − 1 = 30. Hence, n = 30 + 16 + 1 = 47.
(b)
47 − 1
R̄2 = 1 − (1 − R2 ) n−k−1
n−1
= 1 − (1 − 0.943) = 0.913.
48 − 16 − 1
(c) F = 31.04 on (17,30) df. From Table C in Appendix III, the P -value is much
smaller than 0.01. So the regression is highly significant.
(d) The coefficients should be compared with a t 30 distribution. From Table B in
Appendix III, any value greater than 2.04 in absolute value will be significant
at the 5% level. So the constant and variables 4, 8, 12, 13, 14, 15 and 17 are
significant in the presence of other explanatory variables. Note that the signif-
icance level of 5% is arbitrary. There is no reason why some other significance
level (e.g. 2%) could not be used.
(e) The next stage would be to reduce the number of variables in the model by
removing some of the least significant variables and re-fitting the model.

6.2 (a) The fitted model is Ĉ = 273.93−5.68P +0.034P 2 . For this model, R2 = 0.8315.
[Recall: in exercise 5.6, model 1 had R 2 = 0.721 and model 2 had R 2 = 0.859.]
So the R̄2 values for each model are:
46
Model 1 R̄2 = 1 − (1 − 0.721) n−k−1
n−1
= 1 − (1 − 0.721) = 0.715.
45
46
Model 2 R̄2 = 1 − (1 − 0.859) n−k−1
n−1
= 1 − (1 − 0.859) = 0.849.
43
46
Model 3 R̄2 = 1 − (1 − 0.832) n−k−1
n−1
= 1 − (1 − 0.832) = 0.824.
44
These values show that model 2 is the best model, followed by model 3. The t
values for the coefficients are:
Model 1 α : t = 10.22 β : t = −5.47
Model 2 α1 : t = 10.33 β1 : t = −6.61 α2 : t = 4.11 β2 : t = −1.99
Model 3 β0 : t = 8.83 β1 : t = −5.62 β2 : t = 4.57
Of these, only β2 from model 2 is not significantly different from zero. This
suggests that a better model would be to allow the second part of model 2 to
be a constant rather than a linear function.
(b) From the computer output the following 95% prediction intervals are obtained.
116 Part D. Solutions to exercises

Output using Minitab for Exercise 6.2:


MTB > regress ’C’ 2 ’P’ ’Psq’;
SUBC> predict ’newP’ ’newPsq’.

The regression equation is


C = 274 - 5.68 P + 0.0339 Psq

Predictor Coef StDev T P


Constant 273.93 31.03 8.83 0.000
P -5.676 1.009 -5.62 0.000
Psq 0.033904 0.007412 4.57 0.000

S = 14.37 R-Sq = 83.2% R-Sq(adj) = 81.2%

Analysis of Variance
Source DF SS MS F P
Regression 2 17327.0 8663.5 41.95 0.000
Error 17 3511.0 206.5
Total 19 20838.0

Source DF Seq SS
P 1 13005.7
Psq 1 4321.3

Fit StDev Fit 95.0% CI 95.0% PI


173.97 14.29 ( 143.82, 204.13) ( 131.21, 216.74) XX
101.14 4.77 ( 91.08, 111.21) ( 69.19, 133.10)
55.43 4.91 ( 45.07, 65.80) ( 23.38, 87.48)
36.85 4.95 ( 26.40, 47.29) ( 4.77, 68.92)
45.38 7.14 ( 30.31, 60.46) ( 11.52, 79.25)
81.04 18.49 ( 42.02, 120.06) ( 31.62, 130.46) XX
X denotes a row with X values away from the center
XX denotes a row with very extreme X values
Chapter 6: Multiple regression 117

200
150
consumption

100
50
0

20 40 60 80 100 120

price

Exercise 6.2: Quadratic regression of gas consumption against price. 95% prediction inter-
vals shown.

P Ĉ [ 95% PI ]
20 173.97 [ 131.21 , 216.74 ]
40 101.14 [ 69.19 , 133.10 ]
60 55.43 [ 23.38 , 87.48 ]
80 36.85 [ 4.77 , 68.92 ]
100 45.38 [ 11.52 , 79.25 ]
120 81.04 [ 31.62 , 130.46 ]
It is clear from the plot that it is dangerous predicting outside the observed
price range. In this case, the predictions at P = 20 and P = 120 are almost cer-
tainly wrong. Predicting outside the range of the explanatory variable is always
dangerous, but much more so when a quadratic (or higher-order polynomial) is
used.
(c) rP P 2 = 0.990. If we were to use P , P 2 and P 3 , the correlations among these
explanatory variables would be very high and we would have a serious multi-
collinearity problem on our hands. The coefficients estimates would be unstable
(i.e. have large standard errors). Multicollinearity will often be a problem with
polynomial regression.
118 Part D. Solutions to exercises

6.3 (a) From Table 6-15, we obtain the following values


Period Actual Forecast
54 4.646 1.863
55 1.060 1.221
56 -0.758 0.114
57 4.702 2.779
58 1.878 1.959
59 6.620 5.789
Analysis of errors: periods 54 through 59.
ME MAE MSE MPE MAPE ACF1 Theil’s U
0.74 1.11 2.15 34.82 41.32 -0.35 0.34
Strictly speaking, we should not compute relative measures when the data cross
the zero line (i.e., when there are positive and negative values) because relative
measures will “blow up” if divided by zero.
(b) and (c) Optimizing the coefficients for Holt’s method will give better forecasts.
Another approach is to use a simple MA forecast. An MA(2) forecast actually
works better than Holt’s method for both series. Other approaches are also
possible.
Calculate accuracy statistics for your forecasts and compare them with the
forecasts in Table 6-14.

6.4 (a) The fitted equation is

Ŷ = 73.40 + 1.52X1 + 0.38X2 − 0.27X3 .

95% confidence intervals for the parameters are calculated using a t 6 distribu-
tion. So the multiplier is 2.45:

73.40 ± 2.45(14.687) = [37.46, 109.3]


1.52 ± 2.45(0.1295) = [1.20, 1.84]
0.38 ± 2.45(0.1941) = [−0.09, 0.85]
−0.27 ± 2.45(0.1841) = [−0.72, 0.18]

(b) F = 123.3 on (3,6) df. P = 0.000. This means that the probability of results
like this, if the three explanatory variables were not relevant, is very small.
(c) The residual plots on page 120 show the model is satisfactory. There is no
pattern in any of the residual plots.
(d) R2 = 0.984. Therefore 98.4% of the variation in Y is explained by the regression
relationship.
Chapter 6: Multiple regression 119

Output using Minitab for Exercise 6.4:

MTB > Regress ’Y’ 3 ’X1’ ’X2’ ’X3’;


SUBC> Predict 10 40 30;
SUBC> Confidence 90.

The regression equation is


Y = 73.4 + 1.52 X1 + 0.381 X2 - 0.268 X3

Predictor Coef StDev T P


Constant 73.40 14.69 5.00 0.002
X1 1.5162 0.1295 11.71 0.000
X2 0.3815 0.1941 1.97 0.097
X3 -0.2685 0.1841 -1.46 0.195

S = 2.326 R-Sq = 98.4% R-Sq(adj) = 97.6%

Analysis of Variance
Source DF SS MS F P
Regression 3 2001.54 667.18 123.32 0.000
Error 6 32.46 5.41
Total 9 2034.00

Source DF Seq SS
X1 1 1118.36
X2 1 871.67
X3 1 11.51

Fit StDev Fit 90.0% CI 90.0% PI


95.762 1.632 ( 92.590, 98.934) ( 90.239, 101.285)
120 Part D. Solutions to exercises

4
3

3
2

2
residuals

residuals
1

1
0

0
-1

-1
-2

-2
5 10 15 20 30 40 50 60 70

X1 X2
4
3
2
residuals

1
0
-1
-2

10 20 30 40 50 60

X3

Exercise 6.4(c): Residual plots for the cement data.

(e) The signs of the coefficients indicate the direction of the effect of each variable.
X1 increases heat and has the greatest effect (the largest coefficient). The other
variables are not significant, so they may not have any effect. If they do, the
coefficients suggest that X2 might increase heat and X3 might decrease heat.
(f ) For X1 = 10, X2 = 40 and X3 = 30, Ŷ = 73.40+1.52(10)+0.38(40)−0.27(30) =
95.76. 90% Prediction interval: [90.24,101.29]

6.5 The data for this exercise were taken from McGee and Carleton (1970) “Piecewise re-
gression”, Journal of the American Statistical Association, 65, 1109–1124. It might
be worthwhile to get this paper to compare what conventional regression can ac-
complish when there are special features in the data. In this case, the relationship
Chapter 6: Multiple regression 121

between the Boston dollar volume and the NYSE-AME dollar volume underwent a
series of changes over the time period of interest. In this paper, the solution was as
follows:

from Jan ’67 through Oct ’67 Ŷ = 8.748 + 0.0061X


from Nov ’67 through Jul ’68 Ŷ = −20.905 + 0.0114X
from Aug ’68 through Nov ’68 Ŷ = −79.043 + 0.0205X
from Dec ’68 through Nov ’69 Ŷ = 11.075 + 0.0067X

Notice the slope coefficients in these four equations. They are small (because
Boston’s dollar volume is small relative to the big board volumes) but they get
increasingly stronger (from6 1 to 114 to 205) in successive periods of commission
splitting. Then in Dec ’68, the SEC said “no more commission splitting” and it
hurt the Boston dollar volume. The slope went back to 67, which is almost where it
started.

(a) The fitted equation is Ŷ = −66.2 + 0.014X. The following output was obtained
from a computer package.

Value Std. Error t value Pr(>|t|)


(Intercept) -66.2193 39.6809 -1.6688 0.1046
X 0.0138 0.0029 4.7856 0.0000
F statistic: 22.9 on 1 and 33 degrees of freedom
the p-value is 0.00003465
R-sq = 0.4097 Rbar-sq = 0.3918 D-W = 0.694

Clearly, the regression is significant, although the intercept is not significant.


(b) Output from computer package:

Value Std. Error t value Pr(>|t|)


(Intercept) -67.2116 40.2550 -1.6696 0.1047
X 0.0135 0.0030 4.5025 0.0001
time 0.2737 0.6518 0.4199 0.6773
F statistic: 11.25 on 2 and 32 degrees of freedom
the p-value is 0.0001992
R-sq = 0.4129 Rbar-sq = 0.3762 D-W = 0.6814

Here, the regression is significant, but time is not significant. In fact, comparing
these two models shows that adding time to the regression equation is actually
worse than not adding it. See the R̄2 values. And for both analyses, the D-W
122 Part D. Solutions to exercises

250
200
Y

150
100
50

10000 12000 14000 16000 18000

Exercise 6.5(c): Connected scatterplot for the Boston and American stock exchanges.

statistic shows that there is a lot of pattern left in the residuals. A piecewise
regression approach does far better with this data set.
(c) See the plot above.

6.6 (a) and (b) Here are the seasonality indices based on the regression equations
(6.10) and (6.12). They represent the intercept term in the regression for each
of the 12 first differences.
Using (6.10) Using (6.12)
Mar-Feb -2.6 -6.2
Apr-Mar -6.7 -10.6
May-Apr -3.5 -7.4
Jun-May -5.3 -9.2
Jul-Jun -3.6 -7.4
Aug-Jul -5.2 -9.2
Sep-Aug -5.9 -9.7
Oct-Sep -6.9 -10.7
Nov-Oct -4.1 -7.9
Dec-Nov -4.7 -8.5
Jan-Dec -0.8 -4.6
Feb-Jan -2.2 -6.2
These two sets of seasonal indices are not quite the same. In the first equa-
Chapter 6: Multiple regression 123

tion (6.10), all eleven dummy variables for seasonality were allowed to be in
the regression. In the second equation (6.12), the best subsets regression pro-
cedure did not allow the first seasonal dummy into the final equation. The
absolute values are not so important because, in the presence of different sets
of explanatory variables, we expect the intercept terms to be different.
(c) The seasonal indices should be the same regardless of which month is used as
a base.

6.7 (a) Yt = 78.7 + 0.534xt + et


(b) DW = 0.57. dL = 1.04 at 1% level. Therefore there is significant positive
autocorrelation.
124 Part D. Solutions to exercises

Chapter 7: The Box-Jenkins methodology for ARIMA models

7.1 (a) In general, the approximate standard error of the sample autocorrelations is

1/ n. So the larger the value of n, the smaller the standard error. Therefore,
the ACF has more variation for small values of n than for large values of n. All
three series show the autocorrelations mostly falling with the 95% bands. The
few that lie just outside the bands are not of concern since we would expect
about 5% of spikes to cross the bands. There is no reason to think these series
are anything but white noise.

(b) The lines shown are 95% critical values. These are calculated as ±1.96/ n. So
they are closer to zero when n is larger. The autocorrelations vary randomly,
but they mostly stay within the bounds.

7.2 The time plot shows the series as a non-stationary level. It wanders up and down
over time in a similar way to a random walk. The ACF decays very slowly which
also indicates non-stationarity in the level. Finally, the PACF has a very large value
at lag 1, indicating the data should be differenced.

7.3 The five models are

AR(1) Yt = 0.6Yt−1 + et .
MA(1) Yt = et + 0.6et−1 .
ARMA(1,1) Yt = 0.6Yt−1 + et + 0.6et−1 .
AR(2) Yt = −0.8Yt−1 + 0.3Yt−2 + et .
MA(2) Yt = et + 0.8et−1 − 0.3et−2 .

In each case, we assume Yt = 0 and et = 0 for t ≤ 0. The generated data are shown
on the following two pages. There is a lot of similarity in the shapes of the series
because they are based on exactly the same errors.

7.4 (a) The ACF is slow to die out and the time plot shows the series wandering in a
non-stationary way. So we take first differences. The ACF of the first differences
show one significant spike at lag 1 indicating an MA(1) is appropriate. So the
model for the raw data is ARIMA(0,1,1).
(b) There is not consistent trend in the raw data and the differenced data have
mean close to zero. Therefore, there is no need to include a constant term.
(c) (1 − B)Yt = (1 − θ1 B)et .
(d) See the output on page 127. There may be slight differences with different
software packages and even different versions of the same package. The Ljung-
Box statistics are not significant and the ACF and PACF of residuals show no
significant differences from white noise.
Chapter 7: The Box-Jenkins methodology for ARIMA models 125

t AR(1) MA(1) ARMA(1,1) AR(2) MA(2)


1 0.010 0.010 0.010 0.010 0.010
2 1.386 1.386 1.392 1.372 1.388
3 1.362 1.358 2.193 -0.565 1.631
4 2.397 1.898 3.214 2.443 1.590
5 2.758 2.268 4.196 -0.804 2.425
6 2.695 1.832 4.350 2.416 1.622
7 1.947 0.954 3.564 -1.844 0.766
8 0.968 -0.002 2.136 2.000 -0.248
9 2.481 1.780 3.062 -0.253 1.641
10 2.209 1.860 3.697 1.523 2.300
11 1.055 0.162 2.380 -1.564 -0.264
12 -0.797 -1.592 -0.164 0.278 -1.862
13 -1.628 -2.008 -2.106 -1.842 -2.213
14 -1.047 -0.760 -2.024 1.487 -0.561
15 1.062 1.648 0.434 -0.052 1.979
16 0.917 1.294 1.554 0.768 1.653
17 0.560 0.178 1.111 -0.620 -0.273
18 1.276 0.946 1.612 1.666 0.864
19 -1.334 -1.536 -0.569 -3.619 -1.351
20 -0.711 -1.170 -1.511 3.485 -1.872
21 0.484 0.964 0.057 -2.964 1.612
22 2.050 2.306 2.340 5.176 2.461
23 2.070 1.896 3.300 -4.190 1.975
24 0.112 -0.626 1.354 3.775 -0.986
25 0.987 0.242 1.054 -3.357 -0.236
26 2.262 2.222 2.855 5.488 2.745
27 0.327 -0.028 1.685 -6.428 0.030
28 -1.514 -2.328 -1.317 5.079 -3.035
29 0.272 0.154 -0.636 -4.811 0.121
30 -0.427 0.118 -0.264 4.783 0.867

Generated data for Exercise 7.3


126 Part D. Solutions to exercises

AR(1) MA(1)

2
2

1
1

0
0

-1
-1

-2
0 10 20 30 0 10 20 30

ARMA(1,1) AR(2)
4

4
3

2
2

0
1

-2
0

-4
-1

-6
-2

0 10 20 30 0 10 20 30

MA(2)
2
1
0
-1
-2
-3

0 10 20 30

Exercise 7.3: Simulated ARMA series.


Chapter 7: The Box-Jenkins methodology for ARIMA models 127

Output using Minitab for Exercise 7.4:


MTB > ARIMA 0 1 1 ’Strikes’;
SUBC> NoConstant;
SUBC> Forecast 3.

Final Estimates of Parameters


Type Coef StDev T
MA 1 0.3174 0.1886 1.68

Differencing: 1 regular difference


Number of observations: Original series 30, after differencing 29
Residuals: SS = 9256634 (backforecasts excluded)
MS = 330594 DF = 28

Modified Box-Pierce (Ljung-Box) Chi-Square statistic


Lag 12 24
Chi-Square 8.1(DF=11) 34.1(DF=23)

Forecasts from period 30


95 Percent Limits
Period Forecast Lower Upper
31 4164.87 3037.70 5292.04
32 4164.87 2800.11 5529.63
33 4164.87 2598.14 5731.60

(e) The last observation is yt = 3885; the last residual in series is e t = −881.87
(obtained from the computer package). Now
Yt = Yt−1 + et − 0.3174et−1 .
So Ŷ31 = Y30 + ê31 − 0.3174ê30
= 3885 + 0 − 0.3174(−881.87) = 4164.9
Ŷ32 = Ŷ31 + 0 − 0.3174(0) = 4164.9
Ŷ33 = Ŷ32 + 0 − 0.3174(0) = 4164.9

(f ) See the graph on the following page.


7.5 (a) The monthly data show strong seasonality and the seasonal pattern is reason-
ably stable. There is no trend in the data (this is a mature product).
(b) The pattern in the ACF plot shows the dominance of the seasonality. The
autocorrelations at lags 6, 18 and 30 are negative (because we are correlating
128 Part D. Solutions to exercises

Number of strikes in USA


6000
5000
4000
3000

1950 1960 1970 1980

Exercise 7.4(f ): Predicted number of strikes in USA. 95% prediction intervals shown.

the high periods with the low periods) and at lags 12, 14 and 36 they are
positive (because we are correlating high periods with high periods).
(c) The pattern in the PACF plot is not particularly revealing. However, there
is little need to try to interpret this plot when the analysis clearly shows the
dominance of the seasonality. The best approach would be to difference the
series to reduce the effect of the seasonality and then see what is left over.
(d) These graphs suggest a seasonal MA(1) because of the spike at lag 12 in the
ACF and the decreasing spikes at lags 12 and 24 in the PACF. Overall, the
suggested model is ARIMA(0,1,0)(0,1,1) 12 .
(e) Using the backshift operator: (1 − B)(1 − B 12 )Yt = (1 − ΘB 12 )et . Rewriting
gives
Yt − Yt−12 − Yt−1 + Yt−13 = et − Θet−12 .

7.6 (a) ARIMA(3,1,0).


(b) For the differenced data, the PACF has a significant spikes at lags 1, 2 and
3 and a spike at lag 17 which is marginally significant. The spike at lag 17
is probably due to chance. Therefore an AR(3) is an appropriate model for
the differenced data. Consequently, an ARIMA(3,1,0) model is suitable for the
original data.
Chapter 7: The Box-Jenkins methodology for ARIMA models 129

(c) Now

(Yt − Yt−1 ) = 0.42(Yt−1 − Yt−2 ) − 0.20(Yt−2 − Yt−3 ) − 0.30(Yt−3 − Yt−4 ) + et .

Therefore Yt = 1.42Yt−1 − 0.62Yt−2 − 0.10Yt−3 + 0.30Yt−4 + et and

Ŷ1940 = 1.42(1797) − 0.62(1791) − 0.10(1627) + 0.30(1665) = 1778.1


Ŷ1941 = 1.42(1778.1) − 0.62(1797) − 0.10(1791) + 0.30(1627) = 1719.8
Ŷ1942 = 1.42(1719.8) − 0.62(1778.1) − 0.10(1797) + 0.30(1791) = 1697.3

7.7 (a) ARIMA(4,0,0).


(b) The model was chosen because the last significant spike in the PACF was at
lag 4. Note that the spikes at lags 2 and 3 were not significant. This makes
no difference. It is the last significant spike which determines the order of the
model.
(c) The model is

Yt = 146.1 + 0.891Yt−1 − 0.257Yt−2 + 0.392Yt−3 − 0.333Yt−4 + et .

So

Ŷ1969 = 146.1 + 0.891(545) − 0.257(552) + 0.392(534) − 0.333(512) = 528.7


Ŷ1970 = 146.1 + 0.891(528.7) − 0.257(545) + 0.392(552) − 0.333(534) = 515.7
Ŷ1971 = 146.1 + 0.891(515.7) − 0.257(528.7) + 0.392(545) − 0.333(552) = 499.5

7.8 (a) The centered 12-MA smooth is shown in the plot on the next page. The trend
is generally linear and increasing with a flat period between 1990 and 1993.
(b) The variation does not change much with the level, so transforming will not
make much difference.
(c) The data are not stationary. There is a trend and seasonality in the data.
Differencing at lag 12 gives the data shown in the plot on page 131. These
appear stationary although it is possible another difference at lag 1 is needed.
(d) From the plots on page 131 it is clear there is a seasonal MA component of order
1. In addition there is a significant spike at lag 1 in both the ACF and PACF.
Hence plausible models are ARIMA(1,0,0)(0,1,1) 12 and ARIMA(0,0,1)(0,1,1)12 .
Comparing the two models we have the following results
ARIMA(1,0,0)(0,1,1)12 AIC=900.2
ARIMA(0,0,1)(0,1,1)12 AIC=926.9
130 Part D. Solutions to exercises

US electricity generation
300
280
260
240
220
200

1985 1987 1989 1991 1993 1995 1997

Year

Exercise 7.8(a): Total net generation of electricity in USA.

Hence the better model is the first one. Note that different packages will give
different values for the AIC depending on how it is calculated. Therefore the
same package should be used for all calculations.
(e) The residuals from the ARIMA(1,0,0)(0,1,1) 12 are shown in the plots on page
132. Because there are significant spikes in the ACF and PACF, the model is
not adequately describing the series. These plots suggest we need to add an
MA(1) term to the model. So we fit the revised model ARIMA(1,0,1)(0,1,1) 12 .
This time, the residual plots (not shown here) look like white noise. The AIC
is 876.7. Part of the computer output for fitting the revised model is shown
below.
Approx.
Parameter Estimate Std Error T Ratio Lag
MA1,1 0.74427 0.05887 12.64 1
MA2,1 0.77650 0.09047 8.58 12
AR1,1 0.99566 0.0070613 141.00 1
So the fitted model is

(1 − 0.996B)(1 − B 12 )Yt = (1 − 0.744B)(1 − 0.777B 12 )et .


Chapter 7: The Box-Jenkins methodology for ARIMA models 131

Electricity data differenced at lag 12


30

o
o
o o
20

o
o o o
o o oo o
o o
o
o o o o o
oo o
o o
10

o o o
o
o
o o oo oo o o
o
o o
o o oo o o o
o o o o
o o o o o o o
o o o o o o
o o o o
o o o oo o o
oo o o o o o oo o o
o o o o
o
0

o o o o o o
o o
o o o
o o o o
o o o oo
o o o
o o
o
o
o
-10

o
o
o
o
o o
-20

1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997
0.2

0.2
PACF
ACF

0.0

0.0
-0.2

-0.2

0 10 20 30 40 0 10 20 30 40

Exercise 7.8(c): Seasonally differenced electricity generation.


132 Part D. Solutions to exercises

Residuals from ARIMA(1,0,0)(0,1,1) model


3

o
o
o
o
o
2

o o
o o o o
o
o o o o
o o
o
o o o
o
1

o o o o
o o o
o o o o oo
oo o o o o o o o
o o o o o o o o
o o o o o o
o o o o o o o o
o o o o o o
o o o
o ooo
0

o
oo o oo o o o o o
o o o o o o
o oo
o o o o o o
o o o o
o
o o
o
o o o
-1

o
o
o
o o
o
o o
o
-2

o
-3

1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997
0.2

0.1
0.1

0.0
PACF
ACF

-0.1
-0.1
-0.3

-0.3

0 10 20 30 40 0 10 20 30 40

Exercise 7.8(e): Residuals from ARIMA(1,0,0)(0,1,1)12 model fitted to the electricity data.
Chapter 7: The Box-Jenkins methodology for ARIMA models 133

Output using SAS for Exercise 7.8:


Approx.
Parameter Estimate Std Error T Ratio Lag
MA1,1 0.86486 0.06044 14.31 1
MA2,1 0.80875 0.09544 8.47 12
AR1,1 0.27744 0.10751 2.58 1

Variance Estimate = 41.8498466


Std Error Estimate = 6.46914574
AIC = 864.616345
SBC = 873.195782
Number of Residuals= 129

To Chi Autocorrelations of Residuals


Lag Square DF Prob
6 1.60 3 0.659 0.028 -0.036 -0.020 -0.010 -0.014 0.095
12 7.67 9 0.568 0.004 -0.082 0.073 0.128 -0.095 0.072
18 15.32 15 0.429 0.126 0.016 -0.091 0.125 -0.105 -0.020
24 18.67 21 0.607 0.065 -0.048 0.051 0.005 0.069 -0.085

Note that the first term on the left is almost the same as differencing (1 − B).
This suggests that we probably should have taking a first difference as well as a
seasonal difference. We repeated the above analysis and arrived at the following
model: ARIMA(1,1,1)(0,1,1)12 which has AIC=864.6.
The computer output for the final model is shown above. The figures under
the heading Chi Square concern the Ljung-Box test. Clearly the model passes
the test (see Table E in Appendix III).
(f ) Forecasts for the next 24 months are given on the following page.
134 Part D. Solutions to exercises

Month Obs Forecast Std Error Lower 95% Upper 95%


Nov 96 143 240.1614 6.4691 227.4821 252.8407
Dec 96 144 262.5516 6.9981 248.8356 276.2677
Jan 97 145 270.2423 7.1820 256.1659 284.3187
Feb 97 146 244.0064 7.3027 229.6934 258.3194
Mar 97 147 249.8899 7.4074 235.3718 264.4081
Apr 97 148 232.7683 7.5069 218.0550 247.4816
May 97 149 249.3720 7.6042 234.4680 264.2759
Jun 97 150 270.7257 7.6999 255.6341 285.8173
Jul 97 151 295.5439 7.7944 280.2671 310.8207
Aug 97 152 295.6598 7.8878 280.2000 311.1196
Sep 97 153 257.1358 7.9800 241.4952 272.7764
Oct 97 154 246.4526 8.0712 230.6332 262.2719
Nov 97 155 245.0224 8.4340 228.4920 261.5528
Dec 97 156 267.1930 8.6077 250.3222 284.0638
Jan 98 157 274.8228 8.7406 257.6914 291.9541
Feb 98 158 248.5699 8.8622 231.2003 265.9395
Mar 98 159 254.4488 8.9796 236.8491 272.0484
Apr 98 160 237.3258 9.0948 219.5004 255.1513
May 98 161 253.9291 9.2083 235.8811 271.9772
Jun 98 162 275.2828 9.3205 257.0150 293.5506
Jul 98 163 300.1009 9.4313 281.6160 318.5858
Aug 98 164 300.2168 9.5407 281.5173 318.9163
Sep 98 165 261.6929 9.6490 242.7812 280.6045
Oct 98 166 251.0096 9.7560 231.8881 270.1311

7.9 (a) See the plot on the following page. Note that there is strong seasonality and
a pronounced trend-cycle. One way to study the consistency of the seasonal
pattern is to compute the seasonal sub-series and see how stable each month
is. The results are given below.
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
1955: 94.7 94.0 96.5 101.3 102.4 103.7 104.5 104.3 104.1 101.2 98.3 95.4
1956: 94.1 93.5 96.8 103.1 104.1 102.8 103.7 103.6 103.6 101.7 98.6 96.8
1957: 95.9 96.8 99.0 97.7 99.5 101.1 102.0 103.3 105.1 103.2 99.8 96.7
1958: 95.0 94.8 96.1 100.4 101.7 102.1 103.6 104.9 104.4 101.1 96.7 95.3
1959: 93.9 94.5 96.4 100.9 102.1 103.3 104.7 106.0 104.4 100.9 98.7 96.1
1960: 95.7 95.7 95.1 98.9 100.8 102.5 104.6 106.0 104.0 100.1 98.0 96.7
1961: 95.2 94.8 96.5 101.3 101.7 103.7 105.2 105.3 104.3 101.2 97.7 96.5
1962: 93.5 93.7 95.6 100.7 102.3 102.5 104.4 106.4 103.5 101.0 97.6 96.9
1963: 94.6 93.4 95.5 99.1 100.8 104.1 106.1 107.4 104.1 100.7 97.9 97.4
1964: 93.6 93.2 94.6 98.6 100.2 103.5 106.6 107.5 103.6 101.7 97.9 96.9
1965: 95.6 92.7 94.0 96.7 99.4 103.7 108.2 108.0 104.7 100.5 98.4 99.6
1966: 97.0 93.7 95.2 97.0 98.4 104.1 105.9 107.2 104.2 99.7 97.1 96.8
1967: 93.9 93.6 94.1 99.0 102.5 105.7 109.2 109.9 104.9 99.8 98.3 93.8
1968: 91.2 91.7 94.5 99.0 101.9 103.1 105.7 106.0 103.5 100.2 100.7 99.1
1969: 96.0 94.3 94.1 96.8 100.7 104.5 106.3 107.2 103.7 102.5 100.2 99.4
1970: 95.8 93.0 92.0 96.0 100.2 103.7 106.0 105.8 102.7 98.9 97.1 96.5

These detrended data are relatively consistent from year to year with only minor
Chapter 7: The Box-Jenkins methodology for ARIMA models 135

Employment in motion picture industry


240

o
o oo
o
o o
o o
o o oo

o o
o
220

o o
o o
oo o o
o oo
o o
o oo o o
o oo o o
o o o
o
oo o o
o o o
o oo o o o
o o o
o o o
oo o o o
200

o oo o o oo
o o
o o o o o o
o o
o o o o
o oo o o o
oo o o
o o o
o o o o o oo
o o o
o o oo o o o o
o o o o o
o o o
o o o o o
o o o
o oo oo o o o o oo
o o
180

o o o o o o o o
o
o
o o o o oo
o o o o
o o o o oo
o o
o o oo o
o o
oo o
o o
o
o oo

1955 1957 1959 1961 1963 1965 1967 1969 1971


1.0
0.8

0.5
PACF
ACF
0.4

0.0
0.0

-0.5

0 10 20 30 40 0 10 20 30 40

Exercise 7.9(a): Employment in the motion picture industry


136 Part D. Solutions to exercises

110
105
ratios

100
95

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Exercise 7.9(a): Sub-series of detrended data

variations occurring here and there. For example, December 1967 and January
and February 1968 were noticeably lower than surrounding years.
Another way to look at seasonal patterns is via autocorrelation functions. Note
that for the raw data, the ACF shows strong seasonality over several seasonal
lags. This is further evidence of the consistency of the seasonal pattern. The
plot on the previous page shows the detrended data. Again, the seasonal pat-
tern is very consistent although the amplitude of the pattern each year varies.
Unusual results in early 1968 and early 1970 are seen.
(b) For the first 96 months, we identified an ARIMA(0,1,0)(0,1,1) 12 . For the second
96 months, we identified an ARIMA(0,1,0)(1,1,0) 12 : In practice, there is little
difference between these models. This means that once the trend has been
eliminated (by differencing), the seasonal patterns are very similar.
(c) Using the above ARIMA(0,1,0)(0,1,1) 12 model, we obtained the following fore-
casts.
Chapter 7: The Box-Jenkins methodology for ARIMA models 137

Detrended employment in motion picture industry


110

o
o
oo
o o
o o
o o
o
o o o o o o
o o o
105

o o oo o
oo o o o o
o o o o
o o o o o o o
o ooo o o o o o o o o
oo o o
o o
o o
o o oo o o
o o o
o o o o o
o o o o o
o o o o
o o o o o o
100

o o
o o o o o
o o o o
o o o
o o o o o o
o o
o o o o
o o
o o o
o o o o
o o o o o o o
o oo o o o o o o
o o o o
o o
o oo o o
o o o
o o
95

o o o
oo o
o o o o o o oo
o o o o oo
o o
o oo o o o
o
o
o
o
o
o

1955 1957 1959 1961 1963 1965 1967 1969 1971


0.5
0.5

PACF
ACF
0.0

0.0
-0.5

-0.5

0 10 20 30 40 0 10 20 30 40

Exercise 7.9(a): Seasonal differences of employment in the motion picture industry


138 Part D. Solutions to exercises

Month Actual Forecast Upper (95%) Lower (95%) Error


Jan 1963 167.2 167.3286 172.4228 162.2345 -0.1286
Feb 1963 165.0 166.7030 171.7902 161.6159 -1.7030
Mar 1963 168.8 167.7141 172.8012 162.6269 1.0859
Apr 1963 175.0 176.5972 181.6844 171.5100 -1.5972
May 1963 177.9 176.9498 182.0369 171.8626 0.9502
Jun 1963 183.7 179.2212 184.3084 174.1340 4.4788
Jul 1963 187.2 186.1708 191.2579 181.0836 1.0292
Aug 1963 189.3 188.7598 193.8470 183.6727 0.5402
Sep 1963 183.4 186.1678 191.2550 181.0806 -2.7678
Oct 1963 177.3 177.2392 182.3264 172.1520 0.0608
Nov 1963 172.3 170.7711 175.8583 165.6839 1.5289
Dec 1963 171.4 168.8708 173.9580 163.7837 2.5292
Jan 1964 164.9 167.5935 172.6807 162.5063 -2.6935
Feb 1964 164.4 163.9412 169.0247 158.8577 0.4588
Mar 1964 166.9 167.4086 172.4920 162.3251 -0.5086
Apr 1964 174.2 174.2641 179.3475 169.1806 -0.0641
May 1964 177.5 176.4075 181.4909 171.3240 1.0925
Jun 1964 183.6 180.0358 185.1193 174.9523 3.5642
Jul 1964 189.5 186.3499 191.4333 181.2664 3.1501
Aug 1964 191.6 191.2063 196.2898 186.1228 0.3937
Sep 1964 185.1 187.7172 192.8007 182.6338 -2.6172
Oct 1964 181.9 178.9557 184.0392 173.8722 2.9443
Nov 1964 175.4 175.7857 180.8692 170.7022 -0.3857
Dec 1964 174.2 172.6567 177.7402 167.5732 1.5433

(d) For the second half of the data we used the ARIMA(0,1,0)(1,1,0) 12 to obtain
the forecasts at the top of the following page. The actual 1971–1972 figures
are also shown. The source is “Employment and Earnings, US 1909–1978”,
published by the Department of Labor, 1979.
A good exercise would be to take these forecasts and check the MAPE for 1971
and 1972 separately. The MAPE for the first forecast year should be smaller
than the MAPE for the second year.
(e) If the objective is to forecast the next 12 months then the latest data is obviously
the most relevant but to get seasonal indices we have to go back several years
and to anticipate what the next move the large cycle is going to be, we really
need to look at as much data as possible. So a good strategy would be
i. study the trend-cycle by looking at the 12-month moving average;
ii. remove the trend-cycle and study the consistency of the seasonality;
iii. decide how much of the data series to retain for the ARIMA modeling;
iv. forecast the next 12 months and use some judgment as to how to modify
the ARIMA forecasts on the basis of anticipated trend-cycle movements.
Chapter 7: The Box-Jenkins methodology for ARIMA models 139

Month Actual Forecast Upper (95%) Lower (95%) Error


Jan 1971 194.5 196.0141 201.4418 190.5864 -1.5141
Feb 1971 187.9 191.2939 198.9698 183.6180 -3.3939
Mar 1971 187.7 189.9446 199.3456 180.5436 -2.2446
Apr 1971 198.3 197.3595 208.2149 186.5042 0.9405
May 1971 202.7 205.7424 217.8790 193.6057 -3.0424
Jun 1971 204.2 213.1446 226.4396 199.8495 -8.9446
Jul 1971 211.7 217.8789 232.2392 203.5186 -6.1789
Aug 1971 213.4 218.8543 234.2061 203.5025 -5.4543
Sep 1971 212.0 213.2939 229.5769 197.0109 -1.2939
Oct 1971 203.4 208.3371 225.5009 191.1733 -4.9371
Nov 1971 199.5 204.7595 222.7611 186.7580 -5.2595
Dec 1971 199.3 203.4670 222.2690 184.6650 -4.1670
Jan 1972 191.3 196.2469 217.0365 175.4573 -4.9469
Feb 1972 192.1 191.0587 213.6617 168.4557 1.0413
Mar 1972 193.3 189.3618 213.6432 165.0804 3.9382
Apr 1972 203.4 196.9907 222.8417 171.1396 6.4093
May 1972 205.5 205.3066 232.6373 177.9760 0.1934
Jun 1972 218.2 212.5618 241.2960 183.8276 5.6382
Jul 1972 220.3 217.4298 247.5022 187.3575 2.8702
Aug 1972 219.9 218.2314 249.5848 186.8780 1.6686
Sep 1972 211.9 213.0587 245.6428 180.4746 -1.1587
Oct 1972 204.5 207.6473 241.4174 173.8773 -3.1473
Nov 1972 198.5 204.3907 239.3064 169.4750 -5.8907
Dec 1972 200.5 203.2051 239.2300 167.1802 -2.7051

Forecasts for Exercise 7.9(d)

7.10 (a) There is strong seasonality as can be seen from the time plot and the seasonal
peaks in the ACF.
(b) The trend in the series is small compared to the seasonal variation. However,
there is a period of downward trend in the first four years, followed by an
upward trend for four years. At the end the trend seems to have levelled off.
(c) The one large spike in the PACF of Figure 7-34 suggests the series needs dif-
ferencing at lag 1. This is also apparent from the slow decay in the ACF and
the non-stationary mean in the time plot.
(d) You would need to difference again at lag 1 and plot the ACF and PACF of the
new series (differenced at lags 12 and 1). It is not possible to identify a model
from Figures 7-33 and 7-34.
140 Part D. Solutions to exercises

Chapter 8: Advanced forecasting models

8.1 (a) The fitted model in Exercise 6.7 (using OLS) was

Yt = 78.7 + 0.534xt + Nt .

The computer output below shows the results for fitting the straight line re-
gression with AR(1) errors. Hence the new model is

Yt = 79.3 + 0.508xt + Nt where Nt = 0.72Nt−1 + et .

In this case, the error model makes very little difference to the parameters.

Output from SAS for Exercise 8.1:

Approx.
Parameter Estimate Std Error T Ratio Lag Variable Shift
MU 79.27236 0.76093 104.18 0 SALES 0
AR1,1 0.72469 0.14647 4.95 1 SALES 0
NUM1 0.50801 0.02318 21.91 0 ADVERT 0

Constant Estimate = 21.8242442


Variance Estimate = 1.11639088
Std Error Estimate = 1.056594
AIC = 74.2915405
SBC = 77.825702
Number of Residuals= 24

Autocorrelation Check of Residuals


To Chi Autocorrelations
Lag Square DF Prob
6 3.46 5 0.630 0.027 0.099 -0.037 0.111 -0.060 -0.274
12 9.31 11 0.593 0.055 0.126 0.229 -0.227 0.060 -0.095
18 16.39 17 0.497 -0.117 -0.238 -0.080 0.054 -0.108 0.101

(b) The ACF and PACF of the errors is plotted on the following page. An AR(1)
model for the errors is appropriate since there is a single significant spike at
lag 1 in the PACF and geometric decay in the ACF. This is confirmed by the
Ljung-Box test in the computer output above. The Q ∗ values are given under
the column Chi Square. None are significant showing the residuals from the
full model are white noise.
Chapter 8: Advanced forecasting models 141

Errors from regression model

o
2

o
o
o
1

o
o
o

o
0

o o
o
o
o
-1

o o o
o

o o
-2

o
-3

10 20
0.4

0.4
PACF
ACF

0.0

0.0
-0.4

-0.4

2 4 6 8 10 12 2 4 6 8 10 12

Exercise 8.1: Errors from regression model with AR(1) error term.
142 Part D. Solutions to exercises

Output from SAS for Exercise 8.2(a):

Approx.
Parameter Estimate Std Error T Ratio Lag Variable Shift
MU 9.56328 0.40537 23.59 0 HURON 0
AR1,1 0.78346 0.06559 11.94 1 HURON 0
NUM1 -0.02038 0.01066 -1.91 0 YEAR 0

Constant Estimate = 2.07087134

Variance Estimate = 0.51219788


Std Error Estimate = 0.71568001
AIC = 216.450147
SBC = 224.205049
Number of Residuals= 98

Autocorrelation Check of Residuals


To Chi Autocorrelations
Lag Square DF Prob
6 8.35 5 0.138 0.222 -0.100 -0.133 -0.056 -0.007 -0.042
12 15.01 11 0.182 -0.051 0.009 0.175 0.017 -0.121 -0.107
18 16.36 17 0.499 -0.053 0.014 0.019 0.058 0.006 -0.067
24 25.47 23 0.326 -0.071 -0.166 -0.043 0.051 0.160 0.092

8.2 (a) To reduce numerical error, we subtracted 1900 from the year to create an ex-
planatory variable. Hence the year ranged from -25 (1875) to 72 (1972). The
computer output above shows the fitted model to be
Yt = 9.56 − 0.02xt + Nt where Nt = 0.78Nt−1 + et
where xt is the year −1900.
(b) The errors are shown in the plot on the following page. This demonstrates
that a better model would have an AR(2) error term since the PACF has two
significant spikes at lags 1 and 2. The spike at lag 10 is probably due to chance.
The ACF shows geometric decay which is possible with an AR(2) model. So
the full regression model is
Yt = β 0 + β 1 xt + N t where Nt = φ1 Nt−1 + φ2 Nt−2 + et .
Fitting this model gives the output shown on page 144. So the fitted model is
Yt = 9.53 − 0.02xt + Nt where Nt = Nt−1 − 0.29Nt−2 + et .
Chapter 8: Advanced forecasting models 143

Errors from regression model

o
2

o
o o o
o
o o o
oo o
o
o o o
1

o o oo o
o o o
o
o oo
o o
oo o
o
o ooo o
o o o
o o
o o o
0

o o o
o o
o o o o
o
o o o
o o
oo oo o
oo o o
o o
o
-1

o o
o o o oo
o o o
o o o
o o

o
oo oo
-2

oo o

1880 1900 1920 1940 1960


0.8

0.8
0.6
0.6

0.4
0.4

PACF
ACF

0.2
0.2
0.0

-0.2
-0.2

5 10 15 5 10 15

Exercise 8.2(b): Errors from regression model with AR(1) error term.
144 Part D. Solutions to exercises

Output from SAS for Exercise 8.2(b):

Approx.
Parameter Estimate Std Error T Ratio Lag Variable Shift
MU 9.53078 0.30653 31.09 0 HURON 0
AR1,1 1.00479 0.09839 10.21 1 HURON 0
AR1,2 -0.29128 0.10030 -2.90 2 HURON 0
NUM1 -0.02157 0.0082537 -2.61 0 YEAR 0

Constant Estimate = 2.73048107

Variance Estimate = 0.4760492


Std Error Estimate = 0.68996319
AIC = 210.396534
SBC = 220.736404
Number of Residuals= 98

Autocorrelation Check of Residuals


To Chi Autocorrelations
Lag Square DF Prob
6 0.60 4 0.964 0.018 -0.028 -0.003 0.040 0.054 -0.007
12 5.35 10 0.867 -0.032 -0.037 0.167 -0.007 -0.098 -0.055
18 6.21 16 0.986 -0.036 0.005 -0.025 0.035 -0.006 -0.063
24 10.49 22 0.981 -0.003 -0.141 -0.007 0.006 0.116 0.008

8.3 (a) ARIMA(0,1,1)(2,1,0)12 . This model would have been chosen by first identifying
that differences at lags 12 and 1 are necessary to make the data stationary. Then
looking at the ACF and PACF of the differenced data would have shown two
significant spikes in the PACF at lags 12 and 24. There would have also been
a significant spike in the ACF at lag 1 and geometric decay in the early lags of
the PACF.
(b) Since both parameter estimates are positive (and significantly different from
zero), we can conclude that electricity consumption increases with both heating
degrees and cooling degrees. Because b 2 is larger, we know that there is a greater
increase in electricity usage for each heating degree than for each cooling degree.
(c) To use this model for forecasting, we would first need forecasts of both X 1,t
and X2,t into the future. These could be obtained by taking averages of these
variables over the equivalent months of the previous few decades. Then the
model can be used to forecast electricity demand over the next 12 months by
Chapter 8: Advanced forecasting models 145

forecasting the Nt series using the method discussed in chapter 7 and plugging
the forecasts of X1,t , X2,t and Nt into the formula for Yt .
(d) If the model was fitted using a standard regression package (thus modeling N t
as white noise), then the seasonality and autocorrelation in the data would have
been ignored. This would result in less efficient parameter estimates and invalid
estimates of their standard errors. In particular, tests for significance would be
incorrect, as would prediction intervals. Also, when producing forecasts of Y t ,
the forecasts of Nt would be all be zero. Hence, the model would not adequately
allow for the seasonality or autocorrelation in the data.

8.4 (a) b = 3, r = 1, s = 2.
(b) ARIMA(2,0,0)
(c) ω0 = −0.53, ω1 = −0.37, ω2 = −0.51, δ1 = 0.57, δ2 = 0, θ1 = θ2 = 0, φ1 = 1.53,
φ2 = −0.63.
(d) 27 seconds.

8.5 See the graphs on the following page.

8.6 (a) The three series are shown on page 147. For Set 1, four X t values are needed
(since v1 , v2 , v3 and v4 are all non-zero). Therefore 27 Yt values can be produced.
Similarly 26 Yt values for Set 2 and 24 Yt values for Set 3 can be calculated.
(b) The first model is
2.0B
Yt = Xt + N t .
1 − 0.7B
The simplest way to generate data for this transfer function is to rewrite it as
follows

(1 − 0.7B)Yt = (1 − 0.7B)B(2 − 1.4B)Xt + (1 − 0.7B)Nt

so that Yt = 0.7Yt−1 + 2.0Xt−1 − 1.4Xt−2 + Nt − 0.7Nt−1 .


Thus Yt values can only be generated for times t = 3, 4, . . . since we need at
least two previous Xt values. However, for t = 3, we also need Y 2 . To start the
process going, we have assumed here that Y 2 = 0. Other values could also have
been used. The effect of this initialization is negligible after a few time periods.
The second model is easier to generate as we can write it

Yt = 1.2Xt + 2.0Xt−1 − 0.8Xt−2 + Nt .


146 Part D. Solutions to exercises

(a) (b)
2.0

3.0
1.5

2.5
1.0

2.0
weight

weight
0.5

1.5
0.0

1.0
0.5
-1.0

0.0

0 2 4 6 8 10 0 2 4 6 8 10

lag lag

(c) (d)
0.8

1.0
0.6

0.8
0.4

0.6
weight

weight
0.2

0.4
0.0

0.2
-0.2
-0.4

0 2 4 6 8 10 0 2 4 6 8 10

lag lag

Exercise 8.5: Impulse response weights for the four different transfer functions.
Chapter 8: Advanced forecasting models 147

Set 1 Set 2 Set 3 Set 4 Set 5


t Nt Xt Yt Yt Yt
1 -0.8003 50
2 0.8357 90 0.0
3 1.4631 50 110.9 201.5
4 0.7332 30 58.7 51.3 64.7
5 0.3260 80 52.3 58.3 25.7 116.3
6 -0.7442 80 61.3 51.3 135.0 231.3
7 0.7362 30 65.7 62.7 74.7 143.8 132.7
8 1.1931 70 59.2 66.2 88.2 49.3 81.2
9 -1.4681 60 55.5 56.5 87.5 130.2 186.5
10 -0.5285 10 49.5 56.5 79.5 113.7 75.5
11 0.4314 40 37.4 50.4 83.4 16.4 20.4
12 -1.6341 20 27.4 35.4 72.4 75.5 94.4
13 0.8198 40 29.8 29.8 54.8 38.8 56.8
14 0.4183 20 30.4 29.4 47.4 79.0 88.4
15 -0.4065 10 23.6 29.6 41.6 38.6 19.6
16 -0.0615 30 19.9 23.9 40.9 19.3 39.9
17 0.1432 60 29.1 20.1 36.1 59.7 124.1
18 -1.0747 70 46.9 27.9 27.9 118.6 178.9
19 -0.5355 40 56.5 47.5 38.5 139.2 139.5
20 -0.1454 70 56.9 56.9 59.9 79.7 107.9
21 0.2088 10 49.2 57.2 74.2 140.1 120.2
22 -0.6854 30 34.3 48.3 76.3 19.2 -0.7
23 0.1182 30 28.1 35.1 73.1 60.1 88.1
24 0.6971 40 30.7 28.7 54.7 60.7 84.7
25 0.3698 30 34.4 30.4 43.4 80.3 92.4
26 -0.0802 100 46.9 33.9 43.9 59.9 147.9
27 -0.9202 60 64.1 46.1 44.1 199.1 247.1
28 1.1483 90 76.1 66.1 61.1 121.1 149.1
29 -0.1663 60 75.8 74.8 84.8 179.8 203.8
30 -0.5461 100 76.5 75.5 97.5 119.4 167.5

Generated data for Exercise 8.6


148 Part D. Solutions to exercises

8.7 (a) The average cost of a night’s accommodation is C/R.


(b) There are a number of ways this could be done. The simplest is to define the
monthly CPI to be the same as that of the quarter. For example, January,
February and March of 1980 would each have a CPI of 45.2; April, May and
June 1980 would each have a CPI of 46.6; and so on. Other methods might
involve fitting a smooth curve through the quarterly figures and using the curve
to predict the CPI at other points along the time axis.
(c) See figure below.

Consumer price index (Melbourne)


100
80

Average rate per room per night ($)


60
40

1980 1982 1984 1986 1988 1990 1992 1994 1996

Exercise 8.7(c): Time plots of average room rate and CPI.

(d) Our preliminary model is

Yt = a + (ν0 + ν1 B + · · · + ν6 B 6 )Xt + Nt

where Yt denotes the average room rate, Xt denotes the CPI and Nt is an AR(1)
process. The estimated errors from this model are shown in the figure on the
previous page. They are clearly non-stationary and have some seasonality.
So we difference both Yt and Xt and refit the model with Nt specified as an
ARIMA(1,0,0)(1,0,0)12 . The parameter estimates are shown below (as given
by SAS).
Chapter 8: Advanced forecasting models 149

Residuals from dynamic regression with AR(1) errors

o o o
o o o
o o o
o o o
o o o
o o o o
oo
o
0.05

o o oo o
o o
o o o
o oo o o o
o o
o o
oo o
o o o
o o
o o o o
o o o o
o o o
o o
o o oo o
o o o
o o
o oo o o o
0.0

o o o
o o o o o
o oo o o o
oo
o o o o o o
oo o o o o
o o o o o
o o o o o o
o o o o
o o o oo oo oo
o o o
o o o o o o o o o
o o o
o o o o
o o o o
-0.05

o o o
o o
o o o
o o
o o o o
o o o o
o o
o
o o
o o

1980 1982 1984 1986 1988 1990 1992 1994 1996


0.6
0.6

0.4
0.4

PACF
ACF

0.2
0.2

0.0
0.0

-0.4

5 10 15 20 5 10 15 20

Exercise 8.7(d): Errors from regression model with AR(1) error term.
150 Part D. Solutions to exercises

Parameter Estimate s.e. P -value


a 0.20200 0.2848 0.4791
ν0 0.20730 0.2602 0.4267
ν1 -0.41687 0.2634 0.1154
ν2 0.23165 0.2655 0.3842
ν3 0.32048 0.2716 0.2396
ν4 -0.72093 0.2665 0.0075
ν5 0.74707 0.2633 0.0051
ν6 -0.36272 0.2656 0.1739
Thus the intercept and first four coefficients are not significant and can be
omitted. Hence we select b = 4. We shall retain the last three coefficients for
the moment. Since they show no clear pattern, we select r = 0 and s = 3 giving
the model
Yt = (ω0 + ω1 B + ω2 B 2 )B 4 Xt + Nt .
Looking at the ACF and PACF of the error series (not shown) and trying a
number of alternative models led us to the model ARIMA(2,1,0)(2,0,0) 12 for
Nt . That is

(1 − φ1 B − φ2 B 2 )(1 − Φ1 B 12 − Φ2 B 24 )(1 − B)Nt = et .

The parameter values (all significant) were


Parameter ω0 ω1 ω2 φ1 φ2 Φ1 Φ2
Estimate 0.52 0.61 -0.47 -0.49 -0.33 0.37 0.41
The model suggests that there is a lag of four months between changes in the
CPI and changes in the price of travel accommodation. The seasonality inherent
in the model may be due to seasonal price variation or due to the way CPI was
estimated from quarterly data.
(e) Forecasts of CPI were obtained using Holt’s method. These are only needed
from November 1995 because of the time lag of 4 months. Actual data beyond
June 1995 are given in the second column for comparison.
Chapter 8: Advanced forecasting models 151

Month Predicted Actual


Yt Yt Xt−4 Xt−5 Xt−6
Jul 1995 90.4 94.0 115.0 115.0 115.0
Aug 1995 91.8 96.7 116.2 115.0 115.0
Sep 1995 92.0 94.8 116.2 116.2 115.0
Oct 1995 91.6 89.6 116.2 116.2 116.2
Nov 1995 93.4 95.8 116.9 116.2 116.2
Dec 1995 90.3 91.5 117.3 116.9 116.2
Jan 1996 90.4 92.0 117.7 117.3 116.9
Feb 1996 92.6 95.5 118.1 117.7 117.3
Mar 1996 94.7 100.6 118.5 118.1 117.7
Apr 1996 90.7 94.1 118.9 118.5 118.1
May 1996 91.5 97.2 119.3 118.9 118.5
Jun 1996 93.0 102.9 119.8 119.3 118.9
80
Perpetual speed score

60
40

0 20 40 60 80 100 120

Day

Exercise 8.8: Time plot of daily perceptual speed scores for a schizophrenic patient. The
drug intervention is shown at day 61.

8.8 (a) See the figure above.


(b) The step intervention model with an ARIMA(0,1,1) error was used:

Yt = ωXt + Nt where (1 − B)Nt = θet−1 + et


152 Part D. Solutions to exercises

where Yt denotes the perceptual speed score and X t denotes the step dummy
variable. The estimated coefficients were
Parameter ω θ
Estimate -22.1 0.76
(c) The drug has lowered the perceptual speed score by about 22.
(d) The new model is
ω
Yt = Xt + N t where (1 − B)Nt = θet−1 + et
1 − δB
(An ARIMA(0,1,1) error was found to be the best model again.) Here the
estimated coefficients were
Parameter ω δ θ
Estimate -13.21 0.54 0.76
The following accuracy measures show that the delayed effect model fits the
data better.
Model Step Delayed step
MAPE 15.1 15.0
MSE 92.5 91.1
AIC 542.8 538.4
The forecasts for the two models are very similar. This is because the effect of
the step in the delayed step model is almost complete at the end of the series,
60 days after the drug intervention.
(e) The best ARIMA model we found was an ARIMA(0,1,1) with θ = 0.69. This
gave MAPE=15.4, MSE=100.8 and AIC=550.9.
This model gives a flat forecast function (since we did not include a constant
term). The forecast values are 33.9. Because the step effect is almost complete
in the delayed step model, it also gives a virtually flat forecast function with
forecast values of 34.1. Hence there is virtually no difference. If forecasts had
been made earlier (for example, at day 80), there would have been a difference
because the step effect would still be in progress and so the delayed step model
would have showed a continuing decline in perceptual speed. The real advantage
of the intervention model over the ARIMA model is that the intervention model
provides a way of measuring and evaluating the effect of an intervention.
(f ) If the drug varied from day to day and the reaction times depended on dose,
then a better model would be a dynamic regression model with the the quantity
of drug as an explanatory variable.
Chapter 8: Advanced forecasting models 153

8.9 (a)      
Yt − Yt−1 Yt−1 − Yt−2 Yt−2 − Yt−3
= Φ1 + Φ2
Xt − Xt−1 Xt−1 − Xt−2 Xt−2 − Xt−3
 
Yt−12 − Yt−13
+ · · · + Φ12 + Zt.
Xt−12 − Xt−13

(b)
Yt = Yt−1 − 0.38(Yt−1 − Yt−2 ) + 0.15(Xt−1 − Xt−2 )
− 0.37(Yt−2 − Yt−3 ) + 0.13(Xt−2 − Xt−3 ) + · · ·
= 0.62Yt−1 + 0.01Yt−2 + 0.15Xt−1 − 0.02Xt−2 + · · ·

(c)
• Multivariate model assumes feedback. That is, X t depends on past values
of Yt . But regression does not allow this.
• Regression model does not assume X t is random.
• Regression model allows Yt to depend on Xt as well as past values
Xt−1 , Xt−2 , . . .. Multivariate AR only allows dependence on past values
of {Xt }.
• For these data, it is unlikely room rates will substantially affect Y t although
it is possible. Small values in lower left of coefficient matrices suggest that
Xt is not affecting Yt . Yt should depend on Xt . So regression is probably
better.
8.10 (a) An AR(3) model can be written using the same procedure as the AR(2) model
described in Section 8/5/1. Thus we define X 1,t = Yt , X2,t = Yt−1 and X3,t =
Yt−2 . Then write
   
φ1 φ2 φ3 at
X t =  1 0 0  X t−1 +  0 
0 1 0 0

and Yt = [1 0 0]X t .
This is now in state space form with
     
φ1 φ2 φ3 1 0 0 at
F =  1 0 0 ,G =  0 1 0  , H = [1 0 0], et =  0  and zt = 0.
0 1 0 0 0 1 0

(b) An MA(1) can be written as Yt = θat−1 + at where at is white noise. We can


write this in state space form by letting F = 0, G = θ, e t = at−1 , zt = at and
H = 1. Thus
Yt = X t + a t and Xt = θat−1 .
154 Part D. Solutions to exercises

(c) Holt’s method is defined in Chapter 4 as

Lt = αYt + (1 − α)(Lt−1 + bt−1 ),


bt = β(Lt − Lt−1 ) + (1 − β)bt−1 ,

with the one-step forecast as Ft+1 = Lt + bt . Hence the one-step error is


et = Yt − Lt−1 − bt−1 . The first row can be written

Lt = α(Yt − Lt−1 − bt−1 ) + Lt−1 + bt−1


= αet + Lt−1 + bt−1

and the second row can be written

bt = β(Lt − Lt−1 ) + (1 − β)bt−1


= bt−1 + β(Lt − Lt−1 − bt−1 )
= bt−1 + βαet

using the first equation.


Now let Xt,1 = Lt and Xt,2 = bt . Then the state space form of the model is
   
1 1 α
Xt = X t−1 + et
0 1 βα

Yt = [1 1] X t−1 + et .

(d) The state space form might be preferable because


• it allows missing values to be handled easily;
• it is easy to generalize to allow the parameters to change over time;
• the Kalman recursion equations can be used to calculate the forecasts and
likelihood.
Chapter 9: Forecasting the long term 155

Chapter 9: Forecasting the long term

9.1 There is little doubt that the trends in computer power and memory show a very
clear exponential growth while that of price is declining exponentially. It is therefore
a question of time until computers that cost only a few hundred dollars will exist
that can perform an incredible array of tasks which until now have been the sole
prerogative of humans, for example playing chess (a high-power judgmental and
creative process). It is therefore up to our imaginations to come up with future
scenarios when such computers will be used as extensively as electrical appliances
are used today. The trick is to free our thinking process so that we can come up with
scenarios that are not constrained by our perception of the present when computers
are being used mostly to make calculations.

9.2 As the cost of computers (including all of the peripherals such as printers and scan-
ners) is being reduced drastically, and at the same time we will be getting soon to
devices that will perform a great number of functions now done by separate ma-
chines, it will become more practical and economical to work at home. Furthermore,
the size of these all-purpose machines is being continuously reduced. In the next
five to ten years we will be able to have everything that is provided to us now in
an office at home with two machines: one a powerful all-inclusive computer and the
other a printer-scanner-photocopier-fax machine. Moreover these two machines will
be connected to any network we wish via modems so that we can communicate and
get information from anywhere.

9.3 As it was also mentioned in Exercise 9.1, there is no doubt that the trend in computer
and equipment prices are declining exponentially at a fast rate. This would make it
possible for everyone to be able to afford them and be able to have an office not only
at home but at any other place he or she wishes, including one’s car, a hotel room,
a summer vacation residence, or a sail boat.

9.4 Statements like those referred to in Exercise 9.4 abound and demonstrate the short-
sightedness of peoples’ ability to predict the future. As a matter of fact as late
as the beginning of our century people did not predict all four major inventions of
the Industrial Revolution (cars, telephones, electrical appliances and television) that
have dramatically changed our lives. Moreover, they did not predict the huge impact
of computers even as late as the beginning of the 1950s. This is why we must break
from our present mode of thinking and see things in a different, new light. This is
where scenarios and analogies can be extremely useful.
156 Part D. Solutions to exercises

Chapter 10: Judgmental forecasting and adjustments

10.1 Phillips’ problems have to do with the management bias of overoptimism, that is
believing that all changes will be successful and that they can overcome peoples’
resistance to change. This is not true, but we tend to believe that most organisational
changes are successful because we hear and we read about the successful ones while
there is very little mention of those that fail. Introducing changes must be considered,
therefore, in an objective manner and our ability to succeed estimated correctly.

10.2 The quote by Glassman illustrates the extent to which professional, expert invest-
ment managers underperform the average of the market. Business Week, Fortune
and other business journals regularly publish summary statistics of the performance
of mutual funds and other professionally-managed funds, benchmarking them with
the Standard & Poor or other appropriate indexes. The instructor can therefore get
some more recent comparisons than those shown in Chapter 10 and show them in
class.

10.3 Assuming that the length of cycles varies considerably we have little way to say how
long it will take until the expansion started in May 1991 will be interrupted. Un-
fortunately the length (and depth) of cycles varies a great deal making it extremely
difficult to say how long an expansion will last. It will all depend on the specific
situation involved that will require judgmental inputs, structured in such a way as
to avoid biases and other problems.

10.4 There are twenty 8s that one will encounter when counting from 0–100. When given
this exercise most people say nine or ten because they are not counting the eights
coming from 81 to 87, and 89 (they usually count the 8s in 88 often one time).
Chapter 11: The use of forecasting methods in practice 157

Chapter 11: The use of forecasting methods in practice

11.1 The results of Table 1 are very similar to those of the previous M-Competitions. As
a matter of fact the resemblance is phenomenal given the fact that the series used
and the time horizon they refer to are completely different.

11.2 In our view the combined method will do extremely well. More specifically its accu-
racy will be higher than the individual methods being combined while its variance
of forecasting error will be smaller than that of the methods involved.

11.3 It seems that proponents of new forecasting methods usually exaggerate their ben-
efits. This has been the case with methods under the banner of neural networks,
machine learning and expert systems. These methods did not do well in the M3-IJF
Competition. In addition only few experts participated in the competition using
such methods, even though more than a hundred were contacted (and invited to
participate) and more than fifty expressed an interest in the M3-IJF Competition,
indicating that they would “possibly” participate . In the final analysis it seems that
it is not so simple to run a great number of series by such methods resulting in not
too many participants from such methods.
158 Part D. Solutions to exercises

Chapter 12: Implementing forecasting: its uses, advantages, and


limitations

The exercises for Chapter 12 are general and can be answered by referring to the text
of Chapter 12 which covers each one of the topics. Each instructor can therefore form
his/her way of answering these exercises.

You might also like