You are on page 1of 18

Journal of Applied Statistics, 2014

Vol. 41, No. 4, 785801, http://dx.doi.org/10.1080/02664763.2013.856383

An empirical study on technical analysis:


GARCH (1, 1) model
Show-Lin Chena , Nen-Jing Chena and Rwei-Ju Chuangb
a Department

of Economics, Fu-Jen Catholic University, #510, Chung-Cheng Road, New Taipei City 242,
Taiwan; b Department of Statistics and Information Science, Fu-Jen Catholic University, #510,
Chung-Cheng Road, New Taipei City 242, Taiwan
(Received 10 September 2012; accepted 14 October 2013)

One of the deficits of the common Bollinger band is that it fails to consider the fat tails/leptokurtosis
often exists in financial time series. An adjusted Bollinger band generated by rolling GARCH regression
method is proposed in this study. The performance of the adjusted Bollinger band strategy on EUR, GBP,
JPY, and AUD vs. USD foreign exchange trading is evaluated. Results show that in general, the adjusted
Bollinger band performs better than the traditional one in terms of success ratios, net successes, and profit.
In addition, no matter there is transaction cost or not, only adjusted Bollinger strategies are recommended
for investors. Adjusted Bollinger band strategies with MA 5 or 10 are recommended for EUR, GBP, and
JPY. Adjusted Bollinger strategy with MA 20 is the recommended strategies for AUD.
Keywords:

technical analysis; Bollinger band; rolling regression; GARCH(1,1); exchange rate

JEL Classification: F31; G11

1.

Introduction

Exchange rate forecast has been an important subject in the field of international finance research as
well as practice. Traditionally, statistical methodology and empirical analysis based on exchange
rate determination models have been used as the major forecasting techniques. However, exampled
by Meese and Rogoff [16], Cheung et al. [4] and Engle et al. [5], they are not very successful in
forecasts.
The failure of traditional exchange rate models and/or forecasting techniques leads to the use
of microstructure data (e.g. order flow) in exchange rate movement forecasting. According to
Lyons [15], positive order flow in the case of the euro denotes net demand or buying pressure
for euros (following the convention of quoting prices in dollars per euro). Positive order flow in
the case of the yen denotes net demand for dollars or selling pressure for yens (following the
convention of quoting prices in yen per dollar). Evans and Lyons [7,8] indicate the existence of
Corresponding

author. Email: ecos1005@mail.fju.edu.tw

2013 Taylor & Francis

786

S.-L Chen et al.

a close link between daily exchange rate movements and order flow. Evans and Lyons [8,9] find
that order flow has a significant explanatory and forecasting power in exchange rate. It bases on
the view that heterogeneous beliefs are essential to determining prices. The importance of order
flow implies that market information of heterogeneous beliefs should be paid attention in price
forecasting. Therefore, buyer and sellers behavior will be an important aspect to be taken into
account while studying the trend and movement of exchange rate.
It is well known that the order flow data of foreign exchange trade are not as easy to access
as those of the stock trade. Technical analysis, similar to order flow as being a buying/selling
indicator, is widely applied in the price movement forecast in practice exampled by Bollinger
band. Taylor and Allen [22] point out that over 90% of the foreign exchange traders use technical
analysis in their trading decisions. In addition, a number of empirical studies on foreign exchange
rate show profitability of using the technical analysis. Sweeney [21] applies filter rules on the
trading of 10 currencies and finds profitability over 80% of the cases and among which, one
third shows significant returns. Levich and Thomas [14] point out that investment strategy using
moving average and filter rule helps in making a significant profit in GBP, CAD, DM, JPY, and
SF. LeBaron [13] shows that the moving average method helps to gain remarkable return in DM
and JPY trading. Osler and Chang [18] simulate the head and shoulder strategy on six currencies
and obtain significant profit on JY and DM. Silber [20] finds technical analysis to be profitable
in financial market but not in commodity market.
The Bollinger bands, determines overbought and oversold conditions in the market and is one of
the most popular technical analysis techniques for finding investment opportunities, does not seem
to be studied yet in financial literature. It constructs the upper and lower bands to form the buying
and selling signals, therefore contains more statistical implications than other technical analysis.
However, the widely used traditional Bollinger band does not consider the fact that financial series
has a fat tail and/or leptokurtic distribution as indicated in the financial literature. Therefore, this
study uses GARCH model to estimate the standard deviation and form an adjusted Bollinger band
and then carries out a trading simulation based on a widely used Bollinger trading strategy. In
sum, this research proposes an adjusted Bollinger band by using a more sophisticated method and
simulates the trading performance of the adjusted Bollinger band over the traditional Bollinger
band. The contribution of this paper is to combine econometric methodology with technical
analysis and therefore narrow the gap between the academic research and practical application.
The rest of the paper is organized as follows. Section 2 briefly describes the model and
methodology. Section 3 illustrates empirical results. Finally, Section 4 provides conclusions.
2.

Methodology

Traditional and adjusted Bollinger bands are first estimated and then applied in the trading strategy.
2.1

Estimation of traditional Bollinger bands

Three Bollinger bands are constructed to form a trading indicator, they are middle, upper, and
lower bands. The middle band is an n-day moving average of the price series. And traditionally,
the upper and lower bands are two times an n-day standard deviation above and below the middle
band, respectively. It is assumed that Yt is the currency price series at time t. The traditional
Bollinger band calculation procedure is as follows:
(1) find the n-day moving average at day t:
n1

MAt =


Yti

i=0

Journal of Applied Statistics

787

(2) compute the n-day standard deviation:

0.5
2
(Y

MA
)
ti
t

i=0

SDt =
,

(n 1)
n1


(3) calculate the upper band (UB) and lower band (LB):
UBt = MAt + 2 SDt ,
LBt = MAt 2 SDt .
2.2

Estimation of adjusted Bollinger bands

Numerous literatures have indicated that financial series has a fat tail and/or leptokurtic distribution. In time series setting, the proposed model, GARCH(1,1), is useful in describing volatility
clustering nature, which in turn explain, at least partially, the fat-tail phenomenon truly present in
price distributions [1,19]. Empirical work of Bollerslev [3] shows that the kurtosis implied by a
GARCH(1,1) process is greater than 3, the kurtosis of a normal distribution. Thus a GARCH(1,1)
model may replicate some of the fat-tailed behavior observed in financial time series. It has been
applied to a wide range of time series analyses. Its applications in finance have been particularly
successful. Therefore, this study first fits a GARCH(1, 1) model to the foreign exchange rates and
gets the estimated price series.1
The mean equation of a GARCH(1,1) model is
Yt =

Yti i + ut ,

i=0

where  is the difference operator and Yt is a market price series. The largest lag term (p)
considered in this study is 4.2 If i is not significantly different from 0, then Yti will be
removed from the model.
The variance equation is
ut =


ht vt ,

2
ht = + ht1 + ut1
,

where > 0, 0, 0, and + < 1 are required.


The fitted value of exchange rate (Y t = Y t + Y t1 ) generating method and the associated
adjusted Bollinger band are described as follows. The difference between the true and estimated
currency price series is estimated by recursive
pregression from which one obtains the fitted series.
That is, through estimating equation Yt = i=0 Yti i + ut , one obtains Y t , and Y t = Y t +
Yt1 . Y t is then used to calculate the standard deviation for establishing the upper and lower bands.
The upper and lower bands are composed of the points UBt and LBt along the time. A rolling
regression method is applied. That is, the first estimation uses observation 1 through t, and the
second estimation uses observation 2 through t + 1, and so on. If the number of observation is N,
then the number of regression estimated will be N t + 1. The number of bands estimated will
be reduced by the number of days the moving average is based on.

788
2.3

S.-L Chen et al.


Simulation of Bollinger band trading strategy

The trading performances of traditional Bollinger band and the adjusted Bollinger band are simulated and compared. When the trading price goes above (below) the upper (lower) band, the
market is considered to be over bought (sold). Therefore, the investor might gain through selling
(buying). In this study, when the daily high (low) price goes above (below) the upper (lower) band,
the investor sells (buys) at the closing price and then offsets the position at the next closing price
[10]. The signal is said to be successful/profitable if the investor sells (buys) and then liquidates
the position at a lower (higher) price.
The adjusted and traditional bands performance is evaluated by three criteria: success ratio
(success/signal), net success (successfailure), and profit. Here the traditional Bollinger band is
used as a benchmark.

3.

Results and discussion

EUR, GBP, JPY, and AUD are studied for the former three are the most traded currencies and
the last one represents an important commodity currency. EUR, GBP, JPY, and AUD daily high,
low, and closing exchange rates vs. USD from 1 January 2008 to 15 July 2011 are obtained from
FXCM.3
Before performing GARCH model estimation, one must ensure that the variable is stationary.
Firstly, the rolling regression results of ADF unit root test on EUR/USD, GBP/USD, USD/JPY,
and AUD/USD are shown in Figures 14, respectively. The figure illustrates the unit root statistics
of different currencies for level and differential value (diff) as well as critical value (crt). When the
statistics are higher than the critical value (crt), the null hypothesis of nonstationarity will not be
rejected, otherwise variables will be stationary. Figures 14 indicate that almost all the variables
are I(1); therefore, the variables will be taken first difference before model estimation.

4
2
1
28
55
82
109
136
163
190
217
244
271
298
325
352
379
406
433
460
487
514
541
568
595
622
649
676
703
730
757
784

0
-2
-4
-6
-8
-10
-12
level

diff

crt

Figure 1. ADF test statistics on level and difference as well as critical values EUR/USD.

4
2
1
38
75
112
149
186
223
260
297
334
371
408
445
482
519
556
593
630
667
704
741
778
815
852
889
926
963
1000
1037
1074

0
-2
-4
-6
-8
-10
level

diff

crt

Figure 2. ADF test statistics on level and difference as well as critical values GBP/USD.

Journal of Applied Statistics

789

2
1
28
55
82
109
136
163
190
217
244
271
298
325
352
379
406
433
460
487
514
541
568
595
622
649
676
703
730
757
784

0
-2
-4
-6
-8
-10
-12
level

diff

crt

Figure 3. ADF test statistics on level and difference as well as critical values USD/JPY.

2
1
28
55
82
109
136
163
190
217
244
271
298
325
352
379
406
433
460
487
514
541
568
595
622
649
676
703
730
757
784

0
-2
-4
-6
-8
-10
-12
level

diff

crt

Figure 4. ADF test statistics on level and difference as well as critical values AUD/USD.

An ARCH-LM test is carried out on all four series of currency rates. When the test statistics is
less than critical value, the null hypothesis of homogeneity of variance cannot be rejected. The
results indicate that 88.19%, 74.27%, 70.85%, and 84.69% of the rolling series of EUR/USD,
GBP/USD, USD/JPY, and AUD/USD have ARCH effects under 5% level of significance with
degrees of freedom equals 36, respectively. The results are shown in Figures 58. These results
support the use of GARCH model in estimating volatility.
The numbers of signal and success, and the level of success ratio, net success as well as
profit from adjusted and traditional Bollinger bands under various moving averages (5, 10, 20,
and 30) for EUR, GBP, JPY, and AUD are presented in Tables 14, respectively. In each table,
the results obtained from price crossing UB, LB, and both bands are shown and discussed in
sequence. The number of signals generated by Bollinger bands does not mean much unless the
signals have high qualities. In this study, the success ratio, net success, and profit4 are used as
criteria of the quality of the signals. Among the three criteria, the profit is the most relevant one
for the investors. The necessary conditions for a strategy to be recommended to investors are:
(a) success ratio is higher than 50%, (b) net success is non-negative, and (c) profit is positive.
Among the strategies which satisfy the recommendations necessary conditions, the dominant
strategy which has highest success ratio, net success, and profit will be recommended to the
investors. If there is no single dominant strategy, the strategy which generates the best profit will
be the recommended strategy. In the later sections/subsections, the terms adjusted Bollinger and
traditional Bollinger are sometimes used to represent adjusted Bollinger band(s) and traditional
Bollinger band(s), respectively. The performance of Bollinger band strategy will be discussed
firstly without considering transaction cost and secondly with considering transaction cost. The
signal qualities of traditional and adjusted Bollinger bands are compared in the without transaction
cost section. Strategies are recommended to investors in the with transaction cost section.
No discussion of strategy recommendation is provided in the without transaction cost section
because a recommended strategy should be a strategy which has taken transaction cost into

1
55
109
163
217
271
325
379
433
487
541
595
649
703
757
811
865
919
973
1027
1081
1135
1189
1243
1297
1351
1405
1459
1513

1
53
105
157
209
261
313
365
417
469
521
573
625
677
729
781
833
885
937
989
1041
1093
1145
1197
1249
1301
1353
1405
1457
1509

160
140
120
100
80
60
40
20
0
1
53
105
157
209
261
313
365
417
469
521
573
625
677
729
781
833
885
937
989
1041
1093
1145
1197
1249
1301
1353
1405
1457
1509

160
140
120
100
80
60
40
20
0

1
53
105
157
209
261
313
365
417
469
521
573
625
677
729
781
833
885
937
989
1041
1093
1145
1197
1249
1301
1353
1405
1457
1509

790
S.-L Chen et al.

eur

gbp

jpy

aud

Figure 8. ARCH test on AUD/JPY.


crt

Figure 5. ARCH test on EUR/USD.

200

150

100

50

crt

Figure 6. ARCH test on GBP/USD.

crt

Figure 7. ARCH test on USD/JPY.

250

200

150

100

50

crt

Journal of Applied Statistics

791

consideration. However, strategy recommendation without transaction cost will be provided in


the summary of results section for reference purpose.
The only transaction cost involved in the currency margin trade is the bidask spread. The bid
ask spread varies depending on the currency traded, the size of investing, and the trading hour of
the day. As the currencies studied are all major currencies in the world, the bidask spread is low.
For an individual investor depositing about US$10,000 in the margin account, the bidask spread
is mostly 2 pips and is within the range of 24 pips for major currency rates. For EUR/USD,
GBP/USD, and AUD/USD, 1 pip means 0.0001 USD; for USD/JPY 1 pip is 0.01 JPY. Since
the currency rates collected is the mid-point of bid and ask rates, therefore, transaction cost per
round trade of 24 pips are considered in this study. The profit with transaction cost is calculated
by deducting the product of transaction cost per round trade and the number of signals from the
profit without transaction cost.

3.1

The performance of Bollinger band strategy for EUR

3.1.1 Without transaction cost


3.1.1.1 Upper band. Table 1 shows the results on EUR. One can see the results from UB rows
that the adjusted Bollinger band generates higher success ratio and net successes than traditional
Table 1. Simulation results for EUR under various MAs.
MA = 5

MA = 10

Adjusted Traditional Adjusted Traditional


UB
Signal (#)
270
Success (#)
132
Success ratio (%)
48.89
Net success (#)
6
Profit (w/o)
0.0616
Profit (w/0.0002) 0.0076
Profit (w/0.0003) 0.0194
Profit (w/0.0004) 0.0464

MA = 20

MA = 30

Adjusted

Traditional

Adjusted

Traditional

161
73
45.34
15
0.0033
0.0289
0.045
0.0611

121
64
52.89
7
0.0143
0.0385
0.0506
0.0627

30
14
46.67
2
0.0363
0.0423
0.0453
0.0483

185
90
48.65
5
0.0936
0.1306
0.1491
0.1676

105
23
21.90
59
0.0720
0.093
0.1035
0.114

107
56
52.34
5
0.0578
0.0792
0.0899
0.1006

125
22
17.60
81
0.0142
0.0392
0.0517
0.0642

LB
Signal (#)
Success (#)
Success ratio (%)
Net success (#)
Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

234
126
53.85
18
0.1220
0.0752
0.0518
0.0284

139
71
51.08
3
0.0855
0.0577
0.0438
0.0299

100
58
58.00
16
0.1928
0.1728
0.1628
0.1528

25
14
56.00
3
0.0631
0.0581
0.0556
0.0531

115
59
51.30
3
0.0204
0.0026
0.0141
0.0256

69
15
21.74
39
0.0561
0.0423
0.0354
0.0285

75
37
49.33
1
0.0345
0.0495
0.057
0.0645

48
8
16.67
32
0.0345
0.0249
0.0201
0.0153

UB and LB
Signal (#)
Success (#)
Success ratio (%)
Net success (#)
Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

504
258
51.19
12
0.1836
0.0828
0.0324
0.018

300
144
48.00
12
0.0888
0.0288
0.0012
0.0312

221
122
55.20
23
0.1785
0.1343
0.1122
0.0901

55
28
50.91
1
0.0268
0.0158
0.0103
0.0048

300
149
49.67
2
0.0732
0.1332
0.1632
0.1932

174
38
21.84
98
0.0159
0.0507
0.0681
0.0855

182
93
51.10
4
0.0923
0.1287
0.1469
0.1651

173
30
17.34
113
0.0203
0.0143
0.0316
0.0489

Note: Numbers in shade indicate the inferiority of adjusted Bollinger without considering transaction cost.

792

S.-L Chen et al.

Table 2. Simulation results for GBP under various MAs.


MA = 5
Adjusted
UB
Signal (#)
Success (#)
Success ratio (%)

MA = 10

Traditional Adjusted

MA = 20

Traditional Adjusted

MA = 30

Traditional Adjusted

Traditional

279
138
49.46

174
87
50.00

136
71
52.21

33
12
36.36

162
81
50.00

109
28
25.69

94
48
51.06

114
17
14.91

53

80

Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

0.0811
0.1369
0.1648
0.1927

0.0787
0.1135
0.1309
0.1483

0.1181
0.0909
0.0773
0.0637

0.0143
0.0209
0.0242
0.0275

0.0021
0.0303
0.0465
0.0627

0.1476
0.1694
0.1803
0.1912

0.0259
0.0071
0.0023
0.0117

0.0031
0.0259
0.0373
0.0487

LB
Signal (#)
Success (#)
Success ratio (%)
Net success (#)
Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

228
120
52.63
12
0.2975
0.2519
0.2291
0.2063

150
80
53.33
10
0.1434
0.1134
0.0984
0.0834

112
60
53.57
8
0.1912
0.1688
0.1576
0.1464

30
15
50.00
0
0.0351
0.0291
0.0261
0.0231

128
63
49.22
2
0.0941
0.0685
0.0557
0.0429

99
17
17.17
65
0.0376
0.0178
0.0079
0.002

59
29
49.15
1
0.0113
0.0005
0.0064
0.0123

32
5
15.63
22
0.0310
0.0374
0.0406
0.0438

UB and LB
Signal (#)
Success (#)
Success ratio (%)

507
258
50.89

324
167
51.54

248
131
52.82

63
27
42.86

290
144
49.66

208
45
21.63

153
77
50.33

146
22
15.07

Net success (#)


Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

9
0.2164
0.115
0.0643
0.0136

10
0.0647
0.0001
0.0325
0.0649

14
0.3093
0.2597
0.2349
0.2101

9
0.0209
0.0083
0.002
0.0043

2
0.0962
0.0382
0.0092
0.0198

118
0.1101
0.1517
0.1725
0.1933

1
0.0372
0.0066
0.0087
0.024

102
0.0342
0.0634
0.078
0.0926

Net success (#)

Note: Numbers in shade indicate the inferiority of adjusted Bollinger without considering transaction cost.

one. The adjusted Bollinger generates higher profits/lower loss than the traditional one under
MA = 5 and 10.
The minimum success ratio of the adjusted Bollinger band is 48.65%. It is over 50% under
MA = 10 and 30. While the maximum success ratio of the traditional band is 46.67%, they are
as low as 21.9% and 17.6% under MA 20 and 30, respectively. The net successes are all negative
under traditional Bollinger, which of the adjusted Bollinger are positive under MA 10 and 30.
MA 5 is the only MA which generates positive profits for either adjusted or traditional Bollinger.
Under MA 5, the performance of adjusted Bollinger is better than the traditional one based
on all three criteria. However, no strategy passes the necessary conditions for a strategy to be
recommended.
3.1.1.2 Lower band. One can see the results from LB rows that the adjusted Bollinger band
generates higher success ratio and net successes than traditional one. The adjusted Bollinger
generates higher profits than the traditional one under MA = 5 and 10.
The minimum success ratio from adjusted Bollinger strategy is 49.33% under MA 30. It is
above 50% under MA 5, 10, and 20. While which of the traditional one are only 21.74% and
16.67% under MA 20 and 30, respectively. The net successes of traditional Bollinger are -39 and
-32 under MA 20 and 30, while the minimum net success of adjusted Bollinger is -1 under MA

Journal of Applied Statistics

793

Table 3. Simulation results for JPY under various MAs.


MA = 5

MA = 10

MA = 20

Adjusted Traditional Adjusted Traditional Adjusted Traditional

MA = 30
Adjusted

Traditional

71
21
29.58
29
5.9930
7.413
8.123
8.833

58
31
53.45
4
2.4420
3.602
4.182
4.762

48
8
16.67
32
2.8660
1.906
1.426
0.946

UB
Signal (#)
Success (#)
Success ratio (%)
Net success (#)
Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

226
115
50.88
4
9.0820
4.562
2.302
0.042

160
68
42.50
24
7.0120
10.212
11.812
13.412

91
47
51.65
3
7.6280
5.808
4.898
3.988

24
13
54.17
2
2.9700
3.45
3.69
3.93

LB
Signal (#)
Success (#)
Success ratio (%)
Net success (#)
Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

262
131
50.00
0
0.4600
4.78
7.4
10.02

179
86
48.04
7
5.3670
1.787
0.003
1.793

134
70
52.24
6
4.2760
1.596
0.256
1.084

43
16
37.21
11
2.4110
1.551
1.121
0.691

163
114
90
20
55.21
17.54
17
74
69.6770 80.0420
72.937 82.322
74.567 83.462
76.197 84.602

77
44
57.14
11
71.5150
73.055
73.825
74.595

50
12
24.00
26
73.1380
74.138
74.638
75.138

UB and LB
Signal (#)
Success (#)
Success ratio (%)
Net success (#)
Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

488
246
50.41
4
9.5420
0.218
5.098
9.978

339
154
45.43
31
1.6450
8.425
11.815
15.205

225
117
52.00
9
11.9040
7.404
5.154
2.904

67
29
43.28
9
0.5590
1.899
2.569
3.239

269
185
141
41
52.42
22.16
13
103
70.0170 86.0350
75.397 89.735
78.087 91.585
80.777 93.435

135
75
55.56
15
73.9570
76.657
78.007
79.357

98
20
20.41
58
70.2720
72.232
73.212
74.192

106
51
48.11
4
0.3400
2.46
3.52
4.58

Note: Numbers in shade indicate the inferiority of adjusted Bollinger without considering transaction cost.

30. The profits from traditional Bollinger are all positive whereas the profit of the adjusted one is
negative under MA 30.
Adjusted Bollinger with MA 5, 10, and 20 as well as traditional Bollinger with MA 5 and 10
pass the necessary conditions of recommendation strategy. The adjusted Bollinger with MA 20 is
dominated by which with MA 5 and 10. There is no dominant traditional strategy. Therefore, for
both adjusted and traditional Bollinger, MA 5 and 10 are the best choices. Under MA 5 and 10,
the performance of adjusted Bollinger is better than the traditional one based on all three criteria.
As adjusted Bollinger with MA 10 is better than MA 5 in terms of profit, adjusted Bollinger with
MA 10 is recommended to the investors.
Comparing the results from LB and UB, one can see that buying EUR when the prices cross the
LB from above generally gives better results in terms of three criteria with only a few exceptions.
The exceptions include lower success ratio and net success under MA 30 for adjusted Bollinger
as well as lower success ratios under MA20 and 30 for traditional Bollinger.
3.1.1.3 Both bands. From the UB and LB result, the adjusted Bollinger strategy generates
higher success ratio and net success than the traditional one. The adjusted Bollinger generates
higher profits than the traditional one under MA 5 and 10.
The adjusted Bollingers success ratios are above 50% except being 49.67% under MA 20.
However, the success ratio of the traditional Bollinger is above 50% only under MA 10. They

794

S.-L Chen et al.

Table 4. Simulation results for AUD under various MAs.


MA = 5

MA = 10

MA = 20

MA = 30

Adjusted

Traditional

Adjusted

Traditional

Adjusted

UB
Signal (#)
Success (#)
Success ratio (%)

304
134
44.08

155
75
48.39

157
63
40.13

108
40
37.04

138
59
42.75

102
40
39.22

70
30
42.86

138
62
44.93

Net success (#)

36

31

28

20

22

14

Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

0.1880
0.2488
0.2792
0.3096

0.0520
0.083
0.0985
0.114

0.1447
0.1761
0.1918
0.2075

0.0973
0.1189
0.1297
0.1405

0.1231
0.1507
0.1645
0.1783

0.1171
0.1375
0.1477
0.1579

0.0479
0.0619
0.0689
0.0759

0.1011
0.1287
0.1425
0.1563

LB
Signal (#)
Success (#)
Success ratio (%)

241
125
51.87

157
86
54.78

170
88
51.76

134
66
49.25

104
57
54.81

79
44
55.70

40
15
37.5

27
9
33.33

Net success (#)


Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

9
0.0566
0.0084
0.0157
0.0398

15
0.0407
0.0093
0.0064
0.0221

6
0.0542
0.0202
0.0032
0.0138

2
0.0245
0.0513
0.0647
0.0781

10
0.1324
0.1116
0.1012
0.0908

9
0.1023
0.0865
0.0786
0.0707

0
0.0582
0.0662
0.0702
0.0742

9
0.0824
0.0878
0.0905
0.0932

UB and LB
Signal (#)
Success (#)
Success ratio (%)

545
259
47.52

312
161
51.60

327
151
46.18

242
106
43.80

242
116
47.93

181
84
46.41

110
45
40.91

165
71
43.03

Net success (#)


Profit (w/o)
Profit (w/0.0002)
Profit (w/0.0003)
Profit (w/0.0004)

Traditional Adjusted Traditional

27

10

25

30

10

13

20

23

0.1314
0.2404
0.2949
0.3494

0.0113
0.0737
0.1049
0.1361

0.0905
0.1559
0.1886
0.2213

0.1218
0.1702
0.1944
0.2186

0.0093
0.0391
0.0633
0.0875

0.0148
0.051
0.0691
0.0872

0.1061
0.1281
0.1391
0.1501

0.1835
0.2165
0.233
0.2495

Note: Numbers in shade indicate the inferiority of adjusted Bollinger without considering transaction cost.

are as low as 21.84% and 17.34% under MA 20 and 30, respectively. There is only one positive
net success under traditional Bollinger, which is 1 under MA 10. There is only one negative net
success under adjusted Bollinger, which is -2 under MA 20. The profit of the adjusted Bollinger
is below zero under MA = 20 and 30. While which of the traditional one is below zero only
under MA = 20. The best profit of adjusted Bollinger and traditional one is 0.1836 and 0.0888,
respectively. They both happen when MA is 5.
For both adjusted and traditional Bollinger, MA 5 and 10 are the best choices as they have
higher success ratios, net successes, and profits than the other MAs. The adjusted Bollinger with
MA 5 and 10 as well as the traditional Bollinger with MA 10 pass the necessary conditions
of recommendation strategy. However, under both MA 5 and 10, the performance of adjusted
Bollinger is better than the traditional one based on all three criteria. As adjusted Bollinger with
MA 5 is better than MA 10 in terms of profit, adjusted Bollinger with MA 5 is recommended to
the investors while investing in the EUR.
3.1.2 With transaction cost
Like the results from without transaction cost, there is no recommended strategy derived from
UB and the recommended strategy for LB is adjusted Bollinger with MA 10 for 3 levels of

Journal of Applied Statistics

795

transaction cost considered in this study. For UB and LB together, the recommended strategy
changed from adjusted Bollinger with MA 5 to MA 10 after considering transaction cost. This
result is not surprising, because MA 5 only generates a slightly (2.86%) higher profit than which
of MA 10 under zero transaction cost. Since MA 10 generates less than half signals (trades) than
MA 5, it incurs less transaction costs and performs better than MA 5 when transaction costs are
considered.5

3.2 The performance of Bollinger band strategy for GBP


3.2.1 Without transaction cost
3.2.1.1 Upper band. Table 2 is the results on GBP. One can see the results from UB rows
that the adjusted Bollinger band generates higher success ratio, net successes, and profits than
traditional one except under MA 5.
The minimum success ratio of the adjusted Bollinger band is 49.46% under MA 5. It is no less
than 50% under other MAs. The success ratio of traditional Bollinger is 50% under MA 5 and
below 40% under all other MAs.
The net successes are no more than 0 under traditional Bollinger. They are no less than 0 under
adjusted Bollinger except under MA 5 (net success is -3). Traditional band generate negative profits
under all MAs. Adjusted Bollinger produces positive profits except under MA 5. Only adjusted
Bollinger bands under MA 10, 20, and 30 pass the necessary conditions of recommendation
strategy. Among them, the adjusted Bollinger with MA 10 performs best by all three criteria.
3.2.1.2 Lower band. One can see the results from LB rows that the adjusted Bollinger band
generates higher success ratio, net successes, and profits than traditional one, except it has a
slightly (1.33%) lower success ratio under MA 5.
MA 20 and 30 generates positive profits for both adjusted and traditional Bollinger except traditional Bollinger under MA 30. However, under MA 20 and 30, adjusted Bollingers success ratios
are slightly below 50% (49.22% and 49.15%, respectively) and traditional Bollingers success
ratios are 17.17% and 15.63%. Moreover, under MA 20 and 30, the net successes are negative for
both adjusted and traditional Bollinger bands. Therefore, neither adjusted nor traditional Bollinger
has high quality.
Under MA 5 and 10, both adjusted and traditional Bollingers success ratios are above 50%,
net successes are positive, and profits are positive and higher than those under MA 20 and 30 with
only one exception. The exception is that the profit of traditional band under MA 10 (0.0351) is
slightly (7.12%) lower than which under MA 20 (0.0376).
Adjusted and traditional Bollinger with MA 5 and 10 pass the necessary conditions of recommendation strategy. Among these four strategies, the traditional Bollinger under MA 10 is
dominated by the others. Adjusted Bollinger with MA 5 is better than adjusted Bollinger with
MA 10 as well as traditional Bollinger with MA 5 in terms of profit. Therefore, the adjusted
Bollinger with MA 5 is recommended to the investors.
Comparing the results from LB and UB, one can see that Bollinger band strategies with MA 5
and 10 perform better from buying GBP when the prices cross the LB from above than when the
prices cross the UB from below in terms of three criteria.
3.2.1.3 Both bands. From the UB and LB result, the adjusted Bollinger strategy generates
higher success ratio, net success, and profit than the traditional one except a slightly (1.28%)
lower success ratio (50.89% vs. 51.54%) and slightly (11.11%) lower net success (9 vs. 10) under
MA 5.

796

S.-L Chen et al.

The adjusted Bollingers success ratios are above 50% except being 49.66% under MA 20.
However, the success ratio of the traditional Bollinger is above 50% only under MA 5. They
are as low as 21.63% and 15.07% under MA 20 and 30, respectively. Only under MA 5, the
traditional Bollinger has a positive net success. While only under MA 20, the adjusted Bollinger
has a negative net success.
The profit of the traditional Bollinger is below zero under MA = 20 and 30. While which of the
adjusted one is all above zero. The best profit of adjusted Bollinger and traditional one is 0.3093
and 0.0647, under MA 10 and 5, respectively.
Adjusted Bollinger bands under MA 5, 10, and 30 and traditional Bollinger bands under MA 5
pass the necessary conditions of recommendation strategy. The adjusted Bollinger strategy with
MA 10 is a dominant strategy for it outperforms the other three strategies by all three criteria.
Therefore, the investors are suggested to use the adjusted Bollinger band with MA 10 while
investing in the GBP.
3.2.2 With transaction cost
The recommended strategy is exactly the same when transaction cost is considered. That is, for
the upper, lower, and both bounds, the recommended strategy is adjusted Bollinger with MA 10,
5, and 10, respectively.

3.3

The performance of Bollinger band strategy for JPY

3.3.1 Without transaction cost


3.3.1.1 Upper band. Table 3 is the results on JPY. One can see the results from UB rows
that the adjusted Bollinger band generates higher success ratio, net successes, and profits than
traditional one, except a slightly (4.88%) lower success ratio (51.65% vs. 54.17%) under MA 10.
The minimum success ratio of the adjusted Bollinger band is 48.11% under MA 20. It is greater
than 50% under other MAs. The success ratios of traditional Bollinger are below 45% except
under MA 10 (54.17%).
Traditional Bollingers net successes are negative except under MA10. Adjusted Bollingers
net successes are positive except under MA 20. Traditional band generate negative profits except
under MA 30. Adjusted Bollinger produces positive profits except under MA 20.
For adjusted and traditional Bollinger, MA 5 and MA 30 performs best from profitability viewpoint, respectively. But only adjusted Bollinger with MA 5 and 10 pass the necessary conditions
for recommendation strategy. Adjusted Bollinger with MA 5 is better than adjusted Bollinger with
MA 10 in terms of profit. Therefore, the adjusted Bollinger with MA 5 is recommended to the
investors.
3.3.1.2 Lower band. One can see the results from LB rows that the adjusted Bollinger band generates higher success ratio, net successes, and profits than traditional one, except it produces a lower
profit
under MA 5.
Adjusted Bollingers success ratios are above 50% and net success ratios are positive under
MA 20 and 30. However, MA 20 generates negative profits for adjusted Bollinger bands. Its
an indication that the average gain is below average loss. Under MA 20 and 30, traditional
Bollingers success ratios are way below 50%, net successes are negative and so are profits. None
of the strategies under MA 20 and 30 pass the necessary conditions for recommendation strategy.
Under MA 5 and 10, adjusted Bollingers success ratios are no less than 50%, net successes are
non-negative, and profits are positive.

Journal of Applied Statistics

797

Though the traditional Bollinger under MA 5 generates highest profit among all the strategies,
it fails to pass the necessary conditions for recommendation strategy. Adjusted Bollinger with MA
5 and 10 are the only two strategies pass the necessary conditions for recommendation. Among
the two, MA 10 is better than MA 5 by all 3 criteria.
From the results on JPY, one cant judge whether the Bollinger band strategy performs better
when prices cross lower or upper band.
3.3.1.3 Both bands. From the UB and LB result, the adjusted Bollinger strategy generates
higher success ratio, net success, and profit than the traditional one except a slightly (5.24%)
higher loss (73.9570 vs. 70.2720) under MA 30.
The adjusted Bollingers success ratios are all above 50%. However, the success ratios of the
traditional Bollinger are all below 46%. They are as low as 22.16% and 20.41% under MA 20
and 30, respectively.
There is no negative net success under adjusted Bollinger, whereas there is no positive net
success under traditional Bollinger. The profits of the adjusted Bollinger are positive under MA 5
and 10, while which of the traditional Bollinger are all below zero. Adjusted Bollinger with MA
10 generates a best profit of 11.9040.
For both adjusted and traditional Bollinger, MA 5 and 10 are the best choices as they have
higher success ratios, net successes, and profits than the other MAs. Under both MA 5 and 10,
the performance of adjusted Bollinger is better than the traditional one based on all three criteria. And these are the only two strategies pass the necessary conditions for recommendation.
Among these two strategies, MA 10 is the dominant one in terms of the three criteria. Therefore, investors are suggested to use the adjusted Bollinger band with MA 10 while investing
in the JPY.
3.3.2 With transaction cost
For upper bands, the recommended strategy changed from adjusted Bollinger with MA 5 to MA
10 after considering transaction cost. Both adjusted Bollinger with MA 5 and MA 10 are the only
two strategies pass necessary conditions under zero transaction cost. MA 10 generates less than
half signals (trades) than MA 5, it incurs less transaction costs and performs better than MA 5
when transaction costs are considered.6
For lower band as well as both bands together, the recommended strategies are both adjusted
Bollinger with MA 10, they are exactly the same as which of zero transaction cost. Please note
that when the transaction cost per round trade is 4 pips and the rates cross the lower band from
above, no trade is recommended.

3.4

The performance of Bollinger band strategy for AUD

3.4.1 Without transaction cost


3.4.1.1 Upper band. Table 4 shows the results on AUD. One can see the results from UB rows
that the adjusted Bollinger band and the traditional Bollinger bands both generate less than 50%
success ratios, non-positive net success, and negative profits. Traditional strategies do not have a
good quality and the adjusted strategies seem to perform equally bad or even worse. None of the
strategies pass the necessary conditions for recommendation. No strategy can be recommended
to investors.
3.4.1.2 Lower band. One can see results from LB that adjusted Bollinger performs better than
the traditional one with a few exceptions. The exceptions include the adjusted Bollinger generates

798

S.-L Chen et al.

slightly (5.61% and 1.62%) lower success ratios under MA 5 and 20 than the traditional one. And
the net success of adjusted strategy is lower than which of the traditional one (9 vs. 15)
The adjusted Bollinger generates success ratio of above 50% under MA 5, 10, and 20, whereas
traditional Bollinger produces success ratio above 50% only under MA 5 and 20. The adjusted
Bollinger bands success ratios are higher than the traditional one under MA 10 and 30. The
adjusted Bollinger produces non-negative net successes, while traditional one has two negative
net successes. The adjusted Bollingers net successes are higher than the traditional one except
under MA 5. The adjusted Bollinger generates higher profit/lower loss than the traditional one
under all MAs.
The adjusted Bollinger with MA 5, 10, and 20 satisfies the necessary conditions for recommendation. Among them, MA 20 is the dominant one based on three criteria. The traditional Bollinger
with MA 5 and 20 satisfies recommendation conditions. Among those three not dominated strategies, adjusted Bollinger with MA 20 generates best profit. Therefore, adjusted strategy with MA
20 is recommended.
3.4.1.3 Both bands. From the UB and LB results, the adjusted Bollinger generates higher
success ratio, net success, and profit than the traditional one under MA 10, 15, and 20, except a
lower success ratio under MA 30. It is worse than the traditional one under MA 5 by all three
criteria.
None of the strategies generates a success ratio above 50% except the traditional Bollinger
under MA 5. None of the strategies generates a net success above 0 except the traditional Bollinger
under MA 5. And none of the strategies generates a positive profit except the adjusted Bollinger
under MA 20. Traditional and adjusted Bollinger perform best under MA 5 and 20, respectively.
However none of the strategies pass the recommendation conditions.

3.4.2 With transaction cost


Like the results from without transaction cost, there is no recommended strategy derived from
UB as well as both bands together. The recommended strategy for LB is adjusted Bollinger with
MA 10 for three levels of transaction cost considered in this study.

Table 5. Summary of the recommended strategy.


Zero cost
Currency rate
EUR/USD
GBP/USD
USD/JPY
AUD/USD

a Dominant
b No

24 pips cost

Band

Adjusted

Traditional

Adjusted

Traditional

Upper
Lower
Upper and lower
Upper
Lower
Upper and lower
Upper
Lower
Upper and lower
Upper
Lower
Upper and lower

MA 10
MA 5
MA 10a
MA 5
MA 10a
MA 5
MA 10a
MA 10a

MA 20

MA 10
MA 10a
MA 10a
MA 5
MA 10a
MA 10
MA 10a,b
MA 10a

MA 20

strategy among strategies pass necessary conditions.


trade is recommended when the transaction cost per round trade is 4 pips.

Journal of Applied Statistics


4.

799

Summary

The recommended strategies with and without transaction cost are summarized in Table 5. It is
obvious that the adjusted Bollinger bands are overwhelmingly better than the traditional band.
Bollinger band strategy works quite well. It only fails to work in selling EUR or AUD. Adjusted
Bollinger with MA 5 and 10 are suggested for EUR/USD, GBP/USD, and USD/JPY. For these
three currencies, the most recommended strategy is adjusted Bollinger with MA 10 for either
with or without transaction costs. This most suggested 10-day MA is relatively shorter than
the commonly used 20-day MA on other asset trading. This may result from the facts that the
currency market is highly liquid and globalized as well as the news are transmitted in a very fast
pace in the currency market. These two reasons contribute to the high speed that prices reflect
the news.
Adjusted Bollinger with MA 20 is suggested for only buying AUD no matter there is transaction
cost or not. Therefore, the recommended strategy is quite stable. Among the four currency rates
studied, the performance of Bollinger bands on AUD/USD are not as good as which for the other
three majorly traded currencies. This may result from the fact that AUD is one of the commodity
currencies, its value varies by the commodity price. When the commodity price follows a strong
trend and so does the currency rate, the Bollinger band strategy does not work well.

5.

Conclusions, contributions, limitations, and suggestions to future research

Bollinger band has long been used as a contra trend strategy. One of the deficits of the common
Bollinger band is that it fails to consider the fat tail/leptokurtosis often exists in financial time
series. An adjusted Bollinger band generated by rolling GARCH regression method is proposed in
this study. The performance of the adjusted Bollinger band strategy on foreign exchange trading
is evaluated and compared with the traditional one. The daily high, low, and closing exchange
rates of EUR, GBP, JPY, and AUD vs. USD from 2 January 2008 to 15 July 2011 are analyzed
and the Bollinger bands are formed.
Results show that generally speaking, the adjusted Bollinger band performs better than the
traditional one in terms of success ratios, net successes, and profit for EUR, GBP, JPY, and
AUD vs. USD. Adjusted Bollinger with either MA 5 or MA 10 are the most recommended
strategies for EUR, GBP, and JPY. Adjusted Bollinger with MA 20 is the recommended buying strategy for AUD when the currency rate crosses the LB from above. For EUR and AUD,
when the price crosses the UB from below, investors are suggested not to take Bollinger
band strategy.
This paper considers GARCH nature of the financial series and improves on the existing technical analysis. The adjusted technical analysis proposed in this study not only demonstrates
that the consideration of GARCH characteristics of financial series is important, but also provides investors a profitable trading technique. Therefore this paper has contributions from both
academic and practical points of view.
The current study is carried out by using the daily high, low, and closing prices. The future
research can take data of different frequency into consideration. As a final note, it is noticed in
this study that the amplitude of the return varies gradually over time to be described as volatility
shifting. As a result, the GARCH models are chosen properly to deal with such issues. Nevertheless, it is worth special attention, as pointed out by Lastrapes [12], Lamoureux and Lastrapes
[11], and Mikosch and Starica [17], that volatility persistence might be overestimated by ignoring
deterministic regime shifts in the GARCH estimation. Any future research works on Bollinger
band assessment based on this type of fat-tailed data shall keep close investigation on whether
the GARCH time series models include any sudden changes which may be caused by global or
domestic economic/financial events.

800

S.-L Chen et al.

Notes
1. This GARCH (1,1) is the simplest and also the most robust of the family of volatility models [2,3,6]. The generalized
form is a GARCH (p, q), p, q > 1. Empirical evidence shows such higher order models are often useful when a
longer span of data are used, like several decades of daily data or a year of hourly data [6], since it allows long slow
decay of information.
2. The bigger the lag, the longer the estimation time it will take. To simplify the estimation, the largest lag term is set
to be 4.
3. The data covers the subprime and Eurozone debt crises as well as the aftermath of these incidents.
4. In the international foreign exchange margin trade, the calculation of the value of one pip varies dependent on how
currency rate is expressed. For EUR, GBP, or AUD vs. USD, the price is quoted as the USD price of one EUR,
GBP, or AUD. The size of one contract is 100,000 EUR, GBP, or AUD. Therefore, a pip (0.0001 USD per 1 EUR,
GBP, or AUD) is worth US$10 [=(0.0001 USD per 1 EUR, GBP, or AUD) (100,000 EUR, GBP, or AUD)]. For
JPY, the price is quoted as the JPY price of one USD. The size of one contract is 100,000 USD. Hence, a pip (0.01
JPY per 1 USD) is worth [(0.01 JPY per 1 USD/JPY price per 1 USD) 100,000 USD]. When the JPY price per 1
USD is 100, a pip is worth 10 USD.
5. For EUR/USD, the number of signals generated by adjusted Bollinger from top to down is with MA 5, MA 20,
MA 10, and 30.
6. For USD/JPY, the number of signals generated by adjusted Bollinger from top to down is with MA 5, MA 20, MA
10, and 30.

Acknowledgements
We thank the anonymous referees for useful comments. Any remaining errors are ours. Both Chens also thank the National
Science Council (Taiwan) for financial support through grant NSC 99-2410-H-030-020.

References
[1] T.G. Andersen, R.A. Davis, J.P. Krei, and T.V. Mikosch, Handbook of Financial Time Series, Springer, New York,
2009.
[2] A.K. Bera and M.L. Higgins, ARCH models: Properties, estimation and testing, J. Econ. Surv. 7(4) (1993),
pp. 305366.
[3] T. Bollerslev, Generalized autoregressive conditional heterorskedasticity, J. Econometrics 31(3) (1986),
pp. 307327.
[4] Y.-W. Cheung, M.D. Chinn, and A. Garcia Pascual, Empirical exchange rate models of nineties: Are any fit to survive?
J. Int. Money Financ. 24 (2005), pp. 11501175.
[5] C. Engle, N.C. Mark, and K.D. West, Exchange rate models are not as bad as you think, NBER Macroeconomics
Annual 2007 (2008), pp. 381447.
[6] R.F. Engle, GARCH 101: The use of ARCH/GARCH models in applied econometrics, J. Econ. Perspect. 15(4) (2001),
pp. 157168.
[7] M.D.D. Evans and R.K. Lyons, Order flow and exchange rate dynamics, J. Political Econ. 110(1) (2002),
pp. 170180.
[8] M.D.D. Evans and R.K. Lyons, Exchange rate fundamentals and order flow, NBER Working Papers, No. 13151
(2007).
[9] M.D.D. Evans and R.K. Lyons, How is macro news transmitted to exchange rates? J. Financ. Econ. 88(1) (2008),
pp. 2650.
[10] Investopedia, Tales From The Trenches: A Simple Bollinger Band Strategy. Available at http://www.investopedia
.com/articles/trading/07/bollinger.asp (18 February 2009).
[11] C.G. Lamoureux and W. Lastrapes, Persistence in variance, structural change and the GARCH model, J. Bus. Econ.
Stud. 68 (1990), pp. 225234.
[12] W. Lastrapes, Exchange rate volatility and U.S. monetary policy: An ARCH application, J. Money, Credit, Bank. 21
(1989), pp. 6677.
[13] B. LeBaron, Technical trading rule profitability and foreign exchange intervention, NBER Working Papers, No.
5505 (1996), pp. 116.
[14] R.M. Levich and L.R. Thomas, The significance of technical trading-rule profits in the foreign exchange market:
A bootstrap approach, J. Int. Money Financ. 12(6) (1993), pp. 451474.
[15] R.K. Lyons, The Microstructure Approach to Exchange Rates, MIT Press, Cambridge, MA, 2001.

Journal of Applied Statistics

801

[16] R.A. Meese and K. Rogoff, Empirical exchange rate models of the seventies: Do they fit out of sample? J. Int. Econ.
14(12) (1983), pp. 324.
[17] T. Mikosch and C. Starica, Nonstationarities in financial time series, the long-range dependence, and the IGARCH
effects, Rev. Econ. Stat. 86 (2004), pp. 378390.
[18] C.L. Osler and P.H.K. Chang, Head and shoulders: Not just a flaky pattern, Federal Reserve Bank of New York Staff
Paper, No.4 (1995).
[19] P. Posedel, Properties and estimation of GARCH(1,1) model, Metodoloki zvezki 2(2) (2005), pp. 243257.
[20] W.L. Silber, Technical trading: When it works and when it doesnt? J. Derivatives 1(3) (1994), pp. 3944.
[21] R.J. Sweeney, Beating the foreign exchange market, J. Financ. 41(1) (1986), pp. 163182.
[22] M. Taylor and H. Allen, The use of technical analysis in the foreign exchange market, J. Int. Money Financ. 11(3)
(1992), pp. 304314.

Copyright of Journal of Applied Statistics is the property of Routledge and its content may
not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.

You might also like