Algorithmic Finance 5 (2016) 49–68 49

DOI:10.3233/AF-160059
IOS Press

Empirical evaluation of price-based

technical patterns using probabilistic neural
networks
Samit Ahlawat∗
Bank of America, Risk, New York, NY, USA

Abstract. Technical analysis is the art of identifying patterns in historical data with the belief that certain patterns foretell
future price movements. An empirical evaluation of the effectiveness of technical analysis is confounded by the subjectivity
involved in identifying patterns. This work presents a robust framework for pattern identification using probabilistic neural
networks (PNN). The thirty components of the Dow Jones Industrial Average and a set of ten indices are considered.
Fourteen patterns are analyzed. In order to test the possibility that technical patterns are more predictable in certain market
environments, the period under study (1990–2015) is partitioned into bull and bear markets and the statistical significance
of profits earned by identified patterns observed in each environment is analyzed. A range of holding periods from 10 to
50 trading days is considered and a simple model of transaction costs is added. The study reveals that no pattern produces
statistically and economically significant profits for a cross-section of stocks and indices analyzed, though a few patterns are
more successful predictors. Bullish (bearish) patterns are more reliable predictors in bullish (bearish) market environments.
These observations can be explained by the Adaptive Market Hypothesis with certain patterns becoming more accurate
predictors in specific market environments.

Keywords: Neural network, technical analysis, technical trading rules, scatterplot smoothing

1. Introduction rule (buying when shorter period moving average

rises above longer period moving average, and selling
Technical analysis is the art of identifying geomet- when it falls below longer period moving average)
ric patterns in historical prices – often supplemented and trading-range break out (buy when price rises
with volume-based signals – with the belief that above the observed local maximum and sell when
occurrence of patterns are reliable predictors of it falls below the observed local minimum). They
price movement in the immediate future. Academic find statistically significant profits that cannot be
professionals and fundamental analysts typically explained using three null models of efficient market
scoff at technical analysis because of its paucity of hypothesis – random walk, AR(1) and GARCH-M
quantitative justification. Recent works dealing with models. They observe further that volatility of returns
profitability of technical analysis based trading strate- following a buy signal is lower than volatility of
gies have given some credence to the assertion that returns following sell signal, thereby refuting the
technical analysis may not be a complete farce. Brock notion that higher returns for these strategies com-
et al. (1992) test the profitability of moving average pensate higher inherent risks.
∗ Corresponding author: Samit Ahlawat, Bank of America, Osler and Chang (1995) test the profitability of
Risk, 1 Bryant Park, New York, NY 10036, USA. Tel.: +1 914 head and shoulders pattern in foreign-exchange mar-
319 8370; E-mail: [email protected]. kets. They find the strategy yields economically

2158-5571/16/$35.00 © 2016 – IOS Press and the authors. All rights reserved
50 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

significant profit in German Mark and Yen markets technical rules produce statistically and economi-
but not in other exchange rate markets studied in the cally significant profits from early 1970 to late 1980;
work. They use a bootstrap method using random however, by early 1990 such rules no longer pro-
walk null model of efficient market hypothesis to duce statistically significant profits. They explain the
conclude that returns from head and shoulders trad- observation as being consistent with Adaptive Mar-
ing strategy are incompatible with the null hypothesis ket Hypothesis (Lo, 2004). Olson (2004) arrives at a
for German Mark and Yen exchange markets. Rejec- similar conclusion, noting that moving-average trad-
tion of null model could either imply inefficiency of ing rule profits (risk-adjusted) have declined from
those exchange markets or the existence of a different 3.5% during 1970 to around 0% from 1990 to 2000
null model compatible with efficient market hypoth- across 18 exchange rate series. Chavarnakul and Enke
esis (for example, time varying mean). Other works (2008) employ generalized regression neural network
Logue et al. (1978), Sweeney (1986), and Levich (GRNN) to construct two trading strategies based on
and Thomas (1993) also report statistically significant equivolume charting that predict the next day’s price
profits using technical analysis. using volume and price based technical indicators.
Savin et al. (2007) use pattern recognition method They observe that using neural network improves the
presented by Lo et al. (2000) to test if head- profitability of a moving average based trading rule in
and-shoulders pattern has predictive power. After trending markets. However, the time period studied
examining data for S&P 500 and Russell 2000 from is rather limited - one year - and they only consider
1990 to 1999, they conclude that the pattern has S&P 500 index. Furthermore, the difference in prof-
negligible predictive power when used as a stand- itability between neural network based strategy and
alone trading strategy, but has power to predict buy-and-hold strategy is not too large. Other works
risk-adjusted excess returns over market portfolio (Enke and Thawornwong 2005), (Li and Kuo 2008),
returns. They conclude that the period studied in their (Leigh et al. 2005), (Chenoweth et al. 1996) have
work coincided with a bull market, and head-and- studied the application of neural networks in finance.
shoulders being a reversal pattern would not be a Enke and Thawornwong (2005) test the hypothesis
profitable stand-alone trading strategy. that neural networks can provide superior prediction
Researchers have noted the failure of macro- of future returns based on their ability to identify non-
economic models in explaining exchange rate linear relationships. They employ only fundamental
volatility. Neely et al. (1997) and Stephan (2009) measures and do not consider technical ones. Their
have assessed the effectiveness of technical patterns neural network provides higher returns than buy-and-
in explaining exchange rate fluctuations. Neely et al. hold strategy, but they do not consider transaction
(1997) apply genetic algorithm to study the profitabil- costs. Scalar vector regression (SVR) has also been
ity of technical patterns in foreign exchange markets. used in creating automated trading strategies: (Hong
Genetic algorithm is used to design superior trading et al. 2010; Huang 2012; Kazema et al. 2013; Wang
strategies based on filter rules and moving-average and Pardalos 2014).
rules. The rule is tested in an out-of-sample period A challenging aspect for any work attempting to
from 1981 – 1995 and found to generate statisti- perform an empirical assessment of technical analysis
cally and economically significant profits. Neely et al. based trading strategies is the automated identifica-
(1997) further show that the higher profits are not a tion of technical pattern. Osler and Chang (1995)
compensation for bearing higher risks by examining employ a method based on peaks and troughs. They
betas. The significance of technical patterns in foreign define a peak as a local maximum of closing price that
exchange markets has been examined by Stephan is at least χ percent higher than the preceding trough
(2006), Stephan (2008) and Stephan (2009) . Stephan and a trough as a local minimum at least χ percent
(2009) attributes the prevalence of technical analy- lower than the preceding peak. χ is selected based
sis in foreign exchange markets to a virtuous circle on standard deviation; in their work they select a set
whereby traders use it as a tool to form an expec- of values for the cutoff parameter χ. Lo et al. (2000)
tation of current trends, in turn making technical use a novel method based on kernel smoothing. They
analysis a more commonly used tool and increasing use a Gaussian kernel in smoothing, with a constant
its effectiveness as a predictor. Neely et al. (2009) smoothing parameter chosen by visual inspection of
examine the time-varying effectiveness of technical smoothed price curve. Approach used by Lo et al.
patterns as predictors in foreign exchange market. (2000), and Osler and Chang (1995) has the short-
They observe that filter-rules and moving-average coming of using a constant smoothing parameter over
 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 51

the entire price history. Heteroskedasticity in stock Wikipedia (2016) qualifies its description of the
prices is a well documented phenomenon (Bollerslev, pattern using volume: “The left shoulder is formed
1987). Lo et al. (2000) acknowledge this shortcom- at the end of an extensive move during which volume
ing. Methods employed by Osler and Chang (1995) is noticeably high.”. However, Wikipedia (2016) fur-
and Lo et al.(2000) use sequence of successive local ther qualifies the role of volume in left shoulder by
maximum and minimum to identify patterns. In addi- mentioning that the breakout below neckline in that
tion, they use a number of tests to make sure there is region may occur on high or low volume. “The drawn
close fidelity to the technical patterns they recognize, neckline of the pattern represents a support level,
for example, Lo et al. (2000) require the tops in a and assumption cannot be taken that the Head and
double-top pattern to be within 1.5% of their mean. Shoulder formation is completed unless it is broken
It is conceivable for a double-top pattern to occur and such breakthrough may happen to be on more
with successive tops being slightly more than 1.5% volume or may not be”. Because volume does not
of their average. Further, it is distinctly possible that seem to play the defining role in pattern definition
a different smoothing in an area may reveal part of a and there is some disagreement on the exact defini-
pattern. An approach that insists on observing local tion of the volume-based confirming signal, this work
extrema in a specific order while using a constant does not consider volume for the purpose of pattern
smoothing parameter is likely to miss such pattern identification.
occurrences. The remainder of this paper is organized as follows:
This work applies neural networks for recogniz- Section 3 describes the algorithm used in identifying
ing technical patterns in stock prices and evaluates the patterns, Section 4 describes the probabilistic neu-
the performance of patterns as predictors of future ral network used, Section 5 discusses the application
price movements. Neural networks are uniquely of the probabilistic neural network in identifying the
suited to the task of character recognition, and patterns for ten indices and thirty Dow Jones compo-
pattern recognition has distinct similarities to char- nents. Section 6 concludes the work.
acter recognition (Beymer and Poggio 1996). A
class of neural networks called probabilistic neu-
ral networks or PNN is employed. PNN were 2. Algorithm for identification of technical
introduced by Specht (1990). The process of con- patterns
structing a PNN is simpler than that required for a
back-propagation neural network. PNN is used to To recognize a pattern using neural networks, a
identify the following patterns: ascending-triangle, representation of the pattern is required that is robust
descending-triangle, head-and-shoulders, cup-and- to local noise. As a first step, prototypes of price
handle, double-top, double-bottom, triple-top, triple- patterns are first created manually. The prototypes
bottom, broadening-top, down-price-channel, rising- are very geometric, with prices rising and falling
wedge, falling-wedge, up-symmetric-triangle, down- along straight lines. Figures 1–3 show the manu-
symmetric-triangle and down-price-channel for ten ally generated prototypes on a plot. These prototypes
indices and for the thirty components of Dow Jones are referred as prototype patterns in this work. All
Industrial Average. To evaluate the empirical perfor- prototype patterns are twenty days in length. Next,
mance of a trading strategy based on each of the Gaussian noise is added to each day’s price in pro-
patterns, 15 years of history for indices (from 2000 totype pattern. The standard deviation of noise is
to 2015) and 25 years of history for Dow Jones com- selected to be smaller than the maximum daily price
ponents (from 1990 to 2015) is considered. change in the prototype plots. In the present work,
Technical analysts also resort to the use of volume- standard deviation of added noise is taken to be 0.3.
based indictors as confirming signals during pattern 200 realizations of random variable are obtained at
formation phase. However, there is little agreement each point, thereby yielding 200 perturbed plots cor-
between technical analysts on the exact definition responding to each prototype pattern. An example
of confirming signals. The confirming signals rarely of perturbed plot for head-and-shoulders pattern is
constitute the defining aspect of the pattern. To shown in Fig. 4. Next, each point in the unperturbed
illustrate this point, consider the definition of head- price plot is moved to the left by one day. First and
and-shoulders pattern from two sources Investopedia last points corresponding to the first and last day
(2016) and Wikipedia (2016). While Investopedia are kept at their original positions. The third days
(2016) mentions nothing about the role of volume, price displaces the second day’s price and so on.
52 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

Fig. 1. Manually generated pattern shapes.

S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 53

Fig. 2. Manually generated pattern shapes.

54 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

Fig. 3. Manually generated pattern shapes.

200 realizations of Gaussian noise with mean = 0 In order to classify a pattern into one of several
and variance = 0.09 are added to each day’s price classes under consideration, or to determine that it
yielding 200 new perturbed plots. An example of left- does not belong to any of the classes, a representation
shift perturbed plot for head-and-shoulders pattern of pattern is required. This representation is akin to
is shown in Fig. 4. In a similar manner to the left- a fingerprint: patterns belonging to one class should
shift, right shift is performed on unperturbed base produce similar representations. This task is similar
price pattern, keeping the first and last day’s prices to the task confronted in character recognition where
at their original location. After right-shifting by one a handwritten character must be matched to one of
day, 200 realizations of Gaussian noise are added several known characters. However, unlike in char-
to each day’s price yielding another 200 perturbed acter recognition where a character must belong to
plots. Gaussian noise has mean = 0 and variance = one of several classes (alphabets), one also needs to
0.09. An example of right-shift perturbed plot for discriminate the case where a pattern does not match
head-and-shoulders pattern is shown in Fig. 4. This any of the classes. To that end, the algorithm for pat-
procedure produces 601 examples of price patterns tern recognition presented in this work can also be
corresponding to each technical pattern under con- used for character recognition.
sideration. These 601 plots (six hundred perturbed The prototype patterns and their perturbations are
plots and one prototype plot) corresponding to each constructed to have the same length, i.e. they have
pattern comprise the training set for the probabilis- same number of days. This is true of all patterns
tic neural network. PNN are simpler to construct as considered. In this work, the prototype patterns are
compared with multi-layer back-propagation neural chosen to be 20 days in length. This requirement does
networks. For example, back-propagation neural net- not impose any restriction on the length of patterns
work implementation in R package neuralnet (Fritsch that can be classified using this algorithm. Further,
et al. 2012) fails to converge for the data set in because window sizes considered are greater than or
this work (14 × 601 plots). Increasing the num- equal to 20 days, resizing the series down to a length
ber of hidden neurons or the maximum iterations of 20 days does not introduce significant interpola-
does not help to overcome the problem of non- tion inaccuracy. The prices are normalized. Let pmin
convergence during training. Also, increasing the denote the minimum price, p the daily price and pmax
number of hidden layers or hidden neurons increases the maximum price observed over the twenty day
the time taken by back-propagation neural network length of a pattern. Normalized prices are calculated
during learning. This demonstrates the attractive using equation (1).
feature of network simplicity for a PNN. Details
regarding construction of PNN are presented in p − pmin
Section 4. pn = (1)
pmax − pmin
 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 55

Fig. 4. Perturbed head-and-shoulders pattern with left and right shift.

This set of twenty normalized prices is the finger- considered in this work. Details of constructing the
print of the pattern. The perturbations of a pattern will network are presented in the next section. In order to
have fingerprints that are closer to the fingerprints of identify a pattern, a range of window lengths vary-
prototype (unperturbed) pattern. Distance is defined ing from twenty days to sixty days are considered.
as Euclidean distance in the twenty-dimensional Let l denote the window length, L denote the price
space. More formally, distance between two patterns series length and i denote an index in price series. For
is given by equation (2). xi and yi are the normalized each window length, the algorithm checks price pat-
prices. tern between i and i + l days, where index i ranges
from the beginning of price series to L − l − 1 (inclu-
 sive range). The price list observed between days
 20
 [i, i + l] is scaled to a new price list with 20 elements;
distance =  (xi − yi )2 (2) the scaled price list now has the same length as the
i=1
prototype patterns. The price series length scaling is
A probabilistic neural network (PNN) is con- performed using equation (3). Rescaling is analogous
structed to classify patterns belonging to the types to resizing the price series length to 20 days. Price
56 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

for day j in actual price series becomes the price on 1

ym,j = √ D
day dj in rescaled price series where dj is given by 2π||
equation (3). Price for days between dj and dj+1 are 
interpolated. (x − xm )−1 (x − xm )T
×exp −
2
j−i D = 20
dj =  × 20
l (3)
m = 1, 2, . . . 601
j ∈ [i, i + l]
xm = {xm,1 , xm,2 , . . . xm,20 }
Fingerprint of each rescaled price series is calcu-
lated as a tuple of twenty normalized prices using x = {x1 , x2 , . . . x20 } (4)
Equation (1). This fingerprint is presented as input to
In Equation (4), xi refers to the training data point
PNN for classification. In order to identify the case
corresponding to i pattern unit. It is the normalized
where a pattern does not match any of the types con-
price. σ is the variance-covariance matrix defined on
sidered in this work, the PNN uses a threshold. Details
training data belonging to a pattern type. It is calcu-
are presented in the next section.
lated as shown in Equation (5). There are fourteen
variance covariance matrices used in the PNN net-
work, one corresponding to each pattern type.
3. Details on probabilistic neural network

Probabilistic neural networks were first introduced 

601
  
by Specht (1990) as a four-layer neural network capa- i,j = xi,k − xi xj,k − xj
ble of representing non-linear decision boundaries k=1
for a classification problem and offering signifi-

601
cant speedup as compared to the training time of xi,k
k=1
a back-propagation multi-layer feed-forward neural xi =
network. Prababilistic neural networks have four lay- 601
(5)
ers: input layer, pattern layer, summation layer and

601
xj,k
output layer. The first layer is the input layer. The k=1
number of input units is equal to the dimensionality xj =
601
of the problem. In this work, a pattern is represented
by a set of twenty normalized prices, hence the first i ∈ [1, 20]
layer of PNN is comprised of twenty input units. j ∈ [1, 20]
The input units transmit their input as output, with-
out applying any other transformation. The second The third layer of a PNN is the summation layer.
layer of the PNN, known as pattern layer, is defined Summation layer sums up the output from second
by the training set data. In this work, the training data layer’s units belonging to a group. There are fourteen
consists of 14 × 601 inputs – each pattern type has summation layer units in this work, each produc-
600 perturbations and one base price series, giving ing an output corresponding to the likelihood of the
601 training data points for each pattern type, and data matching a pattern type. Input for a summation
there are fourteen patterns types examined. The sec- layer unit is the set of outputs generated by second
ond layer therefore consists of 14 × 601 units. Units layer units belonging to a particular group. Summa-
in second layer of a PNN network are grouped into tion layer unit’s output is shown in equation (6). m
classes the PNN network is meant to classify. In this denotes the index of pattern type, there are 14 types
work, second layer units are grouped into fourteen of pattern considered in this work.
groups. Each group consists of 601 units. Each unit
applies a Gaussian activation function to its input. If
the input is denoted as x, output generated by m unit 
601
zm = ym,j
in second layer pattern j is given by equation (4). xm (6)
j=1
denotes the normalized price vector corresponding to
mth training input. xm and x are vectors of size 20. m ∈ 1, 2, . . . 14
 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 57

The fourth layer of a PNN is the output layer, it Table 1

picks the group having the maximum value and clas- Indices analyzed
sifies the data as belonging to that group. The fourth Ticker Index Name
layer has one unit, its output is given in Equation (7). SPY S&P 500
IWM Russell 2000
IWB Russell 1000
IWR Russell Midcap
Final Output = m ∃ zm > zi ∀m =
/ i (7) IWC Russell Microcap
XLG Russell Top 50
A PNN always classifies a data into one of the IWF Russell 1000 Growth
classes. In order to identify the case where a data IWD Russell 1000 Value
set does not match any pattern, this work requires the IWO Russell 2000 Growth
maximum output to be greater than 100 times the out- IWN Russell 2000 Value
put from other groups. Let zm denote the maximum
output from third layer and zi denote an output from
another unit in third layer. m denotes the third layer patterns. Bulkowski (2005, p. 805) observes average
unit having maximum output and i is another third- length of falling-wedge pattern to be less than two
layer unit. For the data to be classified as matching months, average length of flag pattern to be less than
pattern m, this work requires that Equation (8) hold two weeks (Bulkowski 2005, p. 903), average length
for all i =
/ m. of broadening tops and bottoms to be two months
(Bulkowski 2005, p. 81), average length between left
zm and right shoulder tops of head-and-shoulders pattern
≥ 100 ∀m =
/ i (8) to be two months (Bulkowski 2005, p. 415) and aver-
zi
age length of an island pattern to be just over a month
The threshold value of 100 is an empirical param- (Bulkowski 2005, p. 491). According to Bulkowski
eter, higher values of threshold will produce very (2005, p. 143), pattern length can vary depending on
close matches to the pattern while rejecting potential bull or bear market. A range of holding periods (10,
matches that do not comport with the threshold. Low 20, 30, 40 and 50 trading days) is considered in order
values of threshold parameter will produce greater to test the possibility that some patterns may need
number of matches while producing an occasional longer holding periods for price to move in accor-
false positive by identifying a data to match a pat- dance with the pattern’s prediction.
tern when it does not (i.e. a technical analyst would On each day, the PNN based pattern identifica-
disagree with the classification). tion algorithm is applied to price series of 30 Dow
A diagrammatic representation of the probabilistic components and 10 indices to see if a pattern can be
neural network is shown in Fig. 5. identified over the window length. Dividend and split
adjusted closing prices for the 30 Dow components
and 10 indices are used. After identification, the pat-
4. Application of pattern recognition tern is validated using an independent test (described
in empirical analysis below). Once a pattern is identified and validated, its
return is recorded over the holding period. Return
This work attempts to identify technical patterns is calculated as pi+Npi−pi , where N is the holding
enumerated earlier in prices of thirty Dow Jones com- period length. In order to examine the possibility
ponents and in ten indices: S&P 500 index and nine that technical patterns may be more effective predic-
Russell indices (Table 1). For Dow Jones components tors in certain market environments, the period from
and S&P 500 index, price history from 1990 to 2015 is 1990 to 2015 is partitioned into bull or bear markets
analyzed. For Russell indices, price history from 2000 depending upon the market performance (S&P 500
to 2015 is studied because prices for these indices are index) during the period. The periods were selected
available from 2000 onwards. Patterns with length using publicized dates for the onset of bull and bear
ranging from 20 trading days to 40 trading days are markets widely reported in media. This selection
considered (40 trading days is around two months). entails some in-sample bias because the partitions
Manifestations of patterns with longer duration are are selected ex-post, though the bias is alleviated to a
not identified. This restriction reflects a compromise certain extent by the relatively long duration of peri-
between reducing computation time and considering ods selected compared to the length of patterns and
a window length that covers common occurrences of holding periods considered. A study conducted for
58 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

Table 2 In order to validate identified patterns, a price-

Partition by market trend based test is added. The validation test is based on
Begin Date End Date Classification observing a set of high and low prices in a cer-
1991-01-01 1999-01-01 Bull Market tain order. A technical pattern is characterized as an
2002-01-01 2003-04-01 Bear Market ordered set of high and low prices attained during the
2003-03-01 2007-09-01 Bull Market
2007-10-01 2009-02-01 Bear Market pattern’s observation period. Lo et al. (2000) have
2009-03-01 2015-01-01 Bull Market used kernel smoothing techniques to identify local
price maxima and minima. Locally weighted scatter-
the entire period yields similar qualitative results. plot smoothing algorithm (LOESS) with a quadratic
Since the patterns are either bullish or bearish in polynomial employed for local fitting is used for price
their predictions, it is appropriate to partition the smoothing. Smoothed daily prices are compared to
period between bullish or bearish markets and test their neighboring prices to identify local extrema
the predictive power of all patterns in the two market (maximum or minimum) by comparing the closing
environments. The classification is shown in Table 2. price for a day with the closing price of preceding
Returns observed for different patterns are tabulated and following trading days. If the closing price is
in tables presented. Examples of identified patterns higher than its neighbors, the price is a local maxi-
are presented in Figures 6, 7 and 8. mum. Likewise, if it is lower than its neighbors, it is a
Bullish and bearish environments are selected to local minimum. The identified pattern must have an
coincide with changes in S&P 500 index over a period occurrence of high-low price sequence characterizing
of a year or more, marked by well publicized market the pattern. If an identified pattern fails the validation
events (Table 2). Shorter periods – less than a year test, it is rejected. The high-low sequence for the pat-
– were not used in order to reduce in-sample bias terns is shown in Table 4. Parameters employed in
because the partitions are being selected ex-post. The locally weighted scatterplot smoothing algorithm are
period from 1990 – 1999 is marked by a steady rise in tabulated in 3.
markets, without any major market black-swan event
Table 3
like Black Monday (crash of 1987) or technology Parameters used in LOESS smoothing
crash. 2002 – 2003 partition is widely recognized as
Parameter Value
the period of technology bubble burst. 2003 – 2007
Smoothing 2
partition is recognized as the bull market spawned Polynomial Degree
by brisk growth of mortgage lending; 2007 – 2009 Span 0.4
partition is the ensuing period now referred as the Kernel Gaussian (Least
Great Recession and 2009 – 2015 is the period of Squares Fitting)
recovery from the Great Recession.

Fig. 5. Representation of PNN.

 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 59

AXP Ascending Triangle BAC Descending Triangle

40

20
39

15
Price

Price
38

10
37

5
Jul 15 Aug 01 Aug 15 Sep 01 Sep 15 Oct 01 Nov 01 Nov 15 Dec 01 Dec 15 Jan 01 Jan 15

Date (year: 2004) Date (year: 2008, 2009)

BA Broadening Top CAT Double Bottom

74

33
72

32
70
Price

Price

31
68

30
66

29

Mar 15 Apr 01 Apr 15 May 01 May 15 Jun 01 May 01 May 15 Jun 01 Jun 15 Jul 01 Jul 15

Date (year: 2012) Date (year: 2004)

CSCO Double Top CVX Down Price Channel

23

46
22

44
21
Price

Price

42
20
19

40
18

Jul 15 Aug 01 Aug 15 Sep 01 Sep 15 Oct 01 Feb 15 Mar 01 Mar 15 Apr 01 Apr 15 May 01

Date (year: 2008) Date (year: 2005)

Fig. 6. Specimen patterns identified by the algorithm.

60 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

DD Cup And Handle GS Head And Shoulders

41.0

110
40.5
40.0

105
39.5

100
Price

Price
39.0

95
38.5

90
38.0
37.5

85
Mar 01 Mar 15 Apr 01 Apr 15 May 01 May 15 Oct 01 Oct 15 Nov 01 Nov 15 Dec 01 Dec 15

Date (year: 2007) Date (year: 2011)

HD Symmetric Triangle Down HPQ Symmetric Triangle Up

19.5
33

19.0
32

18.5
Price

Price
31

18.0
30

17.5
29

Apr 01 Apr 15 May 01 May 15 Jun 01 Jun 15 Feb 01 Feb 15 Mar 01 Mar 15 Apr 01 Apr 15

Date (year: 2010) Date (year: 2005)

IBM Triple Bottom INTC Triple Top

16.5
77
76

16.0
75

15.5
Price

Price
74
73

15.0
72

14.5
71

Aug 01 Aug 15 Sep 01 Sep 15 Oct 01 Oct 15 Mar 01 Mar 15 Apr 01 Apr 15 May 01 May 15

Date (year: 2004) Date (year: 2006)

Fig. 7. Specimen patterns identified by the algorithm.

 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 61

JNJ Falling Wedge JPM Rising Wedge

56

56
55
54

54
Price

Price
53

52
52
51

50
Dec 15 Jan 01 Jan 15 Feb 01 Feb 15 Mar 01 Mar 15 Sep 15 Oct 01 Oct 15 Nov 01 Nov 15 Dec 01

Date (year: 2008) Date (year: 2013)

Fig. 8. Specimen patterns identified by the algorithm.

Table 4 that pattern instance. As an example, let i denote the

Validating identified patterns day on which a technical pattern is recognized. Clos-
Pattern High-Low Order ing price is observed for i + 3 day and compared with
Ascending Triangle (H,L,H,L,H) closing price for i day. If the price change is not in
Descending Triangle (L,H,L,H,L) accord with the bullish or bearish price prediction of
Broadening Top (H,L,H,L,H,L)
Double Bottom (L,H,L) the pattern, no trading is done for that pattern occur-
Double Top (H,L,H) rence. Holding period begins from i + 4 day and ends
Down Price Channel (H,L,H,L,H,L) on i + 13 day (both days inclusive) to ensure there is
Cup and Handle (H,L,H,L,H)
no in-sample bias. Three-day period is an empirical
Head and Shoulders (H,L,H,L,H,L)
Symmetric Triangle Down (H,L,H,L,H) parameter; it is meant to test the accuracy of techni-
Symmetric Triangle Up (H,L,H,L,H) cal patterns as a price predictor. Increasing this period
Triple Bottom (L,H,L,H,L) can increase the reliability of patterns at the expense
Triple Top (H,L,H,L,H)
Falling Wedge (H,L,H,L,H,L)
of foregoing trading for more days following pattern
Rising Wedge (H,L,H,L,H,L) identification.
Figures 6, 7 and 8 demonstrate the effectiveness of
the algorithm in identifying patterns.
Recognized and validated patterns are not man- After a technical pattern is identified by the algo-
ually evaluated to ensure correct classification. rithm, price change is recorded for one unit of asset
Technical analysts employ additional tests for clas- for the duration of the holding period. In order
sifying a price series as a technical pattern. To the to ensure that there is no in-sample bias, holding
extent that those validation tests are not employed, period begins after pattern identification, validation
certain price series may have been misclassified. and application of trading rule. A range of holding
Trading strategies based on technical patterns are periods – 10, 20, 30, 40 and 50 trading-days – are
often associated with trading rules. Trading rules considered. Technical analysis categorizes the techni-
are diverse, ranging from simple price increase cal patterns as bullish or bearish – bullish patterns are
or decrease based decisions to complex conditions supposed to presage bullish price movement and bear-
involving volume. In order to study the impact of ish patterns are harbingers of future price declines.
trading rules on the ability of technical patterns to Table 5 shows this classification. In order to assess
forecast future price movements, a simple trading rule the statistical significance of profits for the trad-
is applied: following the identification of the pattern, ing strategy based on respective technical patterns, a
closing price is observed after three days. If the close one-sided t-test is performed. One-sided test is appro-
price has not moved in accordance with the predic- priate to test the bullish or bearish characterization of
tion of the technical pattern, no trading is done for the patterns. 95% significance threshold is used; the
62 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

Table 6
Table 5 Statistically significant patterns observed during bull markets for
Bullish or bearish classification of technical patterns dow components
Pattern Bullish or Bearish Predictor Asset Holding Pattern Obs. t-Stat P Mean
Ascending Triangle Bullish Period Value
Descending Triangle Bearish AXP 30 Double Bottom 13 2.263 0.021 0.977
Broadening Top Bearish AXP 30 Cup and Handle 5 2.045 0.048 2.385
Double Bottom Bullish AXP 40 Cup and Handle 5 2.222 0.038 1.560
Double Top Bearish AXP 40 Falling Wedge 52 1.801 0.039 0.771
Down Price Channel Bearish AXP 50 Double Bottom 13 2.680 0.009 1.703
Cup and Handle Bullish AXP 50 Falling Wedge 52 2.393 0.010 1.280
Head and Shoulders Bearish BAC 30 Falling Wedge 42 1.812 0.039 0.343
Symmetric Triangle Down Bearish BAC 40 Double Bottom 20 2.442 0.012 0.711
Symmetric Triangle Up Bullish BAC 40 Symmetric Triangle Up 35 2.013 0.026 0.361
Triple Bottom Bullish BAC 40 Falling Wedge 42 1.764 0.042 0.420
Triple Top Bearish BAC 50 Symmetric Triangle Up 35 1.957 0.029 0.418
Wedge Falling Bullish BAC 50 Falling Wedge 42 2.209 0.016 0.628
Wedge Rising Bearish BA 10 Falling Wedge 50 1.848 0.035 0.481
BA 20 Falling Wedge 50 2.588 0.006 1.102
BA 30 Falling Wedge 50 2.213 0.016 0.996
BA 40 Falling Wedge 50 2.914 0.003 1.451
statistically significant patterns observed during bull BA 50 Falling Wedge 50 3.652 0.000 2.319
and bear markets are reported in Tables 6 and 7 for CAT 20 Falling Wedge 35 2.453 0.010 1.342
Dow Jones components and in Tables 8 and 9 for CAT 30 Double Bottom 20 1.855 0.039 1.928
CAT 30 Falling Wedge 35 2.326 0.013 1.576
indices respectively. Other patterns were not statisti- CAT 50 Triple Bottom 10 2.424 0.018 2.158
cally significant. Null hypothesis for the test is that CAT 50 Falling Wedge 35 3.336 0.001 3.009
the technical patterns have no predictive power for CSCO 20 Double Bottom 17 1.770 0.047 0.458
future price movements. CSCO 20 Falling Wedge 51 2.388 0.010 0.272
CSCO 40 Double Bottom 17 1.902 0.037 0.628
It can be observed from the tables that only a few CSCO 50 Falling Wedge 51 2.101 0.020 0.455
technical patterns from the set of fourteen patterns CVX 10 Falling Wedge 45 1.958 0.028 0.487
considered in this work produce statistically signifi- CVX 40 Symmetric Triangle Up 39 2.094 0.021 1.120
DD 10 Double Bottom 29 2.316 0.014 0.478
cant profits for a specific ticker. Falling wedge is the
DD 10 Falling Wedge 46 1.682 0.050 0.270
most common pattern occurring in the price charts DD 20 Falling Wedge 46 1.795 0.040 0.437
for the Dow components that produces profits in line DD 30 Falling Wedge 46 2.686 0.005 0.785
with the technical analyst’s predictions. It is followed DD 40 Falling Wedge 46 4.225 0.000 1.339
DD 50 Symmetric Triangle Up 29 1.775 0.043 0.671
by triple bottom and symmetric triangle up patterns, DD 50 Falling Wedge 46 3.775 0.000 1.312
each of which produces statistically significant prof- DIS 10 Double Bottom 15 2.025 0.031 0.511
its for 10 tickers. Cup-and-handle pattern produces DIS 20 Falling Wedge 43 2.528 0.008 0.666
statistically significant profits for 9 tickers. It can be DIS 30 Double Bottom 15 1.768 0.049 0.756
DIS 30 Falling Wedge 43 3.804 0.000 1.069
observed that no pattern produces statistically signif- DIS 40 Double Bottom 15 2.915 0.005 1.425
icant profits for more than half the tickers considered. DIS 40 Falling Wedge 43 4.093 0.000 1.276
More technical patterns produce statistically sig- DIS 50 Double Bottom 15 3.238 0.003 1.421
nificant results during bull markets – this is in part DIS 50 Symmetric Triangle Up 24 2.435 0.011 1.399
DIS 50 Triple Bottom 18 2.104 0.025 1.010
accounted by the longer duration of bull markets. DIS 50 Falling Wedge 43 4.209 0.000 1.627
During bear markets, bearish technical patterns are GE 20 Triple Bottom 10 3.054 0.006 0.523
observed to produce statistically and economically GE 20 Falling Wedge 49 1.929 0.030 0.255
significant profits for Dow Jones components. For GE 30 Triple Bottom 10 2.786 0.010 0.825
GE 40 Double Bottom 19 2.239 0.019 0.601
indices, Table 9 illustrates that both bullish and bear- GE 40 Triple Bottom 10 2.092 0.031 0.878
ish patterns produce statistically significant profits. GE 40 Falling Wedge 49 2.281 0.013 0.283
Cup-and-handle pattern is observed to produce statis- GE 50 Double Bottom 19 1.745 0.049 0.464
GE 50 Triple Bottom 10 2.388 0.019 0.816
tically significant profits more frequently for indices
GE 50 Falling Wedge 49 2.969 0.002 0.490
than for assets, as can be seen by comparing the occur- HD 10 Double Bottom 18 2.277 0.018 0.571
rences of that pattern between Tables 8 and 6. Also, HD 10 Symmetric Triangle Up 39 2.254 0.015 0.320
most patterns require a holding period greater than HD 20 Double Bottom 18 3.497 0.001 1.225
HD 20 Symmetric Triangle Up 39 2.804 0.004 0.651
20 trading days to produce statistically significant
(Continued)
profits.
 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 63

Table 6 Table 6
(Continued) (Continued)
Asset Holding Pattern Obs. t-Stat P Mean Asset Holding Pattern Obs. t-Stat P Mean
Period Value Period Value
HD 30 Double Bottom 18 2.647 0.008 1.336 MRK 50 Symmetric Triangle Up 31 2.705 0.005 0.772
HD 30 Symmetric Triangle Up 39 2.185 0.017 0.743 MRK 50 Falling Wedge 47 1.925 0.030 0.700
HD 30 Falling Wedge 46 2.044 0.023 0.687 MSFT 10 Triple Bottom 14 2.065 0.029 0.299
HD 40 Double Bottom 18 3.284 0.002 1.813 MSFT 20 Triple Bottom 14 2.975 0.005 0.546
HD 40 Symmetric Triangle Up 39 3.045 0.002 1.047 MSFT 30 Symmetric Triangle Up 35 1.717 0.047 0.448
HD 40 Falling Wedge 46 2.284 0.014 0.925 MSFT 30 Triple Bottom 14 2.115 0.026 0.713
HD 50 Double Bottom 18 3.038 0.004 1.858 MSFT 40 Triple Bottom 14 2.165 0.024 0.541
HD 50 Symmetric Triangle Up 39 3.303 0.001 1.376 MSFT 40 Falling Wedge 42 2.049 0.023 0.514
HD 50 Falling Wedge 46 2.969 0.002 1.306 MSFT 50 Symmetric Triangle Up 35 2.166 0.019 0.790
HPQ 30 Triple Bottom 16 2.460 0.013 0.431 MSFT 50 Triple Bottom 14 2.196 0.023 0.668
HPQ 40 Triple Bottom 16 2.267 0.019 0.464 MSFT 50 Falling Wedge 42 2.329 0.012 0.696
IBM 20 Falling Wedge 49 3.170 0.001 1.657 PFE 10 Triple Bottom 11 2.296 0.021 0.361
IBM 30 Falling Wedge 49 3.048 0.002 1.871 PFE 20 Double Bottom 20 2.754 0.006 0.414
IBM 40 Triple Bottom 15 2.229 0.021 2.791 PFE 40 Symmetric Triangle Up 34 1.853 0.036 0.321
IBM 40 Falling Wedge 49 3.066 0.002 2.193 PFE 50 Double Bottom 20 1.795 0.044 0.563
IBM 50 Triple Bottom 15 2.086 0.027 2.698 PFE 50 Symmetric Triangle Up 34 1.778 0.042 0.409
IBM 50 Falling Wedge 49 2.297 0.013 2.570 PG 10 Symmetric Triangle Up 28 3.024 0.003 0.316
INTC 20 Double Bottom 22 2.609 0.008 0.764 PG 20 Symmetric Triangle Up 28 3.406 0.001 0.549
INTC 30 Double Bottom 22 3.049 0.003 0.912 PG 20 Falling Wedge 52 2.924 0.003 0.747
INTC 40 Double Bottom 22 2.595 0.008 1.166 PG 30 Double Bottom 17 1.917 0.036 0.601
INTC 40 Symmetric Triangle Up 28 2.062 0.024 0.537 PG 30 Symmetric Triangle Up 28 3.500 0.001 0.496
INTC 50 Symmetric Triangle Up 28 2.012 0.027 0.635 PG 30 Falling Wedge 52 3.665 0.000 1.068
INTC 50 Falling Wedge 38 1.712 0.048 0.479 PG 40 Double Bottom 17 3.551 0.001 1.100
JNJ 10 Symmetric Triangle Up 33 2.249 0.016 0.320 PG 40 Symmetric Triangle Up 28 2.482 0.010 0.705
JNJ 20 Symmetric Triangle Up 33 1.787 0.042 0.399 PG 40 Falling Wedge 52 3.392 0.001 1.107
JNJ 20 Falling Wedge 59 2.094 0.020 0.463 PG 50 Double Bottom 17 3.382 0.002 1.689
JNJ 30 Symmetric Triangle Up 33 2.303 0.014 0.823 PG 50 Symmetric Triangle Up 28 2.100 0.022 0.680
JNJ 30 Falling Wedge 59 2.385 0.010 0.744 PG 50 Falling Wedge 52 4.687 0.000 1.414
JNJ 40 Double Bottom 20 1.900 0.036 0.856 TRV 40 Double Bottom 15 2.531 0.012 1.112
JNJ 40 Symmetric Triangle Up 33 3.792 0.000 1.311 TRV 40 Falling Wedge 49 1.798 0.039 0.781
JNJ 40 Falling Wedge 59 2.540 0.007 0.764 TRV 50 Double Bottom 15 2.747 0.007 1.656
JNJ 50 Double Bottom 20 2.366 0.014 1.210 TRV 50 Symmetric Triangle Up 28 1.735 0.047 0.952
JNJ 50 Symmetric Triangle Up 33 2.974 0.003 1.538 TRV 50 Falling Wedge 49 2.029 0.024 1.006
JNJ 50 Falling Wedge 59 3.059 0.002 0.922 T 10 Falling Wedge 43 2.102 0.021 0.158
JPM 20 Symmetric Triangle Up 35 2.204 0.017 0.692 T 20 Falling Wedge 43 2.290 0.013 0.259
JPM 30 Symmetric Triangle Up 35 2.899 0.003 1.117 T 30 Falling Wedge 43 2.505 0.008 0.331
JPM 40 Symmetric Triangle Up 35 2.361 0.012 1.079 T 40 Triple Bottom 13 1.794 0.048 0.436
JPM 50 Cup and Handle 5 2.485 0.028 0.900 T 40 Falling Wedge 43 4.291 0.000 0.689
KO 20 Double Bottom 18 1.941 0.034 0.285 T 50 Double Bottom 16 1.845 0.042 0.366
KO 20 Falling Wedge 40 2.254 0.015 0.378 T 50 Triple Bottom 13 1.928 0.038 0.525
KO 30 Falling Wedge 40 2.186 0.017 0.389 T 50 Falling Wedge 43 3.790 0.000 0.735
KO 40 Double Bottom 18 2.109 0.025 0.508 UTX 10 Falling Wedge 40 1.724 0.046 0.339
KO 40 Falling Wedge 40 2.085 0.022 0.428 UTX 20 Double Bottom 20 1.896 0.036 0.873
KO 50 Falling Wedge 40 2.953 0.003 0.608 UTX 30 Double Bottom 20 2.319 0.016 1.274
MCD 10 Double Bottom 17 2.568 0.010 0.631 UTX 30 Symmetric Triangle Up 34 2.864 0.004 1.017
MCD 20 Double Bottom 17 2.161 0.023 0.785 UTX 30 Falling Wedge 40 1.686 0.050 0.920
MCD 30 Double Bottom 17 2.101 0.025 1.038 UTX 40 Double Bottom 20 3.040 0.003 1.523
MCD 30 Symmetric Triangle Up 29 2.867 0.004 0.939 UTX 40 Symmetric Triangle Up 34 2.614 0.007 1.111
MCD 40 Symmetric Triangle Up 29 1.808 0.040 0.680 UTX 40 Falling Wedge 40 1.906 0.032 1.214
MCD 40 Falling Wedge 51 2.530 0.007 0.845 UTX 50 Double Bottom 20 3.624 0.001 2.149
MCD 50 Falling Wedge 51 3.324 0.001 1.286 UTX 50 Symmetric Triangle Up 34 2.236 0.016 1.340
MMM 20 Falling Wedge 35 2.304 0.014 1.105 UTX 50 Falling Wedge 40 2.508 0.008 1.644
MMM 40 Symmetric Triangle Up 29 2.262 0.016 1.424 VZ 40 Falling Wedge 53 2.486 0.008 0.467
MMM 40 Falling Wedge 35 2.679 0.006 1.844 VZ 50 Falling Wedge 53 2.530 0.007 0.610
MMM 50 Falling Wedge 35 3.235 0.001 2.686 WMT 10 Falling Wedge 48 1.812 0.038 0.310
MRK 10 Falling Wedge 47 1.772 0.041 0.198 WMT 40 Double Bottom 26 2.889 0.004 1.310
MRK 20 Symmetric Triangle Up 31 1.800 0.041 0.395 WMT 40 Falling Wedge 48 1.831 0.037 0.877
MRK 20 Falling Wedge 47 2.247 0.015 0.539 WMT 50 Double Bottom 26 2.539 0.009 1.288
MRK 40 Falling Wedge 47 2.022 0.024 0.722 WMT 50 Falling Wedge 48 1.911 0.031 0.839
(Continued) (Continued)
64 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

Table 6 Table 8
(Continued) Statistically significant patterns observed during bull markets for
indices
Asset Holding Pattern Obs. t-Stat P Mean
Period Value Asset Holding Pattern Obs. T-Stat P Mean
XOM 10 Falling Wedge 42 2.814 0.004 0.575 Period Value
XOM 20 Falling Wedge 42 1.693 0.049 0.363 IWB 10 Cup and Handle 7 7.703 0.000 1.131
XOM 40 Double Bottom 18 2.005 0.030 1.489 IWB 10 Symmetric Triangle Up 53 2.045 0.023 0.308
XOM 40 Symmetric Triangle Up 25 2.053 0.025 1.041 IWB 20 Cup and Handle 7 3.877 0.003 1.873
XOM 50 Double Bottom 18 2.656 0.008 2.210 IWB 20 Symmetric Triangle Up 53 4.720 0.000 0.901
XOM 50 Symmetric Triangle Up 25 1.842 0.039 1.164 IWB 20 Falling Wedge 62 2.527 0.007 0.649
IWB 30 Symmetric Triangle Up 53 3.743 0.000 1.181
IWB 30 Falling Wedge 62 2.126 0.019 0.865
IWB 40 Cup and Handle 7 2.523 0.020 2.402
Table 7
IWB 40 Symmetric Triangle Up 53 3.083 0.002 1.264
Statistically Significant Patterns Observed During Bear Markets
IWB 40 Falling Wedge 62 3.165 0.001 1.345
for Dow Components
IWB 50 Symmetric Triangle Up 53 3.687 0.000 1.590
Asset Holding Pattern Obs. T-Stat P Mean IWB 50 Falling Wedge 62 3.146 0.001 1.589
Period Value IWC 10 Double Bottom 10 1.922 0.042 0.329
BAC 50 Rising Wedge 10 -2.425 0.018 -3.166 IWC 20 Cup and Handle 5 2.637 0.023 1.413
BA 30 Rising Wedge 8 -2.262 0.027 -5.180 IWC 30 Double Bottom 10 2.693 0.011 1.583
BA 40 Rising Wedge 8 -2.867 0.010 -5.926 IWD 10 Cup and Handle 9 2.018 0.037 0.494
BA 50 Rising Wedge 8 -3.196 0.006 -6.711 IWD 10 Triple Bottom 7 2.697 0.015 0.674
GE 20 Rising Wedge 6 -2.204 0.035 -0.906 IWD 20 Cup and Handle 9 3.841 0.002 1.405
GE 50 Down Price Channel 7 -2.730 0.015 -1.570 IWD 20 Symmetric Triangle Up 57 3.021 0.002 0.518
HPQ 50 Rising Wedge 8 -2.117 0.034 -1.166 IWD 20 Falling Wedge 69 2.906 0.002 0.937
INTC 30 Rising Wedge 9 -2.247 0.026 -1.541 IWD 30 Cup and Handle 9 6.663 0.000 2.035
INTC 40 Rising Wedge 9 -2.517 0.016 -2.191 IWD 30 Symmetric Triangle Up 57 3.779 0.000 0.898
JPM 30 Rising Wedge 8 -1.928 0.045 -1.879 IWD 30 Falling Wedge 69 2.156 0.017 0.833
JPM 50 Rising Wedge 8 -2.117 0.034 -2.381 IWD 40 Cup and Handle 9 7.449 0.000 2.682
KO 40 Rising Wedge 6 -2.316 0.030 -0.777 IWD 40 Symmetric Triangle Up 57 3.961 0.000 1.042
MRK 20 Rising Wedge 8 -1.900 0.047 -1.974 IWD 40 Falling Wedge 69 2.044 0.022 1.020
MRK 40 Rising Wedge 8 -2.250 0.027 -2.827 IWD 50 Cup and Handle 9 8.812 0.000 3.456
MRK 50 Rising Wedge 8 -2.139 0.032 -3.067 IWD 50 Symmetric Triangle Up 57 3.478 0.000 1.199
MSFT 30 Rising Wedge 7 -2.480 0.021 -1.513 IWD 50 Triple Bottom 7 4.316 0.002 1.602
MSFT 40 Rising Wedge 7 -3.968 0.003 -2.128 IWD 50 Falling Wedge 69 3.046 0.002 1.335
MSFT 50 Rising Wedge 7 -3.805 0.003 -2.117 IWF 10 Falling Wedge 50 1.745 0.044 0.207
PFE 20 Rising Wedge 9 -2.319 0.023 -0.582 IWF 20 Falling Wedge 50 1.704 0.047 0.429
PFE 30 Rising Wedge 9 -2.628 0.014 -0.967 IWF 30 Triple Bottom 22 2.035 0.027 0.664
PFE 40 Rising Wedge 9 -2.698 0.012 -1.008 IWF 40 Triple Bottom 22 3.295 0.002 1.255
PFE 50 Rising Wedge 9 -2.795 0.010 -1.052 IWF 50 Triple Bottom 21 5.870 0.000 1.973
TRV 40 Rising Wedge 7 -2.747 0.014 -2.831 IWM 20 Falling Wedge 77 2.054 0.022 0.725
TRV 50 Rising Wedge 7 -2.284 0.028 -1.913 IWM 30 Double Bottom 27 1.715 0.049 1.271
T 40 Rising Wedge 8 -2.585 0.016 -1.678 IWM 40 Double Bottom 27 2.535 0.009 1.397
T 50 Rising Wedge 8 -4.724 0.001 -2.360 IWM 40 Symmetric Triangle Up 46 2.238 0.015 0.983
UTX 20 Rising Wedge 7 -1.914 0.049 -2.831 IWM 40 Falling Wedge 76 3.992 0.000 1.813
UTX 50 Rising Wedge 7 -1.914 0.049 -3.962 IWM 50 Falling Wedge 76 3.473 0.000 1.729
XOM 10 Rising Wedge 9 -2.175 0.029 -0.990 IWN 10 Cup and Handle 11 2.011 0.035 1.092
XOM 20 Rising Wedge 9 -3.046 0.007 -1.578 IWN 10 Triple Bottom 18 3.379 0.002 0.620
XOM 30 Rising Wedge 9 -2.893 0.009 -1.901 IWN 20 Cup and Handle 11 4.993 0.000 2.220
XOM 40 Rising Wedge 9 -2.223 0.027 -2.021 IWN 20 Symmetric Triangle Up 36 1.693 0.050 0.593
XOM 50 Rising Wedge 9 -2.631 0.014 -3.120 IWN 20 Triple Bottom 18 2.816 0.006 0.874
IWN 30 Cup and Handle 11 2.893 0.007 2.475
IWN 30 Symmetric Triangle Up 36 2.713 0.005 1.157
IWN 30 Triple Bottom 18 2.986 0.004 1.482
Previous works (Allen and Karjalainen 1999; Lo IWN 40 Cup and Handle 11 2.396 0.018 2.907
et al. 2004) have shown transaction costs to be IWN 40 Symmetric Triangle Up 36 2.320 0.013 1.094
an important factor in the profitability of a trading IWN 40 Triple Bottom 18 3.325 0.002 1.822
IWN 50 Cup and Handle 11 3.031 0.006 3.238
strategy, particularly the ones with high turnover.
IWN 50 Symmetric Triangle Up 36 2.241 0.016 1.105
Transaction costs include trading commission, bid- IWN 50 Triple Bottom 18 3.674 0.001 2.093
ask spread and market-impact costs (Lo et al. 2004). IWO 20 Falling Wedge 74 3.214 0.001 1.378
A simple model of transaction costs is introduced in IWO 30 Symmetric Triangle Up 46 2.616 0.006 1.384
IWO 30 Falling Wedge 74 1.950 0.027 1.143
order study how many patterns remain statistically
(Continued)
and economically significant predictors of future
 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 65

Table 8 Table 9
(Continued) Statistically significant patterns observed during bear markets for
indices
Asset Holding Pattern Obs. T-Stat P Mean
Period Value Asset Holding Pattern Obs. T-Stat P Mean
IWO 40 Cup and Handle 5 3.264 0.011 4.053 Period Value
IWO 40 Symmetric Triangle Up 46 3.891 0.000 2.191 IWB 40 Rising Wedge 9 -2.030 0.036 -3.240
IWO 40 Falling Wedge 73 3.101 0.001 1.956 IWC 30 Double Bottom 7 2.389 0.024 1.414
IWO 50 Double Bottom 14 1.766 0.050 2.461 IWD 10 Double Bottom 4 3.470 0.013 0.941
IWO 50 Cup and Handle 5 2.671 0.022 3.586 IWD 40 Rising Wedge 10 -2.376 0.019 -3.500
IWO 50 Symmetric Triangle Up 46 3.333 0.001 2.279 IWD 50 Rising Wedge 10 -2.295 0.022 -3.835
IWO 50 Falling Wedge 73 3.773 0.000 2.235 IWF 40 Down Price Channel 8 -1.884 0.048 -2.679
IWR 10 Cup and Handle 14 2.238 0.021 1.073 IWF 50 Down Price Channel 8 -1.962 0.043 -3.514
IWR 20 Cup and Handle 14 3.571 0.002 2.179 IWM 30 Rising Wedge 12 -1.862 0.044 -2.476
IWR 20 Symmetric Triangle Up 54 3.946 0.000 0.908 IWM 40 Rising Wedge 12 -1.989 0.035 -3.798
IWR 30 Cup and Handle 14 2.930 0.005 1.558 IWM 50 Rising Wedge 12 -2.414 0.016 -4.077
IWR 30 Symmetric Triangle Up 54 4.601 0.000 1.838 IWN 40 Rising Wedge 8 -2.146 0.032 -2.129
IWR 40 Cup and Handle 14 2.330 0.018 2.568 IWN 50 Rising Wedge 8 -2.195 0.030 -2.612
IWR 40 Symmetric Triangle Up 54 7.053 0.000 2.673 IWR 20 Down Price Channel 5 -2.385 0.031 -2.503
IWR 40 Falling Wedge 61 1.725 0.045 1.502 IWR 20 Rising Wedge 5 -2.329 0.034 -3.021
IWR 50 Cup and Handle 14 2.921 0.006 2.691 XLG 10 Symmetric Triangle Up 15 2.043 0.030 0.760
IWR 50 Symmetric Triangle Up 54 5.741 0.000 2.564 XLG 20 Symmetric Triangle Up 15 2.137 0.025 1.220
SPY 10 Symmetric Triangle Up 36 2.143 0.019 0.598 XLG 30 Symmetric Triangle Up 15 3.339 0.002 2.376
SPY 10 Triple Bottom 13 2.482 0.014 0.532 XLG 30 Triple Bottom 11 1.836 0.047 0.945
SPY 20 Cup and Handle 5 2.390 0.031 3.071 XLG 40 Symmetric Triangle Up 15 2.377 0.016 2.086
SPY 20 Symmetric Triangle Up 36 6.377 0.000 1.704 XLG 50 Symmetric Triangle Up 15 2.367 0.016 2.294
SPY 20 Triple Bottom 13 2.129 0.026 0.961
SPY 20 Falling Wedge 39 2.328 0.013 0.982
SPY 30 Symmetric Triangle Up 36 3.564 0.001 1.909
SPY 30 Falling Wedge 39 1.783 0.041 1.117 a bearish pattern produces a profit, transaction costs
SPY 40 Symmetric Triangle Up 36 4.985 0.000 2.933
SPY 40 Triple Bottom 13 1.874 0.042 1.752 will increase the profit. In both of these cases, the
SPY 40 Falling Wedge 39 2.129 0.020 1.670 pattern will contribute to a rejection by the one-sided
SPY 50 Symmetric Triangle Up 36 4.676 0.000 3.078 T-test. When transaction costs are accounted for in
SPY 50 Falling Wedge 39 3.167 0.001 2.488
this manner, and a pattern produces profits that are
XLG 10 Symmetric Triangle Up 31 2.937 0.003 0.760
XLG 20 Cup and Handle 7 2.313 0.027 1.171 opposite to those predicted by the bullish or bearish
XLG 20 Symmetric Triangle Up 31 3.073 0.002 1.220 nature of the pattern, one cannot conclude that taking
XLG 20 Triple Bottom 22 2.115 0.023 0.640 an opposite position in the pattern may open the pos-
XLG 30 Symmetric Triangle Up 31 4.800 0.000 2.376
XLG 30 Triple Bottom 22 2.711 0.006 0.996
sibility of profitable trading using the pattern. This is
XLG 40 Ascending triangle 5 4.587 0.003 2.309 because transaction costs always reduce the profits.
XLG 40 Symmetric Triangle Up 31 3.417 0.001 2.086 As expected, Tables 10, 11, 12 and 13 show that
XLG 50 Ascending triangle 5 3.157 0.013 2.204 fewer patterns cross the threshold of 95% significance
XLG 50 Symmetric Triangle Up 31 3.403 0.001 2.294
XLG 50 Triple Bottom 22 1.868 0.038 1.337
as predictors of future price moves after transaction
XLG 50 Falling Wedge 50 2.010 0.025 1.528 costs are accounted for. Falling wedge pattern is a
more reliable predictor of future price moves for Dow
Jones components during bull markets as compared
price movements. Allen and Karjalainen (1999) with other patterns, as is rising wedge pattern during
model transaction costs as 0.1%, 0.25% or 0.5% of the bear markets. Cup-and-handle pattern is a more reli-
notional. Here, transaction cost is modeled as 0.5% able pattern for indices considered in this work than
of the trade notional. it is for Dow Jones components.
Transaction costs reduce the profits earned (or Adaptive Market Hypothesis proposed by Lo
increase the losses incurred) by a trading strategy. (2004) helps understand the results observed here: it
By virtue of knowing the bullish or bearish nature is possible that certain technical patterns are more
of a technical strategy, a one-sided T-test is used to aggressively traded by market participants leading
assess the statistical and economic significance of a to a greater degree of predictability for those pat-
strategy. For bullish patterns, the transaction costs are terns. No technical pattern is consistently profitable
subtracted from profit (or loss); for bearish patterns, across all securities examined in this work, this could
they are added to loss (or profit). If a bullish pattern point to disappearing opportunities for earning easy
yields a loss, transaction costs will increase the loss; if profits from following one pattern due to increasing
66 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

Table 10 Table 10
Inclusive of transaction costs (0.5%), statistically significant (Continued)
patterns observed during bull markets for dow components
Asset Holding Pattern Obs. t-Stat P Mean
Asset Holding Pattern Obs. t-Stat P Mean Period Value
Period Value PG 40 Falling Wedge 52 2.219 0.015 0.711
AXP 40 Cup and Handle 5 2.084 0.046 1.238 PG 50 Double Bottom 17 3.007 0.004 1.391
AXP 50 Double Bottom 13 2.210 0.023 1.383 PG 50 Falling Wedge 52 3.481 0.001 1.017
AXP 50 Falling Wedge 52 1.759 0.042 0.928 TRV 40 Double Bottom 15 1.898 0.039 0.779
BA 40 Falling Wedge 50 2.051 0.023 1.002 TRV 50 Double Bottom 15 2.375 0.016 1.321
BA 50 Falling Wedge 50 3.007 0.002 1.865 T 40 Falling Wedge 43 3.410 0.001 0.534
CAT 20 Falling Wedge 35 1.938 0.030 1.018 T 50 Falling Wedge 43 3.067 0.002 0.579
CAT 30 Falling Wedge 35 1.882 0.034 1.251 UTX 30 Symmetric Triangle Up 34 1.781 0.042 0.616
CAT 50 Triple Bottom 10 2.351 0.020 1.861 UTX 40 Double Bottom 20 2.384 0.014 1.112
CAT 50 Falling Wedge 35 3.024 0.002 2.677 UTX 40 Symmetric Triangle Up 34 1.728 0.046 0.709
DD 30 Falling Wedge 46 1.770 0.042 0.515 UTX 50 Double Bottom 20 3.083 0.003 1.734
DD 40 Falling Wedge 46 3.421 0.001 1.066 UTX 50 Falling Wedge 40 1.935 0.030 1.249
DD 50 Falling Wedge 46 3.029 0.002 1.039 WMT 40 Double Bottom 26 2.216 0.018 0.987
DIS 30 Falling Wedge 43 2.931 0.003 0.779 WMT 50 Double Bottom 26 1.952 0.031 0.965
DIS 40 Double Bottom 15 2.464 0.013 1.154 XOM 50 Double Bottom 18 2.202 0.020 1.766
DIS 40 Falling Wedge 43 3.293 0.001 0.985
DIS 50 Double Bottom 15 2.743 0.008 1.151
DIS 50 Symmetric Triangle Up 24 1.933 0.033 1.082 Table 11
DIS 50 Falling Wedge 43 3.602 0.000 1.334 Inclusive of transaction costs (0.5%), statistically significant
GE 20 Triple Bottom 10 2.000 0.037 0.350 patterns observed during bear markets for dow components
GE 30 Triple Bottom 10 2.196 0.026 0.651
GE 50 Triple Bottom 10 1.881 0.045 0.642 Asset Holding Pattern Obs. T-Stat P Mean
GE 50 Falling Wedge 49 2.127 0.019 0.351 Period Value
HD 20 Double Bottom 18 2.854 0.005 0.947 BAC 50 Rising Wedge 10 -2.226 0.025 -2.918
HD 20 Symmetric Triangle Up 39 1.809 0.039 0.408 BA 30 Rising Wedge 8 -2.122 0.033 -4.789
HD 30 Double Bottom 18 2.164 0.022 1.058 BA 40 Rising Wedge 8 -2.733 0.013 -5.539
HD 40 Double Bottom 18 2.886 0.005 1.532 BA 50 Rising Wedge 8 -3.043 0.008 -6.328
HD 40 Symmetric Triangle Up 39 2.407 0.010 0.801 GE 50 Down Price Channnel 7 -2.434 0.023 -1.374
HD 50 Double Bottom 18 2.648 0.008 1.577 HPQ 50 Rising Wedge 8 -1.904 0.047 -1.030
HD 50 Symmetric Triangle Up 39 2.794 0.004 1.129 INTC 30 Rising Wedge 9 -2.049 0.035 -1.399
HD 50 Falling Wedge 46 2.359 0.011 1.007 INTC 40 Rising Wedge 9 -2.373 0.021 -2.052
HPQ 30 Triple Bottom 16 1.881 0.039 0.324 JPM 50 Rising Wedge 8 -1.861 0.050 -2.091
HPQ 40 Triple Bottom 16 1.771 0.048 0.356 MRK 40 Rising Wedge 8 -2.040 0.038 -2.546
IBM 30 Falling Wedge 49 1.687 0.049 1.026 MRK 50 Rising Wedge 8 -1.957 0.043 -2.788
IBM 40 Falling Wedge 49 1.895 0.032 1.346 MSFT 30 Rising Wedge 7 -2.161 0.034 -1.310
INTC 20 Double Bottom 22 2.209 0.019 0.627 MSFT 40 Rising Wedge 7 -3.653 0.004 -1.927
INTC 30 Double Bottom 22 2.650 0.007 0.774 MSFT 50 Rising Wedge 7 -3.498 0.005 -1.917
INTC 40 Double Bottom 22 2.326 0.015 1.026 PFE 30 Rising Wedge 9 -2.157 0.030 -0.794
JNJ 40 Symmetric Triangle Up 33 2.820 0.004 0.914 PFE 40 Rising Wedge 9 -2.262 0.025 -0.835
JNJ 50 Symmetric Triangle Up 33 2.299 0.014 1.140 PFE 50 Rising Wedge 9 -2.362 0.021 -0.880
JNJ 50 Falling Wedge 59 1.832 0.036 0.552 TRV 40 Rising Wedge 7 -2.406 0.024 -2.493
JPM 30 Symmetric Triangle Up 35 2.284 0.014 0.875 T 40 Rising Wedge 8 -2.333 0.024 -1.503
JPM 40 Symmetric Triangle Up 35 1.839 0.037 0.838 T 50 Rising Wedge 8 -4.418 0.001 -2.188
JPM 50 Cup and Handle 5 2.232 0.038 0.776 XOM 20 Rising Wedge 9 -2.294 0.024 -1.130
KO 50 Falling Wedge 40 2.034 0.024 0.422 XOM 30 Rising Wedge 9 -2.227 0.026 -1.454
MCD 30 Symmetric Triangle Up 29 1.858 0.037 0.596 XOM 50 Rising Wedge 9 -2.331 0.022 -2.679
MCD 50 Falling Wedge 51 2.440 0.009 0.930
MMM 40 Falling Wedge 35 2.037 0.025 1.336
MMM 50 Falling Wedge 35 2.736 0.005 2.173 competition in the niche. Neely et al. (2009) report
MRK 50 Symmetric Triangle Up 31 1.737 0.046 0.495
MSFT 20 Triple Bottom 14 2.624 0.010 0.424 similar conclusions in foreign-exchange markets
MSFT 30 Triple Bottom 14 1.836 0.044 0.591 where they observe that filter rule based moving-
MSFT 40 Triple Bottom 14 1.782 0.048 0.419 average strategies have largely lost their profitability
MSFT 50 Symmetric Triangle Up 35 1.693 0.050 0.614
by 1990. Todea et al. (2009) report that profitability
MSFT 50 Triple Bottom 14 1.884 0.040 0.546
MSFT 50 Falling Wedge 42 1.778 0.041 0.524 of a moving-average based strategy in six Asian mar-
PFE 20 Double Bottom 20 1.940 0.033 0.285 kets from 1997 to 2008 is episodic. Hsu et al. (2010)
PG 30 Falling Wedge 52 2.385 0.010 0.673 report that technical patterns (moving averages, filter
PG 40 Double Bottom 17 2.863 0.005 0.805
rules) loose much of their predictive power as market
(Continued)
matures.
 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks 67

Table 12 change predictors. It considers a set of fourteen pat-

Inclusive of transaction costs (0.5%), statistically significant terns for the Dow Jones Index components and a set
patterns observed during bull markets for indices
of ten indices, looking at closing prices for the last
Asset Holding Pattern Obs. T-Stat P Mean
25 years for Dow Jones components (and S&P 500
Period Value
index) and 15 years for remaining indices. Techni-
IWB 10 Cup and Handle 7 2.902 0.011 0.441
IWB 20 Cup and Handle 7 2.419 0.023 1.180 cal pattern occurrences ranging from 20 trading days
IWB 30 Symmetric Triangle Up 53 1.698 0.048 0.542 (one month) to 40 trading days (2 months) are ana-
IWB 50 Symmetric Triangle Up 53 2.226 0.015 0.949 lyzed. Period under consideration is partitioned into
IWB 50 Falling Wedge 62 1.742 0.043 0.879
bear or bull markets depending upon the market (S&P
IWC 30 Double Bottom 10 1.877 0.045 1.090
IWD 20 Cup and Handle 9 2.107 0.032 0.756 500) trend during the period in order to examine
IWD 30 Cup and Handle 9 4.554 0.001 1.383 the possibility that certain patterns are more reli-
IWD 40 Cup and Handle 9 6.239 0.000 2.026 able predictors in a particular market environment. A
IWD 50 Cup and Handle 9 7.291 0.000 2.797
IWD 50 Triple Bottom 7 3.015 0.010 1.066
locally-weighted scatterplot (LOESS) based rule is
IWD 50 Falling Wedge 69 1.677 0.049 0.732 employed to validate a pattern once it has been iden-
IWF 40 Triple Bottom 22 1.736 0.048 0.693 tified by the neural network. Holding periods ranging
IWF 50 Triple Bottom 21 4.333 0.000 1.429 from 10 to 50 trading days after a pattern is observed
IWM 40 Falling Wedge 76 2.484 0.008 1.116
IWM 50 Falling Wedge 76 2.092 0.020 1.033
and validated are used. Data analyzed in the study
IWN 20 Cup and Handle 11 3.554 0.002 1.650 does not support the proposition of sustained prof-
IWN 30 Cup and Handle 11 2.152 0.027 1.904 itability following technical trading rules for a cross
IWN 30 Triple Bottom 18 1.858 0.040 0.923 section of stocks comprising the Dow Jones Index and
IWN 40 Cup and Handle 11 1.893 0.043 2.334
IWN 40 Triple Bottom 18 2.303 0.017 1.261 the set of indices. There are a few instances where the
IWN 50 Cup and Handle 11 2.449 0.016 2.664 rules generate statistically and economically signif-
IWN 50 Triple Bottom 18 2.686 0.008 1.531 icant profits that are in accord with the predictions
IWO 40 Cup and Handle 5 2.558 0.025 3.240 of the rule. However these comprise a clear minority,
IWO 40 Symmetric Triangle Up 46 2.467 0.009 1.396
IWO 40 Falling Wedge 73 1.901 0.031 1.200 being outnumbered by the cases where such rules do
IWO 50 Cup and Handle 5 2.129 0.043 2.776 not generate economically or statistically significant
IWO 50 Symmetric Triangle Up 46 2.150 0.018 1.484 profits. Some patterns, like cup-and-handle pattern,
IWO 50 Falling Wedge 73 2.500 0.007 1.477
are more reliable predictors of future price moves
IWR 30 Symmetric Triangle Up 54 2.203 0.016 0.901
IWR 40 Symmetric Triangle Up 54 4.523 0.000 1.732 for indices than they are for Dow Jones compo-
IWR 50 Symmetric Triangle Up 54 3.642 0.000 1.624 nents. Bullish patterns are more reliable predictors
SPY 20 Symmetric Triangle Up 36 2.707 0.005 0.743 in bullish market environments, with falling wedge
SPY 30 Symmetric Triangle Up 36 1.751 0.044 0.947
SPY 40 Symmetric Triangle Up 36 3.393 0.001 1.966
being the most frequently observed pattern. Likewise,
SPY 50 Symmetric Triangle Up 36 3.262 0.001 2.110 bearish patterns (like rising wedge) are more reli-
SPY 50 Falling Wedge 39 1.991 0.027 1.538 able predictors in bearish market environments. This
XLG 30 Symmetric Triangle Up 31 2.901 0.003 1.463 observation suggests that a portion of a technical pat-
XLG 40 Ascending triangle 5 2.608 0.024 1.293
XLG 40 Symmetric Triangle Up 31 1.896 0.034 1.175
tern’s predictability can be attributed to the market
XLG 50 Symmetric Triangle Up 31 2.025 0.026 1.382 environment. Transaction cost of 0.5% reduces the
number of statistically significant patterns observed,
but the foregoing conclusions stand. The results sup-
Table 13 port Adaptive Market Hypothesis (Lo, 2004): some
Inclusive of transaction costs (0.5%), statistically significant
patterns observed during bear markets for indices
technical patterns are more effective predictors of
future price movements in certain market environ-
Asset Holding Pattern Obs. T-Stat P Mean
Period Value
ments and for certain assets.
IWD 40 Rising Wedge 10 -2.081 0.032 -3.057
Role played by volume as a confirming signal for
IWD 50 Rising Wedge 10 -2.036 0.035 -3.394 pattern identification has not been examined in this
IWM 50 Rising Wedge 12 -2.135 0.027 -3.601 work and is a topic for future study. Technical analysts
XLG 30 Symmetric Triangle Up 15 2.018 0.031 1.463 employ additional tests to validate classification of a
technical pattern; including those tests will enhance
5. Conclusion the recognition algorithm. The algorithm outlined is
able to identify the patterns, as confirmed by the
This work presents an empirical assessment of plots of identified patterns; but any claim to accuracy
accuracy of a set of technical patterns as future price must be qualified with the details of pattern definition
68 S. Ahlawat / Empirical evaluation of price-based technical patterns using probabilistic neural networks

employed. Technical analysis literature abounds in approach, Journal of International Money and Finance 12(5),
elaborate identification rules for the patterns – like 451-474.
confirming signals – and rules for when to unwind Li, S.T., Kuo, S.C., 2008. Knowledge discovery in finan-
cial investment for forecasting and trading strategy through
the trade. Adding these elaborate definitions to the wavelet-based SOM networks, Expert Systems with Applica-
algorithm may yield cases where such rules gener- tions 34(2), 935-951.
ate reliably significant profits. This study outlines Lo, A.W., Mamaysky, H., Wang J., 2000. Foundations of tech-
an algorithm that can be effectively used in such an nical analysis: Computational algorithms, statistical inference,
effort. and empirical implementation, Journal of Finance 55(4), 1705-
1765.
Lo, A.W., 2004. The adaptive market hypothesis: Market effi-
ciency from an evolutionary perspective, Journal of Portfolio
References Management 30(5), 15-29.
Lo, A.W., Mamaysky, H., Wang J., 2004. Asset prices and trad-
Allen, F., Karjalainen, R., 1999. Using genetic algorithms to find ing volume under fixed transactions costs, Journal of Political
technical trading rules, Journal of Financial Economics 51, Economy 112(5), 1054-1090.
245-271. Logue, D., Sweeney, R., Willett, T., 1978. The speculative
Beymer, D., Poggio, T., 1996. Image representations for visual behaviour of foreign exchange rates during the current float,
learning, Science 272(5270), 1905-1909. Journal of Business Research 6(2), 159-174.
Bollerslev, T., 1987. A conditionally heteroskedastic time series Neely, C.J., Weller, P.A., Dittmar, R., 1997. Is technical analysis
model for speculative prices and rates of return, Review of in the foreign exchange market Profitable? A genetic program-
Economics and Statistics 69(3), 542-547. ming approach, Journal of Financial and Quantitative Analysis
Brock, W.A., Lakonishok, J., LeBaron, B., 1992. Simple technical 32(4), 405-426.
trading rules and the stochastic properties of stock returns, Neely, C.J., Weller, P.A., Ulrich, J.M., 2009. The adaptive mar-
Journal of Finance 47(5), 1731-1764. kets hypothesis: Evidence from the foreign exchange market,
Bulkowski, T.N., 2005. Encyclopedia of Chart Patterns, 2nd Ed. Journal of Financial and Quantitative Analysis 44(2), 467-488.
New York: John Wiley and Sons. Olson, D., 2004. Have trading rule profits in the currency markets
Chavarnakul, T., Enke, D., 2008. Intelligent technical analysis declined over time? Journal of Banking and Finance 28(1),
based equivolume charting for stock trading using neural net- 85-105.
works, Expert Systems with Applications 34(2), 1004-1017. Osler, C.L., Chang, P.H.K., 1995. Head and shoulders: Not just
Chenoweth, T., Lee, S.S., Obradovic, Z., 1996. Embedding techni- a flaky pattern, Staff Report No. 4, Federal Reserve Bank of
cal analysis into neural network based trading systems, Applied New York, pp. 1-65.
Artificial Intelligence Journal 10(6), 523-541. Savin, G., Weller, P.A., Zvingelis, J., 2007. The predictive power of
Enke, D., Thawornwong, S., 2005. The use of data mining and “head-and-shoulders” price patterns in the U.S. stock market,
neural networks for forecasting stock market returns, Expert Journal of Financial Econometrics 5(2), 243-265.
Systems with Applications 29(4), 927-940. Specht, D.F., 1990. Probabilistic neural networks, Neural Net-
Fritsch, S., Guenther, F., Suling, M., 2012. Neuralnet: Training works 3(1), 109-118.
of Neural Networks. R package version 1.32, http://CRAN.R- Stephan, S., 2006. The interaction between technical currency trad-
project.org/package=neuralnet, 2012. ing and exchange rate fluctuations, Finance Research Letters
Hong, W.C., Dong, Y., Chen, L.Y., Wei, S.Y., 2010. SVR with 3(3), 212-233.
hybrid chaotic genetic algorithms for tourism demand fore- Stephan, S., 2008. Components of the profitability of technical
casting, Applied Soft Computing 11(2), 1881-1890. currency trading, Applied Financial Economics 18(11), 917-
Hsu, P.H., Hsu, Y.C., Kuan, C.M., 2010. Testing the predictive 930.
ability of technical analysis using a new stepwise test without Stephan, S., 2009. Aggregate Trading Behaviour of Techni-
data snooping bias, Journal of Empirical Finance 17(3), 471- cal Models and the Yen/Dollar Exchange Rate 1976-2007,
484. WIFO Working Paper No. 324. Available at SSRN:
Huang, C.F., 2012. A hybrid stock selection model using genetic http://ssrn.com/abstract=1714341.
algorithms and support vector regression, Applied Soft Com- Sweeney, R.J., 1986. Beating the foreign exchange market, Journal
puting 12(2), 807-818. of Finance 41(1), 163-182.
Investopedia, 2016. Head And Shoulders Pattern, viewed Todea, A., Ulici, M., Silaghi, S., 2009. Adaptive markets hypothe-
10 January 2016, http://www.investopedia.com/terms/h/head- sis: Evidence from Asia-Pacific financial markets, The Review
shoulders.asp. of Finance and Banking 1(1), 7-13.
Kazema, A., Sharifia, E., Hussain, F.K., Saberi, M., Hussain, Wang, F.K., Du, T., 2014. Implementing support vector regression
O.K., 2013. Support vector regression with chaos-based firey with differential evolution to forecast motherboard shipments,
algorithm for stock market price forecasting, Applied Soft Expert Systems with Applications 41(8), 3850-3855.
Computing 13(2), 947-958. Wang, X., Pardalos, P.M., 2014. A survey of support vector
Leigh, W., Hightower, R., Modani, N., 2005. Forecasting the New machines with uncertainties, Annals of Data Science 1(3),
York stock exchange composite index with past price and inter- 293-309.
est rate on condition of volume spike, Expert Systems with Wikipedia, 2016. Head and shoulders (chart pattern), viewed
Applications 28(1), 1-8. 10 January 2016, https://en.wikipedia.org/wiki/Head and
Levich, R., Thomas, L., 1993. The significance of technical shoulders (chart pattern).
trading-rule profits in the foreign exchange market: A bootstrap