IFTA Journal 2015 Edition

A Professional Journal Published by The International Federation of Technical Analysts

15

Inside this Issue

6 Optimal f and diversification
by Stanislaus Maier-Paape
10 Feeling the Market’s Pulse With
Google Trends by Shawn Lim, CFTe,
MSTA, and Douglas Stridsberg
89 Know Your System! – Turning
Data Mining From Bias to Benefit
Through System Parameter
Permutation by Dave Walton, MBA

Men are born to succeed, not fail.

—Henry David Thoreau

ISSN 2409-0271
 DO YOU WISH YOU HAD MORE TIME?
Are you spending more time adjusting graphics or
doing menial editing tasks than analysing charts?

Are you frustrated by having no easy way to

rapidly find what most needs your attention?

How much is this wasted time costing you?

We understand.
We’re here to help.
A recurring message we get from our clients is how thrilled they
are that we have taken so much work off their shoulders and given
them back so much time.

Time is what we want most,

but we use worst. William Penn
If you are ready to explore how you can claw back your valuable
time, go to the following address:
mav7.com/time
 IFTA JOURNAL 2014 EDITION

Letter From the Editor

By Aurélia Gerber, MBA, CFA............................................................................................................ 3

EDITORIAL Articles
Aurélia Gerber, MBA, CFA (SAMT)
Editor, and Chair of the Editorial Committee
Optimal f and Diversification
[email protected] By Stanislaus Maier-Paape ..............................................................................................................4
Editor
Feeling the Market’s Pulse With Google Trends
Jacinta Chan, Ph.D. By Shawn Lim, CFTe, MSTA, and Douglas Stridsberg.................................................................8
[email protected]

Elaine Knuth
[email protected]
MFTA Papers
Identification of High Probability Target Zones (HPTZ)
Regina Meani
[email protected] By Andrew J.D. Long, MFTA............................................................................................................. 15
Rolf Wetzer Testing the Effectiveness of Multivariate Wavelet De-noising for Intraday Trading Systems
[email protected] By Adam Cox, MFTA..........................................................................................................................30
Use of Social Media Mentions in Technical Analysis
Send your queries about advertising
By Alex Neale, MFTA. .........................................................................................................................51
information and rates to [email protected] Enhancing Portfolio Returns and Reducing Risk by Utilizing the Relative Strength Index as a
Market Trend Identifier
By David Price BSc, MSc, CFTe, MSTA, MFTA. .............................................................................62
Anatomy of a Living Trend: Swing Charts, High Points and Low Points, Peaks and Troughs and
How Their Underlying Structure May Define Their Forecasting Strength
By Andreas Thalassinos, BSc, MSc, MSTA, CFTe, MFTA..........................................................67
Refining Wilder’s ADX: Adjustment to the Price Actions by Utilizing Closing Prices
By Samuel Utomo, CFTe, MFTA. .................................................................................................... 74
The Alternative Head and Shoulders: A New Perspective on A Pre-Eminent Pattern
By Fergal Walsh, MFTA ...................................................................................................................82
About cover photo:
Wildflowers in meadow during sunset Wagner Award Paper
—Photo by Pawel Gaul
Know Your System! – Turning Data Mining From Bias to Benefit Through System Parameter
Permutation
By Dave Walton, MBA. .....................................................................................................................87

Book Review
Crowd Money: A Practical Guide to Macro Behavioural Technical Analysis—By Eoin Treacy
Reviewed By David Hunt..................................................................................................................97
IFTA Board of Directors..............................................................................................................................98
IFTA Staff......................................................................................................................................................98
Author Profiles.............................................................................................................................................99

IFTA Journal is published yearly by The International Association of Technical Analysts. 9707 Key West Avenue, Suite 100,
Rockville, MD 20850 USA. © 2014 The International Federation of Technical Analysts. All rights reserved. No part of this
publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying for
public or private use, or by any information storage or retrieval system, without prior permission of the publisher.

IFTA.ORG PAGE 1
Hosted by
 IFTA JOURNAL 2014 EDITION

Letter From the Editor

By Aurélia Gerber, MBA, CFA

Dear IFTA Colleagues and Friends:

The theme of this year’s 27th conference in London is “unravelling the

DNA of the market”. Times sure have changed—it used to take several
rooms full of sequencing machines and millions of dollars to decipher a
genome, and now we need only one machine to draft far more complex DNA.
Such developments have also been seen in the investment field.
The principles of technical analysis remain the same, however; price
discounts everything, price movements are not totally random (they move
in trends), and history has a tendency to repeat itself. These form the DNA,
or the genes, of technical analysis.
The IFTA Journal is—through its global distribution to professionals in the field within member
societies from 27 countries—one of the most important forums for publishing leading work in
technical analysis. This year, the Journal has four sections.
In the first section, we have published seven Master of Financial Technical Analysis (MFTA)
research submissions. This body of work offers fresh ways of looking at the behavior of markets
and is testament to the high standing of the MFTA designation. Two articles deal with innovative
ways of refining well-known technical indicators.
Other themes include the anatomy of living trend structure, denoising using multivariate …unravelling the
wavelet algorithm, alternative head and shoulders pattern recognition, integration of multiple
techniques to define high-probability target zones, and the use of social media in investing. DNA of the market
In the second section, one article was submitted by IFTA colleagues from the Society of
Technical Analysts (STA) on the use of Google trends for feeling the market’s pulse, and the
other from Vereinigung Technischer Analysten Deutschlands (VTAD) on optimal f for money
management.
Next, with the permission of the National Association of Active Investment Managers
(NAAIM), we are happy to publish a paper by Dave Walton, winner of the NAAIM Wagner Award
2014. We hope that you find this paper most interesting.
We also had the support of one book proposal reviewer, David Hunt, on crowd behaviour.
This year’s Journal was produced by a returning team for IFTA. I would like to thank, Elaine
Knuth, Jacinta Chan and Regina Meani for their help in editing this Journal. Articles were peer
reviewed by Elaine Knuth and Rolf Wetzer.
We are also able to create this timely and unique journal because of the intellect and generosity
of time and materials from the authors. It was their tremendous spirit and endeavour that
enabled us to achieve the goals of this high-quality journal. We are indebted to all authors for their
contributions and for enabling us to meet our journal submission deadline.
Last but not least, we would also like to thank the production team at IFTA, in particular
Linda Bernetich, Jon Benjamin, and Lynne Agoston, for their administrative, editorial and
publishing work.

IFTA.ORG PAGE 3
 IFTA JOURNAL 2015 EDITION
Many traders and investors know that diversified depots have many benefits compared with single inv
ments. The distribution of risks on many shoulders reduces the risk of a portfolio remarkably while at
same time the return stays unchanged. On the other hand,Prof. the Dr. Stanislaus
return of a Maier-Paape
portfolio can be maxim
[email protected]
subject to a predefined risk level. In the Portfolio Theory of Markowitz (cf. [3]) these facts are form

Optimal f and Diversification derived. Institut für Mathematik

RWTH Aachen, Templergraben 55
As a byproduct, this ansatz yields concrete position sizes for single assets in 52062 Aachen
order to build the opti
By Stanislaus Maier-Paape Germany since the ris
portfolio. This ansatz does, however, not regard the possible drawdown of the portfolio
+49(0)241-8094925
measured solely via the standard deviation. The goal of this paper is to demonstrate, that diversified dep
are also suitable to reduce possible (maximal) drawdowns, while not lowering ones sight on the return.

Optimal f and Kelly betting

Abstract Vince [4] and [5]). In particular when large distributions of possible
For our demonstration we choose the “optimal f ” ansatz, that means position sizing that uses always a fi
In this paper we use the optimal f / Kelly betting approach for trading results are used, this approach quickly gets confusing.
percentage (“fixed fraction trading”) of the actual available investment capital (cf. Vince [4] and [5]).
money management. Optimal f investments yield optimal growth Therefore and for demonstration purposes we want to use optimal f
particular when large distributions of possible trading results are used, this ansatz quickly gets confus
of the equity curve, but at the same time catastrophic drawdowns only in its simplest version also known as “Kelly betting system” (cf.
when used for single investments. Therefore and for demonstration purposes
[1], [2] and we want
for the variant to use optimal
following f p.
below [4], only
30).in its simplest version also kno
as “Kelly betting
In several simulations we show that a simple diversification system” (cf. [1], [2] and for the variant following below
Here a trader can repeatedly place for him favorable[4], p. bets.
30).
of a portfolio helps a lot in order to reduce expected and achieved On each bet he either looses his stake which is a fixed percentage
Here a trader can repeatedly place for him favorable bets. On each bet he either looses his stake which
maximal drawdowns of the equity curve. Therefore the risk f ∈ [0,1] of his capital, or he wins B times his stake. In case we
fixed percentage f ∈ [0, 1] of his capital, or he wins B times his stake. In case we further assume that
measure “maximal drawdown” shows similar behavior with further assume that the winning probability is
winning probability is p ∈ (0, 1) and the loosing probability is q = 1 − p then for the capital Xk after k b
respect to diversification as the commonly used standard p ∈ (0,1) and the loosing probability is q =1 – p then for the capital
we get
deviation in Markowitz portfolio theory. Xk after k bets we get
�
Xk−1 · (1 + Bf ) with probability p
Introduction Xk =
Many traders and investors know that diversified depots Xk−1 · (1 − f ) with probability q .
have many benefits compared with single investments. The
distribution of risks on many shoulders reducesUnderthe risk
theofcondition
a that theUndercapitaltheafter
condition
k − 1 that
bets the capital known
is already after k (equal
– 1 betstoisx),already
the expected value o
Under
portfolio remarkably while at the same time theXreturn the condition
stays that the
knowncapital
(equalafter
to k −
x), 1
the bets is
expected already
value known
of X k
(equal
becomes to x), the expected value of
k becomes � �
unchanged. On the other hand, the return of aXportfolio
k becomes can be � � �
� ��� �
�
1
� �
E Xk� �theXcapital
k−1 = x p · x(1 + q · x(1 − f)
maximized subject to a predefined risk level. In the Portfolio Under the condition E Xthatk � Xk−1 = xafter k= −=1pbets + Bf )known
· x(1is+already
Bf ) + q · x(1 (equal
− f )to x), the(1) expected value of
� �
Theory of Markowitz (cf. [3]) these facts are formally derived. X k becomes � �� � �
= �x · 1 +� Bp − �q f
� �
As a byproduct, this approach yields concrete position sizes � = � x · 1 + Bp − q f
E Xk �� Xk−1 = x = p · x(1 + Bf ) + q · x(1 − f )
for single assets in order to build the optimal portfolio. This Therefore, the bets are only favorable � in �
case �Bp�> q. The
Therefore, the bets are only favorable in case Bp > q. The=expected gain ofqeach of these bets gets maximize
approach does, however, not regard the possible drawdown
Therefore, the ofbets are onlyexpected
favorablegain of each
in case Bp >ofq.these
The bets xgets
expected · 1+ maximized
gainBpof−each f for f = 1. bets gets maximized
of these
for f = 1. This, however, immediately brings about ruin once only one bet gets lost. Clearly this cannot b
the portfolio since the risk is measured solely via the standard
for f = 1. This, however, This, however, immediately brings about ruin once only one
immediately brings about ruin once only one bet gets lost. bet this cannot be
Clearly
meaningful.Therefore,
Hence instead ofare
maximizing theingain, Kelly started to maximize the of
expectation of the natur
deviation. The goal of this paper is to demonstrate, that Hence instead gets
meaningful. the bets lost. Clearly
only
of maximizing thethis
favorable cannot
case
gain, Kellybe
Bp > q. meaningful.
The
started expected
to maximize Hencethe
gain of instead
each of bets of
these
expectation gets maximized
the natural
logarithm
diversified depots are also suitable to reduce possible of the capital instead. Using again the condition that X is already known one obtains
logarithm(maximal)
of for
thef capital
= 1. This, maximizing
however,
instead. Using the gain,
immediately
again the Kelly
brings started
about
condition ruin
thattoonce
maximize
Xk−1 k−1 one the
only bet expectation
is already gets lost. Clearly
known this cannot be
one obtains
drawdowns, while not lowering ones sight on the return. �
meaningful. Hence instead
� of �
of the natural logarithm
maximizing theof the�capital
gain, Kelly instead.
started to Using�again
maximize the the��
expectation of the natural
� � �� ��� �
� � �
� the � �Using again the�condition �� X � � � known � one obtains
E log
logarithm of X�condition
�
capitalXinstead.
= x = p · log �
that X   is already known one obtains
x 1 + Bf �
that +k−1q is
· already
log x 1 − �
f
E log X k � � X k−1= x k – 1 = p · log x 1 + Bf + q · log x 1 − f
Optimal f and Kelly betting �
k
� �
� �
k−1
� log x +� p �log � � � ��
�� �1 − �f���
(1
(1)
� ��� = � 1 + Bf ��� + q log
For our demonstration we choose the “optimal f ” approach, that �
E log Xk � Xk−1 = =x log x= +p · ploglogx 1 ++ Bf
Bf + qq ·loglog 1x −1 −
ff
means position sizing that uses always a fixed percentage (“fixed � (1)
� � � q ��
fraction trading”) of the actual available investment capital
If this = log x +is achieved
(cf. is viewed as a function of f , its maximum
expression p log 1 + Bf
at fopt p −1 q− f > 0
+ q=log
If this expression is viewed as a function of f , its maximum is achieved at fopt = p − B> 0
(Kelly formula). B
(Kelly formula). q
If this expression is viewed as a function of f , its maximum is achieved at fopt = p − > 0
Figure 1: y = log (X k) for fopt (left) and is negative drawdown (right) B
(Kelly formula).
Simulation of single investments
Simulation
Equity log scale: steps = 10000, B = 2, p = 0.4, of single investments
f = 0.1
blood curve(neg drawdown): steps = 10000, B = 2, p = 0.4, f = 0.1
120 In the following we want to do some simulations. Assume for example B = 2 and p = 0.4. The optimal
In the following we want of
Simulation do0 some
to single simulations. Assume for example B = 2 and p = 0.4. The optimal f
investments
according to Kelly then is −0.1
100
according to In
Kelly then is we want to do some qsimulations. Assume
the following 0.6 for example B = 2 and p = 0.4. The optimal f
f = p −q = 0.4 −0.6 = 0.1 = 10% .
according to Kelly−0.2 foptopt
then is = p − B= 0.4 − 2 = 0.1 = 10% .
B q 2 0.6
80 −0.3 fopt = p − = 0.4 − = 0.1 = 10% .
That means, in order to obtain optimal growth ofB the logarithmic 2 utility function in the long run, alway
That means, in order to obtain −0.4 optimal growth of the logarithmic utility function in the long run, always
a stake of 10% of the actual capital has to be used. Using a starting capital of X0 = 1000 a simulation o
60 a stake of 10%
Thatof means,
the actual capital
in order
−0.5 has to
to obtain be used.
optimal growth Using a starting
of the logarithmiccapital
utilityoffunction
X0 = in1000
the along
simulation of
run, always
10000 bets yields the results of Figure 1, left:
10000 bets yields
a stake the
of results
10% of−0.6 of Figure
the 1, left:has to be used. Using a starting capital of X0 = 1000 a simulation of
actual capital
40 10000logbets
Equity scale:yields the results
steps = 10000, B = 2, p =of
0.4,Figure
f = 0.1 1, left:
−0.7B = 2, p = 0.4, f = 0.1
120 Equity log scale: steps = 10000,
120
Equity log scale: steps = 10000, B = 2, p = 0.4, f = 0.1
20 120
−0.8
100
100
−0.9
100
0 80
−1
0 2000 4000 6000 8000 80
10000 0 2000 4000 6000 8000 10000
80
60
60
PAGE 4 IFTA.ORG 60
40
40
40
 ecapital
condition that X
instead. k−1 isagain
Using already
the known onethat
condition obtains
Xk−1 is already known one obtains
� � � IFTA JOURNAL 2015 EDITION
� � � � � �� � � � ��� � �� � � ��
E p log �
· logXxk 1�+ Bf
Xk−1 = + xq · log=x p1 ·−log
f x 1 + Bf + q · log x 1 − f
� (1) (1)
� � � ��� � � � ��
log x + p logIf 1this
+ Bf + q log
= 1
log−x f+ p log 1 + Bf + q log
expression is viewed as a function of f, its maximum is 1 − f Here we set drawdown(k) = 1 – (Xk  /equitymax(k)) ∈ [0,1]
and equitymax(k) = max Xj .
q q 1≤j≤k
ns is
maximum achieved
functionat
viewed asis aachieved atofffopt, its − > 0is (Kelly
= pmaximum formula).
achieved at fopt = p − > 0
B B
Figures 1 (right) and 2 (left) show clearly that for f = fopt severe
drawdowns are to be expected. These drawdowns would have
Simulation of single investments large psychological impact on every trader and investor.
single investments In the following we want to do some simulations. Assume for On the other hand, in case a stake of only f = 1% of the actual
s.
weAssume
want tofor example
doexample BB= =
some simulations. 2 and p =Assume
2 and 0.4.
p =The0.4.optimal
The
for f according
optimal
example Bf= 2toandKellypthen is Thecapital
= 0.4. optimal is used
f (as recommended by many experts) the severe
ly then is drawdowns can be prevented (cf. Figure 2 (right) and Figure 3).
0.6 q 0.6 The expected value (dotted line in Figure 2 (right)) and the result
= 0.4 − = 0.1fopt = = 10%p − . = 0.4 − = 0.1 = 10% .
2 B 2 of this simulation are however considerably lower.
It can be observed that large drawdowns can be avoided for
of thetologarithmic utilitygrowthfunction in the
order obtain optimalThat means thatofinthe
order tolong
logarithmic run, utility
obtain alwaysfunction in the long run, always
optimal growth of the suboptimal f << fopt, but only at the expense of a lower capital
ofsed.
the Using
actuala capital
starting
logarithmic capital
has beofused.
toutility X 0 =Using
function 1000 aa starting
in the simulation a of
long run,capital
stakeofofX
10%
0 =of1000
the a simulation
growth. Whatof remains is the question, whether both, optimal
s the results of actual Figurecapital
1, left:always has to be used. Using a starting capital of capital growth with simultaneously bounded drawdowns, is
X0 = 1000, a simulation of 10000 bets yields the results of Figure reachable? Here the diversification comes into play.
g scale: steps = 10000, B = 2, p = 0.4, f = 0.1
1, left:
On the x–axis the bets k = 1, ..., 10000 are assigned. The Diversified optimal f
dotted line in Figure 1 (left) shows for f = fopt = 10%, k = 1, ..., 10000, The aim of diversification is to load the depot capital risk
the expected value E(log(Xk)) — a line with slope p  log (1 + Bf ) + q on several “shoulders” (virtual depot parts). In case the capital
log (1–f) ≈ 0.0097. This is more or less realized in the simulation. growth on each depot part has positive expected value, the
The right graphic in Figure 1 shows the negative drawdowns whole depot also becomes a positive expected value (through
(– drawdown(k), k = 1, ..., 10000) of this simulation and dotted averaging).
the maximal drawdown (see also the empirical distribution of If the expected returns of the depot parts are of the same
these drawdowns in Figure 2 (left)). order, then the expected return of the whole depot is also of that

00 4000 Figure8000
6000 2: Distribution
10000 drawdowns for fopt (left) and y = log(Xk) for f = 0.01 (right)
verteilung drawdown: steps = 10000, B = 2, p = 0.4, f = 0.1 Equity log scale: steps = 10000, B = 2, p = 0.4, f = 0.01
fopt and is negative � � (right)
7 drawdown 30
Figure 1: y = log Xk for fopt and is negative drawdown (right)
6
25
5

20
2 4
2
3
15

10
1

0 5
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 2000 4000 6000 8000 10000

Figure 3: Negative drawdown for f = 0.01 (left) and distribution (right)

blood curve(neg drawdown): steps = 10000, B = 2, p = 0.4, f = 0.01 verteilung drawdown: steps = 10000, B = 2, p = 0.4, f = 0.01
0 20

18
−0.05
16

−0.1 14

12
−0.15
10
−0.2
8

−0.25 6

4
−0.3
2

−0.35 0
0 2000 4000 6000 8000 10000 −0.2 0 0.2 0.4 0.6 0.8 1 1.2

IFTA.ORG PAGE 5
 � � 10 50 5
Figure 5: distribution drawdown M = 10 (left) and y = log Xk M = 25 with fopt (right)
5 00
5

−0.2 0 0.2 0.4 0.6 0.8 1 1.2 0

0 2000 4000 6000 8000 10000
IFTA JOURNAL 2015 −0.2
EDITION
0
0
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0 2000 4000 6000 8000 10000
The capital after k bets, Xk , is the sum of the capitals
0 0.2 0
of
2000
0.4
the4000depot
0.6 6000
� �
parts
0.8 8000 1 1.2 10000

The capital after k bets, Xk , is the sum of the capitalsFigure

of the5: depot parts
distribution drawdown M = 10 (left) and y =� log �Xk � M = 25 with fopt (right)
Figure 5: distribution drawdown M = 10 (left) and� y = log Xk M = 25 with fopt (right)
Figure 5: distribution drawdown M = 10 (left) and M
y = log Xk M = 25 with fopt (right)
magnitude, i.e. we give away nothing. Nevertheless, soThe the hope,M
�
�
capital after k bets, Xk , is the sum of the capitals of the k depot parts
the fluctuation of the equity curve of the wholeThe depot The capital after k bets, X , is the sum X of the
k = capitals ofYthe ,depot parts
Xwill be k bets,
Yik ,Xk , is the sum of the capitals depoti parts
k
k =
capital after of the M
�
reduced by the gains and losses of the partial depots. We want to i=1 M
i=1 MX = � Yikk ,
� Xkk = Yi ,
apply this idea to fractional trading with optimal f. where the i–th depot Xkpart = is capitalized
Yi , i=1 before the X
k i=1 k–th bet
where the i–th depot part where the i–th
is capitalized depot
before the part
k–th is
bet with Xk−1 and
capitalized
with and the
before
the capitalthe
capital
i=1 after
k–ththe k–th
after bet
the bet
with
k–thisk−1
X denoted
bet
k−1 k
Y . In the capital
Mistheiand
where the i–th depot part Mis capitalized before the k–th bet with XMk−1 and capital after the k–th b
Simulationk with
denoted Yi . In this case the partial
expected
denoted
depots
Y klog–growth
. In this
where the ki–th depot
is
case
denoted given
Ythe by this
expected
i k. part
In this
case
partthe expected
is capitalized
log–growth
case the expected
log–growth
before the k–th
log–growth is
is
isgiven
given
bet with
X
given k−1 byby
by M and the capital after the k–th b
where the i–th depot is capitalized before the k–th bet with and the capital after the k–th bet is
Thereto let us again consider the Kelly ibetting variant with Yi . In this
denoted
k . In this case the
case the expected log–growth is givenMby
� expected� log–growth is�given by
� �
denoted Y � �
� each�bet � �� =
� however, before
B =2, p = 0.4, fopt = 10% . This time, i the � � � ��
� ��� � � � �� EE � log �Xk � ���� �Xk−1
� �� ==xx
capital will be splitted
E loguniformly �
Xk � onXM= k−110=(orxM= 25)= virtual �� � �
log
� X k � X� k−1 =
depot parts. Then each partial depot bets (stochastically E log X
E klog
�X k X
� k−1
� Xk−1 == xx � = =
� �
� � ��
�M
�M � M � j� �M −j � B+1 ��
independent) with an fopt fractionM of its�partial�depot. � log(x) + M p �
1 �− p � M −jlog 1 + f · j (2)
B + 1 − 1 .
log(x)
� �+ � j p j
� 1 − p� log � 1 + f · j M � � − 1 . (2)
� � ��
M
M
The lower dotted line in the left graphic of Figure � �
j 4 shows as log(x) M
M −j
�
� M B
j=0
j=0� +j
�1 j �M −j B + 1 M
�
log(x) + p 1−p log � +
1 + f · jj p 1 − p
− 1 �. log 1 + f · j − 1 .
in Figure 1 the expected value for a single investment
j per bet. j=0 M M j � M −j M B + 1
j=0
The upper dotted line (which is very dose to the equity curve)
Forlog(x)
convenience + we
For convenience For
give the argument
convenience
we give the jargument
p for
we give
− pBy
1(2).
for the
log
construction
(2). argument
By construction
1+f· j
for (2). By construction M
−1
j=0  �
shows the expected value of log(Xk) when M partial depots arewe give the
For convenience argument for   (1By
(2). + Bf ) · Xk−1 �M
construction with probability p
For used (cf. (2) below).
convenience we give the argument for (2). By construction Y k =  (1 + Bf ) · Xk−1 �
M with probability p .
 Yi k =  i � �M .
)(1
 (1 + Bf ·X− f) ·M X with probability q
k−1 (1 − f ) · Xk−1 with
k−1 M probability p
with probability q
Observations 
For convenience we give Yik =
� the argument for (2). �By construction �.
 (1 + Bf ) · XSince Muse Mwith 
probability
− f )are ·Xp k−1 q �Y k , . . . , Y k �� but basically only
ƒƒ The capital growth is evenYfaster
k as expected for the
k−1 we
single depot parts(1there M
2MM different with probability
possible results for
Since we use M depot parts there are 2 different possible results for Y11k , . . . , YMM
k but basically onl
i = �
number of winners and losers counts. Therefore, . if�
we assume that j of the M trades are winners and M
investment.  (1 − fSince number
) · Xwek−1 of winners
M depot withand losers
(1 2+counts.
M Bf
probability q )Therefore,
· Xk−1if we Massume that
with of thek �M trades are
� k j probability p winners and M
use M parts there
ƒƒ The drawdown (Figure 4 right and Figure 5 left) is reduced k we use different
Sinceare M depot possible results
parts there arefor Y1 , . . . , YM
2M different but basically
possible only the

remarkably. number of winners and Yi losers

= counts.
results Therefore, if we assume
� that of the M trades are winners and. M − j
only jthe
 for (Y
(1 −�1
k
,... , YMk) but
f ) .·.X
basically
� M
number of winners
with probability q
Since we use M depot parts there are 2M different possible results and losers
forcounts.
Y1k , .Therefore, k if
, Yk−1 we assume
but basically thatonly
j of the
the M
M
The capital after k bets, Xk, is the sum of the capitals of the trades are winners and M – j trades are losers, then under the
number of winners and losers counts. Therefore, if we assume that j of the M trades are winners and M − j
depot parts condition that Xk - 1 = x is known we5obtain 5 � �
Since we use M depot parts there are 2M different possible results for Y1k , . . . , YM
k bu

number of winners and losers counts. Therefore, 5if we assume that j of the M trades are
Figure 4: y= log (Xk) for M = 10 partial depots with fopt (left) and negative drawdown (right)
Equity log scale: steps = 10000, B = 2, p = 0.4, M = 10, f = 0.1 blood curve(neg drawdown): steps = 10000, B = 2, p = 0.4, M = 10, f = 0.1
200 0

180 5
−0.05
160

140 −0.1

120
−0.15 5
100
−0.2
80

60 −0.25
40
−0.3
20

0 −0.35
0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000

Figure 5: Distribution drawdown M = 10 (left) and y = log(Xk) M = 25 with fopt (right)

verteilung drawdown: steps = 10000, B = 2, p = 0.4, M = 10, f = 0.1 Equity log scale: steps = 10000, B = 2, p = 0.4, M = 25, f = 0.1
50 250

45

40 200

35

30 150

25

20 100

15

10 50

0 0
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 2000 4000 6000 8000 10000

PAGE 6 IFTA.ORG
 IFTA JOURNAL 2015 EDITION
trades are losers, then under the condition that Xk−1 = x is known we obtain
re losers, then under the condition that Xk−1 = x is known we obtain
�M �
�� k
log Xk �
= Mlog Yi statistical tests which raise the alarm when correlations are
� i=1
log Xk = log Yik
� �
= log
i=1 x
j(1 + Bf ) M + (M − j)(1 − f ) M x changed (cf. Wied [6]).
� � � �
x
� ��
j) x
� In Figure 5 (right) and Figure 6 we can observe that for M = 25
= log = j(1 + Bf
log(x) )M
+ log 1+
+ (Mj
Bf− j)(1
+ 1 −
− fM M f −
� � M
� � � � � � �
� depot parts the drawdown is furthermore reduced remarkably
= log(x) + log+ log
= log(x) 1 +1 M j
+ fBf B+1
j + 1 − M −f
M − 1
j
while the expected equity growth is extended a little.
� � �� To be applicable for real investments, the easy Kelly betting
B+1
= log(x) + log 1 + f j M −1 example has to be substituted by a realistic returns distribution
�
Now (2) follows easily because Y1k , . . . , YM
�
k is binomially distributed. and as investment fraction in the depot parts optimal f from
Now (2) follows easily because (Y1k,... , YMk  ) is binomially distributed. Vince (cf. [4]) would have to be used. Since Kelly betting is just
�
� k k
� an easy case of optimal f, we expect that more complex return
follows easily because 1 , .M
Remark:YFor M is binomially
. .=, 1Yformula (2) is equaldistributed.
to the old formula from (1): distributions would yield similar results. A drawdown control,
Remark: For M = 1 formula (2) is equal to the old formula from (1):
� � ���
as suggested� in the “leverage space trading model” in [5], would
� ��
E log Xk � Xk−1 = x = log(x) + p · log(1 + Bf) + ( 1 − p) log(1 − f) not be necessary.
k: For M = 1 formula (2) is equal to the old formula from (1):
In particular
� � f =� �f�
Inoptparticular f =function
of the utility fopt of the
(1)utility function
is in general (1) is in
no longer for maximizing Conclusion
general
optimal the utility
no �longer optimal �� for maximizing the utility function (2). With the help of simulations it was possible to verify
E log Xk � Xk−1 = x
function (2). Nevertheless, we obtain a win–win situation:
= log(x) + p · log(1 + Bf) + ( 1 − p) log(1 − f)
Advantages:Nevertheless,
� The severewe obtain aare
drawdowns win–win situation.
controlled. that the use of optimal f position sizing in connection with
� The expected gain grows remarkably compared to a single investment. diversified partial depots yields a remarkable reduction of the
cularNevertheless, Advantages
f = fopt ofthere theare
utility • The
function severe
(1) is indrawdowns
general
also disadvantages which should be mentioned:no are controlled.
longer optimal for maximizing maximal drawdown compared to single investments while
the utility
(2). Disadvantages:
Nevertheless, we � obtain
• The
a win–win
More signals
expected
situation:
are needed
gain grows remarkably
for each bet
concurrently the expected equity growth is raised. Suboptimal

ntages: � The severe (preferably drawdowns
compared
stochastically toindependent
are controlled.
a single investment.
or at least uncorrelated). fixed fraction trading approaches are literally declassified.
� The fees are multiplied. The dificulty of applying this method is, however, to provide
� The expected gain grows remarkably compared to a single investment.
Nevertheless, there are also disadvantages which should be many uncorrelated investment possibilities simultaneously. A
The disadvantages seem to be of technical nature. They are, however, in fact restrictive or at least difficult
eless, there are alsomentioned
disadvantages which should be mentioned:
to realize. The assumption that the investments in partial depots is possible stochastically independent,
consistent implementation of such a strategy results in a win–
vantages:
is probably� notMore signals
realizable areglobally
in our neededconnected
for eachfinancial
bet markets. As easing of this assumption, win situation
one and may be viewed as a further prove why many
Disadvantages
could demand(preferably • More
of the signals
stochastically
that the correlation aretheneeded
independent
returns of or at
depot for each
least
parts betor at least in absolute
uncorrelated).
is zero experts valuefor a long time call diversification the only “free lunch”
small. This� can
The be fees
monitored (preferably
by usual
are multiplied. correlation stochastically
estimators. One, independent
however, has to be on alerton WalltheStreet. This seems to be a valuable complementation of
when
correlations grow dramatically as it or at least
happens uncorrelated)
regularly the classical portfolio theory where the only risk measure used
in financial crises (so called “correlation meltdown”).
To be warned
advantages to bethere
seemearly, • The
are powerful
of technical fees are
statistical
nature. multiplied.
tests
They which
are, raise the alarm
however, in factwhen correlations
restrictive was
or are
at thedifficult
changed
least standard deviation and therefore drawdowns were not
e. The assumption that the investments in partial depots is possible stochasticallyat independent,
(cf. Wied [6]). all addressed.
bly In
not realizable
Figure 5 (right)in The
and disadvantages
our globally
Figure seem tothat
connected
6 we can observe be for
of technical
financialM =markets. nature.
25 depot As They
parts easing of thisis assumption,
the drawdown furthermore one
emandreduced are,
thatremarkably however, in fact
while theofexpected
the correlation restrictive
equity growth
the returns or at least
of the isdepot
extended difficult
a little.
parts is zero or at least in Notes
to realize. absolute value
This can be monitored The assumption
by usual that the investments
correlation estimators. in partial depots ishas to be on 1. alert
One, however,
T. Ferguson, The Kelly Betting System for Favorable Games, Statistics
when the
Department, UCLA.
possible stochastically independent, is probably
ons grow dramatically as it happens regularly in financial crises (so called “correlation not realizable 2. meltdown”).
J. L. Kelly, Jr. A new interpretation of information rate, Bell System Technical J.
in our globally connected financial markets. As easing of this
arned early, there are powerful statistical tests which raise the alarm when correlations
35:917-926, (1956).
are changed
assumption, one could demand that the correlation of the 3. Henry M. Markowitz, Portfolio Selection, FinanzBuch Verlag, (1991).
d [6]). 4. R. Vince, The Mathematics of Money Management, Risk Analysis Techniques for
returns of the depot parts is zero or at least in absolute value Traders, A Wiley Finance Edition, John Wiley & Sons, Inc., (1992).
6
small. This
e 5 (right) and Figure 6 wecan canbeobserve
monitored thatby usual
for M =correlation
25 depot partsestimators.
the drawdown5. is furthermore
R. Vince, The Leverage Space Trading Model: Reconciling Portfolio Management
One, Strategies and Economic Theory, Wiley Trading, (2009).
remarkably while thehowever,
expectedhas to begrowth
equity on alertiswhen the correlations
extended a little. grow 6. D. Wied, Ein Fluktuationstest auf konstante Korrelation, Doktorarbeit,
dramatically as it happens regularly in financial crises (so called Technische Universit¨at Dortmund, (2010).
“correlation meltdown”). To be warned early, there are powerful

Figure 6: M = 25 drawdown (left) and distribution (right)

blood curve(neg drawdown): steps = 10000, B = 2, p = 0.4, M = 25, f = 0.1 verteilung drawdown: steps = 10000, B = 2, p = 0.4, M = 25, f = 0.1
0 6 80

−0.02 70

−0.04 60

−0.06 50

40
−0.08

30
−0.1

20
−0.12

10
−0.14
0
−0.16 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
0 2000 4000 6000 8000 10000

IFTA.ORG PAGE 7
 IFTA JOURNAL 2015 EDITION

Shawn Lim Douglas Stridsberg

[email protected] [email protected]
Feeling the Market’s Pulse Society of Technical Analysts’ UK Flat 326, Dashwood Studios

With Google Trends Dean House, Vernham Dean

ANDOVER, Hampshire SP11 0JZ
United Kingdom
126 Walworth Road
SE17 1JL London
United Kingdom
By Shawn Lim, CFTe, MSTA, and Douglas Stridsberg
+44 20 7125 0038 +44-778-4664939

Abstract that Google Trends can be utilized by the financial technician

This article explores how Google Trends has been applied as a source of data in addition to price and volume. Google
in different disciplines and the relevance of search query data enjoys an almost 70% share of the search engine market
to financial markets. We contend that if Google Trends can be (Netmarketshare.com, 2014), and its data has been shown to
used to recover retail investor interest in a particular security, reveal investor sentiment and interest. We will introduce Google
market or issue, it can provide valuable information to a Trends and lay the foundation for its inclusion in a technical
technical analyst. We propose possible applications of Google analyst’s toolkit by investigating what Google Trends in fact can
Trends to improve existing concepts in technical analysis (price tell us about the stock market and, in particular, future stock
movements, trend analysis, oscillators and trading bands) and price movements. We aim to propose new ways to incorporate
demonstrate how it can act as a useful tool to improve signal Google Trends data into mainstream technical analysis in order
reliability. In addition, we suggest other areas that could benefit to enhance profitability.
from the use of this data source (volume analysis, sentiment
analysis and event studies). Understanding Google Trends
Introduction What is Google Trends
The term ‘Big Data’ was born around the turn of the last Google Trends is a service by Google that offers users the
decade to describe the exponential growth in the availability ability to, among other things, visualize the relative popularity
and size of enormous, largely unstructured sets of data, such of a keyword (i.e. the number of searches done for it) over time.
as those generated by Google users. The analysis and use of It also offers the opportunity to compare one keyword with
this data has gained traction in recent years as a means of another, as well as to rank the most popular search terms in
understanding and predicting consumers’ needs in order to gainvarious categories and in various geographical regions. Perhaps
a competitive advantage. One of the first sources of Big Data to
the most interesting aspect of the data is the fact that it reveals
open to the public for analysis was Google Trends, which gained
the intentions of a user, often long before they act (Da et al.,
particular momentum when Choi & Varian (2009) demonstrated 2011). The data is not presented in its raw form—rather, it is
that various business metrics, such as sales volumes, could benormalized to avoid problems with changing Google popularity
predicted from Google Trends data. In this article, we argue and changing Internet usage. The data is then scaled from 0 to
100 in order to be comparable
Figure 1: An example search for ‘german cars’ in Google Trends, showing the SVI to other keywords, where
between 2004 and 2014. 100 represents the maximum
popularity during the time
period chosen. In our paper,
we will use the term search
volume index (SVI) for the data
provided by Google Trends on
the relative popularity of a
search term.
In Figure 1 we see an
example of the SVI for the
search term ‘german cars’.

Literature review
The increasing importance
of the Internet as a primary
source of information has been
one of the key themes that has
characterized the last decade.
With our recent ability to
uncover the revealed interest

PAGE 8 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

of individuals through the SVI that has been made publicly as technical analysts think about this potential new source of
available via Google Trends, researchers have taken advantage data? Da et al. (2011) provide some useful suggestions in their
of that source of data to explore potential applications. The paper that explores the link between Google Trends and other
results from these studies have been fairly encouraging, with proxies of investor attention and concludes that Google Trends
studies from a range of disciplines posting positive results from is likely to measure the attention of retail investors. Beyond
the use of Google Trends in various forms of inquiry. empirical evidence, the intuition behind that interpretation is
In Medicine, for example, researchers have explored the also fairly convincing—when we think about financial market
possibility of using Google Trends to detect the spread of participants and the avenues through which they access
disease. Ginsberg et al (2009) conducted a study on the ability information, professional investors are likely to have access to
to detect influenza epidemics with search query data and found additional paid data sources, such as Bloomberg, and typically
that flu trends could be predicted from search data, while use that as their primary source when searching for security
Chan et al. (2011) conducted a similar study to detect Dengue specific information. Hence, the participants who use Google
epidemics with similar positive results. These findings have to search for security information are likely to be those who do
led to the development of the Google Flu Trends application not have access to any specialist data sources, a group probably
based on aggregated search data and demonstrate the power of best described collectively as retail investors.
Google Trends in enabling us to gain an insight into a range of Besides attributing the identity of the group tracked by
issues from the revealed interest of people identified through the search volume index to retail investors, Da et al. (2011)
what they search for on Google. also provide some evidence on the behavioural characteristics
The ability of Google Trends to predict the outcome of of this group of investors. In particular, they find that an
human-determined processes has also been the subject of increase in the SVI predicts higher stock prices in the next
a number of studies. Stephens (2013) explored the ability of two weeks and an eventual price reversal within the year. This
Google Trends to predict election turnout and found that it is consistent with other literature (Barber and Odean, 2011)
can be used to proxy voting intention in various parts of the that has documented the tendency of retail investors to be
United States. There has also been an interest in the ability influenced by various behavioural biases that contribute to
of Google Trends to help us better understand the underlying such short-term overreaction. Beer et al. (2012) found evidence
state of the economy. The ability of Google Trends to ‘nowcast’ for similar dynamics of short-term overreaction in the French
macroeconomic data has been studied in Choi and Varian market and provide additional evidence of the ability of Google
(2009), which found a positive correlation between initial Trends to capture retail investor interest by studying the
unemployment claims and searches related to jobs, welfare and relationship between the SVI and mutual fund flows.
unemployment. Vosen and Schmidt (2011) conducted a study Throughout this paper, we employ these two key insights
to evaluate the use of Google Trends as an indicator for private from existing academic literature in our exploration of its
consumption and found that it offers significant benefits to potential relevance to technical analysis. We view Google
forecasters of private consumption over traditional consumer Trends as a proxy for retail investor interest and as a potential
confidence indices. tool to identify the short-term overreaction often displayed
Besides macroeconomic variables, there has been an interest by this group of investors. We utilize two terms to refer to
in the use of Google Trends to predict the economic performance this hypothesized relationship between the Google Trends
of industries and corporations. Carrier and Labe (2010) tested indicator and retail investor interest. Firstly, we refer to the
the ability of Google Trends to predict automobile sales in Chile situation where the Google Trends indicator (SVI) is increasing
with positive results, while Azar (2009) investigated how oil while the general price is moving in an uptrend or a downtrend
prices react to search volumes related to electric cars and found as short-term retail interest to capture the dynamic described
a positive connection between them. This presents a potential in Da et al. (2011). Following the results from that study, we
means to improve our fundamental forecasts of industry and contend that observing such movements in the SVI captures the
company performance. growing interest of retail investors and is likely to be followed
by a reversal in trend. In addition, we define a security to be
Relevance to financial markets oversearched when the SVI has been steadily increasing and
The ability of Google Trends to reveal information about shows a strong indication of short-term retail interest.
financial markets has also been of recent interest. Dimpfl Secondly, we refer to the situation where the SVI is falling
and Jank (2011) investigated the link between search queries while the price is moving in an uptrend or a downtrend as
and stock market volatility and found a positive result, as the sustainable smart money to capture the implied dynamic of a
inclusion of Google Trends helped to improve in-sample and price trend driven by non-retail interest. We define sustainable
out-of-sample volatility forecasts. Bordino et al. (2012) explored smart money to be broader than institutional interest and to
the link between Google Trends and stock volume with positive refer to all investor interest driven by investors with access to
findings for the NASDAQ, while Joseph et al. (2011) conducted an more sophisticated sources of information. We contend that
investigation of the link between Google Trends and abnormal such interest is likely to be more informed and less subject to
returns with positive findings for the S&P 500. the biases often described in Behavioural Finance literature
The existing academic literature seems to suggest that there and hence we would expect such a trend to be more sustainable
is some information contained in search volume data that can than a similar price movement characterized by short-term
help improve various financial forecasts, but how should we retail interest.

IFTA.ORG PAGE 9
 IFTA JOURNAL 2015 EDITION

Applications to Technical Analysis Analysis of trends

The degree to which Google Trends data will be useful and The concept of trend plays a critical role in the analysis of
applicable depends mainly on the keywords chosen. For other price movements within technical analysis. Through various
applications, previous studies have used keywords ranging tools and techniques, technical analysts attempt to decipher
from generic economic terms such as ‘economics’, ‘jobs’ and two components of price movements: the direction of the
‘unemployment’, to sentiment words such as ‘fear’, ‘hope’ and current trend and the likelihood that it will continue. As a source
‘worry’, to generic product names such as ‘electric car’ and of information on the intensity of retail interest in a particular
‘travel’. When looking at a particular security, however, the security, Google Trends can play a critical role in informing
two most common types of keywords have been the stock technical analysts on the likely sustainability of a current trend.
ticker and company name. It is assumed that retail investors, The first way Google Trends can do that is through a direct
when seeking information on a particular company, will tend comparison of the evolution of trend strength with search
to use either of the two types, as opposed to generic economicvolume interest. One indicator that could be used for such
terms or product names. A quick glance at some company a comparison is the Average Directional Indicator (ADX),
names and their tickers suggests that company names are first proposed in Wilder (1978). The ADX is a directionless
much more widely used; however, this may be because users indicator that measures trend strength, with a reading below
are using the company name to search for products produced 25 suggesting no trend in the market and a reading above 25
by the company. Both types of keywords could potentially being indicative of a trending market, with a higher reading
work, but for the purpose of this paper, we will use tickers suggesting a stronger trend. A comparison of the ADX with
exclusively. Google Trends can thus allow us to better understand which
types of investors are driving the market and hence, better
evaluate the likely sustainability of the current trend.
Two possible scenarios could
be observed through a careful
Figure 2: A demonstration of how the SVI (bottom graph) remains relatively analysis of price action in
aflat during a strong trend indicated by the ADX (middle graph) rising above 25. conjunction with changes in the
The keyword searched for was ‘SBUX’.
SVI. The first possible scenario
exists when the current trend is
being driven primarily by retail
investors. We can detect the
presence of such short-term retail
interest when the ADX is rising and
moves above 25 and the security
is oversearched, as detected by
an SVI that increases along with
it. This oversearched scenario
suggests that the trend is likely to
be short-lived, and we are likely to
see a trend reversal soon, as the
current price appreciation is more
likely to be driven by short-term
sentiment than the improvement
of long-term fundamentals.
The second possible scenario
exists when the current trend
is being driven primarily by
professional investors who are
likely to have a longer term
horizon and to react less violently
to short-term sentiment. We
can detect the presence of such
sustainable smart money when
the ADX is rising and moves
above 25 but the security is not
oversearched and the SVI remains
constant or falls. This suggests
that there has not been an
increase in retail investor interest,
and hence, we would expect

PAGE 10 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

the trend to be more sustainable and therefore more likely to conjunction with oscillators as a secondary indicator to flag
continue to run its course. potential false signals. Oscillators refer to the class of trend
Figure 2 demonstrates the application of this tool to security indicators that allow the analyst to identify short-term
SBUX (Starbucks Corp.). At point A, we see the ADX beginning extremes, commonly referred to as overbought and oversold
to increase and rise above 25, indicative of a market beginning conditions. Common oscillators employed by technical analysts
to trend. The SVI remains fairly flat over this period, suggesting include Relative Strength Index (RSI), Rate of Change (ROC) and
the presence of sustainable smart money and, as expected, the the Stochastic oscillator.
trend continues over the period and does not reverse quickly. Oscillators can be used to signal that a trend may be nearing
its end when there is a divergence between the oscillator and the
Analysis of price price action. These divergences are commonly termed bullish
Price movement is the ultimate deciding factor in the success and bearish divergence and are indicative of an impending
of an executed trade. It is also the most relevant and direct reversal. Bearish divergence refers to the situation where the
source of data, along with trading volume, that a technical market is trending upwards and the price makes a new high, but
analyst has access to in his or her analysis. As such, the analysis the oscillator does not make a new high and instead falls lower
of price together with Google Trends data is the crudest and than the initial high. Bullish divergence refers to the opposite
perhaps the first one should undertake when deciding on a situation, but in the case of a downwards trending market.
trade. As a proxy for retail investor attention, Google Trends can Oscillators attempt to capture the momentum of a trend and use
give us clues about the nature and likely medium-term outcome that in conjunction with price action to evaluate when a trend
of a price movement. is starting to lose momentum and therefore likely to reverse.
The first of two scenarios exists when a short-term price Google Trends can play a useful role in that endeavour by
movement occurs but is not either closely preceded by or closely providing information that either confirms the signal generated
followed by an increase in the SVI for the company. Assuming by such divergent movements or signals the need for further
Google Trends measures retail investor attention, this scenario investigation.
would suggest the price movement is not fueled by such retail The divergence between an oscillator and price action
investors and thus, may instead be fueled by professional signals that the market is likely to have run too far ahead of
investors with a longer term horizon and, perhaps, with better itself and that we are likely to see a reversal soon. One possible
market knowledge. This theory would suggest the reason reason for that could be the short-term nature of investors that
behind the price movement should be investigated, as it may be react violently to news and information and cause the price to
one of importance and may be a
sustainable trend.
The second scenario exists Figure 3: A demonstration of a sudden fall in the price with a corresponding spike in the
when a significant short-term SVI (bottom graph). The keyword searched for was ‘BP’.
price movement occurs and the
security is oversearched. By the
same argument as in the previous
section, an increase in the SVI for
the firm or its stock ticker would
indicate that the price movement
is fueled by short-term retail
interest. This, in turn, may suggest
the price movement is not part of
a sustainable trend.
Figure 3 demonstrates the
situation above in the case of
security BP (BP Amoco PLC). At
region A we notice a sudden fall
in the price of BP, while in this
case, we see an increase in the
SVI around the same point in
time. This indicates that the fall
is fueled by retail investors and
should revert shortly, which it
partly did.

Analysis of oscillators
The third way that Google
Trends can help us in trend
analysis is when used in

IFTA.ORG PAGE 11
 IFTA JOURNAL 2015 EDITION

deviate too far from its fundamental value—hence, the resulting Analysis of trading bands
trend reversal. Amongst the various market participants, Google Trends can also be used as a secondary indicator
retail investors are the most likely to be influenced by various
in conjunction with trading bands to improve their signal
behavioural biases that might affect their investing decision and
reliability. Trading bands are bands plotted above and below the
thus exhibit such short-term behaviour. price line that try to provide a relative definition of high and low
We can use Google Trends to help us identify such a phenomenon
prices. Common examples of trading bands include the Keltner
by investigating the evolution of retail investor interest prior to the
Channel and Bollinger Bands.
oscillator signal. If the security is oversearched and the SVI has been
Trading bands can be used to identify short-term price
increasing sharply prior to the signal, this suggests that the trend
overreaction, typically detected when the price moves above
has been driven primarily by retail interest and thus confirms the
the upper band or below the lower band. As with the previous
signal generated by the oscillator. However, if we see the security is
indicators, Google Trends can provide useful information on
not oversearched and the SVI has remained fairly flat or falls prior
the underlying participants that are driving the observed price
to the signal, this suggests that the trend is likely to be driven by
action and thus provide insights into the validity of the signal
institutional interest and thus suggests that we should further
generated.
investigate the validity of the signal generated by the oscillator.
If the price moves above the upper band or below the lower
Figure 4 demonstrates the application of this idea to security
band while the security is oversearched, this confirms the signal
LYG (Lloyds TSB Group PLC) using the RSI. At A, we see the case
of likely short-term overreaction and hence, we would expect
of a bullish divergence, while the SVI has been fairly flat prior to
the trend to reverse soon. However, if the signal is generated
that. This indicates the need for further investigation, and we see
while the SVI is decreasing or staying relatively flat, this
the trend continuing to trend higher, making a new high later,suggests that the bands may have identified the start of a new
trend and not a short-term overreaction and is indicative of the
suggesting that the signal generated by the oscillator was correct
need to further analyse the underlying drivers of the recent
and thus, the potential utility of Google Trends as a secondary
indicator to improve signal reliability. price movements.
Figure 5 demonstrates the
application of this idea to security
Figure 4: A demonstration of a bullish RSI divergence (middle graph) with a
relatively flat SVI (bottom graph). The keyword searched for was ‘LYG’.
HPQ (Hewlett-Packard Co.), using
Bollinger Bands. At point A, we see
the price move below the lower band
while the SVI increases sharply,
indicating that the security is
oversearched. As expected, this is
indicative of a short-term retail
overreaction, and we see the trend
reverse quickly at point B as returns
to the upper half of the band. Later,
we see the price move above the
upper band while the SVI remains
relatively flat. This is the start of
a new trend, as we see the price
continue to trend higher, and the use
of Google Trends would have been
helpful in identifying the need to
further investigate this observed
phenomenon.

Extensions
Volume
Volume is an important
additional source of data and should,
if available, be used to confirm or
question the signals given by Google
Trends data. Volume can be seen
as the source of the total investor
attention, and as such, comparing
it with SVI data can give investors
an idea of the ratio between the
volume of retail investors and other

PAGE 12 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

investors. In order to successfully analyse this, one would first we could get the SVI for recession or depression as a proxy for
need to find a reference volume level during a period of no or low negative sentiment or use recovery or expansion as a proxy
SVI activity. This reference volume would then represent the for positive sentiment. Next, we could combine keywords that
current average interest from non-retail investors. If the SVI refer to the security or market of interest with keywords that
then increases, one can get a very rough idea about the number transmit a particular sentiment. For example, one could search
of retail investors that have entered the market. This method for X profit warning or X earnings disappointment as a proxy
is, of course, prone to a lot of uncertainty since there could be for negative sentiment or X earnings surprise as a proxy for
an inflow of smart money at the same time as an increase in SVI positive sentiment. Through a careful choice of search terms,
data. we can use the SVI to detect small hints of the underlying
sentiment. However, the success of such an endeavour would
Event studies depend on the suitability and relevance of the keywords chosen
Besides helping us to detect the amount of retail investor for the security of interest and is likely to require an iterative
interest in a security or market, the SVI can also be used as a process for each security to find the most suitable keywords and
source of information as part of an event study. By varying the combinations for sentiment analysis.
keywords used to calculate the SVI, we can gather information
on general interest in those keywords which can prove Google Trend indicators
particularly useful for certain types of event studies. Some A possible extension of the use of the SVI could be to produce
possible applications would include program evaluation, where some sort of indicator from its data. This could aid analysts in
we can analyse the SVI of related keywords to try to understand identifying signals more easily and may give them the ability
the likely take up of a new government policy or to get an to quantify the strength and credibility of the signal they are
early indicator of the success of a company’s new marketing observing. For the many signals identified in this paper that
strategy. The SVI could also help us to better understand the are based on movements in the SVI, suitable indicators could
level of interest around corporate actions and announcements include ROC and RSI, both measuring the speed of change in
and could reveal unique insights about the interest and impact the data. Such indicators should, of course, only be used in
of such news. Through the SVI, we have a powerful tool to conjunction with the SVI and the price and not as a replacement
conduct such inquiry and test hypotheses on the interest that for the SVI. Caution must also be exercised, as the SVI itself is
news and events might have generated, which could then not guaranteed to be correlated to price movements. We leave
provide additional insights to inform our trading decisions. this branch of Google Trends analysis open to the reader to
One example of a potential application can be seen in Figure 3, explore.
which plots the price and SVI of the BP stock and captures the
impact on search interest in the wake
of the BP Deepwater Horizon oil spill
Figure 5: A demonstration of a bullish Bollinger Bands signal with a steadily
that occurred in 2011. By combining
increasing SVI (bottom graph). The keyword searched for was ‘HPQ’.
the insights gained from Google Trends
with the observed price movements,
analysts might have been able to draw
conclusions on the likely sustainability
of the price correction and might have
been in a better position to predict the
subsequent reversal.

Recovering sentiment through

choice of search terms
One limitation to using a security
ticker or a company or market name
for a search is that, while it provides
information on investor interest, it
offers only a directionless indicator
with no information on the underlying
sentiment. However, we can still use
the SVI to uncover sentiment-related
information through a careful choice
of search terms. There are two possible
ways that we could vary keywords.
First, we could select keywords that
imply a certain type of sentiment. For
example, to try to identify the general
retail market sentiment of the economy,

IFTA.ORG PAGE 13
 IFTA JOURNAL 2015 EDITION

Conclusion
In this paper, we have explored various ways to use
the information made available through Google Trends to
complement established methods in Technical Analysis.
When applying these ideas and thinking of other potential
applications, we would encourage analysts to think creatively
about the information that could be recovered, while keeping
in mind the potential limitations of the tool as a coarse
measure of retail interest. In addition, the applicability of the
information to any particular security is likely to be influenced
by its underlying ownership structure and the potential
influence of retail investors on price movement. However,
in spite of its potential limitations, we believe the dynamic
nature of the information source and its potential versatility
make it a tool with much promise to be developed further by
Technical Analysts. As more knowledge in this area emerges
and its applicability becomes more widespread through greater
adoption of the Internet and Google around the world, we
believe Google Trends could become a valuable instrument in
every Technical Analysts’ toolkit.

Software and Data

SVI data courtesy of Google Trends (www.google.com/
trends). Stock charts courtesy of StockCharts (www.
stockcharts.com).

References
Azar, J. 2009. Electric cars and oil prices.
Barber, B. M. and Odean, T. 2011. The behavior of individual investors. Handbook
of the Economics of Finance, 2 pp. 1533–1570.
Beer, F., Hervé, F. and Zouaoui, M. 2013. Is Big Brother Watching Us? Google,
Investor Sentiment and the Stock Market. Economics Bulletin, 33 (1), pp. 454–466.
Bordino, I., Battiston, S., Caldarelli, G., Cristelli, M., Ukkonen, A. and Weber, I.
2012. Web search queries can predict stock market volumes. PloS one, 7 (7), p.
40014.
Chan, E. H., Sahai, V., Conrad, C. and Brownstein, J. S. 2011. Using web search
query data to monitor dengue epidemics: a new model for neglected tropical
disease surveillance. PLoS neglected tropical diseases, 5 (5), p. 1206.
Carrière-Swallow, Y. and Labbé, F. 2011. Nowcasting with Google Trends in an
emerging market. Journal of Forecasting.
Choi, H. and Varian, H. 2009. Predicting initial claims for unemployment
benefits. Google Inc.
Choi, H. and Varian, H. 2012. Predicting the present with Google Trends.
Economic Record, 88 (s1), pp. 2–9.
Da, Z., Engelberg, J. and Gao, P. 2011. In search of attention. The Journal of
Finance, 66 (5), pp. 1461–1499.
Dimpfl, T. and Jank, S. 2011. Can Internet search queries help to predict stock
market volatility?
Ginsberg, J., Mohebbi, M. H., Patel, R. S., Brammer, L., Smolinski, M. S. and
Brilliant, L. 2009. Detecting influenza epidemics using search engine query data.
Nature, 457 (7232), pp. 1012–1014.
Joseph, K., Babajide Wintoki, M. and Zhang, Z. 2011. Forecasting abnormal
stock returns and trading volume using investor sentiment: Evidence from online
search. International Journal of Forecasting, 27 (4), pp. 1116–1127.
Netmarketshare.com. 2014. Search engine market share. [online] Available
at: <http://www.netmarketshare.com/search-engine-market-share.
aspx?qprid=4&qpcustomd=0> [Accessed 20 Apr. 2014].
Stephens-Davidowitz, S. 2013. Who will vote? Ask Google.
Vosen, S. and Schmidt, T. 2011. Forecasting private consumption: survey-based
indicators vs. Google trends. Journal of Forecasting, 30 (6), pp. 565–578.
Wilder, J. W. 1978. New concepts in technical trading systems. Greensboro, N.C.:
Trend Research.

PAGE 14 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Andrew J.D. Long

Identification of High Probability Target [email protected]
488-A Redfox Rd.
Zones (HPTZ) Waterloo, Ontario
Canada N2K 2V8
By Andrew J.D. Long, MFTA +1(519) 746-9666

Abstract Data Collection

Technical Analysis is used to assess the current market Since July 2012, the process of identifying HPTZ has been
through historical data in an attempt to forecast future ongoing through the publication of real-time forecasts to a
market potential. There are multiple varying techniques and subscriber base. The results of these forecasts are used for the
methodologies that are employed to attempt this, each with purposes of “proof of concept”. As of September 30, 2013, 539
their own strengths and weaknesses. forecasts had been made across seven different markets and
Each technical tool or method offers its own perspective and three timeframes. These markets are SPX, US$, EUR/JPY,EUR/
usually has several different options or potential scenarios for USD, VIX, GOLD and OIL. The timeframes for the forecasts occur
future market movement. across the weekly, daily and hourly charts.
In some cases, more than one technique is used together in More detail on data collection and final results for the
order to attempt to forecast the market with greater accuracy. methodology are given in the Performance section.
For example, Elliott Wave and Fibonacci Theory are often paired
up to assist in determining the size and locations of future waves. Practical Application
Following along with this concept, it will be shown that using A purely technical trading strategy will be discussed that is
multiple techniques, properly integrated together, can increase a natural progression of the HPTZ process to demonstrate how
the probabilities of an accurate analysis. a trader may use the method practically. This discussion can be
While it is possible to demonstrate the theory through back- found in the Technical Trading Method section.
testing data, the more effective means of real-world published
forecasts will be used to showcase the methodology and its Methodology
effectiveness. The concept of Fibonacci is prevalent in the HPTZ
methodology and is used to tie different tools together through
Introduction one commonality. The idea that the market moves in waves
Through the integration of multiple technical analysis and is fractal in nature, and that Fibonacci expansion and
methodologies, it will be shown that it is possible to increase contraction occurs throughout, is part of the underlying base
the probability of market forecasts. Target areas identified are or “belief system” of the process. Even if the analyst is not sold
referred to as High Probability Target Zones (HPTZ). on these ideas, they should be assumed while applying the
methodology.(1)
Methods Used Whenever possible, any variables or settings that can be
modified by the analyst should be done so with a Fibonacci
Fibonacci Bollinger Bands number or ratio. For example, moving averages are set to 13 and
Trend Lines and Channels Patterns Elliott Wave Theory 34; the Williams %R uses a 13 period setting.
Moving Averages Indicator (Williams %R) The general concept is to overlay several different technical
methods and tools. Where these tools appear to “converge” or
Each method on its own has had extensive research and tell a similar story is where we look for High Probability Targets.
testing. Multiple books and research papers can be found on
any one of the individual techniques. These are the standard Passive and Active Technical Analysis
tools taught and required by the Society of Technical Analysts The method is divided into two steps.
to earn its diploma and should be familiar to anyone who is a
technical analyst. 1. Passive TA: This is a reference to the initial setup of the charts.
I make no claim to any of the individual processes, and their For example, Fibonacci clustering is used and set up on the
applications follow standard practice. Any exceptions (e.g., chart prior to trading. Any long-term studies that are assessed
specific indicator settings) are detailed and explained in the and applied to the chart are referred to as Passive TA.
Methodology section. 2. Active TA: This is a reference to any analysis that is
It is the integration and overlapping of several common tools performed on the current market wave(s) and is projecting a
and where their sum creates areas of interest that is the focus potential future outcome.
and my contribution.
More detail on each tool and how it is applied will be The general process has the analyst set up several Passive
demonstrated in the Methodology section. tools prior to the market open. During the trading day, as the

IFTA.ORG PAGE 15
 IFTA JOURNAL 2015 EDITION

Chart 3: Dropping down to the daily, the process is

market is unfolding, Active analysis are then applied, and places continued; T-M (pink), M-U (orange) and V-W (green) are
where there are high concentrations of overlapping tools or added.
convergences are noted.

Fibonacci Clustering
This technique identifies significant levels of support and
resistance for the market, and I first read about the method
from Constance Brown. Although she does not coin the phrase
“Fibonacci Clustering” in her book, Technical Analysis for the
Trading Professional (Chapter 6 : Adjusting Traditional Fibonacci
Projections for Higher-Probability Targets), she discusses using
multiple Fibonacci analyses and noting the places where they
group.(2)
This is a very effective technique, and the concept is used in
both the Passive setup and the Active analysis.

Application Capturing the significant waves from the weekly and daily
The concept and application is simple and straightforward. timeframes with Fibonacci studies sets up levels the market
Starting with a higher timeframe, set a Fibonacci Retracement clearly respects. From W (July 2012) forward, all studies could
Study from the lowest price to the highest price. Any other have been in place and offering guidance to the present day.
obvious high-low moves or waves should also have a Fibonacci This example has demonstrated the application of several
Study applied to them. Repeat this process for lower time scales. Fibonacci studies with Price Retracements. The same process is
I will usually assess the weekly, daily and hourly timeframes and also done for both Price and Time Extensions.
recommend that these are the minimum that should be used. Areas or levels of interest are those where we see several
studies overlapping or grouping together.
Chart 1: OIL Weekly: A Fibonacci study is added from the Once you have set up Fibonacci studies for several
low X to high Y.
timeframes, the next passive methods are added.

Trendlines, Channels and Patterns

Similar to the application of the Fibonacci studies, significant
supports and resistances are located and added across several
timeframes. Parallel trendlines set up channels, and non-
www.triggers.ca
parallel trendlines identify patterns or wedges.
Chart 4: Supports and Resistances (s/r’s): Horizontal
(blue), Angled (black), Patterns (green, purple). The same
should be done for multiple timeframes, as we have for
the Fibonacci studies.

Chart 2: Further studies are added to obvious waves;

Y-L (blue) and L-M (purple).

In the weekly Oil chart (chart 4), significant supports and

resistances have been added. Note that the black trendlines are
not horizontal. In the example above, the lowest black trendline
was placed (A), and the other parallel lines were then added,
creating a series of channels, one on top of the other. The green

PAGE 16 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 6: Blue Box: Samewww.triggers.ca

as blue box on the weekly chart:
and purple trendlines identify different wedge patterns. can see the 34ma & 13ma LBB’s provide support; Yellow
Note the three red arrows. These mark places where there Box: Area of hourly Chart 7 below.
www.triggers.ca
are several supports and resistances coming together or
crossing, and where the market has a pivot. These are examples
of what we are looking for in the future. Areas on the chart,
ahead of the current market, where several supports and
resistances group together, are places of interest for HPTZ. Just
as the market has pivoted where we can see the red arrows, we
can expect that there is a potential for this to occur again, at the
next areas of convergence.
As with all the tools we use for the HPTZ methodology,
trendlines, channels, and patterns are basic technical analysis
techniques, and most introductory books on the subject will
discuss these. Technical Analysis Simplified, by Clif Droke, gives
clear direction on supports, resistances, channels and basic
pattern recognition for further reading.(3) Chart 7: With no other technical tools to provide
perspective, it is unclear why the market turned at A,B
and C. Looking to the daily timeframe holds some answers.
Moving Averages and Bollinger Bands
Moving averages and Bollinger Bands, like supports and
resistances, are more basic building blocks for technical
analysis.

Settings: • 34 Moving Average (MA) with a 2 standard

deviation Bollinger Band (BB)
• 13 Moving Average (MA) with a 2 standard
www.triggers.ca

deviation Bollinger Band (BB)

Note that the settings are Fibonacci numbers.

Just as Fibonacci levels and trendlines can provide support Market turns at A, B and C occur without any obvious
and resistance, moving averages and Bollinger Bands are also reasons on the hourly chart. If we look back to the daily
respected in a similar manner. timeframe, we can see that: pivot at A has the 13ma lower BB;
pivot at B has 13ma upper BB; and pivot at C has the 13ma LBB
Chart 5: Red Arrows: Market reverses off of BB; Blue and the 34ma.
Arrows: Market bounces from 34ma; Green Arrows:
For additional reading, consider the source of Bollinger
Market bounces from 13ma [continues in Charts 6].
Blue Box: Same location as blue box on the Daily Chart 6 Bands, John Bollinger. (4)
(below); Weekly bars appear to spike to “nothing”, Daily
Chart shows MA and BB. What We Have So Far…
Before we take a look at what we have so far with our OIL
example, I want to give one example from the EUR/US$ to
demonstrate what it is we are watching out for with trendlines
and Bollinger Bands. (No moving averages are included in the
example below, just Bollinger Bands.)
Chart 8: Where we see Bollinger Bands near or crossing
significant trendlines, we have areas of interest for
target locations.

We continue to use multi-timeframes here as well. Keep track

of the MA’s and BB’s on the weekly, daily and hourly timeframes
at a minimum. The market respects the MA’s and BB’s across
timeframes; monitoring several can offer perspective as well as
targets and triggers.

IFTA.ORG PAGE 17
 IFTA JOURNAL 2015 EDITION

Chart 10: Trendlines added; solid red and blue are the
In the example above, follow along with the EUR/US$ from original trendlines, with the dashed lines running
left to right and note the convergences of BB’s with the trend parallel.
lines. At W, we can see the market is moving toward a trendline,
and we can also see the 13ma lower BB. At X, we can see the
34ma lower BB and another trendline. If the market does not
hold at W, then the next target to look to would be X.
Here is what our Oil chart looks like with all the tools we have
placed so far:
Chart 9: All tools from examples so far are included.
Green boxes highlight potential target areas for the
market.

To the right of S we can see the lower target we had placed on

previously. This is an excellent example of what we are looking
for to identify HPTZs. We can see the blue channel support, a
Fibonacci cluster that has previously held the market, and the
34ma lower Bollinger Band all coming together in one area. As
the chart develops and tools are added, places of convergence
become of interest.

Elliott Wave
Two authors for consideration discuss Elliott Wave Theory:
It is important to locate targets both above and below the Robert Prechter gives a classical approach (5); Glen Neely offers
market. With what we have so far, we can see the next likely his own spin in efforts to improve wave counts. (6)
target(s) when the current move is complete (green boxes). Both For the purposes of the HPTZ method, however, we are
the upper and lower Bollinger Bands are at significant levels not concerned about trying to achieve a complete or overall
marked by the Fibonacci studies. As we do not know for certain accurate wave count. The method and concepts are used for
which way the market will break, we look for potential targets wave projections and to get a sense of what could be expected
both above and below the current market. next from the market in the near future.
The methods discussed for the tools so far have been We are primarily focused on the concepts of Alternation,
for Passive Technical Analysis setup. They are to be in the 3’s and 5’s, Fractals, and Wave Extensions (Fibonacci). All of
background and offer a potential road map of supports and these are used for an Active TA to give us guidance and locate
resistances. potential targets.
These tools and methods offer targets as they are; however, Alternation: Consolidation waves 2 and 4 in a 5 wave count
we can increase probabilities and find more HPTZs by applying will usually alternate “form”. Where one wave has a sharp
Active Technical Analysis to the current market wave(s). move, the other will be mild. This can be helpful in trying to
Where we see Active TA forecasts overlapping with Passive determine what to expect from the next consolidation. As well,
TA we have HPTZ considerations. if the current structure is unrecognizable, the next should have
a familiar pattern.
More Trendlines 3’s and 5’s: As the market moves in waves of 3’s and 5’s, we
The first time we applied trendlines we were looking for can get a sense of where and when the next turn or pivot could
significant support and resistances from the past that may occur when the market completes a 3 or 5 wave count. This
be respected in the future. This also included channels and shows us how far the current wave could extend and where to be
patterns. looking for potential target locations.
We are going to add trendslines again, only this time, the Fractals: The basic application of Elliott Wave Theory
focus is on the current market movement or waves. We want to demonstrates the fractal nature of the markets. Each wave
identify the significant supports/resistances/channels/patterns within a 5 or 3 wave count can be further subdivided into a
for the most recent action. 5 or 3 wave count. Each wave in those counts can be further
In Chart 10, the red and blue trendlines could have been in subdivided… and so on. This is helpful as you move across
place prior to the lift we see at S. Note the blue arrow pointing timeframes and try to locate the current market position
to the cross of the blue and red trendlines, to where the markets within the overall structure. Also to consider: the subdivisions
initial lift from S moves. If we look to the next lower red dashed of a fractal are in proportion to the whole. If one wave
trendline, we can see the market lifted up through this, where it structure can be identified to have Fibonacci ratios, then
was crossed by a blue dashed trendline (just above S). they all do.

PAGE 18 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 11: Alternation, 3’s & 5’s, wave extension (wave 3 =

2 x wave 1), and the fractal nature of the market. In this case, we take the extension of wave 1 (AA – BB) and
place the start at the end of wave 2. We would be expecting
wave 3 to be a Fibonacci extension of wave 1.
The clustering at DD would have our initial attention,
but the market continues through the levels with strength.
Note however this cluster area appears to hold the overall
consolidation pattern.
At CC we have the extension aligning with the red
extension from Chart 12. The market pushed through the
level slightly, moving to the next ratio, but does pull back, and
the level at CC also provides resistance for the consolidation
pattern.
Had the market continued with strength through the ratios
at CC, we would then have been looking toward the cluster at
EE as the next likely levels to find resistance.

Wave Extensions (Fibonacci): Apply Fibonacci extensions Chart 13: Fibonacci extensions with a potential Elliott
Wave count.
(and/or retracements) to the more recent wave structures. Look
for several waves across timeframes, as we have done for the
Fibonacci clustering. While extensions can be used with Elliott
Wave to forecast future waves, it isn’t necessary to have an
Elliott Wave count to apply extensions (retracements).

Example 1: No Count
Starting with a higher timeframe again, we look for obvious
wave structures associated with the current market.
To the right, we can see L – XX is the wave from which the
extension originates.
The exact Elliott Wave count for the structure is not known.
However, we assume that the when the market lifts past XX and
continues the wave, it will do so with a Fibonacci ratio in relation
to L-XX.
We can see the market lifts to ZZ where the extension aligns
with a retracement study from a larger wave structure.
Chart 12: Fibonacci extension, no Elliott Wave count. The current count may not be entirely correct. At this point,
wave 3 may actually be a third wave of a lesser degree from
the original 1-2 count from which we started the extension.
If this is the case, then we are looking towards EE for the end
of wave 3 (possibly higher, not shown on chart). Regardless
of whether we can figure out the correct count or not, as long
as we can identify wave structures, we can apply Fibonacci
extensions (retracements).

Indicator: Williams %R
While this tool does not print on the price/time graph to
assist with HPTZ identification directly, I do use it for timing
considerations in relation to potential HPTZs and other
significant technical levels/tools.
As with the other variables throughout the methodology, I
use a Fibonacci number for the period setting: 13.
The common practice, as given by Investopedia.com, is
to use the indicator as an overbought or oversold tool. The
Example 2: Elliott Wave Count extreme levels, above -20 and below -80, show when this
Note: the view is the last wave structure from Chart 12 occurs and suggest that the market may be ready for a turn.7
above. AA is marked on both charts for reference. It has been my experience to use the tool a little differently:
As the market lifted past point X, we may have been able to periods of extreme levels indicate positive or negative
identify a potential count unfolding and labelled the 1-2 count. pressure on the market, and it is during these times that the

IFTA.ORG PAGE 19
 IFTA JOURNAL 2015 EDITION

market makes its greatest advances/declines. I have found Multiple areas of convergence should be apparent both above
that when the W%R is at an extreme level it supports and and below the current market. This is desirable and is further
“carries” a trend, and not until the indicator pulls back from explained in the Technical Trading Method section.
the extreme level (drops below -20, lifts above -80) does the Targets that contain tools the market is currently respecting
trend end. should be looked at more closely than those that don’t.
It is not the intent of this paper to prove this, but rather I am
explaining how I use it in relation to the HPTZ method. Technical Trading Method
The indicator can be helpful when approaching technical As well as identifying target locations for potential market
levels/tools and HPTZs. If the indicator is at an extreme level moves, the tools set potential technical trigger considerations
as the market approaches these, it will need to move out of the for the market as it moves through them.
extreme level in order for the market to reverse. Otherwise, the This allows for a purely technical trading strategy. As HPTZs
indicator is suggesting the market may continue and move to are identified both above and below the current market, the
the next significant tool, level or HPTZ. As the tools and levels technical tools between the current location and the potential
are also potential trigger considerations, using the indicator in targets give us technical trigger considerations. This allows us
this manner also assists with market timing for potential entry/ to set up a strategy without any bias. Regardless of whether the
exit considerations. market moves up or down, we can follow along as it crosses the
Larry Williams first introduced his Williams %R indicator in technical tools.
1966. Read more about the indicator and how Williams used it to As the market moves toward a HPTZ, the technical tools it
turn $10,000 in to $1 million in a year.8 crosses (trigger considerations) can be evaluated by the trader/
investor for risk based on their own personal preferences.
Additional Tools There are usually several tools and triggers to choose from.
Following along with the logic we are using to identify HPTZ I classify any trigger consideration that occurs prior to the
(convergence of tools), other tools the analyst is familiar with market breaking out of its current boundaries as an aggressive
may also be layered on top of the methodology as given in trigger consideration. Those tools that currently hold the
attempts to offer more perspectives and increase the overall market, and when broken would be considered a “breakout” by
accuracy of the targets. the market, are trigger considerations with less risk.
While I have used other tools myself, I have not done further While setting up the charts to identify HPTZ, a technical
research on any specifically to see if they are in fact increasing trading method is also naturally set up. All the significant
the odds or not. market levels, supports and resistances should have been
The tools and methods outlined above are the base or located while applying the process. We originally set these up to
backbone of the methodology. All of the HPTZ recorded for the look for convergences through Passive and Active TA.
results are a product of the method outlined. The methodology then allows any trader/investor style to
Very rarely, an additional tool was used. However, I found use these as technical triggers for entry and exit considerations,
that they would just confirm targets already identified and based on their own risk tolerances. This is a methodology, not a
chosen, rather than giving me anything new to consider. trading plan. An individual’s own trading plan/strategy can be
Gann: I have tried out Gann levels and price/time squared built with the methodology as the base.
targets. The thought on using Gann is that the method is Further examples and explanation can be found in the next
completely outside the Fibonacci concept we are employing section, Performance.
throughout the process, and it could offer an “independent”
perspective, increasing odds when the Gann analysis aligns Performance
with the rest. I have noticed, from time to time, that the I will first go over the data collection and performance of the
Gann levels and targets will correspond, but I have not done methodology. I will then provide examples taken from published
enough investigation into this specifically to provide any real calls to demonstrate the HPTZ methodology as well as discuss
conclusions or data. trigger considerations as they pertain to a technical trading
Fibonacci Circles: This is another tool that I have started method.
using on a couple of the instruments from time to time. I
have found it to be successful in identifying levels, similar to Data Collection
standard Fibonacci extensions and retracements, but again, the The time period for the data collection begins July 2012 and
tool only helped confirm what was already being shown. ends September 30, 2013. The tables below list the markets
I encourage analysts to further study these and other tools used, number of forecasts made, and hits and misses. Targets
in addition to the HPTZ methodology given to see if the process were identified on the weekly, daily and hourly timeframes.
can be improved upon further. As targets are identified both above and below the market,
only those targets that are “activated” are included in the
Identifying High Probability Target Zones results. “Activation” occurs when the market makes a breakout
When we have completed the Passive TA, and performed of its current bounding supports/resistances and makes a move
an Active TA on the current wave structures, areas where we toward the target. See Chart 14 for examples of targets and
see high concentrations of tools are places of interest for HPTZ “activation”.
considerations.

PAGE 20 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 14: Arrow bases sit on significant technicals that need to be broken for a “breakout” of the current market to
occur and to activate targets above or below the market.

Table 1: Data Collection From Specific Instruments

July 2012- July 31st 2014 HPTZ Forecast SPX USD EUR/JPY EUR/USD VIX GOLD OIL
Total # of HPTZ$ Forecasted 117 132 119 106 74 96 80
# Targets Hit: Both Price & Time* 89 109 92 88 60 73 69
# Targets Hit: Price Level Only* 18 13 16 8 7 13 4
Total # All Targets Hit** 107 122 108 96 67 86 73
# Targets Missed 10 10 11 10 7 10 7

Table 2: Result Totals Across All Markets

Clarification On Hit Determination
July 2012–July 31st 2014 HPTZ Forecast Totals / % Both Price & Time means exactly that, and the market moved
Multiple Markets Combined Totals All Markets into the zone indicated. Price Only is counted if the price level
Total # of HPTZ$ Forecasted 724 was reached but just outside the time parameters given. This
# Targets Hit: Both Price & Time* 580 must be close, off by only a few bars to be counted as a hit for
# Targets Hit: Price Level Only* 79 price level but a “just miss” for time. While the market does
not technically land in the HPTZ exactly as identified, it does so
Total # All Targets Hit** 659
close enough that it would have been deemed “tradeable”.
# Targets Missed 65
Hit % : Both Price & Time Only* 80% Examples
Hit %: All Targets Hit** 91% All examples below except for the first are taken from
published calls as of the date given for the chart.
Through the examples, I will demonstrate the HPTZ
In our example above (Chart 14) the supports at target Z held and methodology and discuss trigger considerations as they pertain
the market lifted. Target X is not included in the results. We know to the Technical Trading Method.
the market lifts from the current location shown in the chart and
lands in A. Once the market pushes past the upper Bollinger Bands Example 1: Oil Continued
and the Fibonacci level sitting at the base of the green up arrow, Finishing off the oil example I have been using, here is the
target A becomes “activated” and target B is not included. The completed chart with all the tool sets applied.
market would need to break down through the 34ma, sitting at the Chart 15 is ahead in time from the original Chart 9, and we
base of the red down arrow, to activate B and include it in the results. can see the market spiked to the identified HPTZ at A.
(Note: Data has been updated from the original paper submitted The market never makes a move towards target B and it would
in 2013 and was current until August 2014) not be included in the calculations for hit/miss percentage.

IFTA.ORG PAGE 21
 IFTA JOURNAL 2015 EDITION

Chart 15: Our oil example with all tools applied. Image
captured further on in time than original Chart 9; note Example #2: US$
target at A was hit. The following examples are taken from published calls,
dates are shown.
Charts are divided into two identical charts (time periods)
with different tool sets. They are presented to the subscriber
base in this format for easier reading and identification of
tools and trigger considerations.
Chart 16: Market at A—where will the market move to
next? Lifting from the purple trend s/r and making new
highs would offer trigger considerations for more lift, with
the next targets above the market the green dashed trend
line and then the solid black. Dropping below the grey
support & resistance area that is supporting the market
offers a trigger consideration for a move to the lower target
at C. Targets D and E can be seen to be identified well ahead
of the market.
Chart 17 picks up the market for Friday, August 23.
Yellow and blue highlights show the market makes the most Resistance at A holds, and the market dropped through
of its move in the blue highlighted area, when the W%R is above C, landed on top of (and slightly inside) target D, and then
the -20 level. gradually made its way over to target E.
Several targets other than A and B have been added (green Arguments for either a lift or more down can be made from
boxes with a red border). They have been chosen because we can here… how do you know what the market will do?
see a high concentration of tools converging in the area(s). Technical triggers for consideration are the 13 and 34ma’s;
Although there are many tools and methods that can be top of the grey support/resistance zone; dropping below
used to make an educated guess as to where the market will the current green dashed trend support. Waiting for these
move next, there is no way to ever know with 100% certainty technicals to break to use as triggers allows you to enter the
which way the market will break. The tools we have set up are market with a target in mind (G above, F below), not caring
identifying significant supports and resistances, and we can use which direction the market breaks because the method allows
them as trigger considerations (when broken or respected) in a you to follow along with the market, not just guess at what it
trading plan/strategy. might do.

Chart 16: US$ Daily, Wednesday July 10. Note, target boxes hit prior to current market location (green and yellow
boxes); Letters mark target locations and reference through charts 17 and 18.

PAGE 22 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 17: Friday, August 23; Market drops to C, grazes D and lands in E. Where is it going next? Can you know for sure?

In our final chart for this example, Chart 18 shows the market lifting from E, moving through the significant technical trigger
considerations previously listed, and landing in target G.

Chart 18: Friday, September 6, 2013; market finds support at E and lifts to G.

Example 3: S&P
The series of charts that follows for Example 3 takes a look at the S&P market through the period from Tuesday, March 9, 2013, to
Thursday, May 9, 2013.

IFTA.ORG PAGE 23
 IFTA JOURNAL 2015 EDITION

Chart 19: Base of arrows sits on significant technicals for trigger considerations; arrows point to next likely market
support/resistance, targets.

Chart 20: Market lifts to A; technical trigger considerations adapt to new conditions (e.g., patterns, supports,
resistances); red arrow moved to new significant technical for trigger consideration.

PAGE
PAGE2424 IFTA.ORG
IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 21: Market drops from A but does not break the technical trigger for a down as marked on Chart 20; market lifts
and does break the technical trigger for a lift (Chart 20) and moves to C; it then drops and spikes to HPTZ-D; Fibonacci
circles can be seen added as a new experimental tool, unknown at this time if they are increasing or decreasing odds
(no noticeable change in statistics).

Chart 22: Lifting again, the market breaks the technical trigger consideration (Chart 21) and moves to E; dropping from
there, the market moves toward lower targets, finding support on the 34ma (pink moving average).

IFTA.ORG PAGE 25
 IFTA JOURNAL 2015 EDITION

Chart 23: Market broke the 34ma and dropped to G; bounced back up to F, continued to lift and now sits close to H;
where is the market going next? Can you ever know for sure? Arrows mark technical trigger considerations to follow,
regardless of where it moves.

Chart 24: Market moves over to touch H; lifts through technical trigger; new target at K added based on developing market.

PAGE 26 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 25: Market lifts over the next two days, landing in HPTZ –K.

The example demonstrates: the targets, they can be used as the backbone for a purely
1. The identification of targets at areas of technical technical trading plan or strategy.
convergence: note on the charts that the target locations While we would ultimately like a larger sampling size for data
contain several tools together in one location. collection, what has been currently observed for over a year now
2. The use of the technical tools as trigger considerations for merits further investigation.
market entries and exits. The overall hit/miss percentages remained consistent
3. The development of the process and method as the market throughout the data collection time period, regardless of market
moves forward: triggers are dynamic and relative to current conditions.
tools; new targets are identified as we analyse the market’s
movement in real time. Chart 26: While in an overall positive trend, different
market conditions exist: grey non-trending; red
4. Initial market movement from A to G was in a non-trending,
pullbacks and reversals. Various conditions existed
consolidating market. The lift from G to K is a 16-day for weeks at a time; hit/miss ratio remained consistent
uptrend: targets are hit regardless of market conditions. regardless of market condition.
5. Addition of “non-standard” tools (Fibonacci Circles): other
tools can also be added to the current methodology in the
attempt to increase odds.

Conclusion
Through the use of several common technical analysis
techniques, I have demonstrated that it is possible to locate
areas on the chart to where the market could move with some
degree of accuracy.
The locations identified are never a given, and multiple
potential target locations make it difficult to know which target
it will be, similar to trying to know which direction the market
is going to break. However, I have shown that the market will
move to an area of technical convergence.
Using the HPTZ methodology and setting up the charts
to find these areas identifies significant technical levels,
supports and resistances. These, in turn, can be used as trigger
considerations for market entry and exits; and combined with

IFTA.ORG PAGE 27
 IFTA JOURNAL 2015 EDITION

Chart 27-B: Market drops, lands in identified HPTZ.

July 2012 marks the time period when data collection began.
Although we can see a definite overall upward trend, different
market conditions existed throughout the lift.
As well, targets were identified on several timeframes, each
potentially having different market conditions relative to their
time period.
Given all the variable market conditions gone through in the
last year, and that the statistics for the hit/miss ratio remained
consistent throughout (i.e., no change during non-trend
periods), we can be confident that the ratios should continue to
remain consistent in future market conditions.
HPTZs and data collection are ongoing, and results are
updated at www.triggers.ca for future reference and to monitor
progress.

Recent Examples
The following charts are more recent examples of the HPTZ Recent Example 2: Oil, April 2014
methodology that have occurred since this paper was originally www.triggers.ca
written. Chart 28-A: Oil Daily. Consolidating between two grey
support and resistance zones. These offer technical
Note that these charts are from TradingView.com. They allow
triggerwww.triggers.ca
considerations for a break of the market in
you to publish an “idea”, whereby the chart is saved with your either direction. Targets above and below offer places to
technical analysis at the time. This published chart cannot be look once the market shows its direction.
altered in any way, except for the loading of new data to see how
your “idea” panned out.
The following charts follow this format, and as such, they
come in sets. The first image is the original published (saved)
idea; and the second image is what occurred. The green area
on the second chart identifies the new market data, or what
occurred, after the chart was published.

Recent Example 1: Gold, April 2014

www.triggers.ca
Chart 27-A: Gold 60 min. Potential targets above and
below market; technical trigger considerations could be
current channel(s) support for a drop, lift past current
consolidation high for a continuation of the trend.

Chart 28-B: Market breaks to the downside, several

technicals are moved through that could have offered
trigger considerations as the market fell to the HPTZ.

PAGE 28 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Recent Example 3: S&P, June 2014

www.triggers.ca Software and Data
All market charts used are courtesy of eSignal.com
Chart 29-A: S&P 500, 60 min. Having just moved out of a
HPTZ, two more—one below and one above—can be seen,
All 539 calls that were used in the data collection over the
as well as nearby technicals for trigger considerations,
www.triggers.ca
regardless of the direction of the next move. time period specified are available upon request.

Notes
1. Prechter, Robert R. and A.J. Frost, Elliott Wave Principle: Key to Market Behavior,
10th edn. (Gainesville, Georgia: New Classics Library, 2005)
2. Brown, C., Technical Analysis for the Trading Professional: Strategies and
Techniques for Today’s Turbulent Financial Markets (New York, NY: McGraw-Hill,
1999) pg. 85
3. Droke, C., Technical Analysis Simplified (Columbia, MD: Marketplace Books,
2000) pgs. 21, 67 & 73
4. Bollinger, J., Bollinger on Bollinger Bands (New York, NY: McGraw-Hill, 2002)
5. Prechter, Robert R. and A.J. Frost, Elliott Wave Principle: Key to Market Behavior,
10th edn. (Gainesville, Georgia: New Classics Library, 2005)
6. Neely, G., Mastering Elliott Wave (Brightwaters, NY: Windsor Books, 1990)
7. Investopedia.com, Definition of “Williams%R”, http://www.investopedia.com/
terms/w/williamsr.asp
8. Williams, L., How I Made One Million Dollars… Last Year…, 3rd edn.
(Brightwaters, NY: Windsor Books, 1979)

References
Barnes, R., Trading in Choppy Markets: Breakthrough Techniques for Exploiting
Chart 29-B: Market lifts through the blue dashed Non-Trending Markets (Chicago, IL: Times Mirror Higher Education Group, 1997)
Fibonacci level and solid red trendline to the next HPTZ, Bollinger, J., Bollinger on Bollinger Bands (New York, NY: McGraw-Hill, 2002)
breaks out of the solid black channel, and continues on Brown, C., Technical Analysis for the Trading Professional: Strategies and
to the second HPTZ. Techniques for Today’s Turbulent Financial Markets (New York, NY: McGraw-Hill,
1999)
Droke, C., Technical Analysis Simplified (Columbia, MD: Marketplace Books, 2000)
Elder, A., Trading for a Living (New York, NY: John Wiley & Sons, Inc., 1993)
Eng, W., The Day Traders’ Manual: Theory, Art, and Science of Profitable Short-Term
Investing (New York, NY: John Wiley & Sons, Inc., 1993)
Etzkorn, M., Trading with Oscillators: Pinpointing Market Extremes – Theory and
Practice (New York, NY: John Wiley & Sons, Inc., 1997)
Farley, A., The Master Swing Trader: Tools & Techniques to Profit from Outstanding
Short-Term Trading Opportunities (New York, NY: McGraw-Hill, 2001)
Mandelbrot, B., The (Mis)Behavior of Markets: A Fractal View of Risk, Ruin &
Reward (New York, NY: Basic Books, 2006)
Miner, R., Dynamic Trading: Dynamic Concepts in Time, Price and Pattern Analysis
with Practical Strategies for Traders and Investors (Tucson, AZ: Dynamic Traders
Group Inc.)
Neely, G., Mastering Elliott Wave (Brightwaters, NY: Windsor Books, 1990)
Pesavento, L., Fibonacci Ratios with Pattern Recognition (Greenville, SC: Traders
Press, Inc.)
Prechter, Robert R. and A.J. Frost, Elliott Wave Principle: Key to Market Behavior,
10th edn. (Gainesville, Georgia: New Classics Library, 2005)
Williams, L., How I Made One Million Dollars… Last Year…, 3rd edn. (Brightwaters,
NY: Windsor Books, 1979)

IFTA.ORG PAGE 29
 IFTA JOURNAL 2015 EDITION

Adam Ivan Michael Cox

[email protected]
Testing the Effectiveness of Multivariate PO Box 38-182

Wavelet De-noising for Intraday Trading Systems Howick 2144

Auckland
New Zealand
By Adam Cox, MFTA
649 5308-854

Abstract a. VIDYA—variable index dynamic average (Chande

Testing the effectiveness of multivariate wavelet de- 2001, pp. 53–56)
noising for intra-day trading systems; A New Zealand dollar b. Butterworth Filters (Ehlers 2004, p. 191)
study studies the effects that various multivariate wavelet c. TEMA–triple exponential moving average (Malloy 1994)
denoising schemes have on trading system using popular
technical analysis trading rules. The NZDUSD intraday 2. Examples of other technical indicators employing de-
currency returns sampled at 10 minute monotonic intervals noising concepts include:
were selected for this research. a. RAVI–range activation verification index (Chande
Technical trading rules assessed include: Elders 2001, p. 7)
Impulse Repose system, Momentum crossover rule and b. TRIX (Hutson 1982 )
a moving average rule. Each trading rule was assessed c. KST–know sure thing (Pring 1993).
against unprocessed data, a univariate wavelet denoising,
technique as well as two multivariate wavelet denoising The premise of utilising de-noised time series in trading
techniques. The univariate denoising case, selected Symlet applications lies in the fact that the de-noised series allows the
12, with denoising applied at level five (hard) thresholding. trader to identify sustained directional trends by mitigating
Multivariate denoising included Symlet 12, with denoising the effects of short-term noise. Positive expectations of a
applied at level 5 (hard) thresholding. Both five and trading system lie in a system’s ability to distinguish between
six principal component examples were applied in the noisy and trending states (Chande 2001, p. 43).
multivariate examples. Each wavelet denoising technique De-noising, however, may often be confused with time
applied symmetrical padding. series smoothing. A de-noised time series does not need to
The research methodology applied unit root testing, tests look smooth. De-noising in this research proposal pertains
for heteroskedasticity, cointegration and granger causality to the removal of data points of high frequency in order to
to identify the data generation process of the NZDUSD 10 deal with the effects of heteroskedasticity (Barclay et al.
exchange, as well as confirm the effects (if any) on trading 1997). Heteroskedasticity, or time-varying volatility, is a
system results by applying each of the denoising techniques characteristic of financial time series data, particularly at
selected in this study. higher frequencies (Batten and Hogan 2001). De-noising
System testing results include a wide range of trading techniques using averaging and filtering methodologies can be
system assessment techniques and measures including: problematic for traders because smoothed time series often
expectations analysis, drawdown analysis, the probability lag the underlying price time series (Ehlers 2004, pp. 13–15).
of non-randomness of system returns, and the relationship This research examines whether trading system results
between drawdown values and trade entry. This analysis were enhanced by first de-noising intraday New Zealand
was complemented by trading system metric reports exchange rate data by the multivariate wavelet smoothing
generated from TradeStation 2000. methodology.
The research results were mixed. Multivariate denoising Wavelet algorithm design used in online or trading
did not always improve system performance, particularly applications must, however, ensure that data is generated
where trade entry timing was delayed, particularly where in a strictly causal basis. Smoothing algorithms such
a ‘filter’ device had already been applied within the trading as splines and weighted moving averages that employ
rule. The results of trading techniques which demonstrated forward data points cannot be utilised in real-time trading
a negative overall expectation were not greatly improved by applications. A scrolling window (See Figure 1) approach
denoising alone. has subsequently been applied, but by doing so, edge
effects can be introduced (Misiti et al. 2012). The window
Introduction is 1,000 data points wide (n), and scrolls through the time
Various technical analysis indicators employ data smoothing series monotonically (x) from left to right. No forward
or utilise smoothed price data. information can be applied, and only points from Tx-n to Tx
are applied.
1. Many examples exist of using various smoothed or filtered
versions of the underlying security price time series,
including:

PAGE 30 IFTA.ORG
forward data points cannot be utilised in real-time trading applications. A scrolling window (Refer Fig 1)
IFTA JOURNAL
approach has subsequently been applied, but by doing so edge effects can be introduced (Misiti et al, 2012). 2015 EDITION
The window is 1,000 data points wide (n), and scrolls through the time series monotonically (x) from left to
right. No forward information can be applied, and only points from Tx-n to Tx are applied.
Figure 1. Rolling window
i. Shrink the values (soft thresholding method), or
ii. Set them equal to zero (hard thresholding).
Step 3: Join the modified high-pass portion to the original low-
pass filtered version, creating the modified vector z.
Step 4: Compute i iterations of the inverse wavelet transform
on z to obtain u, thus obtaining the original but de-
noised version of the signal.

Thresholding
Figure 1. Rolling window. For a threshold level λ (lambda) applied to x of d, two basic
Only the right-hand side of the data window can be utilised types of thresholding are available: hard and soft.
as updated price data enters the equation. A method such as
3|P a g e
symmetric padding may be utilised to extend the width of the Hard Thresholding
data window used by the wavelet algorithm inResearch
MFTA order toAdam
mitigate If |d|< λ, reduce d to 0.
Cox (2014) - Abbreviated
right-hand side edge effects. A right-hand side extension If |d|> λ, keep d.
utilising symmetric padding was applied in this research.
An abbreviated version of the original research paper is Soft Thresholding
presented herein. For brevity’s sake, appendices and associated If |d|< λ, reduce d to 0.
tables have been removed. However, references have been left If |d|> λ, shrink or reduce d by λ, in accordance with a
intact to assist any Q&A sessions following from this research. shrinking/threshold method selected by the user.

Wavelets In this research, the heurusure1 method was used to apply

soft thresholding.
Why Wavelets?
Wavelets have a possible useful part to play in trading Multivariate Wavelet Filtering
applications because of their ability to provide multi-resolution Conceptually, the multivariate wavelet de-noising applies
analysis for time-based signals. Application of the Fast Fourier an univariate wavelet approach, but takes into consideration
Transform (FFT) infers a time invariant deterministic signal. the correlation structure between the variables. The principal
Once cycle and magnitude of the single are identified, it may be components technique is applied in accordance with the
decomposed ( + and – infinity) into a series of Sine and Cosines number of principal components defined by the user, noting
(Hamilton 1994, p. 152), thus allowing future states of the signal that the number of principal components must lie in the range
to be predicted. of 1 >= x < number of data series (e.g., six data series may have
Many signals, such as financial time series, are characterised no more than six principal components and no fewer than 1). In
by heteroskedasticity. In short, these data may not be this way, detailed data to be filtered may initially be modified
successfully represented in the frequency domain by the to take into account the noise correlation matrix provided
combination of Sines and Cosines. The Short Term Fourier by principal component analysis. The smaller the number of
Transform (STFT) deals with this issue by applying a sliding principal components specified, the more noise removed.
window approach. However, this approach is based on the
notion of time invariance. In the context of financial data, the Research Methodology
question arises “what window length should be selected?” Use of
a window may provide perfect frequency domain representation Overview
but poor temporal information (Polikar 2006). A narrow window Research methodology aims to identify differences
provides good time resolution but poor frequency resolution, between the trading system candidates using raw prices and
and a wide window provides the opposite effect. Heisenberg’s identical systems using price time series de-noised by the
uncertainty principle applies equally to signal analysis. multivariate wavelet transformation. Specifically:
Wavelets address the time vs. frequency dichotomy, which a. Understand the data generation process (dgp) of each
is particularly important in context to non-stationary financial currency along with summary statistics to identify
time series data characterised by heteroskedasticity. how benefits (if any) may arise from applying wavelet
smoothing algorithms in a trading system context.
De-noising Testing includes non-stationarity (unit root testing),
At a conceptual level, wavelet de-noising entails a three-step serial autocorrelation, runs (non-randomness), and
approach. For a vector y: heteroskedasticity hypothesis testing.
b. Identify whether lead-lag relationship exists between
Step 1: Compute i iterations of the wavelet transform on y, various currency pairs expressed in New Zealand terms
obtain the new vector z, made of low pass filtered against the New Zealand dollar, that is, do various
portion l, and a high-pass filtered portion d. currencies granger-cause the New Zealand dollar or do
Step 2: Apply threshold rule to the high-pass filtered portion d these currencies respond to common events and the
of z, either: inclusion of a multivariate algorithm may be limited.

IFTA.ORG PAGE 31
 IFTA JOURNAL 2015 EDITION

Figure 2. Currencies and currency codes

c. Identify if the same currency pairs are cointegrated with Symbol Commodity Currency Terms Currency
the New Zealand dollar to identify whether any long-
NZD/USD New Zealand Dollar United States Dollar
term relationships exist between these pairs. Whilst a
correlation matrix has been produced to provide insight NZD/AUD New Zealand Dollar Australian Dollar
into correlation between currency pairs, cointegration NZD/EUR New Zealand Dollar Euro
testing can confirm whether currencies are held together in NZD/GBP New Zealand Dollar British Pound
longer term relationships, and if so, benefits may arise by NZD/JPY New Zealand Dollar Japanese Yen
including such currencies because of this factor. Moreover,
whilst cointegration is not a necessary condition for NZD/CAD New Zealand Dollar Canadian Dollar
granger-causality, general causality exists in a cointegrated
relation. These data are similar to that of the constituents of the New
d. Understand the data generation process, serial Zealand Trade Weighted Index (TWI);4 however, in this research
autocorrelation, heteroskedasticity, variance and cross- paper, NZD/CAD exchange rate has been included. Currencies
correlation of the de-noised New Zealand time series using selected are free floating, and each respective country’s reserve
a sample of wavelet de-noising methods. If currency pairs bank operates open market operations in a not dissimilar
do not granger-cause the New Zealand dollar, and the de- manner. Although the Japanese Yen is characterised in
noised time series does not display a causal relationship numerous interventions by Japanese Ministry of Finance, the
that leads the New Zealand dollar in the cross-correlation Japanese Yen cross-rate has been retained, as it is a constituent
plot, benefits arising from applying multivariate methods of the NZD TWI. The Chinese Yuan has been excluded on the
may be conjectured as arising from the benefits of de- basis that it is not allowed to trade freely (including the dual
noising (subject to lag) alone. Yuan system).
e. Identify trading measures that may identify benefits
(if any) of applying the multivariate wavelet denoising Swap Points
method. Apply stationary measures to assess comparative Examining the effect of swap points on trading results, swap
trade performance. Where data samples are sufficiently points published by Dukascopy were reviewed. Because the
large, such statistics would asymptotically converge to the effect is assumed small in context of the number of data points
appropriate distribution where means and variance do not used in trading (63,004, noting that a scrolling window of 1,000
depend the temporal element, but on sample size. data points was used for wavelet de-noising purposes), and
i. Stop loss policy has been homogenously applied to each substantial opportunity was provided by stochastic trends in
trading system. Hence, trade entry efficiency and the the data, swap points were not calculated. Rather, capital gains
relationship between Mean Adverse Excursion and or losses were the sole contributor of system performance.
Closed Trade Drawdown are compared for each trading
system and data enumeration to confirm or reject the TradeStation Data
null hypothesis. Mean Favourable Excursion, being a Exchange rate data was multiplied by 10,000, with 0.01 as
function of Stop Loss policy does not form the basis of the minimum movement and same value. CSV files post-data
hypothesis testing. validation were imported into TradeStation 2000, including data
ii. Summary statistics, including Number of Trades, Entry files related to de-noised datasets. A single USD unit is purchase
Efficiency etc., are used to supplement conclusions or sold in USD terms against the NZD. Profits and losses are
arrived at from hypothesis testing, and used to therefore stated in USD. Four data series were therefore
conjecture the reason for differences or similarities in required for each de-noised trading system, including:
trading results.
f. Identify each hypothesis to be tested, and confirm testing a. NZD/USD mid-closing data
results. b. NZD/USD bid closing data
g. Confirm conclusions complemented by TradeStation2 c. NZD/USD offer closing data
reports and equity curve analysis, as outlined above, d. NZD/USD de-noised mid-closing data
to understand benefits (if any) of applying multivariate
wavelet de-noising in respect to trading system candidates The rationale of applying mid prices and the uses of two-way
and the New Zealand Dollar (NZD/USD). pricing are explained in Data Analysis.

Data Selection
The analysis is based on 10-minute monotonic time series
closing spot currency data, as provided by Dukascopy.3 Input
data used in the model is based on NZD-based cross-rates. This
ensures that all data are stated on a homogenous basis (i.e., the
commodity currency is the NZD (Bin 2011). These data include:

PAGE 32 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Software
Figure 3 shows the Software utilised in this research proposal
Figure 3. Software and functions applied in this research.
Software Module ƒƒApplication
Matlab Applied Econometrics using ƒƒCo-integration tests the Augmented Dickey Fuller two step testing procedure (Alexander 329),
Matlab, James LeSage.
Department of Economics
ƒƒGranger causality testing via the Johansen Procedure for multivariate applications (op cit, pp 357-358).
University of Toledo 1999.
Matlab Wavelet ƒƒDe-noised time series data is created using Matlab’s1 multivariate wavelet smoothing function wmulden function which applies
toolbox, Mathworks. Principal Components in a multivariate wavelet denoising function,
ƒƒUnivariate denoising is provided by the wpdencmp wavelet packet denoising function.
Statistics Toolbox ƒƒTime series analysis statistics including: Augmented Dickey Fuller (Matlab’s adftest function) statistic in relation to unit root
testing (LeSage, 1999),
ƒƒ LBQ Test to test serial autocorrelation (Matlab’s lbqtest),
ƒƒRuns test to validate data series randomness (Matlab’s runtest),
ƒƒAutocorrelation, partial autocorrelation (Matlab autocorr and parcorr functions),
ƒƒCross-correlation (Matlab’s crosscorr function) and Quintile-Quintile plot (Matlab’s QQPlot function),
ƒƒThe Jarque Berra (Matlab’s jbtest function) test confirmation normality of distributions,
ƒƒPrincipal component analysis and principal component correlation matrix to conjecture the number of principal components the
multivariate wavelet denoising function to be utilised (Matlab’s princomp function),
ƒƒA simply spectral plot (Matlab’s periodogram function) is used to confirm, at a general level, if any currency series are periodic. If
not, then benefits of trading system adopting the wavelet de-noised data, may therefore originate from denoising and/or lead-lag
effects as confirmed by any granger causality and/or co-integrated relationships (if any) with the New Zealand Dollar.
NB. Matlab Functions used to create scrolled de-noised data in batches were designed and programmed by the author.
Microsoft ƒƒAnalysis trading data provided by TradeStation,
Excel 2010 ƒƒApply F-Test to confirm whether variances of the data sets are identical and confirm which T-Test must be appropriately applied
(Levine et al 2002, pp 350-352),
ƒƒApply Student’s T-Test to confirm whether the mean of the samples are statistically different or identical (Ibid, pp 385-387),
ƒƒ Confirm the regression slope of two regressions (Maximum Adverse Excursion v. Closed Trade Drawdown) are statistical similar or not.
ƒƒA range of basic summary statistics including: arithmetic mean, sample standard deviation, coefficient of variation, skew,
kurtosis, minimum and maximum,
ƒƒProduce a correlation matrix of currency returns and as well as correlation matrices of equity curves for each trading system,
ƒƒCreation of Visual Basic for Applications to interface wavelet smoothing functions from Excel to Matlab and visa-versa.
NB. MS Excel analysis sheets were designed by the author.
Trade Station ƒƒDesign trading system candidates, TradeStation Code to export trading data (Stridsman, 2000 pp 126-129),
2000i2 ƒƒImport Bid, Offer, Mid and Wavelet smoothed data (data series #4 in TradeStation) series for each trading system to apply to that
trading system uses bid and offer prices to execute trades, and mid prices in respect to indicator calculations ,
ƒƒThe SafeZone stop loss technique (Elder 2002, pp 173-180), is employed for each trading system enumeration,
ƒƒAnalysis carried out in MS Excel and was complemented by TradeStation Performance Reports.
NB. Trading Systems were written by the author.

Wavelet Selection
A number of mother wavelets are available; specifically, in this analysis, the following mother wavelets were considered by first
exploring univariate examples using sampled NZD/USD data. A univariate wavelet has been selected in order to provide the initial
base case by which multivariate examples could be assessed against.
This is a heuristic approach that aims to review (by visual inspection) the level of timeliness and denoising characteristics of
various wavelet types. Wavelet selection methods may be the subject of another research exercise. The following wavelet types were
initially examined.
Figure 4. Mother wavelets considered in this research
Symlet(Sym) Sym 4 Sym 8 Sym 12

Coiflets Coiflets 1 Coiflets 5

Daubechies(Db) Db4 Db8 Db12

IFTA.ORG PAGE 33
 IFTA JOURNAL 2015 EDITION

Wavelets Selected for Research

The following wavelets were selected and applied in this research paper:

Figure 5. Wavelet settings used in this research paper

Wavelet Window De-noising Thresh- Extension Extension Principle
Wavelet Type Level olding Type Length Components
Univariate Sym 12 Level 5 (Sym125) 1000 5 Soft Symmetrical 500 N/a
Multivariate Sym 12 Level 3,6,6 PC (Sym125_66PC) 1000 3 Soft Symmetrical 500 6
Multivariate Sym 12 Level 5,5,5 PC (Sym125_55PC) 1000 3 Soft Symmetrical 500 6

The multivariate wavelets selected provide the basis of hypothesis testing as outlined in the research, whilst the univariate wavelet
analysis provided the base case by which complexity and possible utility of utilising multiple data may be gauged.

Multivariate Wavelet Examples

Each of the wavelet families were visually inspected at various levels using all principal components to understand the
characteristics of each.

Figure 6. Comparison of multivariate wavelet examples and each Mother Wavelet

Data Analysis
Data Analysis Results
Bid-Offer Spread Analysis
Bid-offer spread was examined for the following currency included in this research (refer to Data Selection). Mean bid-offer
spread and standardisation of spread was examined on a day of week basis to understand both the relative quantum of spread, but
also the day-of-week effect. Data analysed in this exercise included 262,479 10-minute monotonic data points for each currency pair,
as provided by Dukascopy. In each case, the period of assessment was from 7/10/2005 5:50:00 to 3/10/2012 0:00:00. This sample
assumed to be representative of the sample of data used in this research.
Figure 7. Bid-Offer (pips) mean and standard deviation of bid-offer spread

PAGE 34 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

The pattern of bid-offer spread and standard deviation over the course of the trading week is roughly similar for all currencies.
Whilst de-noising could be applied to separate bid-offer series for each currency pair, the complexity of doing so is beyond the scope
of this research. Bid-offer spread histograms and spread quintiles stated as a percentage of mid-prices is set out in Figure 8. Bid-offer
spread lies in the range of 0.0055–0.0065 of each currency’s mid-price.

Figure 8. Histogram and cumulative probability of bid-offer spread as a percentage of mid-price for each currency
used in this research

As a compromise, mid prices have been selected as the basis by which research has been applied, including the application of
denoised time series, whilst two-way pricing has been retained for trading purposes.

Trading Volume Analysis

Liquidity patterns throughout the trading week, as indicated by the mean and standard deviation of trading volume, are similar for
each currency pair. Trading volume, whilst a proxy of true currency volume, is consistent with research by Dacorogna et al. (2001), who
As a compromise, mid prices have been selected as the basis by which research has been applied, including the application of de-noised
time series, whilst two-way pricing has been retained for trading purposes. identified weekly ‘seasonalites’, as well as research results
set out by Aldridge (2010).
Figure 9. Mean volume and standard deviation of each currency by day of week

IFTA.ORG PAGE 35
 IFTA JOURNAL 2015 EDITION

As a consequence of this analysis, trading system design has implemented various trade entry time and mandatory exit rules (refer
to Trading System Selection).

Summary Statistics (Mid price close data)

Summary and other statistics are prepared on mid-rate closing prices for each currency:

Figure 10. Summary statistics of each currency

Time Series Tests

Figure 11 sets out the hypothesis summary.
Figure 11. Hypothesis results carried out on unprocessed mid-rate close data
Test of Normality The Jarque-Bera tests complemented by the Quintile-Quintile Plots (refer Fig. 16) confirm a lack of normality of each of each of the
exchange rate used in this research. Significant excess kurtosis (substantially greater than a figure of 3 as for a normal distribution)
is also indicated. NZD/JPY kurtosis was re-examined with a larger data set of 264,000 data points which confirmed similar findings.
Standard errors of kurtosis and skewness have not been computed due to difficulty of accurate assessment in absence of normality.
Excess kurtosis signifies the probability of large negative or positive values is greater than under the corresponding normal density
function. For trading, this provide trading opportunities, but may also negatively impact trading by producing whipsaws, or large start
trade drawdowns if trade entry efficiency is not high, or stop loss levels are tightly set.
The lack of normality, and kurtosis in intraday financial data has also been identified by Alexander (1993, p 287), which examines
different data sampling times for USD/DEM data. She explains, “This is one of the stylised facts of high frequency financial returns,
which is particularly pronounced in foreign exchange markets” (ibid).
Each of the samples displays skew greater than zero, with the NZD/JPY, NZD/CAD, NZD/EUR having negative skew and remaining
currencies skewed to the right (or positive skew). Summary data results confer general findings set-out by Aldridge (2010, p 94) who finds
excess kurtosis and negative skew in 5 minute, 15 minute and 1 hour spot EUR/USD exchange rate data.
Stationarity Testing Exchange rate data was tested for the presence of a unit root at levels (exchange rates), and in accordance with the Augmented Dickey
Fuller (ADF) test statistic was non-stationary ( refer Table 1.1).
Stationarity was achieved by way of first differences (LN returns) of the data. Whilst the test was not ‘augmented’ in respect to
heteroskedasticity, the test results are expected for financial time series data and infer that statistical analysis including regression
analysis should be conducted on first differenced data. That is to say that the hypothesis of no unit root when tested on first differenced
data was not rejected and confirmed that such transformed data was stationary. As a consequence statistics in this research is based
on return data and not on exchange rates at levels, but it has been assumed the data is integrated order 1 that is I(1) and not I(2). Second
differences have not been employed in this research.

PAGE 36 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Tests for Randomness Runs Test indicated that runs above or below zero1 (left and right tails) are greater than expected in a random series for each exchange
rate (refer Table 1.2).
Whilst the runs tests conducted in this research have been carried out at a basic specification, from a technical analysis perceptive, the
presence of runs indicates that trading opportunities such as those identified by momentum or trend trading systems have validity. Too
many runs may, however, identify oscillating behaviour, however, in accordance with the periodograms; deterministic periodicity in the
frequency domain has not been identified.
Serial Correlation Each of the time series indicated the presence of serial autocorrelation both in respect to correlograms (ACF and PACF plots set out in
Figure 10) as well as the Ljung Box Q-Test results on LN returns (Table 1.3) and has significance for technical trading methods employing
short term momentum e.g. Elder’s Impulse System.
Heteroskedasticity Heteroskedasticity of time varying volatility was present in each of the time series tested as confirmed by Engle’s ARCH tests (Table 1.6),
which rejected the null hypothesis on no arch effects. The Ljung Box Q-Test on Squared Returns (Table 1.5) further confirms with the
null hypothesis on autocorrelation rejected for squared returns. Autocorrelation plot on squared returns (Table 1.4) also confirms which
significant ARCH effects being noticeable by the persistence of autocorrelation of squared returns. The use of wavelets may therefore lie
in their ability to denoising time series, such as short term exchange rate data with heteroskedastic characteristics.
Periodicity Inspection of the periodogram for each currency indicates that exchange rate is not periodic i.e. a Nyquist frequency is not noted. (Figure.
10) As a consequence, the wavelet transform may have real utility in context of denoising non-stationary, heteroskedastic non-periodic
data such as 10 minute exchange rate data tested in this research as opposed to exploiting periodicity. The lack of genuine periodicity has
been experienced by other trades, Elder (1993, p 124) stating, “Trouble is, cycles keep changing their length and disappearing”, he goes on
to say “…shows that noise is greater than cycle amplitude most of the time”.
Cointegration Whilst currencies were found to be highly correlated (refer Correlation Matrix Fig 19.) with the New Zealand dollar, particularly
currencies other than the NZD/AUD which exceed 65% correlation (in the sample period examined), cointegration was not clearly
demonstrated however. Whilst cointegration at 95% confidence level was signalled for NZD/GBP–NZD/USD and NZD/AUD currencies
pairs, the results were reversed at 99% confidence level. As such, the conclusion is that Null hypothesis of no cointegration cannot be
accepted, rather than outright rejection. In all cases cointegration tests were conducted on levels (refer Tables 1.7 and 1.8)
Whilst the lack of cointegration in research examines triangular relationships between currency pairs and the associated cross-rates to
determine if arbitrage or trading opportunities may exist (Wong & Ling), cointegration testing in respect to this research was carried out
simply to identify how, if any, advantages arising from the wavelet approach may have arisen, i.e. were various currencies cointegrated
with the NZD/USD at the 10 minute interval?
Granger Causality Whilst Granger-causality or lead-lag relationship between various currency pairs and the USD/NZD would arise if cointegration was
demonstrated, cointegration is not a necessary condition. Granger-causality testing conducted in this research to ascertain if any
currency leads the NZD/USD? If so, it may be conjectured, in the absence of a forecast model, that any advantages provided by trading
systems using data de-noised by the multivariate wavelet denoising function, may have arisen from the lead-lag relationships present
in the multivariate dataset. Granger causality testing was conducted on pairwise basis using up to 3 lags (refer Table 1.9) However, the
granger causality tests only indicated that:
1. A lead-lag relationship from NZD/EUR, NZD/CAD and NZD/AUD to the NZD/USD is not evident at lags 1 and 3.
2. However, a unidirectional lead relationship is noted from NZD/GBP to NZD/USD (but at only 90% confidence interval) using up to three lags.
3. Bidirectional or contemporaneous relationship is noted between NZD/GBP and NZD/JPY, however, a stronger unilateral lead
relationship from NZD/EUR to NZD/AUD.

Trading System Selection loss or profit-taking policies. For example, complexity grows at
the rate of n*(n-1) / 2 5 bidirectional relationships or n!/r! * (n-r)!
Introduction combinations without repetition. So, for a total of seven rules
Three trading systems were selected in order to test in combinations of three rules, 35 different combinations are
whether differences exist between each trading system using available altogether that would require analysing. Furthermore,
raw exchange rate data (NZD/USD mid prices) and alternative simplicity also ensures avoidance of multicollinearity, which
trading systems using multivariate wavelet de-noising. A in context of technical analysis, Edwards and Magee6 (2001,
univariate wavelet de-noising method was also selected in order Glossary) explains as, “…the incorrect procedure of using
to compare multivariate results with those of the multivariate identical data to supply different indicators...”
cases to determine the utility (if any) of applying the more
complex multivariate approach. System Descriptions
Trading systems were selected on the basis of providing
the most comparable challenge to the de-noised system Momentum 100 system
alternatives. The trading systems include: The momentum effect has been widely researched and
documented in both the international, equity, and currency
a. A 100-day momentum system markets by both technicians (Pring) and academics alike.
b. A moving average cross-over system For example, Rouwenhorst tested momentum strategies
c. Elder’s Impulse System in 12 European markets using data from 1980 to 1995 and
concluded that momentum returns were present in every
These systems were selected on the basis of providing country. Okunev (2003 ) found that momentum effects were
the most challenge to the de-noised alternatives. Moreover, found in currencies, whilst Fama (1993) refers to momentum
trading rule simplification and use of an homogenous stop loss as “the premier unexplained anomaly”. In other words, the
policy allowed comparative system analysis to clearly identify success of momentum-based investing is regarded by many
benefits (if any) from applying the multivariate system without as an exception to the efficient market hypothesis, whereas
being confounded by rule complexity or heterogeneous stop the principle of momentum is fundamental to many technical

IFTA.ORG PAGE 37
 IFTA JOURNAL 2015 EDITION

analysis-based methodologies. Bernstein (2001, p. 17) stated expanded to include several moving averages and/or respective
“It provides you an idea of trend strength”. “Momentum is a cross-overs (e.g., the Bill William’s Alligator System) (Williams
measure of internal strength or weakness of a market”, ibid, p. 1995). Guppy elaborates on the moving average system by
43. In addition, texts such as Pring (1993, p. 55) debate use of introducing the GMMA, or Guppy Multiple Moving Average
arithmetic momentum vs. rate of change in respect to long-term System (Guppy 2004) (also known as ribbon averages), whilst
inter-security analysis; Pring stated “The division method is Chande (2000, p. 71) explains, “the moving average cross-over
far superior since proportional price moves are represented by system (MAXO) system is the simplest trend-following strategy,
an equal risk or fall in the oscillator regardless of actual level of but it can also be an anti-trend strategy”. Elder (2002, p. 94)
price”. Momentum has therefore been selected in this research recommends, “Use a longer term EMA to indicate trend, and the
as a comparative base case for hypothesis testing. However, in shorter to find entry points”.
this analysis, inter-currency analysis has not been undertaken, As a consequence, a moving average cross-over system has
and the “subtraction method” (ibid) is therefore applied. been included in this analysis and complements the momentum
Momentum, even at longer periods, is subject to the effects 100 system. Where the momentum 100 system mitigates
of heteroskedasticity, and often, momentum-based systems whipsaw trades or false positive trades by requiring no less
employ additional methods to complement the initial long than two consecutive crossovers (Stridsman 2000, p. 274),
or short signals provided by the momentum indicator. Such the moving average crossover system applies a longer term
methods may include any or all of the following: use another exponential moving average and a shorter term simple moving
indicator to confirm the trade setup, use an indicator or method to capture longer term trends.
to identify a price pull-back to identify the trigger for the A minimum number of consecutive cross-overs or cross-
trade and as a filter (e.g., Elder’s Force Index) (Elder 1993, pp. unders are not featured in this system, but moving averages
229–234), or validate a trading setup by enforcing minimum provide unprocessed price data and the benefits of a low-pass
consecutive cross-over periods before a trading signal is filter, which may mitigate the effects of heteroskedasticity,
provided. albeit lag (Ehlers 2004, pp. 13–15). Whilst Elder asserts that a
To provide the most challenge in a comparative sense to the moving average should be applied in respect to its lag, (e.g., a
de-noised enumerations being examined, as well as to allow 10-period SMA should be lagged and plotted under the fifth and
benefits of the de-noised system, the following settings were sixth bar (Elder 1993, p. 127), lags are not applied to the moving
used: averages employed in the cross-over system in this research.
Momentum length. Momentum length was set to 100 Two arbitrary moving average lengths were selected.
periods. This would allow the non-de-noised series to identify The momentum system used 100 periods as the basis of its
larger trends and provide additional challenge to the de- calculation. Therefore, in order to provide additional challenge,
noised system alternatives. but also insight in comparative system performance, the
Consecutive cross-over rule. To mitigate false positive moving average cross-over system employs:
signals (Type 1 error) two consecutive cross-overs or cross-
unders of the zero line are required by the system before a Long-term average. A 200 (10-minute) period exponential
trading signal is generated. This policy favours the de-noised period was used. This provides a longer look-back window
series and is aimed at challenging the de-noised alternatives. than that of the momentum 100 system, but utilising an
This type of rule has been adopted by Chande (2000), pp. exponential moving average to ensure signal timeliness
121–134 in the 65-EMA 3CC system, which requires three as an exponential moving average assigns more weight to
consecutive closes under or over the 65-day simple moving incoming prices while exponentially decaying prior average
average before a trade is triggered and explains “...then the values (Elder 2000, p. 90).
number of closes you will use will act like a filter in reducing Short-term average. A 28 (10 minute) period simple
the number of trades” (ibid, p. 122). Whilst this method may moving average has been arbitrarily selected. The shorter
add delay into the momentum system, Bernstein (2001, p. averaging period ensures response timeliness. Hence, a
45) asserts, “… momentum tends to lead price most of the simple moving average has been selected rather than an
time”. Therefore, it is assumed that any lag created by the exponential average. The simple moving average acts as a
cross-over rule would be offset by the leading nature of finite response filter having lag equivalent to half the size
the momentum indicator. Elder’s Force Index has not been of its window (Ehlers, 2004, p. 47). An exponential moving
applied in this research because the Force Index requires average of alpha = 1/(L+1) would have a lag in simple moving
access to volume data. A single definitive source of volume average terms of 2/(P+1), where P is the period used for the
data, unlike that provided by a stock exchange, is not simple moving average (Hutson 1984).
available in the spot exchange rate market. Often, volume
data provided relates only to a specific data vendor or venue. Elder’s Impulse System
The momentum 100 system and the moving average
Moving Average Cross-Over System cross-over system are aimed at identifying and subsequently
The moving average cross-over system appears frequently trading persistent trends in the NZD/USD. As identified in Data
in trading systems and technical texts. Terminology such as, Analysis, currency rates examined showed substantial runs (the
golden cross-over or bull cross-over, have become mainstream Runs Test) and trend-like behaviour (serial autocorrelation),
technical analysis terminology. Cross-over systems have been but are expected to be subject to potential false positive trades

PAGE 38 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

and suffer from the effects of whipsaws due to the high level of d. De-noised time series are imported into TradeStation as a
heteroskedasticity. Hence, differential results between the de- fourth data series. Indicators such as momentum or MACD
noised vs. raw exchange rates could be driven from the ability are subsequently based on this smoothed data series.
of each system to identify trend commencement whilst dealing
effectively with heteroskedasticity. Limit orders have been selected in order to avoid
To provide additional challenge, Elder’s Impulse System has TradeStation’s bouncing tick rule. The parameter for this rule
been applied. Elder describes the Impulse System in context has also been set to 0% in TradeStation. The purpose of this
of a momentum, or impulse trading system (Elder 2002, pp. analysis is to provide measures by which comparative system
157–158). Adopting the shorter term exponential moving performance can be gauged in accordance with hypothesis
average (refer below to Impulse System settings), the Impulse tests set out in this research. The motivation of using separate
System becomes a swing trading-like approach, given the extent bid and offer prices follows the analysis of bid offer prices and
of heteroskedasticity (refer to Data Analysis) measured in the associated standard deviation (refer to Data Analysis).
NZD/USD exchange rate. Testing of shorter term trade settings,
as employed by Elder’s Impulse System, complements medium Entry and Exit Times
and longer term trading systems such as the momentum 100 The following time-based rules are applied for each trading
and moving average cross-over systems because shorter term system and data enumeration in accordance with bid-offer and
systems tend to be more susceptible to heteroskedasticity volume data analysis:
(noise). However, this research does not assume that de-noising Monday entries. Trades may only occur post 7:00 GMT time
would be beneficial to all trading methods. If lag is introduced in (7 pm or 8 pm, depending on daylight saving).
context of a shorter term trading system, trading signals may Friday exits. Any open trades as of 19:30 GMT time each
be become mistimed, with overall system performance being Friday are exited in accordance with bid offer rules stated above.
degraded. The motivation for these rules follows trading volume
Settings employed in this system include those set out in analysis of the currency pairs (refer to Data Analysis) and
Elder (2001, p. 159). Optimal length has not been selected, nor ensures that trading system results are generated at the most
has a triple screen approach, as recommended by Elder, been liquid times during the trading week.
adopted. For example, Elder recommends, “Before you rush off
to apply the Impulse System, … Remember how Triple Screen Trading System Exits
analyses markets in more than one time frame” (ibid, p. 158). A stop loss policy is used homogenously for all trading
systems and is the primary exit method unless a reverse
The Impulse System settings include the following: signal is provided by a system. A common methodology has
been utilised so that one system may not be advantaged or
MACD histogram. MACD histogram employs the following settings: disadvantaged by a stop loss methodology uniquely applied
to it. For example, Guppy (1999, p. 152) states, “Traders who
a. MACD short moving average–12 periods use moving averages for entry do not often use them as exit
(exponential average) signals”.
b. MACD long moving average–26 periods The stop loss method is the SafeZone stop loss (Elder 2002,
(exponential average) pp. 173–180). A setting of 10 periods and three standard
c. Signal line (MACD histogram)–9 periods deviations has been applied consistent with Elder’s settings. The
(exponential average) stop loss is based on mid-price raw data.

Moving average. A 13-period exponential moving average Trading System Hypothesis Testing
was employed as per Elder.
Hypothesis
Elder’s Impulse System has also been selected because it was This research proposal is not aimed at creating great trading
based on closing prices like momentum and the moving average systems or determining whether a good trading system can
system, thereby providing a homogenous base by which the be created using de-noised data provided by the multivariate
different trading system results can be compared. wavelet algorithm. Rather, in respect to trading systems applied
to intraday New Zealand dollar data (NZD/USD) the following
Trading Rules hypothesis is addressed:

Bid Offer Prices The null hypothesis:

For each trading system, the following rules are applied: H0 Technical trading indicator performance is not improved by
using NZD time series data de-noised by a multivariate wavelet
a. Long trades, including short exit trades, are executed at transform, as measured by the selected metrics set out below.
current offer closing prices as a limit order.
b. Short trades, including long exit trades, are executed at The alternative hypothesis:
current bid closing prices as a limit order. H1 Technical trading indicator performance is improved by
c. Trading signals use mid-price closing data. using NZD time series data de-noised by a multivariate wavelet

IFTA.ORG PAGE 39
 IFTA JOURNAL 2015 EDITION

transform, as measured by the selected metrics set out below. a homogenous stop loss policy is applied and the relative
Trading results must be interpreted in context of the system’s ability to deal with trade entry efficiency (deal
hypothesis stated above and the data selected in respect to the with volatility) and identify sustained (profitable) trends is
generation of the de-noised data series. Selection of different inferred from this measure.
input data may yield different results, including the acceptance c. Is the variance and mean size of MAE statistically different?
or rejection of the currently stated null hypothesis. Whilst stop loss policy is homogenously applied, the core
differentiator between systems reviewed is driven from
Research Measures entry efficiency. MAE variance and means may provide
Measures that support or reject the null hypothesis are based further insight to the TradeStation’s trade efficiency
on the need to measure relative trading performance of each statistics when Start Trade Draw Down is not assessed
trading system using raw exchange rate vs. system performance directly.
using de-noised data. Stridsman speaks of the need to de-clutter d. Is the relationship between MAE and CTD (slope of the
analysis to exclude unnecessary information (Stridsman 2000, two regressions) statistically different? Similar to the
p. 135), whilst also adopting an approach that analyzes winning above rationale, but the relationship to MAE and CTD
and losing trades (ibid). The essence of this approach has been is compared, that is to say, the regression: MAE% = B0 +
adopted, not for system design, but for hypothesis testing. B1 (CTD$) + e are compared. The null hypothesis of: no
As a consequence, hypothesis testing has been undertaken in differences between the slopes of the regressions tests
terms of the following win-loss enumerations: de-noised data trading results against the results of
unprocessed data. Stridsman (2000, p. 152) uses a similar
a. Winning long trades approach but tests the relationships between MAE and
b. Winning short trades CTD. In this research, this approach has been adopted as
c. Losing long trades proxy to infer trade entry efficiency, rather than relying
d. Losing short trades on TradeStation performance summaries and absence
of specific STD figures. Whilst the regression of: MAE% =
Hypothesis testing has been undertaken in the context B0 + B1 (CTD %) + e may have been measured, the actual
of data characteristics such as trend behaviour and specification of the regression is not being primarily
heteroskedasticity, including: sought; rather, the goal here is to test the null hypothesis
Trade entry efficiency. In the context of volatile market as set out above.
conditions, does the system enter prematurely, only to
experience high levels of draw-down (MAE), noting that Start Each of the points a through d above are tested for statistical
Trade Drawdown (STD) is not measured in Stridsman (2000 differences between the raw exchange rate system and an
pp. 126–129). Where a system comparison is applied with a alternative de-noised system. The pair-wise procedure is
homogenous stop loss policy (and no opening trade stop loss), outlined below:
differences in MAE are determined by entry timing. Statistically
different MAE figures, therefore, provide inferences about STD i. In the first instance, F-Tests were conducted to confirm
and trade entry efficiency. whether variances are identical; the null hypothesis of
False positive trade signals. Due to exchange rate volatility, no difference of the variance of returns generated by each
a system may be plagued by whipsaw trades which result in system and data enumeration is tested. Lack of normality of
overtrading and/or with increased magnitude of drawdown. distribution has not been remediated, and first differences
Number of trades. A system may overtrade if not able to using natural logarithms are assumed to be adequate for
differentiate trading opportunities it was designed to identify. data remediation and stationarity purposes.
Average winning and losing trade return, as well as the
number of consecutive winning and losing trades are also ii. Student’s T-Tests were conducted to confirm the null
provided to complement the aforementioned metrics, as well as hypothesis of no difference of the means of returns generated
develop expectations for each system and data enumeration. from the de-noised data and unprocessed data for each
The core measures used to differentiate relative trading trading is tested. T-Test type was selected in accordance with
performance and which form the basis of hypothesis testing the results of F-Test hypothesis testing, as described above.
include:
In addition to this analysis, the following has also been completed:
a. Is the variance of trading returns statistically different?
Trade size variability is driven from the system e. Are returns generated by the respective systems random or
performance, both in terms of system stability. This non-random? Stridsman (2000, p. 274) discusses the need
measure is complemented by the number of consecutive to ascertain the dependencies among trades, stating “…
winning and losing trades and average trade win and loss7 do a runs test, which tells you if your system has more or
as provided by the TradeStation reports. fewer streaks of consecutive wins or losses than what
b. Whether the mean size of returns is statistically different could have been expected if the trades had been randomly
between trading systems and data enumerations. Trading distributed”,
loss and gain are a function of trade entry efficiency when

PAGE 40 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

f. Does the trading system generate positive expectations? Trading System Results
Chande (2001, p. 43) discusses the need of trading Trading results and hypothesis testing are presented for each
systems to have positive expectations, stating “A trading trading system:
system that has a positive expectation is likely to be
profitable in the future”. Hence, relative expectation, and Momentum 100 System
whether expectations are positive or negative, assists Hypotheses Test Results
qualitative analysis of each trading system enumeration. There was limited statistical difference between the various
The Trading Strategy Accuracy (TSA) approach as set out trading systems with statistically significant differences as
by Aldridge (2010, pp. 228–231), has not been adopted in indicated in green (refer to Figure 12). Differences, however, were
this research. Whilst TSA is interesting, this approach constrained to losing long and short trades, where mean MAE
does not provide stationary analysis as required by figures for losing long and short trades were statistically larger
statistical testing. To apply TSA, regression slope analysis than those produced by unprocessed NZD/USD exchange rate data.
would need to be undertaken; however, this approach has Variance of MAE figures were not statistically different when
already been adopted by analysing MAE vs. CTD directly. compared to the system utilising unprocessed NZD/USD exchange
rate data. The relationship between MAE and CTD was not statistically
Complementing this analysis, trade diagnostics are provided different when the slope of the regressions were compared to that
by TradeStation reports and in particular: of the system employing unprocessed exchange rate date.

a. Number of trades, complemented by winning and losing Expectations Analysis

trade statistics The de-noised series provided the highest expectations,
b. Trade entry efficiency statistics particularly that using Symlet 12, Level 5 with five principal
components and the univariate base case.
These measures are aimed at providing an objective,
measurable and stationary approach to trade system Trading Systems Returns Analysis
performance assessment. Whilst equity curves have been Variability of returns around the mean is noticeably higher for
provided in each system test (as well as a correlation matrix of the system using unprocessed exchange rate data with the highest
the equity curves of each system tested), measures such as slope kurtosis. This variability must also be noted in context of the larger
of each system’s equity curves itself has not been tested, because right-hand skew, which illustrates a number of positive outliers
this type of measure would depend on when testing commenced. using unprocessed data, but on average (refer to Figure 13), mean
The tests conducted in this research assume trading system returns ($) were smaller than other model candidates—but were
results are stationary for statistical testing purposes. not statistically significantly different (refer to Figure 15).

Figure 12. Comparative trading system hypothesis testing results

Statistical difference between wavelet results
and results generated on price?

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
Momentum 100 System 95% Cconfidence level applied to F-Test and T-Tests. Level 3 Level 5
Level 5
6 Principal 5 Principal
Components Components
Maximum adverse excursion - winning long trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet models is

Null Hypothesis: Mean Ttest NO NO NO
identical to that of the non-wavelet system

Variance and mean of MAE are not identical between Comparison of 2

Alternative Hypothesis: Regression: P&L-MAE NO NO NO
the wavelet system and non-wavelet systems regression slopes

Maximum adverse excursion - winning short trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet models is

Null Hypothesis: Mean Ttest NO NO NO
identical to that of the non-wavelet system

Variance and mean MAE are not identical between the Comparison of 2
Alternative Hypothesis: Regression: P&L-MAE NO NO NO
wavelet system and non-wavelet systems regression slopes

Maximum adverse excursion - losing long trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet models is

Null Hypothesis: Mean Ttest Yes Yes Yes
identical to that of the non-wavelet system

Variance and mean MAE are not identical between the Comparison of 2
Alternative Hypothesis: Regression: P&L-MAE NO NO NO
wavelet system and non-wavelet systems regression slopes

Maximum adverse excursion - losing Short trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet models is

Null Hypothesis: Mean Ttest Yes Yes Yes
identical to that of the non-wavelet system

Variance and mean MAE are not identical between the Comparison of 2
Alternative Hypothesis: Regression: P&L-MAE NO NO NO
wavelet system and non-wavelet systems regression slopes

IFTA.ORG PAGE 41
 IFTA JOURNAL 2015 EDITION

Figure 13. Comparative expectations analysis results

Multivariate Multivariate Univariate
Symlet 12 Symlet 12 Symlet 12 Price
Level 3 Level 5 Level 5
(6 Principal Components) (5 Principal Components)
Number–Wins 668 666 712 725
Number–Losers 1158 1147 1197 1422
Avg. Loser ($) -15.23747841 -15.17523976 -14.78279031 -12.94655
Avg. Winner ($) 28.04041916 28.24024024 27.26264045 27.065172
% Winner 0.365826944 0.367346989 0.372970141 0.3376805
% Loser 0.634173066 0.632653061 0.627029859 0.6623195
Expectation 0.594742607 0.773303916 0.898899948 0.5612483

Figure 14. Comparative trading system returns analysis

Returns Distribution
Multivariate Multivariate Univariate
Symlet 12 Symlet 12 Symlet 12 Unprocessed Price
Level 3 Level 5 Level 5
(6 Principal Components) (5 Principal Components)
Mean ($) 0.58 0.76 0.88 0.55
Skew 2.6731 2.6604 2.7263 3.0367
Kurtosis 12.1706 12.0205 12.8018 15.2621
Histgrams

Mean and Variance Analysis of Returns

There were no statistical differences between the mean and variances of comparative returns ($) of the de-noised systems when
compared to the system using unprocessed NZD/USD exchange rate data (refer to Figure 15).
Figure 15. Trading system return mean and variance of returns hypothesis testing results
Hypothesis Tests ( Variance identical)

Multivariate Multivariate Univariate

Symlet 12 Symlet 12 Symlet 12
Level 3 Level 5 Level 5
6 Principal Components 5 Principal Components

Two-sample F-test for variances (two-tailed) Two-sample F-test for variances (two-tailed) Two-sample F-test for variances (two-tailed)

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
CTD Level 3 CTD Level 5 CTD
Level 5
6 Principal 5 Principal
Components Components
Mean 0.549977 0.581994 Mean 0.549977 0.756611 Mean 0.549977 0.878648
Variance 846.3645103 971.8262493 Variance 846.3645103 973.8321984 Variance 846.3645103 936.7009424
Observations 2191 1866 Observations 2191 1853 Observations 2191 1953
df 2190 1865 df 2190 1852 df 2190 1952
F 0.87090106 F 0.869107133 F 0.90355894
alpha 0.05 alpha 0.05 alpha 0.05
p-value 1.998112103 p-value 1.998356411 p-value 1.978893999
F-crit 1.091460388 F-crit 1.09164909 F-crit 1.090254943
Significant? NO Significant? NO Significant? NO

Hypothesis Tests ( mean identical)

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
CTD Level 3 CTD Level 5 CTD
Level 5
6 Principal 5 Principal
Components Components
t-Test: Two-Sample Assuming Equal Variances t-Test: Two-Sample Assuming Equal Variances t-Test: Two-Sample Assuming Equal Variances

Variable 1 Variable 2
Mean 0.549977179 0.581993569 Mean 0.549977179 0.756610901 Mean 0.549977179 0.878648233
Variance 846.3645103 971.8262493 Variance 846.3645103 973.8321984 Variance 846.3645103 936.7009424
Observations 2191 1866 Observations 2191 1853 Observations 2191 1953
Pooled Variance 904.0676282 Pooled Variance 904.7688048 Pooled Variance 888.9373532
Hypothesized Mean Difference 0 Hypothesized Mean Difference 0 Hypothesized Mean Difference 0
df 4055 df 4042 df 4142
t Stat -0.033802312 t Stat -0.217663553 t Stat -0.354232429
P(T<=t) one-tail 0.486518228 P(T<=t) one-tail 0.413851125 P(T<=t) one-tail 0.361591367
t Critical one-tail 1.645229489 t Critical one-tail 1.645230698 t Critical one-tail 1.645221592
P(T<=t) two-tail 0.973036456 P(T<=t) two-tail 0.82770225 P(T<=t) two-tail 0.723182734
t Critical two-tail 1.96054918 t Critical two-tail 1.960551063 t Critical two-tail 1.960536885

NO NO NO

PAGE 42 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Tests for Randomness of Trading System Returns

The null hypothesis that returns generated by each system are no different from a random process is not rejected according to the
results of the runs tests set out in Figure 16.

Figure 16. Comparative trading system runs test results

Runs Test Number of Runs above and below 0.
Multivariate Multivariate Univariate
Symlet 12 Symlet 12 Unprocessed
Symlet 12
Level 3 Level 5 Exchange Rate
Level 5
(6 Principal Components) (5 Principal Components)
Hypothosisi 0 0 0 0
Probability 0.7739 0.0627 0.0905 0.6029
Number of Runs 865 881 941 960
N1ii 668 666 712 725
NO 1198 1147 1241 1455
Z Statistic 0.2906 1.8608 1.6925 -0.5162
Conclusion The Null hypothesis of randomness in respect to returns from the momentum system cannot be rejected
i
The result is H=0 if the null hypotheses (“sequence ud random”) cannot be rejected at the 5% significance level, or H=1 if the null hypothesis can be rejected at the 5% level.
ii
‘n1’ –number of values above 0 (or up) –number of values below 0 (or down).

Comparative Equity Curves

Drawdown to net negative returns is exacerbated in the system employing unprocessed NZD/USD exchange rate data. Peak equity
levels of the univariate case surpassing those of the other systems where equity curves for the multivariate system being highly
correlated. Reviewing equity peaks, it appears the de-noised system leads the unprocessed version, although systems oscillate in a
similar way with a peak to trough cycle around identical linear growth rates being visually evident.

Comparative Drawdown Curves

Comparative drawdown curves reveal a similar message as stated above, with unprocessed data displaying a larger and more sustained
drawdown period; with peak drawdown lagging the de-noised series, as identified in the equity curve analysis (refer to Figure 17).

Figure 17. Comparative equity curves Figure 18. Comparative drawdown curves

Moving Average Cross-Over System Results

Hypothesis Test Results
Trading system MAE variance means that hypothesis testing for each of the four categories demonstrates there was no statistical
difference between the system except in the case of the univariate wavelet de-noising method, which concluded that variance of MAE
was statistically smaller to that of the system based on unprocessed price; however, examining TradeStation trading performance
reports, trading statistics were practically identical.

IFTA.ORG PAGE 43
4.7), however, examining TradeStation trading performance reports (Table 4.1 and Table 4.2), trading statistics
were practically identical.
IFTA JOURNAL 2015 EDITION

Figure 19. Comparative trading system hypothesis testing results

Statistical difference between Wavelet
Results and results generated on price?

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
Moving Average Cross Over System 95% Cconfidence level applied to F-Test and T-Tests. Level 3 Level 5
Level 5
6 Principal 5 Principal
Components Components

Maximum adverse excursion - winning long trades Variance Ftest NO NO NO

The variance and mean MAE of the

Null Hypothesis: wavelet models is identical to that of Mean Ttest NO NO NO
the non-wavelet system

Variance and mean of MAE are not

Comparison of 2 regression
Alternative Hypothesis: identical between the wavelet system Regression: P&L-MAE NO NO NO
slopes
and non-wavelet systems

Maximum adverse excursion - winning short trades Variance Ftest NO NO NO

The variance and mean MAE of the

Null Hypothesis: wavelet models is identical to that of Mean Ttest NO NO NO
the non-wavelet system

Variance and mean MAE are not

Comparison of 2 regression
Alternative Hypothesis: identical between the wavelet system Regression: P&L-MAE NO NO NO
slopes
and non-wavelet systems

Maximum adverse excursion - losing long trades Variance Ftest NO NO Yes

The variance and mean MAE of the

Null Hypothesis: wavelet models is identical to that of Mean Ttest NO NO NO
the non-wavelet system
Variance and mean MAE are not
Comparison of 2 regression
Alternative Hypothesis: identical between the wavelet system Regression: P&L-MAE NO NO NO
slopes
and non-wavelet systems

Maximum adverse excursion - losing Short trades Variance Ftest NO NO NO

The variance and mean MAE of the

Null Hypothesis: wavelet models is identical to that of Mean Ttest NO NO NO
the non-wavelet system

Variance and mean MAE are not

Comparison of 2 regression
Alternative Hypothesis: identical between the wavelet system Regression: P&L-MAE NO NO NO
slopes
and non-wavelet systems

Expectations Analysis
Figure 33. Comparative trading system hypothesis testing results.
Expectation analysis demonstrates the system based on unprocessed price had higher expectations than that of the de-
noised systems tested, and this is due to marginally higher average win ($) as well as the probability of winning. Two categories of
expectations are evident, with the multivariate case occupying the least performing group.

Figure 20. Comparative trading system expectations analysis.

Multivariate Multivariate Univariate

Symlet 12 Symlet 12 Symlet 12 Price
Level 3 Level 5 Level 5
(6 Principal Components) (5 Principal Components)
Number–Wins 159 153 139 135
Number–Losers 260 258 219 219
Avg. Loser ($) -17.84615385 -17.49626856 -17.39726027 -17.79452055
Avg.
30 |Winner
P a g($) e 35.02515723 36.37906497 37.20863309 38.8962963
% Winner 0.37947494 0.363420428 0.388268156 0.381355932
% Loser 0.62052506 0.636579572 0.611731844 0.618544058
MFTA Research Adam Cox (2014)
Expectation 2.217183771 2.083135392 3.804469274 3.824858757

Trading System Returns Analysis

Trading system return variability was improved only for the systems utilising multivariate wavelet de-noising, as illustrated by
lower kurtosis, although return distribution for the univariate case has the largest kurtosis and skew. This illustrates that outlier
winning trades were characteristic of each system.

PAGE 44 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Figure 21. Comparative trading system returns analysis.

Multivariate Multivariate Univariate
Symlet 12 Symlet 12 Symlet 12 Unprocessed Price
Level 3 Level 5 Level 5
(6 Principal Components) (5 Principal Components)
Mean ($) 2.21 2.05 3.76 3.76
Skew 1.9091 1.9762 2.1594 2.0689
Kurtosis 5.6907 5.8023 7.2464 6.3383
Histogram

Mean and Variance Analysis of Returns

There was no statistical difference between the variances and means of the returns ($) produced by the system employing
unprocessed exchange rate data and systems that utilise de-noised data series as measured by the F-Tests and T-Tests, respectively
(refer to Figure 22).
Figure 22. Trading system returns mean and variance of returns ($) hypothesis testing results.
Hypothesis Tests ( Variance identical)

Multivariate Multivariate Univariate

Symlet 12 Symlet 12 Symlet 12
Level 3 Level 5 Level 5
6 Principal Components 5 Principal Components

Two-sample F-test for variances (two-tailed) Two-sample F-test for variances (two-tailed) Two-sample F-test for variances (two-tailed)

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
CTD Level 3 CTD Level 5 CTD
Level 5
6 Principal 5 Principal
Components Components
Mean 3.761111 2.206651 Mean 3.761111 2.049065 Mean 3.761111 3.762431
Variance 1405.864779 1247.49767 Variance 1405.864779 1250.56199 Variance 1405.864779 1397.616535
Observations 360 421 Observations 360 428 Observations 360 362
df 359 420 df 359 427 df 359 361
F 1.126947819 F 1.124186398 F 1.00590165
alpha 0.05 alpha 0.05 alpha 0.05
p-value 0.238357755 p-value 0.246144515 p-value 0.955435657
F-crit 1.219861917 F-crit 1.218845895 F-crit 1.229855341
Significant? NO Significant? NO Significant? NO

Hypothesis Tests ( mean identical)

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
Price Level 3 Price Level 5 Price
Level 5
6 Principal 5 Principal
Components Components
t-Test: Two-Sample Assuming Equal Variances t-Test: Two-Sample Assuming Equal Variances t-Test: Two-Sample Assuming Equal Variances

Mean 3.761111111 2.206650831 Mean 3.761111111 2.049065421 Mean 3.761111111 3.762430939

Variance 1405.864779 1247.49767 Variance 1405.864779 1250.56199 Variance 1405.864779 1397.616535
Observations 360 421 Observations 360 428 Observations 360 362
Pooled Variance 1320.480715 Pooled Variance 1321.495452 Pooled Variance 1401.729201
Hypothesized Mean Difference 0 Hypothesized Mean Difference 0 Hypothesized Mean Difference 0
df 779 df 786 df 720
t Stat 0.595909596 t Stat 0.658555742 t Stat -0.000473612
P(T<=t) one-tail 0.27570436 P(T<=t) one-tail 0.255186973 P(T<=t) one-tail 0.499811122
t Critical one-tail 1.646812027 t Critical one-tail 1.646794565 t Critical one-tail 1.646972715
P(T<=t) two-tail 0.55140872 P(T<=t) two-tail 0.510373946 P(T<=t) two-tail 0.999622244
t Critical two-tail 1.963013919 t Critical two-tail 1.962986715 t Critical two-tail 1.963264257

NO NO NO

IFTA.ORG PAGE 45
 IFTA JOURNAL 2015 EDITION

Tests for Randomness of Trading System Returns

The null hypothesis that returns generated by each system are no different from a random process is not rejected according to the
results of the runs tests set out in Figure 23.

Figure 23. Comparative trading system runs test results.

Runs Test Number of Runs above and below 0.
Multivariate Multivariate
Univariate
Symlet 12 Symlet 12 Unprocessed
Symlet 12
Level 3 Level 5 Exchange Rate
Level 5
(6 Principal Components) (5 Principal Components)
Hypothosisi 0 0 0 0
Probability 0.770 0.279 0.960 0.952
Number of Runs 195 185 172 169
N1i 159 153 139 135
NO 260 268 219 219
Z Statistic -0.2906 -1.8608 0.0489 0.0526
Conclusion The Null hypothesis of randomness in respect to returns from the momentum system cannot be rejected

Comparative Equity Curves

Visually, the equity curves are very similar in growth rate and drawdown, although the multivariate cases are once again grouped
together, and the unprocessed price and univariate case appear in a distinct group of their own. This dichotomy is caused by initial
drawndown on trading commencement (although mean and variance of returns are not statistically different). Lead-lag behaviour of
the price and univariate group and the multivariate group is evident, although trade entry and exit efficiency are similar.

Comparative Drawdown Curves

Comparative drawdown curves reveal a similar message as stated above, noting extended drawdown period for the multivariate cases.
Figure 24. Comparative equity curves noting first Figure 25. Comparative drawdown curves noting first
400 trades 400 trades

Elder’s Impulse System Results

Hypothesis Test Results
Statistically significant differences between the systems was limited to mean MAE figures as indicated in green. For long and short
losing trades, statistical differences in mean MAE figures are noted, with the de-noised series having larger negative MAE figures on average
than those of the unprocessed exchange rate version. This result is very similar to that of the Momentum 100 system and demonstrates a
commonality when using de-noised series (i.e., trade entry timing may be been negatively impacted, and/or the incidence of false positive
trades has not been effectively mitigated by using the de-noising method or specific settings).
Overall, MAE figures for winning and losing trades are more negative for those systems that employed de-noised data. However, in
pragmatic terms, the differences are negligible.

PAGE 46 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Figure 26. Comparative trading system hypothesis testing results

Statistical difference between Wavelet Results and
results generated on price?

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
Impulse System 95% Cconfidence level applied to F-Test and T-Tests. Level 3 Level 5
Level 5
6 Principal 5 Principal
Components Components

Maximum adverse excursion - winning long trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet

Null Hypothesis: models is identical to that of the non- Mean Ttest Yes Yes Yes
wavelet system

Variance and mean of MAE are not identical

Comparison of 2 regression
Alternative Hypothesis: between the wavelet system and non- Regression: P&L-MAE NO NO NO
slopes
wavelet systems

Maximum adverse excursion - winning short trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet

Null Hypothesis: models is identical to that of the non- Mean Ttest NO NO NO
wavelet system

Variance and mean MAE are not identical

Comparison of 2 regression
Alternative Hypothesis: between the wavelet system and non- Regression: P&L-MAE Yes NO NO
slopes
wavelet systems

Maximum adverse excursion - losing long trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet

Null Hypothesis: models is identical to that of the non- Mean Ttest Yes Yes Yes
wavelet system

Variance and mean MAE are not identical

Comparison of 2 regression
Alternative Hypothesis: between the wavelet system and non- Regression: P&L-MAE NO NO NO
slopes
wavelet systems

Maximum adverse excursion - losing Short trades Variance Ftest NO NO NO

The variance and mean MAE of the wavelet

Null Hypothesis: models is identical to that of the non- Mean Ttest Yes Yes Yes
wavelet system

Variance and mean MAE are not identical

Comparison of 2 regression
Alternative Hypothesis: between the wavelet system and non- Regression: P&L-MAE NO NO Yes
slopes
wavelet systems

Expectations Analysis
Each system produced negative expectations, with the multivariate cases not having dissimilar results. The smoother univariate
produced the least negative expectations. This is reflected in the comparative equity curves and correlation matrix.
In this research, multivariate systems have more negative expectations than those of the alternative systems. Moreover,
statistically, each de-noised series had limited differential characteristics when compared to unprocessed exchange rate data.
For example, mean and variance of returns were similar, albeit slightly reduced; however, heteroskedasticity (refer to Appendix  2)
was still present in each de-noised series. As a result, a shorter term system like Elder’s Impulse System, which, if sensitive to
heteroskedasticity and whipsaw trades, may not always experience dramatic improvements when multivariate de-noising methods
are employed.
Figure 27. Comparative trading system runs test results

Multivariate Multivariate Univariate

Symlet 12 Symlet 12 Symlet 12 Price
Level 3 Level 5 Level 5
(6 Principal Components) (5 Principal Components)
Number–Wins 1167 1141 1174 1453
Number–Losers 2088 2112 2108 2927
Avg. Loser ($) -12.7098 -12.5433 -12.0384 -10.4151
Avg. Winner ($) 20.3882 20.7555 20.3220 18.9869
% Winner 0.3585 0.3506 0.3577 0.3377
% Loser 0.6415 0.6494 0.6423 0.6683
Expectation -0.8433 -0.8672 -0.4628 -0.6614

Trading System Returns Analysis

A common theme that has developed from the momentum system is that the de-noised systems offer tighter profit and loss ranges.
Lower kurtosis is experienced in the Elder’s Impulse System results using the de-noised versions, with the univariate case having the
smallest negative mean return lowest kurtosis. Whilst each system produced overall negative mean returns, positive skew indicates
the presence of large positive outliers. This effect being most evident in the unprocessed exchange rate version.

IFTA.ORG PAGE 47
 IFTA JOURNAL 2015 EDITION

Figure 28. Comparative trading system returns analysis

Returns Distribution
Multivariate Multivariate Univariate
Symlet 12 Symlet 12 Symlet 12 Unprocessed Price
Level 3 Level 5 Level 5
(6 Principal Components) (5 Principal Components)
Mean ($) -0.8281 -0.8518 -0.4529 -0.6493
Skew 2.1819 2.2415 2.1750 2.5443
Kurtosis 10.8470 11.1960 10.5180 13.3637
Histgrams

Mean and Variance Analysis of Returns

There were no statistical differences noted between the means and variances of returns ($) of the de-noised system and the system
utilising unprocessed NZD/USD exchange rate data.

Comparative Equity Curves

The correlation between the univariate and other versions is evident, with the de-noised series producing overall fewer trades. Equity
curve look and feel confirms the lack of statistical difference of mean and variances of returns as set out in Figure 45.

Figure 29. Trading system return mean and variance of returns hypothesis testing results
Hypothesis Tests ( Variance identical)

Multivariate Multivariate Univariate

Symlet 12 Symlet 12 Symlet 12
Level 3 Level 5 Level 5
6 Principal Components 5 Principal Components

Two-sample F-test for variances (two-tailed) Two-sample F-test for variances (two-tailed) Two-sample F-test for variances (two-tailed)

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
CTD Level 3 CTD Level 5 CTD
Level 5
6 Principal 5 Principal
Components Components
Mean -0.649260 -0.828054 Mean -0.649260 -0.851796 Mean -0.649260 -0.452892
Variance 368.5689515 481.1997559 Variance 368.5689515 483.946326 Variance 368.5689515 465.10798
Observations 4462 3315 Observations 4462 3313 Observations 4462 3354
df 4461 3314 df 4461 3312 df 4461 3353
F 0.765937528 F 0.761590556 F 0.792437385
alpha 0.05 alpha 0.05 alpha 0.05
p-value 2 p-value 2 p-value 2
F-crit 1.065800553 F-crit 1.065812673 F-crit 1.065566737
Significant? NO Significant? NO Significant? NO

Hypothesis Tests ( mean identical)

Multivariate Multivariate
Univariate
Symlet 12 Symlet 12
Symlet 12
Price Level 3 Price Level 5 Price
Level 5
6 Principal 5 Principal
Components Components
t-Test: Two-Sample Assuming Equal Variances t-Test: Two-Sample Assuming Equal Variances t-Test: Two-Sample Assuming Equal Variances

Variable 1 Variable 2
Mean -0.649260421 -0.828054299 Mean -0.649260421 -0.851795955 Mean -0.649260421 -0.452892069
Variance 368.5689515 481.1997559 Variance 368.5689515 483.946326 Variance 368.5689515 465.10798
Observations 4462 3315 Observations 4462 3313 Observations 4462 3354
Pooled Variance 416.5764738 Pooled Variance 417.7301331 Pooled Variance 409.9940043
Hypothesized Mean Difference 0 Hypothesized Mean Difference 0 Hypothesized Mean Difference 0
df 7775 df 7773 df 7814
t Stat 0.382037305 t Stat 0.432094391 t Stat -0.424362285
P(T<=t) one-tail 0.351222068 P(T<=t) one-tail 0.332842398 P(T<=t) one-tail 0.335656652
t Critical one-tail 1.645049634 t Critical one-tail 1.645049684 t Critical one-tail 1.645048655
P(T<=t) two-tail 0.702444135 P(T<=t) two-tail 0.665684796 P(T<=t) two-tail 0.671313304
t Critical two-tail 1.960269147 t Critical two-tail 1.960269225 t Critical two-tail 1.960267623

NO difference NO difference NO difference

PAGE 48 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Tests for Randomness of Trading System Returns

The null hypothesis that returns generated by each system are no different from a random process cannot be rejected according to
the results of the runs tests set out in Figure 30.
Figure 30. Comparative trading runs test results
Runs Test Number of Runs above and below 0
Multivariate Multivariate
Univariate
Symlet 12 Symlet 12 Unprocessed
Symlet 12
Level 3 Level 5 Exchange Rate
Level 5
(6 Principal Components) (5 Principal Components)
Hypothosis 0 0 0 0
Probability 0.417 0.4516 0.389 0.770
Number of Runs 1489 1478 1460 1934
N1 1180 1167 1141 1453
NO 2096 2088 2115 2927
Z Statistic -0.8128 -0.7509 -0.897 -0.289
Conclusion The Null hypothesis of randomness in respect to returns from the momentum system cannot be rejected

Comparative Drawdown Curves

Comparative drawdown curves (Figure 32) reveal a similar message.

Figure 31. Comparative equity curves. Figure 32. Comparative drawdown curves

Conclusion No particular data series indicated a cointegrated relationship

This research aimed at identifying whether benefits could with the NZD/USD exchange rate over the short run using 10
be derived from employing multivariate wavelet de-noising to minute data, and therefore, any system improvement could not
various trading systems. be attributed to the effects of a cointegrated relationship with
An identical stop loss method, the Elder’s SafeZone stop the New Zealand dollar. There was no evidence that the exchange
loss technique was employed, thereby reducing the number of rates Granger-cause or lead the NZD/USD exchange rate, and
dimensions requiring analysis. To provide an objective basis a bidirectional or contemporaneous information transmission
by which improvement (if any) could be identified, a series of across/between markets was assumed, although weak evidence
statistical tests were undertaken for each trading system, was provided to confirm lead-lag relationship between other
where trading results of each de-noising method were compared currency pairs. However, these tests were inconclusive and it has
to the unprocessed data version. been assumed that no effective Granger-causing relationship
This process commenced with an analysis of the individual could be exploited by utilising the multivariate de-noising
exchange rates (6) used in de-noising, in order to understand method as specified in this research.
the possible effect de-noising would have on the New Zealand Trading system results analysis and hypothesis testing did
dollar and conjecture why trading system improvements could not conclude overwhelming evidence that trading system results
be experienced if multivariate de-noised data was employed. were improved using de-noised exchange rate data; specifically:
Each exchange rate was found to have significant runs,
serial autocorrelation and heteroskedasticity, indicating the 1. For the momentum 100 system, the multivariate cases
possibility of substantive trends and therefore technical trading provided significant reduction in trading activity, marginally
opportunities. higher expectation level, but with similar overall trading
results, with entry efficiency being roughly equivalent to

IFTA.ORG PAGE 49
 IFTA JOURNAL 2015 EDITION

the system that employed unprocessed NZD/USD exchange References

rate data. Each of the de-noised series provided higher Aldridge, Irene. High Frequency trading – a practical guide to algorithm strategies
expectations than the unprocessed price version. However, and trading systems. John Wiley& Sons, (2010).
Alexander, Carol. Market models – a guide to financial data analysis.John Wiley
statistically different means and variance of returns could & Sons Ltd, (1993).
not be confirmed, whilst the multivariate de-noised versions Barclay, Bonner and Hamilton. Application of wavelet transforms to
had lower kurtosis and skew. Statistically, more negative experimental spectra: smoothing, denoising and data set compression. Analytical
chemistry, Vol 69 No 1 (1997), pp 78-90.
MAE figures were confirmed for the multivariate systems, but
Batten and Hogan. The spot AUD/USD Foreign Exchange market: evidence from
this did not impact comparative overall returns. high frequency data.University of Technology, Sydney. 2001.
2. For the moving average cross-over system, two distinct Bernstein, J. Momentum stock selection. Using the momentum method for
groups of trading results have emerged; unprocessed price maximum profit.McGraw-Hill, (2002).
Bin, Bianchi and Su. Monthly seasonality in currency returns: 1972-2010.The FINSIA
and univariate de-noising, and those trading results produced Journal of Applied Finance, (2011), Issue 3, pp 6-11.
by multivariate de-noising. Whilst hypothesis testing and Campbell, Lo and MacKinlay. The econometrics of financial markets.Princeton
returns analysis has not identified significantly (statistically) University Press, (1997).
different results, the lead-lag of the equity curves being a Crowder. Foreign exchange market efficiency and common stochastic trends.
Journal of International Money and Finance, (1994), Vol 13, No 5, pp 551-564.
more striking difference between the respective systems. Chande,Tushar. Beyond technical Analysis. How to develop and implement a
3. In terms of the Elder’s Impulse System, significant winning trading system. John Wiley and Sons, 2nd Edition 2001.
improvement was not evident when de-noised data series Dacorogna, Gencay, Muller, Olsen and Pictet. An introduction to high-frequency
finance. Academic Press, (2001).
were employed. Expectations were worse for the multivariate
Elder, Alexander. Trading for a Living: Psychology, Trading Tactics, Money
case than that of the unprocessed exchange rate data version, Management John Wiley and Sons, (1993).
although trading activity had been reduced, signifying an Elder, Alexander. Come into my trading room. John Wiley and Sons, (2002).
improvement in trading discretion. It may be conjectured Edwards, D & Magee, J Technical Analysis of Stock Trends, 8th Edition CRC Press
that poorer expectations of the multivariate systems could LLC (2001)
Ehlers, John. Moving averages and smoothing filters Stocks & Commodities
be attributed to poorer trade entry efficiency. More negative Magazine Vol. 7:3 pp 87-90, 1988.
MAE figures were also noted for the de-noised versions and Ehlers, John. Cybernetic analysis for stock and Futures. Cutting edge DSP
in the context of an homogenous stop loss policy (Elder’s technology to improve trading. John Wiley & Sons, (2004).
SafeZone), trade entry timing may have been degraded due to Frances, Kofman and Moser. Garch effects on a test of cointegration. Review of
quantitate financial and accounting (1994), pp 19-26.
the use of de-noised data.
Guppy, Daryl. Trend Trading. A seven step approach to success. Wrightbooks, (2004).
Guppy, Daryl. Chart Trading. An approach for investors and traders. Wrightbooks, (1999).
Results have been mixed. However, the null hypothesis that Fama, E. and K. French, Common Risk Factors in the Returns on Stocks and
de-noised data using multivariate wavelet de-noising (as utilised Bonds. Journal of Financial Economics, 1993. 33(1).
by the trading systems set out in this research) does not improve Hakkio, Craig. Co-integration: how short is the long run? Journal of International
Money and Finance (1991), pp 571-581.
trading system performance cannot be rejected. Hamilton, James D. Time series analysis. Princeton University Press, (1994).
Further research is proposed to address matters not Hutson,J. Good TRIX. Technical analysis of stocks and commodities magazine,
addressed in this research including: (1982) Vol 1, No 5, pp 105-180 .
LeSage, J. Applied Econometrics using Matlab Department of Economics University
of Toledo (1999).
1. Lead-lag relationships between input data or even the de-
Levine, Stephen, Krehbeil, Berenson. Statistics for managers using Microsoft
noised data and the underlying price have not been fully Excel Prentice Hall (2002)
explored. Lo and MacKinlay. Stock markets prices do not follow random walks: evidence
2. Where additional data series are selected, use of principal from a simple specification test. The review of financial studies (1998), Vol 1, No 1,
pp 41-66.
components may be further explored to reduce dimensions Joiner,A and Braddick,P. ANZ Fair Value Exchange Rate Model: An Update ANZ
in the dataset employed. The effects of using principal Research, ANZ Bank Melbourne Australia (2006).
components in context of six data series has been limited; Malloy,P. Triple exponential moving average. Technical analysis of stocks and
commodities magazine, Feb 1994.
however, utility of this methodology may be further explored
Meharena,Z. The information transmission across unites states and New Zealand
if a larger dataset were employed. spot exchange rate Massey University, (1996).
Misiti, Misiti, Oppenheim, and Poggi. Wavelet toolbox User’s Guide The Mathworks,
Notes (2012).
Matlab. GARCH Toolbox User’s Guide The Mathworks, (2002).
1. Heuristic sure thresholding method.
2. Trademark of Omega Corporation Munro,A. What drives the New Zealand dollar?, Vol 67, No 2 pp 21-33 Reserve Bank
3. www.dukascopy.com of New Zealand. (2004).
4. The constituent New Zealand cross rates and weighting factors for the New Okunev, J. and D. White, Do momentum-based strategies still work in foreign
Zealand dollar TWI as of 15 October, 2013 are set out below: currency markets? Journal of Financial and Quantitative Analysis, 2003. 38.
United States Dollar USD 0.2990 Polikar, R. The Wavelet Tutorial. Iowa State University of Science and Technology,
Euro EUR 0.2670 (2001).
Japanese Yen JPY 0.1482 Pring, Martin, J. Martin Pring on Market Momentum. McGraw-Hill, (1993).
Australian Dollar AUD 0.2225
UK Sterling GBP 0.0633 Stridsman, T. Trading systems that work. Building and evaluating effective trading
systems. McGraw-Hill (2000).
5. n*(n-1)/2 is equivalent to n!/(((n-2)!*2), which may also be stated in some texts.
Williams. Trading Chaos: Applying Expert Techniques to Maximize Your Profits.
6. Magee actually define the word, ‘Mulitcolincarity’, but it is assumed that this John Wiley & Sons (1995).
is a spelling mistake and the word ‘multicollinearity’ is actually meant. Wong, H & Ling K. A cointegrated heteroskedastic model with financial applications.
Department of Applied Mathematics, The Hong Kong University, and Department
of Economics, University of Western Australia. Preprint 2000-5 (2000).

PAGE 50 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Alexander Neale
[email protected]
Use of Social Media Mentions in 139 Westgate Apartments

Technical Analysis 14 Western Gateway

London E16 2EN
United Kingdom
By Alex Neale, MFTA
+44 (0) 7871 302 845

Abstract These numbers can be derived from the company itself, the
This paper investigated the growing trend of using social particular industry’s sector, the overall economy, or any
media mentions in technical analysis to aid in investment combination.
decisions and whether the On Balance Volume (OBV) Technical analysis is rooted in economic theory, and
methodology can be amended to include social media mention fundamental analysis by default, as market prices are the result
volumes to create a new indicator On Balance Sentiment Volume of supply and demand. Charles Dow is considered by many to
(OBSV). Volatility in social media mention volumes was also be the father of technical analysis, and his beliefs have come to
compared to traditional share volume volatility to investigate be known as Dow Theory, which is central to technical analysis
whether the technical analysis tenet that “volume precedes study. Charles Dow and his partner, Edward Jones, formed Dow
price” could also be applied to social media mentions. Stocks Jones & Company in 1882. Dow was the founder of the Wall Street
traded on the Dow Jones Industrial Average (DJIA) were the Journal. In July 1885, he published the first ever stock market
focus of the paper. The investigations comparing OBV and OBSV average, and in 1897, he created the Dow Industrials Average
conclude that social media mention volumes can act as a useful Index (DJIA). The constituents of the DJIA are the focus of this
tool in equity trading, indicating that social media mentions paper.
will continue to play an increasingly important role in technical
analysis. Behavioural Finance
Despite Charles Dow’s major contribution to the formation
Introduction of Wall Street, and despite the numerous works on technical
This paper investigated the growing trend of using social analysis written over the past century, technical analysis has
media mentions in technical analysis to aid in investment not been widely accepted in academia. However, the field of
decisions. I investigated whether the On Balance Volume (OBV) behavioural finance is increasingly finding evidence to support
methodology can be amended to include social media mention some of the long-held beliefs of market technicians. Behavioural
volumes to create a new indicator On Balance Sentiment Volume finance studies the effects of cognitive, social, and emotional
(OBSV). I also compared social media mention (SMM) and factors on the economic decisions of both individuals and
traditional share volume volatility to investigate if the technical institutions, and the resultant consequences on market prices.
analysis tenet that “volume precedes price” applies to social In 1979, Kahneman and Amos Tversky published work on the
media mentions. Stocks traded on the Dow Jones Industrial “Prospect Theory”, citing that “choices amongst risky prospects
Average (DJIA) were the focus of my work, and covered the exhibit pervasive effects inconsistent with the basic tenets of
period from February 2, 2012, to August 30, 2013. Utility Theory”,3 indicating that standard economic teaching on
In this paper, I will briefly cover the history of technical utility theory does not explain all the actions taken by investors.
analysis, behavioural finance, and sentiment in investing, as The field of behavioural finance has described numerous other
this background is required to appreciate the underlying logic cognitive biases. Cognitive biases occur through heuristics,
of using social media mentions. Then, I detail the investigations which are often simple but largely efficient rules, either learnt
comparing OBSV and OBV data, this paper is not intended to or hard coded through evolution, that allow us to quickly solve
detail a trading system. complex problems. These rules work under most circumstances,
but not all, and can lead to systematic errors and biases. A
What Is Technical Analysis? number of these documented cognitive biases have similarities
In order to consider the proposed OBSV indicator, it is with the beliefs of technical analysts, such as Anchoring.4
necessary to briefly describe the technical approach. A straightforward way to demonstrate that we do not
“Technical analysis is the study of market action for the always think or act logically can be described with simple visual
purposes of forecasting future price trends”1 and “in its basic illusions. Despite our hunter-gatherer evolutionary focus on
form, technical analysis is the study of past market data, sight, we can still be tricked with five very simple lines, such as
primarily price and volume data; this information is used to with the Muller-Lyer illusion in Figure 1. The end arrows act as
make trading or investing decisions”.2 depth cues, creating the appearance that the horizontal line in
Fundamental analysts however believe that the value of a Image A is wider than the horizontal line in B; they are in fact
security can be determined through its set of financial numbers. the same width.

IFTA.ORG PAGE 51
 IFTA JOURNAL 2015 EDITION

Figure 1: Simple Muller-Lyer Perspective Illusion

In this example (Figure 4), moving the four colour shapes,
without otherwise resizing or altering them, seems to create
a new blank square in the lower version. Clearly, this creation
of new space is not possible. The illusion appears as the blue
triangle’s hypotenuse has a ratio of 2:5, while the red triangle’s
hypotenuse has a ratio of 2:5.333, so despite appearances, the
upper shape is not a triangle.
All these illusions occur because we process images in ways
In her book Technical Analysis for The Trading Professional, that are useful to see, rather than seeing the world as it actually
Constance Brown dedicates a whole chapter to the principles of is.6 These examples describe visual biases, as what we think we
depth perspectives applied to two-dimensional charting.5 She see in the above examples is illogical and appears incorrect, yet
states that “our minds will unconsciously look for subtle clues to intelligent readers still see them, often still seeing them even
create a solution for the missing plane that has been omitted in after being told of the effects being demonstrated. Behavioural
our [2 dimensional] charts.” finance describes numerous examples where intelligent
individuals can make seemingly illogical and incorrect
Figure 2: Photo-Based Illusion statistical and investing decisions. This is important, as a
Image source: Richardwiseman.wordpress.org.
criticism of technical analysis over the years has been its basis
in seemingly illogical investor behaviour. Behavioural finance
details how some of this investor behaviour is not entirely
illogical (it only appeared so), and that investing biases can be
systematic, and can therefore be forecasted.
For example, it has been demonstrated by Dr. Hill that people
regularly underestimate the likelihood of runs in coin toss
experiments, and that people also incorrectly assume that the
first digit on diverse data, such as the length of a river chosen
at random, should have an equal probability of being a 1, 2, 3, 4,
Figure 2 is a photo-based example of a Ponzo Illusion. The 5, 6, 7, 8 or 9. Benford’s Law, however, describes that number 1
effect was named by Mario Ponzo in 1911 and describes that should actually appear as the first digit around 30% of the time,
images are viewed in context with their surroundings, not by due to the natural logarithmic characteristics of the Arabic
precisely examining each element by itself. Due to the context numbering system.7
of perspective, most viewers see the top car as larger than the Readers may assume that this general misunderstanding of
lower; however, they are the same size. basic statistics and logic only applies to the general public and
does not apply to professionals. It has been shown that doctors
Figure 3: Tonal Illusion can also make seemingly illogical decisions.8
Image source: Lottolab Studios
It has been demonstrated that investors buy more and sell
less when the critical chart is characterised by a salient high
rather than a low.9 This is due to the ‘Anchoring’ cognitive bias,
which is a psychological heuristic that influences the way people
intuitively assess probabilities.4
Interesting work on monkey behaviour by Dr. Agnieszka
Tymula has led him to state that “Human sensitivity to wealth
levels developed before the advent of money”,10 and that “It
seems likely that the biological mechanisms that mediate
The images in Figure 3 detail how our perception of tone can changes in risk attitudes with wealth evolved around satiety
also be confused. To many viewers, it appears that A is darker mechanisms rather than around mortgages”, thus indicating
than B. However, they are the same shade, as can be seen in the that the human outlook on wealth, and therefore investing,
masked version on the right. may be much more deeply rooted in our evolution than simple
economic utility theory would suggest. Indeed, most financial
Figure 4: Geometry Illusion market participants would now agree that, while the markets
are largely efficient and follow the standard utility theory
some of the time, they are not efficient all of the time. Technical
analysis attempts to highlight these periods of inefficiency in
order to forecast future price moves.

Sentiment in Technical Analysis

Technical analysis on share trading has traditionally used
market action, which is made up of price and volume. Due to the
large amount of information available on social media, a third

PAGE 52 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

arm of technical analysis is increasingly being investigated— markets. On 13 August 2013, the billionaire activist investor
namely, sentiment. Carl Icahn posted on Twitter that he had taken a “large position”
Sentiment is a general term covering the degree to which in Apple, and that “We believe the company to be extremely
a group of market participants, in aggregate, are bullish or undervalued.” Apple’s share price gained over 5% intraday on
bearish on the market. Sentiment at its broadest, therefore, is this tweet.
fundamental analysis plus technical analysis plus the effects of Twitter was only formed seven years ago, and already it has
any additional cognitive, social and emotional factors. 218 million monthly users; 500 million tweets are sent each day
Nobel Laureate Robert J. Shiller in his book Irrational on average. Twitter provided these details in October 2013 as
Exuberance wrote that when certain key structural, cultural part of its plans for a $7bn IPO.
and psychological factors combine, the individual sentiment Bloomberg is a financial services group, and its Bloomberg
effects can act in unison to move markets into bubbles,11 such terminals are used by around 300,000 trading professionals.
as the 1711 South Sea Bubble, 1920s Florida Land Boom, and In May 2013, it launched the ‘Bloomberg Social Velocity’ service
the strong stock market bull trend through 1999, which was (Bloomberg terminal code: BSV <GO>), which tracks the number
concentrated on the telecom, media and technology stocks. of social media mentions (SMMs) on individual stocks and
Numerous measures of investor sentiment have been created highlights movements out of the norm based on Bloomberg
over the years, such as Market Vane, Mutual Funds Cash Assets algorithms. Currently, Bloomberg does not publish individual
Ratio, and Put/Call Ratios, among many others. Sentiment SMM volume data; it only summarises the information in what it
measures are becoming increasingly used in finance. For calls Social Media Velocity alerts.
example, in 2007, Malcolm Baker, then the associate professor Numerous websites have been created to cater to the
of finance at Harvard Business School, wrote “Now, the question growing demand from investors in this space, which has become
is no longer, as it was a few decades ago, whether investor known as ‘Social Trading’—websites such as StockTwits, eToro,
sentiment affects stock prices, but rather how to measure stockstreams.net and Knowsis, and to a lesser degree, estimize.
investor sentiment and quantify its effects.”12 com, Socialmention.com, Backtweets.com and bottlenose.com.
In finance, the market can be separated into informed These developments highlight how social media continues to
players and noise players. The term ‘noise trader’ was first mature into an increasingly important source of information for
described by Fisher and Black in 1985.13 Another term is investors.
‘uninformed player’, and it relates to the apparent random The amount of data available online, and the computing
activity around the equilibrium price. The finance assumption expertise required to analyse it, has led to a new science being
is that the equilibrium price is largely set by the informed created: computational social science. Numerous social media
investors. Bradford De Long, also at Harvard, wrote that ‘noise data mining opportunities have been exploited, such as Chicago
traders’ can affect prices and that “The unpredictability of noise authorities scanning tweets in 2013 for possible food poisoning
traders’ beliefs creates a risk in the price of an asset that deters outbreaks,20 Asur and Huberman in 2010 finding that social
rational arbitragers from aggressively betting against them. media content can be used to forecast box-office revenues for
As a result prices can diverge significantly from fundamental movies,21 and Daniel Gruhl in 2005 finding similar results for
values.”14 book sales,22 thus providing evidence that social media mentions
There have been numerous examples of studies on can be used for forecasting in general. In 2011, Johan Bollen
sentiment, such as Da, Engelberg and Gao, 2011, detailing that and Huina Mao provided data that was “strongly indicative of
monitoring the totals of Google searches on U.S. companies a predictive correlation between measurements of the public
acts as a proxy for investor interest and can be used to forecast mood states from Twitter feeds and DJIA values” and suggested
higher stock returns in the following two weeks.15 Gidófalvi16 it was an interesting area for additional research into financial
in 2012 and Schumaker17 in 2009 separately demonstrated that forecasting.23
financial news articles affect share returns, and Edmans and
Garcia in 2007 even described an effect when negative sports Social Media Mention Data
sentiment affects stock returns.18 These papers are mentioned There is not an official source for SMM data, in the way
to describe how evidence has been found that noise trader that exchanges publish official OHLC and volume data. For
sentiment can affect stock returns. The next section details how this paper, I used SMM volume data provided by Knowsis.
social media mentions are increasingly being used to describe Knowsis is a financial services provider, and it identifies
and measure noise trader sentiment. and quantifies underlying behavioural trends from a broad
range of online sources to generate alpha. It uses proprietary
Social Media in Investing technology to identify and amalgamate financially relevant
In April 2, 2013, the Securities and Exchange Commission online conversation (from social media, blogs, forums) into
(SEC) announced that “companies can use social media outlets quantifiable and actionable output to help with trading,
like Facebook and Twitter to announce key information in investment and risk management. The algorithm includes
compliance with Regulation Fair Disclosure (Regulation FD), as searching for the use of company stock ticker codes in social
long as investors have been alerted about which social media media (e.g., WMT, Wal-Mart; KO, Coca-Cola), and it also
will be used to disseminate such information.”19 searches for company related text, such as Coca-Cola, Coke,
This announcement has been seen as the date at which social Cola. Knowsis can provide two datasets per asset, the volume
media matured into a genuine data source for the financial of SMMs it collates, and a normalized data series of the SMMs it

IFTA.ORG PAGE 53
 IFTA JOURNAL 2015 EDITION

Figure 5: Walt Disney, OBV Graph Including Simplified

collates. The volume of SMMs was used for this paper. Net Field Trends (Feb 2012–Jan 2013)
The focus of my work was to examine if using Knowsis’ SMM
volume data on DJIA-listed stocks would demonstrate results
comparable to, or better than, traditional share volumes. To
test this, I started the investigations on the On Balance Volume
methodology, as it has a wide following in the financial markets,
and the underlying logic seemed applicable to SMM volume data.

On Balance Volume
On Balance Volume (OBV), as the names implies, uses share
volume data in its calculations. Numerous methods have been Figure 5 shows the net field trends of Walt Disney for 2012.
designed over the years to analyse volume data. These include OBV net fields can be rising, neutral or falling. This graph
Demand Index, Volume ROC, Chaikin Money Flow; however, highlights how rising fields cluster in the bullish phases of the
among the most widely used on stocks is OBV, and it was market, and falling fields cluster in bearish phases. Figure 6
largely brought to the market by Joseph Granville (August shows the first weeks of this graph in more detail.
20, 1923–September 7, 2013) in his book, New Key to Stock
Market Profits,24 first published in 1963. The idea, however, Figure 6: Walt Disney—Net Field Trends (February 2012–
March 2012)
was originally called cumulative volume and had been written
about Woods and Vignolia as early as 1946, and also presented
to the American Statistical Association in 1932 by Paul Clay of
Moody’s Investors Services.25 In his book, Granville explained
OBV by using DJIA constituent stocks in his calculations, so his
methodology was first designed to be used on well-capitalised
and liquid stocks.
OBV is a simple calculation; starting with a nominal OBV
value, the day’s volume is either added or subtracted to the
previous day’s OBV, depending on the direction of the stock
on that day. (The Excel formula is detailed in the Appendix.)
The numerical levels in OBV graphs are not significant— only
the direction and its relationship to price is considered. OBV is
usually displayed as line graph under the graph of the stock. To create net field trends, two additional lines need to be
As OBV is such a simple calculation, it has been added to by created from the basic OBV—namely, peak and trough lines.
numerous technicians over the years, such as Marstein’s Volume These are shown in Figure 6. The Appendix outlines this
Price Trend in 1966, which adds emphasis to days with larger procedure, and Table 1 details the data used to create the graph
changes in price.26 Granville himself also suggested that OBV in Figure 6. Once these peak and trough signal lines are created,
can be calculated on each of the open, high, low and close prices the OBV line itself is not used in signal generation, so the OBV
for extra weighting. For this paper, however, I focused on the line is included for reference only in Figure 6.
simple OBV methodology. The net field signal methodology is very simple: Go long when
One major aspect of OBV interpretation has been the both peak and trough lines have moved higher, and close long
argument that informed investors are better capitalised than positions if one of the signal lines then moves lower. Go short if
noise players, and that as a result, accumulation/distribution both signal lines have moved lower, and close shorts if one of the
from informed players may become visible in volume data ahead signal lines then moves higher.
of changes in price. For this reason, OBV has come to be known Using the data in Table 1, and referencing the lines in Figure
primarily as a divergence indicator. 6, we can see that at Point A, the previous move in the peak line
“The advantage of recording the OBV is in observing when was marginally negative, but the last move in the trough line
the trend of the prices diverges from the OBV values.”27 OBV was positive; therefore, there is no net field trend. At Point B, the
divergence monitoring is interesting but difficult to test. Less trough line turns negative. This matches the negative direction
well known is that Granville also devised net field trends from of the peak line and creates a net falling field; short positions
the OBV, and these fields create more testable trading signals. are opened at the open the next day. At Point C, the trough line
direction has turned positive, turning the falling field off; at the
OBV/OBSV Net Field Trends open the next day, the short positions are closed. At Point D, the
For this paper, I calculated OBV and the related net field peak line also turns positive, posting the first net rising field;
trends. The formulas used in these calculations are detailed in long positions would be opened at the open the next day. This
the Appendix. The methodology employed is an interpretation rising field remains in place until Point E, when the trough line
of Granville’s net field trend calculations. Then, I repeated the moves lower; where the net field turns neutral, long positions
calculations using SMM volumes to create an OBSV indicator and would be closed at the open the next day. These calculations
compared the results. were used for the full data series, on each stock, using share
volumes, and then run again with SMM volumes.

PAGE 54 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Table 1: Data Used to Calculate Wal-Mart Graph in Figure 6

Date Close Volume OBV Peak Line Trough Line Peak Trough Field Trades
13/02/2012 41.79 9065800 77,294,000 78,344,500 68,228,200 - +
14/02/2012 41.6 8210300 69,083,700 77,294,000 68,228,200 - +
15/02/2012 41.25 7891200 61,192,500 77,294,000 68,228,200 - +
16/02/2012 41.54 7464500 68,657,000 77,294,000 61,192,500 - - Falling
17/02/2012 41.75 7758700 76,415,700 77,294,000 61,192,500 - - Falling Sell Short
21/02/2012 41.57 5675100 70,740,600 76,415,700 61,192,500 - - Falling
22/02/2012 41.27 6176200 64,564,400 76,415,700 61,192,500 - - Falling
23/02/2012 41.48 6271000 70,835,400 76,415,700 64,564,400 - +
24/02/2012 41.31 6150600 64,684,800 70,835,400 64,564,400 - + Close Short
27/02/2012 41.64 9132300 73,817,100 70,835,400 64,684,800 - +
28/02/2012 41.93 6895200 80,712,300 70,835,400 64,684,800 - +
29/02/2012 41.99 14858700 95,571,000 70,835,400 64,684,800 - +
01/03/2012 42.39 7995900 103,566,900 70,835,400 64,684,800 - +
02/03/2012 42.36 5070500 98,496,400 103,566,900 64,684,800 + + Rising
05/03/2012 42.7 9288800 107,785,200 103,566,900 98,496,400 + + Rising Buy Long
06/03/2012 42 10434500 97,350,700 107,785,200 98,496,400 + + Rising
07/03/2012 41.75 7793500 89,557,200 107,785,200 98,496,400 + + Rising
08/03/2012 42.02 6163800 95,721,000 107,785,200 89,557,200 + -

For this paper, I used SMM volume data on 29 of the 30 of the stocks that constituted the DJIA as of January 2013. Dow Jones &
Company periodically makes changes to the constituents of its DJIA Index. The five most recent changes are detailed in Table 2.
UnitedHealth is not included in my calculations due to a lack of company-specific SMMs.
Table 2: Last Five DJIA Constituent Changes
Date Exclusions Inclusions
September 20, 2013 Alcoa, Bank of America, HP Goldman Sachs, Nike, Visa
September 24, 2012 Kraft Foods UnitedHealth Group
June 8, 2009 General Motors, Citigroup Travelers Companies, Cisco Systems
September 22, 2008 American International Group (AIG) Kraft Foods
February 19, 2008 Altria Group, Honeywell Chevron, Bank of America

Table 3 compares the average share and SMM volume data per stock over the period covered. From this it is clear that SMM volumes
on average have a much higher standard deviation than traditional share volumes. This table also details the correlation between
share and SMM volumes on each stock. The average was 0.375, with a standard deviation of 0.243, detailing a good relationship
between the two datasets (e.g., the correlation between Microsoft and Intel stock volumes over the period was 0.276).
Table 3: DJIA Constituents, Volume and SMM Volume Data
Ticker Name Volume Volume SMM Volume SMM Volume Volume/
Standard Standard SMM Volume
Deviation Deviation Correlation
MMM 3M Company 2,759,219 893,174 29 45 0.390
AA Alcoa 19,407,210 7,947,085 268 583 0.590
AXP American Express Company 5,459,920 1,920,215 48 115 0.200
T AT&T, Inc. 25,701,315 11,352,331 232 252 0.091
BAC Bank of America 165,920,269 84,596,651 341 224 0.531
BA The Boeing Company 4,888,393 2,940,233 189 376 0.753
CAT Caterpillar Inc. 6,692,147 2,410,151 173 296 0.587
CVX Chevron Corporation 6,112,829 1,938,063 313 573 0.092
CSCO Cisco Systems, Inc. 40,053,697 18,985,608 287 415 0.688
KO The Coca-Cola Company 14,788,585 6,690,808 147 131 0.233
DD E. I. du Pont de Nemours and Co. 5,603,296 2,870,308 41 109 0.526
XOM Exxon Mobil Corporation 13,965,967 5,425,609 654 1,266 -0.003
GE General Electric Company 42,689,608 17,290,412 282 428 0.387
HPQ HP 22,660,240 15,455,822 160 333 0.788
HD The Home Depot, Inc. 8,506,140 3,460,825 97 279 0.411
INTC Intel Corporation 41,860,113 18,044,523 461 620 0.397
IBM International Business Machines 3,970,718 1,738,014 244 366 0.599

IFTA.ORG PAGE 55
 IFTA JOURNAL 2014 EDITION

JNJ Johnson & Johnson 11,164,481 7,627,398 87 148 0.073

JPM JPMorgan Chase & Co. 28,351,060 17,429,625 1,803 3,195 0.745
MCD McDonald’s Corp. 5,551,451 2,304,350 208 459 0.142
MRK Merck & Co. Inc. 14,430,342 7,916,679 420 279 0.303
MSFT Microsoft Corporation 47,852,733 21,459,149 912 1,159 0.630
PFE Pfizer Inc. 36,716,645 24,116,103 644 549 0.046
PG The Procter & Gamble Company 9,939,426 5,023,999 192 201 0.389
TRV The Travelers Companies, Inc. 2,380,454 902,890 8 16 -0.019
UTX United Technologies Corp. 3,805,662 1,341,419 66 118 0.212
VZ Verizon Communications Inc. 13,268,574 7,004,712 200 327 0.181
WMT Wal-Mart Stores Inc. 8,528,863 4,272,934 104 109 0.510
DIS The Walt Disney Company 8,435,038 3,304,364 216 282 0.403
Average 21,429,807 10,574,602 304 457 0.375
Standard Deviation 30,765,137 15,733,252 350 595 0.243

OBV, OBSV, Net Field Trend: Profit and Loss Comparisons

The net field trend methodology described in Table 1 was applied to volume data for the 29 stocks covered and repeated using SMM
volume data. Table 4 details the results.

Table 4: Number of Trades, % Wins and % Profit and Loss, for OBSV and OBV Net Field Trading Signals
OBSV Net Field Signals OBV Net Field Signals
Code Name Trades % Win % P&L Trades % Win % P&L
MMM 3M Company 55 38.2% -0.28% 57 43.9% -0.20%
AA Alcoa 44 38.6% -0.06% 42 42.9% -0.16%
AXP American Express Company 59 42.4% 0.07% 57 49.1% -0.11%
T AT&T, Inc. 36 47.2% -4.40% 53 41.5% -3.75%
BAC Bank of America 44 38.6% -1.53% 54 42.6% 1.61%
BA The Boeing Company 50 52.0% 8.08% 55 45.5% 0.54%
CAT Caterpillar Inc. 47 51.1% 11.31% 48 37.5% -14.21%
CVX Chevron Corporation 49 40.8% -5.66% 50 40.0% -17.37%
CSCO Cisco Systems, Inc. 44 38.6% -5.05% 41 34.1% 0.77%
KO The Coca-Cola Company 54 50.0% -0.04% 61 44.3% -5.35%
DD E. I. du Pont de Nemours and Co. 54 37.0% -7.63% 55 43.6% -8.40%
XOM Exxon Mobil Corporation 51 47.1% -0.05% 59 37.3% -25.64%
GE General Electric Company 52 38.5% -0.16% 45 42.2% -2.35%
HPQ HP 53 37.7% -0.65% 43 44.2% -7.61%
HD The Home Depot, Inc. 52 59.6% 0.38% 50 52.0% 5.82%
INTC Intel Corporation 55 38.2% -0.28% 55 34.5% -0.30%
IBM International Business Machines 46 41.3% -0.12% 46 43.5% 0.03%
JNJ Johnson & Johnson 46 54.3% 0.17% 45 57.8% 0.14%
JPM JPMorgan Chase & Co. 48 47.9% 0.21% 51 35.3% 0.31%
MCD McDonald’s Corp. 48 37.5% -0.30% 49 40.8% -0.09%
MRK Merck & Co. Inc. 42 57.1% 0.23% 44 38.6% 0.02%
MSFT Microsoft Corporation 50 42.0% -0.21% 53 37.7% -0.32%
PFE Pfizer Inc. 49 34.7% -5.17% 58 37.9% -0.25%
PG The Procter & Gamble Company 40 35.0% 8.53% 44 31.8% 0.02%
TRV The Travelers Companies, Inc. 20 60.0% 9.86% 50 46.0% -0.11%
UTX United Technologies Corp. 54 44.4% 13.23% 50 48.0% 0.24%
VZ Verizon Communications Inc. 44 54.5% 0.03% 47 42.6% -0.08%
WMT Wal-Mart Stores Inc. 48 52.1% 0.14% 47 44.7% 0.17%
DIS The Walt Disney Company 47 51.1% 0.34% 52 38.5% 0.01%
Average 47.6 45.1% 0.72% 50.4 42.0% -2.64%
Standard Deviation 7.3 7.7% 4.96% 5.4 0.06% 6.47%
Totals 21.00% -76.64%

The average percentage profit and loss on SMM volume trades was 0.72% compared to -2.64% for share volume trades. SMM trades
had a lower average standard deviation on the P&L returns at 4.96% from 6.47%. However, these standard deviations are still high,
the data in Table 4 details how much of the P&L for both OBV and OBSV comes from just a handful of stocks. OBV and OBSV trading

PAGE 56 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

signals are essentially trend-following in nature. Figure 7 is an Climax Indicator

overlay graph of United Technologies Corp (UTX) and Exxon Granville suggested that once the net fields of the stocks in
Mobil (XOM) graphs. XOM was the worst performing stock in the index had been calculated, they could be totalled to form
OBV signals. The graph details how XOM range-traded around a net cumulative ‘Climax Indicator’, which in a single figure,
$90 for much of 2013. This caused serial whipsaws in net field gave a reading on the overbought/oversold nature of the DJIA.
calculations. In comparison, UTX posted a strong bull trend A reading of +30, for example, would occur when all 30 stocks
from the lows of June and was one of the strongest performers are in rising fields. He suggested that a cluster of levels over
in OBV, and was the strongest in the OBSV. +18 signalled an approaching market top, and a cluster of levels
The correlation of average returns between OBV and OBSV under -16 signalled an approaching market bottom.
signals was -0.12088. Also of interest is how the maximum Table 6 details the results. The OBV climax index saw 13 days
loss in OBSV was -7.63 (DD), with just three other stocks at -5 at -16 or under, with each signal followed by a 1% gain on average
(CSCO, CVX and PFE). Traditional OBV, in comparison, had three in the DJIA 10 days later, compared with only three days under
stocks with worse than 10% in losses (CAT, CVX and XOM) and a -16 for the OBSV climax indicator, with an average of 0.43%. The
maximum loss of -25.64%. This data shows that OBSV compares OBSV also trailed the OBV for the highs in the climax indictor. Of
well to OBV in terms of both returns and standard deviations. interest is how, over this period, after both extremes, the market
Table 5 displays the information for the DJIA Index itself, moved higher. According to Granville’s suggested overbought
where DJIA Index closing prices were used and volume data on guidelines, we would have expected to see the DJIA on average
the individual constituents was cumulated into one volume fall after readings of above +18. However, the average climax
data series. This generated 27 OBV trade signals with a win/ indicator readings were positive for both data series, 1.89 and
loss percentage of only 29.6% and cumulative loss of 18.22%. 1.18, reflecting the bullish underlying market over the period.
For OBSV, however, the loss improved to -1.26%, and with
improvements in the win/loss percentages to 45.5%. However, Table 6: OBSV and OBV Climax Indicator Results
this performance was less than the average seen in the OBSV OBV
constituents. To calculate the DJIA SMM volume data daily, I Average Climax Indicator 1.89 1.18
totalled all of the individual stock daily SMM volumes into a Standard Deviation of Climax Indicator 7.52 8.82
single cumulative data series.
Max Climax Indicator Over Period 21 22
Table 5: OBSV and OBV Net Field Signals for DJIA Index Min Climax Indicator Over Period -17 -20
OBSV Net Field Signals OBV Net Field Signals Days Climax Indicator 16 or Under 3 13
Trades % Win % P&L Trades % Win % P&L
Days Climax Indicator 18 or Over 2 3
DJIA
Constituents 47.6 45.1% 0.72% 50.4 42.0% -2.64% Total % Moves 10 Days After 16 or Under 1.28% 13.06%
Average
Average % Moves 10 Days After 16 or Under 0.43% 1.00%
DJIA Index 55.0 45.5% -1.26% 27 29.6% -18.22%
Total % Moves 10 Days After 18 or Over 1.93% 3.31%
Average % Moves 10 Days After 18 or Over 0.96% 1.10%

Figure 7: Exxon Mobil, United Technologies, Comparison Graph

Table 4 detailed that OBSV
had a better period than
OBV for the individual
constituents using net field
trends. Table 6 details that
averaging the constituent’s
data underperforms. This
details that the methodology
of collating SMMs on specific
individual stock mentions
filters out some of the wider
macro market influences,
giving SMM data a higher
correlation to the company’s
specific risk.
Specific risk is the risk
relating directly to the
company, such as the risk of
bankruptcy, whereas market
risk is the risk that share

IFTA.ORG PAGE 57
 IFTA JOURNAL 2015 EDITION

prices can be affected by macro market events, such as wars. characterised by more significant spikes of activity. These SMM
Traditional volumes will be affected by market risks to a higher spikes understandably often coincide with news events, which
degree than SMM volumes. News of U.S. missile strikes in Syria, can either be directly related to share prices, such as profit
for example, would affect stock market prices, and as a result, warnings, or related to more generic chatter, such as Wal-Mart
traditional stock volumes, and through index tracking and Thanksgiving Day sales promotions.
other market correlations, this event would impact prices and It was necessary to investigate whether SMM volume spikes
volumes at Walt Disney, for example. However this macro event compare with share volume spikes and precede price moves, as
is unlikely to greatly alter SMMs directly tied to Walt Disney. So, Gervais suggested, or whether SMM volume data moves more
while the OBV net fields may be cumulated to form a cumulative in line with traditional noise player sentiment surveys and acts
climax index, the outlook for the OBSV is less compelling. as a contrarian measure. To filter for volume spikes, a 30-day
To create a wider cumulative market sentiment reading, the volatility filter was created. Volumes above this filter signalled
collection of the SMMs would need to be widened beyond stock- a spike.
specific mentions and focus on more general terms, such as Volume Volatility Extreme = (30-Day Moving Average) + (30-
bullish, optimistic, bearish and nervous. Day Standard Deviation x 2)
Table 7 details that after SMM volume spikes the price over
the following 10 days on average gained 5.02%, compared to
The High-Volume Return Premium 3.8% for traditional share volume spikes, albeit with a slightly
In The High-Volume Return Premium (2001), Simon Gervais higher standard deviation, in line with the findings of Da,
found that stock prices for stocks with unusually high trading Engelberg and Gao, 2011.15
volumes over a day or week tended to appreciate over the course Also of interest is the performance of volume spikes on the
of the following month.28 The following section investigates DJIA itself (bottom of Table 7). On the individual constituents,
whether this is true for SMM volumes. Share volumes have the average number of SMM volume spikes is 7.1; however,
traditionally been seen as a summary of informed investor this drops to 2 for the parent index, while the share volume
sentiment; SMM volume could be described as summarising spike on the DJIA was 9 against an average of 10.1 for the
noise player sentiment. Noise player sentiment survey constituents (no significant change). This is in line with the
measurements have traditionally been used as contrarian data in Table 6, showing that SMM data has a higher focus
indicators.2 The next section looks into volume spikes in both on the stock-specific risk over market risk. Stock specific
volumes and SMM volumes. risks by definition are reduced by diversification. This helps
Figure 8 shows normalised Wal-Mart share volumes and to explain why the profitability on OBSV net field signals is
SMM volumes. The correlation for Wal-Mart share and SMM higher than for OBV, as the data is more focused on sentiment
volumes over the period was 0.51. (The correlation for Wal- tied to the underlying security and less influenced by broader
Mart’s share price to the DJIA over this period was 0.679.) market forces.
The average volume and SMM volume correlation across the
constituents was 0.38, with a standard deviation of 0.24. Figure
8 illustrates this correlation and how the SMM volumes are

Figure 8: Wal-Mart Normalised Share Volumes and SMM Volumes Overlay Graph

PAGE 58 IFTA.ORG
 IFTA JOURNAL 2014 EDITION

Table 7: 10-Day Price Change Following OBV, OBSV Price Spikes

Volume Volatility Spike Price Change 10 Days After Spike
Joint Shares SMM Joint Shares SMM
MMM 3M Company 1 10 8 -1.05% 12.40% 5.18%
AA Alcoa 4 9 5 -6.71% -3.20% -6.60%
AXP American Express Company 1 13 11 -3.54% 8.42% -5.08%
T AT&T, Inc. 1 11 10 1.18% -9.55% 18.84%
BAC Bank of America 1 9 2 10.29% 26.27% 3.05%
BA The Boeing Company 3 9 9 -4.35% -5.70% 3.44%
CAT Caterpillar Inc. 5 10 7 -5.15% -5.21% -4.26%
CVX Chevron Corporation 0 7 1 0.00% -4.53% 0.03%
CSCO Cisco Systems, Inc. 5 12 6 12.03% -10.01% 28.95%
KO The Coca-Cola Company 2 8 6 6.37% 7.96% 10.10%
DD E.I. du Pont de Nemours and Co. 4 10 7 -1.88% -1.31% 14.53%
XOM Exxon Mobil Corporation 0 8 7 0.00% 6.15% -6.37%
GE General Electric Company 5 10 8 7.21% 9.28% 18.73%
HPQ Hewlett Packard 8 11 9 0.66% 6.81% 23.13%
HD The Home Depot, Inc. 5 11 9 7.90% 20.02% 27.32%
INTC Intel Corporation 4 10 8 -8.92% -18.41% -24.42%
IBM International Business Machines 4 11 6 7.48% 18.37% 3.92%
JNJ Johnson & Johnson 0 10 2 0.00% 8.07% 2.74%
JPM JPMorgan Chase & Co. 3 8 7 -8.87% -1.52% -14.84%
MCD McDonald’s Corp. 7 13 11 1.11% 6.59% 8.00%
MRK Merck & Co. Inc. 0 8 2 0.00% 4.58% -6.05%
MSFT Microsoft Corporation 2 9 7 -1.07% -2.94% -4.48%
PFE Pfizer Inc. 1 13 4 -2.56% 14.04% 0.47%
PG The Procter & Gamble Company 4 13 6 2.09% 8.29% 0.77%
TRV The Travelers Companies, Inc. 0 9 10 0.00% -1.73% 9.03%
UTX United Technologies Corp. 1 10 12 -0.97% -13.64% 1.91%
VZ Verizon Communications Inc. 2 12 8 6.11% -2.23% 8.03%
WMT Wal-Mart Stores Inc. 7 10 10 19.90% 25.25% 16.97%
DIS The Walt Disney Co. 2 10 9 -3.77% 7.76% 12.53%
Average 2.8 10.1 7.1 1.15% 3.80% 5.02%
Standard Deviation 2.3 1.7 2.9 6.52% 11.07% 12.38%

DJIA Dow Jones Industrial Average 1 9 2 -2.85% -3.07% -1.60%

Analysis volumes, as the trend-following and company-specific nature

The use of SMM data compares well to traditional volume, of SMM volume data was reduced when the data was combined
albeit with a higher standard deviation. When using the OBV net into a single dataset for the whole market.
field methodology, the number of trades signalled by SMM data The calculations on volume spikes detailed did not take
has a higher standard deviation in profit and loss returns, but account of the direction of price moves on the day of the
investors are rewarded for taking on this additional risk with volume increases, so market impact of positive and negative
higher returns. sentiment may be asymmetric.29 This potential bias would need
Volume precedes price theories were not contradicted; to be investigated further. Over the period covered, the DJIA
however, the idea that noise player sentiment is best used as gained 16.5%. I would also like to test OBSV in neutral market
a contrarian indicator was not confirmed by the SMM volume, conditions and in a bearish phase of the market. However,
as it did not act as contrarian indicator. Instead, it mirrored due to the infancy of the field, this is not yet possible, as the
traditional volumes by confirming and preceding price moves. data provided by Knowsis is all the data from the inception of
The investigation into volume spikes suggests that SMM and their services. The relatively short time series will impact the
volume data had positive correlations in spike activity, and reliability of the results. However, the numbers using social
that gains following spikes were 10.1% and 7.1% on average media have increased exponentially, so comparisons made over
for volumes and SMM volumes. Cumulative volumes have longer timeframes would also raise issues of their own.
traditionally been used to help predict market moves on the A number of questions on the source and calculation of
parent index. Totalling SMM volumes to form a climax indicator, SMM data remain. SMMs on the DJIA are not regulated by the
however, did not create superior returns over traditional SEC; therefore, any data derived from them is vulnerable to

PAGE 59 IFTA.ORG IFTA.ORG PAGE 59

 IFTA JOURNAL 2015 EDITION

more errors, lobbying and outright fraud than traditional share SMM information occurs, an opportunity exists for data miners
trading volumes. For example, on 23 April 2013, more than $135 to capture relevant market information on well capitalised
billion was wiped off the S&P 500 for a short period intraday stocks. As a result, the level of sophistication and focus on SMM
after the Associated Press Twitter account was hacked and a analysis will continue its rapid expansion. Volume analysis has
fake tweet about an explosion at the White House was posted. been widely used in financial markets for over 100 years, and On
Also, Mary C. Joyce in 2010 described “Twitterbombing”, where Balance Volume for over 50 years. As social media technology
campaigners bombard social media with phrases to raise continues to evolve, the likelihood is that SMM data will also
awareness of their particular cause, which can artificially inflate emerge to become a significant new field in technical analysis.
SMM data.30
Investors looking to use SMMs also need to decide how they Notes
search for text. Some data providers only search for SMMs that 1. Murphy, John J. Technical Analysis of the Financial Markets. NYIF, 1999. p. 1.
include the ticker code of the stock in the posting; this method is 2. Kirkpatrick, Dahlquist, Technical Analysis: The Complete Resource for Financial
Market Technicians. FT Press. 2007. p. 3, p. 92.
employed at Bloomberg. The tweet from Carl Icahn on Apple, for 3. Kahneman, Dl; Tversky, A. “Prospect Theory: An Analysis of Decision under
example, did not include the Apple ticker as a hashtag so would Risk.” Econometrica March 1979. (47:2): p. 1.
have been missed using this method. Also, careful attention 4. Kahneman, D: Tversky, A. “Judgment under uncertainty: Heuristics and
biases”. Science, 1974 (V185): 1124–1130.
need to be made on which social media platforms to include, as 5. Brown, C. Technical Analysis for the Trading Professional, McGraw Hill, Second
the marketplace is rapidly evolving. Edition, 2012. p. 405.
The constituents of the DJIA are among the most stable of 6. Lottolab, Illusions: http://www.lottolab.org/articles/illusionsoflight.asp.
Viewed October 10, 2013.
the major global indices, but the index does still have occasional 7. Hill, Theodore P. “The First Digit Phenomenon”. American Scientist, July–
constituent changes, making cumulative SMM data difficult to August 1998 (V86:4): 358.
compare over longer time periods without adjusting for these 8. Redelmeier, Donald A., M.D.; Shafir, Eldar, Ph.D., “Medical Decision Making
in Situations That Offer Multiple Alternatives.” The Journal of the American
changes. Medicine Association (JAMA). January 25, 1995 (V273:4): 302-305.
Most DJIA constituents are household names (e.g., 9. Mussweiler, Thomas; Schneller, Karl. “What Goes Up Must Come Down: How
McDonalds, Intel, Hewlett Packard, IBM), and the power Charts Influence Decisions to Buy and Sell Stocks” The Journal of Behavioral
Finance 2003, (Vol. 4, No. 3): 121–130.
of brand awareness and brand identity is an issue, as SMM 10. Tymula, Agnieszka, et al. “Thirst-dependent risk preferences in monkeys
volumes on Hewlett Packard are considerably higher than those identify a primitive form of wealth”. Center for Neural Science Proceedings of
on UnitedHealth, for example. UnitedHealth, as of October 2013, the National Academy of Sciences of the United States of America, PNAS, 2013
(1308718110v1-201308718).
had a market capitalisation of $73bn and 133,000 employees 11. Shiller, Robert J. Irrational Exuberance. Princeton University Press. 2000.
and is the largest healthcare coverage provider in the United 17–148.
States, yet its brand is barely known among the general public. 12. Baker, Malcolm; Wurgler, Jeffrey. “Investor Sentiment in the Stock Market.”
Journal of Economic Perspectives. Spring 2007 (V21:2): 129–151.
As a result, it has few SMMs. This effect will create a brand 13. Fisher, Black. “Noise.” The Journal of Finance, December 30, 1985 (V41:3):
awareness bias to SMM data if not correctly adjusted and may 529–543.
result in analysis of SMM data in the future focusing on the 14. De Long, J. Bradford, et al. “Noise Trader Risk in Financial Markets.” The
Journal of Political Economy, August, 1990 (V98:4): p. 703.
company names with strong brand awareness. 15. Da, Zhi; Engelberg, Joseph; Gao, Pengjie. “In Search of Attention.” The Journal
SMM data calculations are not standardised. If Bloomberg of the American Finance Association, September 2011 (V66:5):1461–1499.
does start to publish SMM time series data, it may become the 16. Gidófalvi, Győző. “Using News Articles to Predict Stock Price Movements”,
University of California, Department of Computer Science and Engineering,
most widely used data source due to its prevalence amongst the June 15, 2001.
financial professionals, but not all in the industry agree with its 17. Schumaker, Robert P; Chen, Hsinchun. “A quantitative stock prediction
approach. Even with some form of standardisation on how SMM system based on financial news.” Information Processing & Management.
September 2009 (V45:5): 571–583.
data is calculated, it is unlikely ever be regulated to the same 18. Edmans, Alex; Garcia, Diego. “Sports Sentiment and Stock Returns.” The
extent as share volumes. As a result, SMM data is set to remain Journal of Finance, August 2007 (V62:4): 1967–1998.
secondary to official price and volume data. 19. US Securities & Exchanges Commission, SEC, News Release: 1365171513574.
April 2nd, 2013 http://www.sec.gov/News/PressRelease/Detail/
PressRelease/1365171513574#.UlKRWFDWR51.
Conclusion 20. Foodborne Chicago: http://foodborne.smartchicagoapps.org/. Viewed
Despite all the issues with the collection of SMM data, the October 10th, 2013.
21. Asur, Sitaram; Huberman, Bernardo. “Predicting the Future with Social
outlook for this emerging field in technical analysis looks Media.” HP LABS, 2010.
intriguing. The intention of this paper was to investigate the use 22. Gruhl, D; Guha, R; Kumar, R; Novak, J; Tomkins, A. “The Predictive Power of
of social media mentions in equity selection. This was achieved Online Chatter.” Conference on Knowledge Discovery and Data Mining (KDD),
2005.
by comparing the proposed On Balance Sentiment Volume 23. Bollen, Johan; Mao, Huina. “Twitter Mood as a Stock Market Predictor”. IEEE
indicator with the established On Balance Volume, and also Computer, October 2011. (V44:10): 91–94.
analysing the extremes in volume and SMM volume data. Social 24. Granville, Joseph E. New Key to Stock Market Profits. Martino Publishing, 2010.
25. Dormeier, Buff Pelz. Investing with Volume Analysis. FT Press. 2011.
media mention volumes performed well against share volume 26. Martsein, David. How to Chart Your Way to Stock Market Profits. Arco, 1966.
data in terms of correlations, net field signals and volume 27. Kaufman, Perry J. New Trading Systems and Methods, 5th Edition, 2005. p. 537.
volatility extremes. Perhaps most interesting was that SMM 28. Gervais, S; Kaniel, R; Mingelgrin D. “The High-Volume Return Premium” The
Journal of Finance, (V56:3): 2001.
volumes closely matched share volumes in periods of volume 29. Akhtar, Shumi; Faff, Robert; Oliver, Barry; Subrahmanyam, Avanidhar. “Stock
extremes, preceding price moves and not acting as a contrarian salience and the asymmetric market impact of consumer sentiment news.”
indicator as seen with other sentiment measures. Journal of Banking & Finance, December 2012, (V36:12): 3289–3301.
30. Joyce, Mary C. Digital Activism Decoded: The New Mechanics of Change.
The results indicate that before market standardisation of International Debate Education Association, 2010.

PAGE 60 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Software and Data Appendix – Excel Formulas

Image from http://richardwiseman.wordpress.com/2013/09/26/amazing- Column A: Date
perspective-illusion/. Author Richard Wiseman, Professor of the Public Column B: Close
Understanding of Psychology at the University of Hertfordshire; Posted
September 28, 2013. Image copied October 8, 2013. Column C: Volume/ SMM Volume
Image from http://www.lottolab.org/illusiondemos/Demo%2016.html. Lottolab Column D: On Balance Volume/On Balance Sentiment
Studio, “Brightness Illusions 10”; Author Beau Lotto. Image copied October 8, 2013. Cell D2: =C2 Cell D3: =IF((B2=B3),D2,IF(B3>B2,(D2+C3),(D2-C3)))
If the close today equals the close yesterday, then do nothing, as the
OBV/OBSV value does not change because the day’s volume cannot
References be attributed to a direction. If the price today closed higher, then
add the day’s volume to yesterday’s OBV/OBSV value; Otherwise,
Asur, Sitaram; Huberman, Bernardo A. “Predicting the Future with Social the price was down on the day, and subtract the day’s volume from
Media.” HP LABS, 2010. yesterday’s OBV/OBSV value.
Akhtar, Shumi; Faff, Robert; Oliver, Barry; Subrahmanyam, Avanidhar. “Stock To calculate simplified OBV/OBSV Net Field Trends, these additional columns are
salience and the asymmetric market impact of consumer sentiment news.” required:
Journal of Banking & Finance, December 2012, (V36:12): 3289–3301. Column E: OBV/OBSV Peak Line: Cell E4: =IF(IF(AND((D3>D4),(D3>D2)),1,0)=1,
Baker, Malcolm; Wurgler, Jeffrey. “Investor Sentiment in the Stock Market.” D3,E3)If D3 is greater than D4, and D3 is greater than D2, then D3
Journal of Economic Perspectives. Spring 2007 (V21:2): 129–151. is a three day peak, in which case the OBV/OBSV value of that day
Bollen, Johan; Mao, Huina. “Twitter Mood as a Stock Market Predictor”. IEEE is recorded in Column E, if not then the previous OBV peak value is
Computer, October 2011. (V44:10): 91–94. maintained.
Brown, Constance. Technical Analysis for the Trading Professional, McGraw Hill, Column F: OBV/OBSV Trough Line: Cell F4: =IF(IF(AND((D3<D4),(D3<D2)),-1,0)=-
Second Edition, 2012. 1,D3,F3)Similar calculation to Column E, but it looks for trough
values.
Da, Zhi; Engelberg, Joseph; Gao, Pengjie. “In Search of Attention.” The Journal of
the American Finance Association, September 2011 (V66:5):1461–1499. Column G: Peak Line Moving Higher or Lower: Cell G4: =IF(E3=0,0,IF(E4<E3,”-”,
IF(E4>E3,”+”,G3))) Measures the last direction of the peak line,
Dormeier, Buff Pelz. Investing with Volume Analysis. FT Press. 2011. positive or negative.
Edmans, Alex; Garcia, Diego. “Sports Sentiment and Stock Returns.” The Journal Columns H: Trough Line Moving Higher or Lower: H4 =IF(F3=0,0,IF(F4<F3,”-
of Finance, August 2007 (V62:4): 1967–1998. ”,IF(F4>F3,”+”,H3))) Measures the last direction of the trough line,
Evans, Stuart. “On Balance Volume.” Stocks & Commodities, (V17:5): 234–240. positive or negative.
Faber, Bruce. “On Balance Volume Indicator.” Stock & Commodities, (V12:7): Column I: Net Field Trend: Cell I4: =IF(AND(H4=”+”,G4=”+”),”RISING”,IF(AN
303–305. D(H4=”-”,G4=”-”),”FALLING”,””)) If both peak and trough lines are
positive, this equals a Net Rising Field. If both peak and trough
Fisher, Black. “Noise.” The Journal of Finance, December 30, 1985 (V41:3):
lines are negative, it equals a Net Falling Field. Different peak and
529–543.
trough directions gives no Net Field These net field calculations
Gervais, S; Kaniel, R; Mingelgrin D. “The High-Volume Return Premium.” The are a simplified version of the methodology described by Granville.
Journal of Finance, (V56:3): 2001. Readers wishing to precisely replicate Granville’s net field
Gruhl, D; Guha, R; Kumar, R; Novak, J; Tomkins, A. “The Predictive Power of Online methodology should refer to his instructions.
Chatter.” Conference on Knowledge Discovery and Data Mining (KDD) (2005). Joint Stock and SMM Volume
Gidófalvi, Győző. “Using News Articles to Predict Stock Price Movements”, Column A: Share Volumes:
University of California, Department of Computer Science and Engineering, June
Column B: SMM Volumes
15, 2001.
Column C: Joint Volumes Cell C1: =A1+(B1*(AVERAGE(A:A)/ AVERAGE(B:B)))
Granville, Joseph E. New Key to Stock Market Profits. Martino Publishing, 2010.
Volume Volatility Extremes
Hill, Theodore P. “The First Digit Phenomenon”. American Scientist, July–August
1998 (V86:4): 358. Column A: Volumes:
Joyce, Mary C. Digital Activism Decoded: The New Mechanics of Change. Column B: Cell B30: =IF(A30>(AVERAGE(A1:A30)+(2*(STDEV(A:A30)))),1,0)
International Debate Education Association, 2010.
Kahneman, Dl; Tversky, A. “Prospect Theory: An Analysis of Decision under
Risk.” Econometrica, March 1979. (47:2): 263–291.
Kahneman, D; Tversky, A. “Judgment under uncertainty: Heuristics and biases”.
Science, 1974 (V185): 1124–1130.
Kaufman, Perry J. New Trading Systems and Methods, 5th Edition, 2005.
Kirkpatrick, Dahlquist, Technical Analysis: The Complete Resource for Financial
Market Technicians. FT Press. 2007.
De Long, J. Bradford, et al. “Noise Trader Risk in Financial Markets.” The Journal
of Political Economy, August, 1990 (V98:4): 703–738.
Martsein, David. How to Chart Your Way to Stock Market Profits. Arco, 1966.
Murphy, John J. Technical Analysis of the Financial Markets. New York Institute of
Finance, 1999.
Mussweiler, Thomas; Schneller, Karl. “What Goes Up Must Come Down How
Charts Influence Decisions to Buy and Sell Stocks” The Journal of Behavioral
Finance, 2003, (V4:3): 121–130.
Pring, Martin. Technical Analysis Explained, McGraw Hill, Fourth Edition, 2002.
Pring, Martin. Investment Psychology Explained Classic Strategies to Beat the
Market. Wiley, 1993.
Redelmeier, Donald A. M.D.; Shafir, Eldar Ph.D., “Medical Decision Making in
Situations That Offer Multiple Alternatives.” The Journal of the American Medicine
Association (JAMA). January 25, 1995 (V273:4): 302–305.
Schumaker, Robert P; Chen, Hsinchun. “A quantitative stock prediction system
based on financial news” Information Processing & Management. September 2009
(V45:5): 571–583.
Shiller, Robert J. Irrational Exuberance. Princeton University Press, 2000.
Tymula, Agnieszka, et al. “Thirst-dependent risk preferences in monkeys
identify a primitive form of wealth”. Center for Neural Science Proceedings of
the National Academy of Sciences of the United States of America, PNAS, 2013
(1308718110v1-201308718).

IFTA.ORG PAGE 61
 IFTA JOURNAL 2015 EDITION

Enhancing Portfolio Returns and Reducing David Price

[email protected]

Risk by Utilizing the Relative Strength Index 31 Gresham Street

London EC2V 7QA
as a Market Trend Identifier United Kingdom
+44 (0)20 7658 6940
By David Price BSc, MSc, CFTe, MSTA, MFTA

Note: This paper is covered by Creative Commons License The traditional concept of RSI
Attribution-NonCommercial 3.0. The RSI indicator, developed by J. Welles Wilder, is one
of the most popular indicators within the field of technical
Abstract analysis. This oscillator measures relative strength by
The Relative Strength Index (RSI), developed by J. Welles comparing the magnitude of recent gains to recent losses,
Wilder, is one of the most popular tools within technical which makes it a momentum indicator. The RSI value will
analysis. RSI is conventionally understood to represent when always oscillate between 0 and 100; the RSI value will be 0 if
a market is either overbought or oversold, providing sell or the market has fallen in all the periods in the calculation and
buy signals, respectively, to investors in the process. Used 100 if the market rises in all of those periods.
in this way, the RSI appears to function best as an indicator Should you pick up any technical analysis textbook and
when a market is ranging and is less effective as an indicator look at its explanation of RSI, it will almost certainly focus
when a market is trending. solely on how the indicator creates investment signals at the
Continuing on from Wilder’s original model, in more overbought end of the range (a reading of 70 or over) or at the
recent times, it’s been proposed that the RSI can also identify oversold end (a reading of 30 or below), which was Wilder’s
the trend of a market (e.g., whether it is in a bullish phase original consideration. A limitation of the RSI indicator used
or a bearish one). If this is indeed the case, it could present in this way is that it functions with more clarity in ranging
relevant buy and sell signals during market trends, enabling markets (where it can provide buy or sell signals at bottoms
investors to better preserve capital during bear markets and and tops, respectively) than it does in trending markets.
achieve greater capital appreciation during bull markets. Michael Kahn1 provides a useful analogy for
This research paper, therefore, seeks to test the practical understanding RSI in the traditional manner. A car has both
benefits to investors of utilizing RSI as a market trend an accelerator pedal and a brake. When you are driving and
identifier. Through backtesting the implementation of the RSI start applying the brakes, the car will slow down. However,
in this way as an investment strategy, this paper will analyze unless there is enough force applied to the brakes the car
what benefits this approach provides to a portfolio’s levels will still continue to move forward. The same goes for the
of return and risk, compared to a portfolio that was fully markets, when buying pressure outweighs selling pressure
invested during the same time period. The performance of RSI the accelerator is applied and RSI increases. When selling
as a market trend identifier will be summarized in this paper. pressure outweighs buying pressure the brakes are engaged
and RSI decreases. We will return to this analogy shortly.
Introduction
Since the turn of the millennium, there has been no An alternative use of RSI
shortage of eventful episodes that have affected the Constance Brown2 challenged the traditional concept
world’s markets. Investors within equity markets during of RSI, arguing that there was an alternate way of reading
this turbulent time will generally have seen their portfolio its values. By reading RSI values in her proposed manner,
perform poorly. Institutional investors who had sought the indicator would reveal two specific market phases. She
to generate returns on their assets are likely to have asserted that where a market was experiencing a downtrend
experienced a “lost decade” or more, had they invested on a (a bear market phase) the RSI level of 60 would act as a
buy and hold basis, as markets have transitioned between level of resistance. Should the RSI value rise up to reach this
bull and bear phases and ultimately not appreciated much level, it would then weaken, confirming the resumption of
(if indeed at all) in that time. More tactical investors who the downtrend in the market. In an uptrend (a bull market
will have tried to time their investments may have fared phase), she believed the inverse was true, with the RSI
better; however, with the high level of volatility in markets, level of 40 proving to be a level of support. Should the RSI
this approach will have also generated opportunities for value fall to this level, it would then strengthen, confirming
uncompensated risk. the resumption of the market’s uptrend. If this is correct,
What indicator within the technical analysis panoply then the significance of the traditional 30 and 70 oversold/
would have enabled those investors to better navigate the overbought marks would be limited, as those levels would not
turbulent market conditions? Let us consider the utilization be reached, depending on what phase the market was in.
of the Relative Strength Index (RSI) as one such tool. Based on this understanding, the illustration that follows
demonstrates how RSI values can be interpreted as to

PAGE 62 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

whether they are in a bull market or a bear market phase. market when RSI is identifying that it is not in a definitive
You’ll see from the illustration that where the RSI value stays bear phase (based solely on its own indication), we can build
above 40, the market would be considered to be in a bull upon Kahn’s analogy that was referenced earlier. When RSI is
market phase, whereas if the RSI is below 60, this signifies equal to or above 40, RSI is able to fluctuate between using the
that the market is in a bear market phase. accelerator and brakes, and the invested portfolio will gain the
return of the market. When RSI is below
40 we use an additional instrument in our
metaphorical car: the handbrake. Adding to
Kahn’s analogy then, an override mechanism
kicks in that has much the same effect
as a handbrake. The market investment
comes to a complete halt, neither benefiting
from a rise in the market nor deteriorating
where markets are falling. It bides its time
until the market gives a non-bear market
phase indication (i.e., an RSI value of 40)
Evidently, these two market phases are not mutually before releasing the handbrake and letting the combination of
exclusive. What do we make of an RSI value of 50, which would accelerator and brakes have their freedom back.
be within both phases? To understand that value and which
phase the market is in, we would need to know where the RSI Methodology for the RSI market
has come from and to where it is going. The latter part of that identifier backtest
statement is (theoretically) impossible to know with certainty;
the market will go where it goes, and only future RSI readings To test the hypothesis that RSI can be used utilized as a
will tell us for sure which phase it is in. By that point, it might be market identifier, and whether its readings provide an investment
too late to invest accordingly. approach that enhances portfolio returns and reduces risk, the
What we can take from this model is that there are two strategy is backtested in three different markets:
mutually exclusive parts: 1) an RSI value over 60 signifies a
clear bull market phase; and 2) an RSI value below 40 signifies a ƒƒ The S&P 500 Index – the U.S. stock market index of the 500
clear bear market phase. By interpreting Brown’s alternate RSI leading companies by market capitalization.
proposition in this way, we can establish more clarity as to the ƒƒ The FTSE 100 Index – the UK stock market index of the 100
phase of the market and probable direction (assuming the trend leading companies by market capitalization.
has not come to an end). ƒƒ The Nikkei 225 Index – the Japanese stock market index of
Investors who wish to generate capital appreciation by being the 225 leading companies by market capitalization.
long the market, we can assume, will be content for RSI to be
60 or over as this should be signaling a clear bull market phase The methodology applied for the market identifier strategy
and appreciation in market levels. The issue therefore is: what is as follows. Using the standard 14-period settings for the RSI,
happens when RSI is below 40 and they are invested in a clear a weekly value was calculated at the end of the week. Where the
bear market and depreciating market levels? value was 40.0 or above, the portfolio invested into the stock
The answer to this question, and the use of RSI as a market market at the next opening price. The portfolio therefore gained
identifier, is lacking from technical analysis publications. A the returns of the markets. Where the value was below 40.0, the
thorough review of all available IFTA Journals and Society of portfolio exited the stock market at the next opening price and
Technical Analysts’ Market Technician journals since Brown’s did not receive the returns of the market (or returns from any
book was published in 1999, and many of the established reading other source).
books, reveal that RSI has not been further researched and In the remainder of this paper, the strategy based on the RSI
considered in the way she prescribes. Much research has been market identifier will be called the ‘RSI MI40+’ for brevity, while
conducted regarding RSI, including the impact of applying the investment strategy for comparison purposes is a buy and
it with other technical indicators and considering the use of hold approach in each stock market.
the mid-level 50 RSI value as a complementary investment The backtest in each market was made with a theoretical
signal, but I could find no research that utilized RSI as a market starting balance of US$100 million. This would be a reasonable
identifier as Brown provides, nor of any research into the effect amount for an institutional investor to allocate into a single
of avoiding investing in clear bear market phases as identified country’s largest stocks and provides a sufficiently large size
by the alternate use of RSI. to demonstrate the deviance in performance between the
The objective of this paper is therefore to understand if RSI MI40+ strategy and the buy and hold strategy. Where the
the RSI indicator could be used in a systematic way that would portfolio is invested in foreign markets, the effect of changes
improve capital appreciation and reduce risk, by ensuring the in foreign exchange rates was not taken into account, as this is
portfolio was not invested during clear bear market phases, as outside the scope of what this paper tested for.
identified by the alternate use of RSI. The time period for the backtests is from 1 January 2000 to 31
By using RSI to keep a portfolio only invested within the December 2012. This time period provides a rich sample of volatile

IFTA.ORG PAGE 63
 IFTA JOURNAL 2015 EDITION

markets, with many natural and market-related events. There hold approach suffered from being fully invested during the
has been the ‘dotcom bust’, 9/11, the market rebound following pronounced market slide.
the Iraq War, the London bombings, Hurricane Katrina, the global Table 1a reveals the extent of the outperformance of the RSI
financial crisis and collapse of Lehman Brothers, Quantitative MI40+ strategy. Whereas the buy and hold approach overall saw
Easing and the Fukishima disaster. These events have provided a deterioration in the initial portfolio, the RSI MI40+ strategy
threat and opportunity alike to investors’ portfolios. produced a cumulative return in excess of 13%. At the same
time, the volatility of the returns was smaller at 1.95% compared
The results of the RSI market identifier to 2.68% for the buy and hold approach. This improvement
backtests in risk reflected the benefit of the portfolio not always being
invested during significant changes in the market index, most
U.S. Equities: The S&P 500 notably between 2007–2009. Overall, the RSI MI40+ strategy
As can be seen in Chart 1, the RSI MI40+ strategy applied to was invested in 84.6% of the time periods in the backtest,
the S&P 500 outperformed the buy and hold approach over the meaning it outperformed and reduced risk despite being
period under review. By the time the market decline following invested within the market for a substantial amount of the
the dotcom bust had bottomed out and turned upwards in 2003, possible time.
the two strategies were achieving much the same portfolio Table 1b illustrates the effect that the RSI MI40+ strategy
value. In the boom markets of 2003–2007, the buy and hold had on the S&P 500 portfolio in monetary terms. The minimum
approach was generating an enhanced return due to the strong absolute value (the ‘trough’) reached by the RSI MI40+ strategy
uptrend. However, where the global financial crisis started was $6.5 million higher than for the buy and hold approach,
in 2007 and experienced a significant fall until 2009, the RSI while the maximum value (the ‘peak’) it reached was almost
MI40+ strategy began to outperform. The strategy preserved $8 million higher. By the end of the period under review, the
capital better during this timeframe, whereas the buy and RSI MI40+ strategy yielded an increased value of $14.1 million
relative to the buy and hold
approach, a significant
Chart 1: Evolution of performance for the two strategies improvement.

UK Equities: The FTSE 100

In Chart 2, we can see
that the RSI MI40+ strategy
applied to the FTSE 100
outperformed the buy and
hold approach over the
period under review. The
two strategies evolved with
little differentiation in the
first couple of years. By
mid-2002, as the market’s
downturn became more
pronounced, the RSI MI40+
Table 1a: Performance of the two strategies strategy began to outperform
by preserving capital. In the
Since inception Since inception subsequent boom market of
Investment Volatility of Time invested
cumulative annualized
strategy returns (%) (%) 2003–2007, the RSI MI40+
return (%) return (%)
S&P 500 RSI strategy returned much the
13.04 0.95 1.95 84.6 same as the buy and hold
MI40+
S&P 500 buy and approach, albeit with a higher
-1.06 -0.08 2.68 100.0
hold level of capital with which
to build upon. As the global
financial crisis started in
Table 1b: Monetary performance of the two strategies 2007, both approaches were
Minimum Maximum experiencing deteriorations
Final value
Investment absolute absolute Base capital Base capital in capital, although the RSI
reached
strategy value value to trough (%) to peak (%)
($ million) ($ million) ($ million) MI40+ strategy avoided the
S&P 500 RSI
worst of the falls and caught
53.9 116.2 -46.11 16.18 113.0 the market’s subsequent turn
MI40+
S&P 500 buy back up in 2009.
47.4 108.3 -52.59 8.35 98.9
and hold

PAGE 64 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Chart 2: Evolution of performance for the two strategies

Table 2a highlights the outperformance
of the RSI MI40+ strategy. Whereas the
buy and hold approach saw a deterioration
in the initial portfolio of 9.3%, the RSI
MI40+ strategy produced a cumulative
return in excess of 6.5%. At the same time,
the volatility of the returns was smaller at
1.88% compared to 2.60% for the buy and
hold approach. This improvement in risk
reflected the benefit of the portfolio not
being always invested during prolonged
deteriorations in the market index,
specifically 2002 and between 2008–2009.
Table 2a: Performance of the two strategies Overall, the RSI MI40+ strategy was
Since inception Since inception invested in 83.6% of the time periods in the
Investment Volatility of
cumulative annualized Time invested (%)
strategy return (%) return (%) returns (%) backtest, meaning it outperformed and
FTSE 100 RSI reduced risk despite being invested within
6.51 0.49 1.88 83.6 the market for a substantial amount of the
MI40+
FTSE 100 buy
-9.33 -0.75 2.60 100.0
possible time.
and hold Table 2b illustrates the effect that the
RSI MI40+ strategy had on the FTSE 100
Table 2b: Monetary performance of the two strategies portfolio in monetary terms. The minimum
Minimum Maximum Final value absolute value (the ‘trough’) reached by
Investment absolute value absolute value Base capital Base capital reached
strategy ($ million) ($ million) to trough (%) to peak (%) ($ million) the RSI MI40+ strategy was $14.5 million
FTSE 100 RSI higher than for the buy and hold approach,
68.2 123.9 -31.76 23.89 106.5 while the maximum value (the ‘peak’) it
MI40+
FTSE 100 buy reached was almost $20 million higher. By
53.7 104.4 -46.32 4.46 90.7
and hold the end of the period under review, the RSI
Chart 3: Evolution of performance for the two strategies
MI40+ strategy yielded an increased value
of $15.8 million relative to the buy and hold
approach, a significant improvement.

Japanese Equities:
The Nikkei 225
In Chart 3, the RSI MI40+ strategy
applied to the Nikkei 225 outperformed
the buy and hold approach over the period
under review. As the index deteriorated
from mid-2000 onwards, the RSI MI40+
strategy preserved capital relatively. As the
market bottomed in 2003, the RSI MI40+
Table 3a: Performance of the two strategies strategy was able to gain the appreciation
Since inception Since inception that the buy and hold approach was
Investment Volatility of benefiting from, albeit with a higher level
cumulative annualized Time invested (%)
strategy returns (%)
return (%) return (%) of preserved capital. Outperformance
Nikkei 225 RSI was maintained until 2007 and then,
-18.60 -1.57 2.29 72.6
MI40+ when the market began to deteriorate
Nikkei 225 buy
-46.42 -4.67 3.08 100.0 substantially between 2007–2009, the RSI
and hold
MI40+ strategy significantly preserved
capital. As the market entered a trading
Table 3b: Monetary performance of the two strategies
range between 2009–2012 the RSI
Minimum Maximum Final value
Investment absolute value absolute value Base capital Base capital MI40+ strategy maintained its overall
reached
strategy to trough (%) to peak (%) outperformance.
($ million) ($ million) ($ million)
Nikkei 225 RSI
53.4 117.4 -46.62 17.38 81.4
Table 3a demonstrates the extent of
MI40+ the outperformance of the RSI MI40+
Nikkei 225 strategy. The buy and hold approach saw a
39.4 112.3 -60.57 12.32 57.1
buy and hold marked deterioration of 46.4% in the initial

IFTA.ORG PAGE 65
PAGE 65 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

portfolio while the RSI MI40+ strategy The overall improvements in return and risk profiles for each of the markets is
only fell 18.60%. At the same time, the indicated by the green arrow shown in Charts 4a, 4b and 4c below.
volatility of the returns was smaller at
2.29% compared to 3.08% for the buy and Chart 4a: Improvement in return and risk for the S&P 500 backtest
hold approach. This improvement in risk
reflected the benefit of the portfolio not
being always invested during significant
changes in the market index, especially
the periods 2000–2003 and 2007–2009.
Overall, the RSI MI40+ strategy was
invested in 72.6% of the time periods in the
experiment, meaning it outperformed and
reduced risk despite being invested within
the market for a substantial amount of the
possible time.
Table 3b illustrates the effect that the
RSI MI40+ strategy had on the Nikkei 225
portfolio in monetary terms. The minimum Chart 4b: Improvement in return and risk for the FTSE 100 backtest
absolute value (the ‘trough’) reached by
the RSI MI40+ strategy was $14 million
higher than for the buy and hold approach,
while the maximum value (the ‘peak’) it
reached was over $5 million higher. By
the end of the period under review, the
RSI MI40+ strategy yielded an increased
value of $24.3 million relative to the buy
and hold approach, which is over half of
the remaining value of the buy and hold
portfolio and a substantial improvement.

Conclusion
The RSI MI40+ strategy outperformed Chart 4c: Improvement in return and risk for the Nikkei 225 backtest
the buy and hold approach within each of
the three markets that were the subject of
the backtest, experiencing less volatility
of its returns compared to a buy and hold
approach in the process. In both the case
of the S&P 500 and the FTSE 100, the
RSI MI40+ strategy produced absolute
gains for the initial portfolio size while
the buy and hold approach experienced
a loss of capital. In the case of the Nikkei
225, both the RSI MI40+ strategy and
the buy and hold approach experienced
losses in capital; however, the RSI MI40+
strategy significantly preserved capital in Software and Data
comparison. Bloomberg, MS-Excel
Based on the results of these backtests
there is clear value to long-term investors in Notes
utilizing the market identifier possibilities 1. Kahn, Michael.Technical Analysis: Plain & Simple. Financial Times Prentice Hall, 1999, p. 83.
that the RSI presents. By avoiding clear 2. Brown, Constance. Technical Analysis for the Trading Professional. McGraw-Hill, 1999, pp. 4–18.

bear market phases, as identified by the References

RSI, investors would be able to improve Brown, Constance. Technical Analysis for the Trading Professional. McGraw-Hill, 1999.
the risk/return profile of their portfolios, Edwards, Robert. D. and John Magee. Technical Analysis of Stock Trends. BN Publishing, 2008.
Kahn, Michael. Technical Analysis: Plain & Simple. Financial Times Prentice Hall, 1999.
enhancing capital preservation, gaining Kaufman, Perry J. A Short Course In Technical Trading. Wiley Trading, 2003.
capital appreciation and reducing volatility Murphy, John J. Technical Analysis of the Financial Markets. New York Institute of Finance, 1999.
of their returns in comparison to a buy and Person, John L. A Complete Guide to Technical Trading Tactics. Wiley Trading, 2004.
Pring, Martin J. The Investor’s Guide To Active Asset Allocation. McGraw-Hill, 2006.
hold approach.

PAGE 66 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Anatomy of a Living Trend: Swing Charts, High Andreas Thalassinos

Points and Low Points, Peaks and Troughs and 1, Damaskinou Street
Office 101
How Their Underlying Structure May Define Their Nicosia 1070
Cyprus

Forecasting Strength +357 22 762618

By Andreas Thalassinos, BSc, MSc, MSTA, CFTe, MFTA

Abstract Technical Analysis for the Financial Markets, the failure and non-
The paper is based on the fundamental premise of Technical failure swings. In the quest to develop a profitable automated
Analysis that prices move in trends. The basic structures (Gann trading system, I discovered through extensive reading that
Swing charts, HIP and LOP, Peaks and Troughs) that comprise a minimum of 1:1 risk reward is needed for a system to avoid
trends will be revisited with proposed variations in an attempt, negative balance. Of course, money management is out of the
through algorithmic tests, to diagnose the healthy (promising) scope of this paper, but I am extremely obliged to urge all new
vs. weak setups. The tests will mostly focus on the definitions traders to devote the relevant time to study the subject.
of peaks and troughs (e.g., higher high and higher low, higher In the rest of this paper, I will try to prove through
close and higher low), the length of the swing leg, and the swing’s algorithmic tests that the length of the swing, its position in
position in the life of the trend. The paper will not be limited to the life of the trend, and the structure of peaks and troughs are
swings only, but the study will be extended to cover the basic directly related to the strength of the unfolding trend—that is,
structures of a trend reversal (i.e., failure and non-failure swings). how far the trend will travel.

Introduction Trend Building Blocks

In my early trading steps, I stumbled upon a few books on Many theories have been developed regarding trend. The
technical analysis, some of them more in-depth than others, most popular technical tool used is of course the renowned
while some referred to experienced traders and the rest to moving average. If prices are above the moving average, then
novice traders. I was amazed then to discover that all books the market is in an uptrend, and consequently, if prices are
gyrated around trend! Some of them even mentioned the below the moving average, then the market is in a downtrend.
principles or tenets of technical analysis, which of course refer Naturally, more questions follow. What is the best moving
to trend. One of the three tenets states that prices move in average period ? What is the best moving average method?
trends, and more specifically in another tenet, that prices do not Which prices should be averaged? After investigating moving
move in a straight line but conversely follow a zigzag path. We averages a bit further, a trader soon discovers that regardless
all read or heard repeatedly throughout our technical analysis of the type and price averaging, they are all lagging. Moreover,
education the phrases “trend is your friend”, “follow the trend” they are a sure prescription for negative balance during a
and “never go against the trend”. Many years have passed sideways trendless range.
since my first trading steps and now, as a certified technical Going back to the 19th century, the father of technical
analyst, I can assure you that this is one of the best pieces of analysis, Charles Dow, observed that it is the direction of peaks
advice one can get in his/her early trading career. Do not let the and troughs that defines the trend of the market. Thus, when
simplicity behind it obscure its value or importance. After all, the market experiences two peaks and troughs successively
it’s general advice that applies to all markets and all financial higher than each other, then the market is said to be in an
instruments. As simple as it sounds, though, the advice, I have uptrend (Fig. 1).
to admit, is much more complicated or sophisticated, if I may
use this word, or even better, it barely reveals the tip of the Figure 1
iceberg. Try to adopt it while trading a specific market or a
specific financial instrument, and a lot of questions will arise. It
encompasses many questions that need to be answered before
adopting it to trade the markets. What is a trend? A simple
question with supposedly a simple answer. How can we spot
the early beginning of a trend and the end of it? How far will
this trend travel? This is the question. What are the building
blocks of a trend and how do they affect its future life? Welles
Wilder mentioned the HIPs and LOPs as building blocks; Bill
Williams mentioned fractals as a recursive entity that makes
up the trend; and of course, John Murphy mentions in his book,

IFTA.ORG PAGE 67
 IFTA JOURNAL 2015 EDITION

Of course, the opposite is true as well. When the market five consecutive bars, where the highest bar is preceded by
shows two peaks and troughs successively lower than each two lower highs and followed by two lower highs. Lows are not
other, then the market is said to be in a downtrend (Fig. 2). important for this pattern. Up fractals form peaks throughout
the life of a trend. On the other hand, a down fractal is a series of
Figure 2 at least five consecutive bars, where the lowest bar is preceded
by two higher lows and followed by two higher lows. Highs are
not important for this pattern. Down fractals form troughs
throughout the life of a trend (Fig. 5).
Figure 5

Unfortunately, trend does not move in only two directions.

When markets depict equal peaks and equal troughs then the
market is trendless or in a range or moving sideways (Fig. 3).
Figure 3 Naturally, more questions arise. Which price is best to use?
Closed or live? Charles Dow preferred the closed price, and more
specifically, he considered the daily close the most significant
price. Well, in today’s markets, most traders follow lower
than daily timeframes and do not have the patience for a bar/
candlestick to complete.
Last, but not least, is W.D. Gann, one of the best traders of
all times. He formulated the swing charts. According to Clif
Droke in the book Gann Simplified, “he had various ways of
constructing a swing chart, which is basically a chart in which
extreme tops and bottoms of each time increment (whether a
So far, so good. But allow me to ask another simple question. day, week, month, or year, depending upon the type of chart
What are peaks and troughs? Well, since the direction of trend used) is connected with a line for every two or three period
depends on the direction of successive peaks and troughs, then move in a single direction.” 3 In other words, a Gann swing chart
it is imperative that we define the building blocks of trend. can be constructed by following the highs and lows of the bars.
Welles Wilder mentions HIPs and LOPs (i.e., high points and low For example, a two-bar swing chart will define a bottom after
points) (Fig. 4) in his book, New Concepts in Technical Trading the market has made two consecutive higher-highs. On the
Systems. other hand, a two-bar swing chart will define a top after the
market has made two consecutive lower-lows (Fig. 6).
Figure 4
Figure 6

LOP is an abbreviation used for LOW POINT. A LOP is any time

bar that has a time bar immediately before it and immediately Swing charts may be developed by using three-bar or even
after it with a higher low. A HIP signifies a HIGH POINT and is one-bar swing charts. Also, higher-close and lower-close bars
defined as any time bar immediately before it and immediately may be used, as Gann himself mentions in his rule of five.
after it with a lower high.1 Trend may be defined as either of the above building blocks.
In his book Chaos Theory: Applying Expert Techniques to Now, the question has to do with the swing leg. What is the
Maximize Your Profits, Bill Williams states that “market or minimum swing-leg length to adopt? Does the length of the
behavioural fractals indicate a significant behaviour change.”2 swing have to do with how far the underlying trend will travel?
He goes on to say that fractals constitute the underlying In the Elliott Wave Principle, Frost and Prechter state that
structure of the Elliott Wave. Fractals come in two flavours. Up “It is our practice to try to determine in advance where the next
fractals and down fractals. An up fractal is a series of at least move will likely take the market.”4

PAGE 68 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Peaks and Troughs—Gann Charts A failure swing occurs after a trend has been in effect,
Throughout the rest of the paper and algorithmic tests, the exhibiting successively higher peaks and higher troughs during
following definitions, parameters and assumptions are adopted: an uptrend (lower peaks and lower troughs in a downtrend)
Peaks and troughs are defined by employing Gann’s two-bar until prices fail to make a higher peak and instead swing
swing charts. The first variation is based on higher high/higher direction, falling below the last trough. Peak Y fails to move
low for peaks and lower high/lower low for troughs utilizing live higher than previous Peak X and instead, prices swing direction
closing prices (Fig. 7). and fall below Trough Z (Fig. 10).
Figure 7 Figure 10

The second variation is based on higher close/higher low

for peaks and lower high/lower close for troughs, but this time
utilizing closing prices after the bar/candlestick had closed (Fig. 8). During a downtrend, the opposite is true. A series of
successively lower peaks and lower troughs is interrupted by
Figure 8 the failure of prices to make a new lower trough, and instead,
they swing direction upwards, breaking the last peak. Trough
Y fails to move lower than Trough X, and instead, prices swing
direction and break Peak Z (Fig. 11).

Figure 11

Another parameter is the swing’s length—that is, the

distance between peak and trough (Fig. 9).
Figure 9
A non-failure swing occurs after a trend has been in effect,
exhibiting successively higher peaks and higher troughs during
an uptrend (lower peaks and lower troughs in a downtrend)
until prices do make a higher peak but swing direction, falling
below the last trough. Peak Y moves higher than previous Peak X
and prices swing direction and fall below Trough Z (Fig. 12).
Figure 12

Swing’s Position in the Life of the Trend

The third parameter used is the position of the swing in the
life of the trend. Two cases are under study: the first swing in
the opposite direction (i.e., the reversal of a prior trend) and all
swings after the reversal. A distinction is made between failure
swing and non-failure swing reversals—that is, the first swing in
the opposite direction.

IFTA.ORG PAGE 69
 IFTA JOURNAL 2015 EDITION

Table 1
In the course of a downtrend, the opposite is true. A series
of successively lower peaks and lower troughs is interrupted
by the prices making a new lower trough and swing direction
upwards, breaking the last peak. Trough Y moves lower than
Trough X and prices swing direction and break Peak Z (Fig. 13).
Figure 13

Algorithmic Tests
Due to the vast popularity that foreign exchange enjoys
nowadays among traders, I decided to use the most liquid and
popular currency pair, EUR/USD, to run all tests.
Also, I decided to set a realistic fixed spread of 2 pips for all
tests. Apart from the above definitions and parameters, I used
the 30-minute timeframe, as I strongly believe that the majority
of today’s traders utilize timeframes below the daily. All tests The second algorithmic test is based on the following parameters:
were run on EUR/USD for the period of 1/9/2011–8/9/2013. ƒƒ Peaks are defined as Higher High/Higher Low (HH/HL) using
live closing prices.
The first algorithmic test is based on the following parameters: ƒƒ Troughs are defined as Lower High/Lower Low (LH/LL) using
ƒƒ Peaks are defined as Higher High/Higher Low (HH/HL) using live closing prices.
live closing prices. ƒƒ Swing lengths of 8.5, 16, 20 and 25 pips are used.
ƒƒ Troughs are defined as Lower High/Lower Low (LH/LL) using ƒƒ Price targets of 130%, 160%, 200%, 250%, 300% and 350% of
live closing prices. the swing length are examined.
ƒƒ Swing lengths of 8.5, 16, 20 and 25 pips are used. ƒƒ Only non-failure swings are used.
ƒƒ Price targets of 130%, 160%, 200%, 250%, 300% and 350% of ƒƒ EUR/USD is the financial instrument used.
the swing length are examined. ƒƒ Periodicity of 30 minutes is employed.
ƒƒ Only failure swings are used, ignoring the rest of the swings. ƒƒ Spread is set to 2 pips.
ƒƒ EUR/USD is the financial instrument used. ƒƒ The algorithmic test will test the period of 01/09/2011 until
ƒƒ Periodicity of 30 minutes is employed. 29/09/2013.
ƒƒ Spread is set to 2 pips.
ƒƒ The algorithmic test will test the period of 01/09/2011 until The second test is based on non-failure swings, with the
29/09/2013. rest of the parameters remaining the same. Looking at Table 2,
we observe that the results are again inversely proportional,
As you can see below (Table 1) the results are clearly inversely as expected. Once more, the 130% level exhibits the highest
proportional. That is, the higher the price target, the lower probability to be reached, but this time it is even improved,
the probability to achieve it. Also, as expected, the 130% level with a success rate close to 76%, regardless of the swing length.
exhibits the highest probability to be reached—about 60%– Furthermore, the 160% level shows a 50% probability to be
75%—regardless of the swing length. Furthermore, the 160% reached. On the other hand, price targets of 300% and 350%
level enjoys about 50% probability to be achieved. Surprisingly, have lower probabilities to be achieved compared to the first
the rest of the target levels fall below the 50% success rate. test of failure swings. The probability of reaching a 1:1 ratio,
Another important finding has to do with risk-to-reward ratio. 200% target level is in the range of 37%–40%. The probability
The probability of reaching a 1:1 ratio, 200% target level is in the of reaching a 2:1 ratio, 300% target level is in the range of
range of 30%–40%. Also, the probability of reaching a 2:1 ratio, 10%–16%. The length of the swing is inversely proportional
300% target level is in the range of 13%–20%. In addition, the to the percentage of the trades that reached the target of the
swing length of 8.5 pips shows the best performance compared important levels of 200% and 300%.
to the rest of the swing lengths. The length of the swing is
inversely proportional to the percentage of the trades that
reached the target of the important levels of 200% and 300%.

PAGE 70 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Table 2 Table 3

The third algorithmic test is based on the following parameters: Summing up the first three tests, we realize that the price
ƒƒ Peaks are defined as Higher High/Higher Low (HH/HL) using target of 130% of the swing length has about 70% probability
live closing prices. of reaching it, regardless of the length of the swing. More
ƒƒ Troughs are defined as Lower High/Lower Low (LH/LL) using importantly, a 1:1 reward-to-risk ratio has a probability of less
live closing prices. than 50%. The next important higher price target of 300%, 2:1
ƒƒ Swing lengths of 8.5, 16, 20 and 25 pips are used. reward-to-risk ratio has a probability of less than 35%. Of course,
ƒƒ Price targets of 130%, 160%, 200%, 250%, 300% and 350% of the swing length and the percentage of the trades that achieved
the swing length are examined. the target of 200% and 300% are inversely proportional. Overall,
ƒƒ All swings are considered except reversals. the third test revealed that all swings in the direction of the
ƒƒ EUR/USD is the financial instrument used. unfolding trend, excluding reversals, have a higher probability of
ƒƒ Periodicity of 30 minutes is employed. reaching the important price targets of 200% and 300%.
ƒƒ Spread is set to 2 pips.
ƒƒ The algorithmic test will test the period of 01/09/2011 until The fourth algorithmic test is based on the following parameters:
29/09/2013.
ƒƒ Peaks are defined as Higher Close/Higher Low (HC/HL) using
The third test excludes failure and non-failure swings and only closed closing prices.
takes into account swings in the direction of the trend. The rest ƒƒ Troughs are defined as Lower Close/Lower Low (LC/LL) using
of the parameters remain the same. As you can observe, Table closed closing prices.
3 reveals that the results are again inversely proportional. The ƒƒ Swing lengths of 8.5, 16, 20 and 25 pips are used.
130% level exhibits the highest probability to be reached, with ƒƒ Price targets of 130%, 160%, 200%, 250%, 300% and 350% of
a success rate in the range of 73%–76%, regardless of the swing the swing length are examined.
length. Additionally, the 160% level shows a further improvement ƒƒ Only failure swings are used.
of the probabilities to be achieved, reaching 60%. The probability ƒƒ EUR/USD is the financial instrument used.
of reaching a 1:1 reward-to-risk ratio, 200% target level is in the ƒƒ Periodicity of 30 minutes is employed.
range of 36%–45%. Also, the probability of reaching a 2:1 ratio, ƒƒ Spread is set to 2 pips.
300% target level is in the range of 19%–24%. Once more, the ƒƒ The algorithmic test will test the period of 01/09/2011 until
swing length of 8.5 pips shows the best performance compared 29/09/2013.
to the rest of the swing lengths on the important levels of 200%
and 300%. The inverse proportionality between the length of the This time, as will you will notice in the table below (Table
swing and the percentage of the trades that achieved the targets 4), the results follow the same patterns. The higher the price
of 200% and 300% has been confirmed. target, the lower the probability to achieve it. More specifically,
the 130% holds the highest probability to be reached—about
63%–73%—regardless of the swing length. Furthermore, the
160% level shows a probability to be achieved of close to 50%.

IFTA.ORG PAGE 71
 IFTA JOURNAL 2015 EDITION

Table 5
The rest of the target levels have less than a 50% probability to
succeed. The probability of reaching a 1:1 ratio, 200% target level
is in the range of 47%–55%. Also, the probability of reaching a
2:1 ratio, 300% target level is in the range of 17%–26%. For the
fourth time, the inverse proportionality remains intact.

Table 4

The sixth algorithmic test is based on the following parameters:

ƒƒ Peaks are defined as Higher Close/Higher Low (HC/HL) using

closed closing prices.
ƒƒ Troughs are defined as Lower Close/Lower Low (LC/LL) using
The fifth algorithmic test is based on the following parameters: closed closing prices.
ƒƒ Swing lengths of 8.5, 16, 20 and 25 pips are used.
ƒƒ Peaks are defined as Higher Close/Higher Low (HC/HL) using ƒƒ Price targets of 130%, 160%, 200%, 250%, 300% and 350% of
closed closing prices. the swing length are examined.
ƒƒ Troughs are defined as Lower Close/Lower Low (LC/LL) using ƒƒ All swings are considered except reversals.
closed closing prices. ƒƒ EUR/USD is the financial instrument used.
ƒƒ Swing lengths of 8.5, 16, 20 and 25 pips are used. ƒƒ Periodicity of 30 minutes is employed.
ƒƒ Price targets of 130%, 160%, 200%, 250%, 300% and 350% of ƒƒ Spread is set to 2 pips.
the swing length are examined. ƒƒ The algorithmic test will test the period of 01/09/2011 until
ƒƒ Only non-failure swings are used. 29/09/2013.
ƒƒ EUR/USD is the financial instrument used.
ƒƒ Periodicity of 30 minutes is employed. The last test (Table 6) follows the same result patterns. The
ƒƒ Spread is set to 2 pips. higher the price target, the lower the probability to achieve it.
ƒƒ The algorithmic test will test the period of 01/09/2011 until The 130% holds the highest probability to be reached—about
29/09/2013. 73%—regardless of the swing length. Furthermore, the 160%
level shows a probability to be achieved of close to 56%–59%.
The fifth test, as seen in Table 5, confirms the ongoing The rest of the target levels have less than a 40% probability
pattern of the 130% price target. It holds a probability of about to succeed. The probability of reaching a 1:1 ratio, 200%
73, regardless of the swing length. Furthermore, the 160% level target level, is in the range of 33%–38%. Also, the probability
shows probabilities to be achieved close to 56%. The rest of the of reaching a 2:1 ratio, 300% target level is in the range of
target levels have less than a 50% probability to succeed. The 15%–18%. The fact that the swing length of 16 pips has a
probability of reaching a 1:1 ratio, 200% target level is in the slightly higher probability (18.02%) than the length of 8.5 pips
range of 36%–39%. Also, the probability of reaching a 2:1 ratio, (17.70%) is not enough, I believe, to cancel the pattern of inverse
300% target level is in the range of 13%–17%. As expected, the proportionality.
length of the swing and the percentage of trades that reached
the targets of 200% and 300% is inversely proportional.

PAGE 72 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Table 6
higher performance when combined with higher high/higher
low structures for troughs and lower high/lower low for peaks
using live closing prices. In a nutshell, closed closing prices act
as filters on the reversals, thus decreasing the number of false
swings in the opposite direction, where the live closing prices
perform better when the trend is already in progress.
Of course, I do not claim that the methodology that I followed
is flawless. There is always room for improvement. For example,
in the future, during a new set of tests, I will include more
financial instruments to cover a wider spectrum of the market.
Also, I will experiment with a variety of time spans once the
hurdle of accurate data availability is overcome.
Chart 1

Summing up the last three tests, we realize that the

price target of 130% of the swing length has about 63%–75% Chart 2
probability of reaching it, regardless of the swing length and
the position. More importantly, a 1:1 reward-to-risk ratio has a
probability of less than 50%. The next important price target
of 300% that is a 2:1 reward-to-risk ratio has a probability of
less than 30%. The swing length and the percentage of the
trades that achieved the target of 200% and 300% are inversely
proportional. Furthermore, the last test revealed that reversals
of failure swings have a higher probability of reaching the
important price targets of 200% and 300% when the HC/HL and
LC/LH structures are used utilizing closed closing prices.

Conclusion Software and Data

In conclusion, I have noticed that the size of the swing’s leg MetaTrader 4 Version 4.00 Build 509 from MetaQuotes Software Corporation,
Limassol, Cyprus.
is important to the health of the unfolding trend—that is, how
far the trend will travel. The smaller the length of the swing, the
higher the probability of the prices to reach the target. It has Notes
been proven through algorithmic tests that I ran on MetaTrader 1. Wilder, Welles J., New Concepts in Technical Trading Systems, Hunter, 1978. p. 7.
4 that the swing’s length is inversely proportional to the 2. Williams, Bill, Trading Chaos: Applying Expert Techniques to Maximize Your
Profits, Wiley, 1995. p. 134.
percentage of the targets reached, with emphasis on the crucial 3. Droke, Clif, Gann Simplified, Marketplace Books, 2001. p. 56.
levels of 200% (1:1 reward-to-risk ratio), as seen below in Chart 1 4. Prechter, Robert R. and A. J. Frost, Elliott Wave Principle,New Classic’s Library,
and 300% (2:1 reward-to-risk ratio) in Chart 2. 1998. p. 88.
More specifically, the 200% target level has less than a 50%
probability to be reached, where the 300% level barely reaches References
the 27% success rate. In addition, the 130% target level enjoys Droke, Clif, Gann Simplified, Marketplace Books, 2001.
a constant 70% success rate, regardless of the length and the Edwards Robert D. and John Magee, Technical Analysis of Stocks and
Commodities, 9th Edition, Amacom, 2007.
position of the swing.
Gregory-Williams, Justine and Bill Williams, Trading Chaos, 2nd Edition,
The hypothesis, that the position of the swing in the life of Wiley, 2004.
the trend, is important for the forecasting strength of the swing Murphy, John J., Technical Analysis of the Financial Markets, NYIF, 1999.
it was verified through the algorithmic tests as well. Failure Prechter, Robert R. and A.J. Frost, Elliott Wave Principle, New Classic’s
Library, 1998.
swings boast higher performance when combined with the
Pring, Martin J., Technical Analysis Explained, 4th Edition, McGraw-Hill, 2002.
building blocks of peaks and troughs based on higher close/
Wilder, Welles J., New Concepts in Technical Trading Systems, Hunter, 1978.
higher low and lower close/lower high using closed closing Williams, Bill, Trading Chaos: Applying Expert Techniques to Maximize your Profits,
prices. On the other hand, all swings after a reversal boast Wiley, 1995.

IFTA.ORG PAGE 73
 IFTA JOURNAL 2015 EDITION

Samuel Utomo
Refining Wilder’s ADX: Adjustment to the [email protected]
Studento L11/15, Foresta, BSD City
Price Actions by Utilizing Closing Prices Tangerang Banten
15345 Indonesia
By Samuel Utomo, CFTe, MFTA +6287852244440

Abstract Unfortunately, it has been so obvious that many, if not all,

Average Directional Movement Index, or just ADX, is one of technical traders who use technical indicators do not even
very few known indicators to gauge the strength of the trend bother to take a glimpse of the inputs and assumptions from
within the price action. Unfortunately, there is a flaw within the the developers. But, we should be able to realize the significant
inputs of the Directional Movement, which rely solely on the use impact to the nature of an indicator that can result from a
of highs and lows instead of using closing prices to determine slight change within the indicator’s input. Although many have
the directional moves. This causes many misinterpretations of regarded ADX as great support to filter trading signals, few
the proper direction of the price movements, as highs and lows have realized that this indicator is built upon the foundation of a
only represent trading range while closing prices always show trading range rather than the actual price direction.
the direction of price in the end of each period, especially when How can a “directional movement” indicator be valued with
the market becomes choppy and highly volatile. trading range and not the actual price direction?
Therefore, this paper mainly focuses on the development This paper intends to answer the question above by
of the new type of Average Directional Movement Index that redeveloping ADX using closing prices as its Directional
utilizes closing prices as its input, so it can remove the biases Movement input, and we call it the Modified ADX. We will use
that result from using highs and lows in original ADX input. the Analysis of Covariance (ANCOVA) method to compare the
Then, it will be followed with thorough statistical analysis slope of the ADXs relative to the volatility, then use student
to compare both the sensitivity to the volatility and trading t-test to validate and compare the trading result objectively.
performances. The objective of this paper is to show the Thus, two hypotheses are tested in this paper:
evidence that using closing prices is more statistically reliable
than using highs and lows to determine the original Directional 1. The slope of Modified ADX has significantly less significance
Movement value. compared to the slope of original ADX relative to the price
volatility.
Introduction 2. Trading systems that use the Modified DM input have better
As a trend follower and technical trader, we should trading performances than systems that use the original DM
acknowledge that Average Directional Movement Index, input.
developed by J. Welles Wilder in 1978, has provided us with an
unbelievably powerful tool to face the worst enemy of every The Development of Original ADX
trader—whipsaws. This indicator has been specifically designed Wilder’s first intention to develop Average Directional
to determine whether the current trend has reliable odds to be Movement Index (ADX) was based on the idea to be able to rate
traded profitably or not—in other words, gauging the strength the directional movement of any or all commodities or stocks
of the trend in an objective manner. Thus, if one reads this paper on a scale of 0 to 100, while the objective of this indicator is
carefully, then he or she should be curious enough to ask this to define markets relative to technical trading systems. In
question: What is truly wrong with this indicator? other words, ADX is a technical tool to measure the trend
I believe that the finding of ADX indicator by Wilder is similar strength regardless of the direction of the trend. The higher the
to when someone discovers precious stones from the earth he ADX value, the stronger the tendency of the trend to persist.
mines, perhaps stones like diamond stones. The first time he Conversely, the lower the ADX value, the less likely for the trend
sees the diamond stones, it will absolutely not be in the form to persist.
that makes many people desire them the most, because they are There are several calculations to be done before obtaining
the rough diamond stones. It takes times to refine the diamond the ADX value:
stones until they can show themselves again in a way that puts
sparks in the eyes of the observers. It happens to be the same 1. Directional Movement (DM)
with ADX indicator; it needs to be refined in some ways to ƒƒ +DM = Current High − Previous High
enhance its capabilities. Thus, as a researcher, I have to examine ƒƒ –DM = Previous Low − Current Low
in a scientific way and break down each part of this indicator in ƒƒ If + DM > –DM and +DM > 0, then DM = + DM, or else, + DM = 0.
order to rebuild it in a possibly better form. ƒƒ If –DM > +DM and –DM > 0, then DM = –DM, or else –DM = 0.

PAGE 74 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Figure 1: Scenarios for the calculation of original DM value

highs and lows of two periods, this type of DM will only use
closing prices to determine the DM value, whether it is positive
or negative.

ƒƒ + DM = Current Close – Previous Close

ƒƒ –DM = Previous Close – Current Close
ƒƒ If Current Close > Previous Close, then DM = + DM
ƒƒ If Current Close < Previous Close, then DM = –DM
ƒƒ If Current Close = Previous Close, then DM = 0

There are also several scenarios that should be considered

as a rule of thumb to determine the value of the DM. These
scenarios will be the major part that differs between the
There are several scenarios that we should consider as a rule Modified ADX and the original ADX. They are shown to expose
of thumb to determine whether a DM will be valued as +DM or the use of closing prices to determine DM.
–DM (Figure 1). The DM determinations within scenarios A, B, G,
and H are so obvious that we can judge them virtually because Figure 2: Scenarios for the calculation of Modified DM
they represent an impulsive type of price action. It means that
the current bar has a higher (lower) high and a higher (lower) low
compared to the previous bar. Conversely, we need to be more
thorough when dealing with both scenario C and D because we
need to calculate each +DM and –DM, then value DM using the
largest one. The two last scenarios, E and F, are special cases that
we will not assign any DM value, or just put 0 as its value.

2. Directional Indicator (DI)

ƒƒ + DIn = + DMn/TRn (True Range)
ƒƒ –DIn = –DMn/TRn

‘n’ represents the period that we decide to use. In his book,

Wilder recommends a 14-day period because it represents the Referring to Figure 2 above, The bar patterns in scenarios
time span of an average half-time cycle period. A, B, E, and F seem to be very similar to scenarios E and F in the
original DM, where the bars have the same high and low value.
3. Average Directional Movement Index (ADX) Instead of valuing the DM with 0 value, I prefer to value the DM
ƒƒ ADXn = ( + DIn - –DIn ) / ( + DIn + –DIn ) based on the comparison with the closing price. Scenarios A and
E are assigned as –DM because the current closing price is lower
As you can see above, the basic input of the ADX is the than the previous closing price, and both B and F are assigned
measurement of the DM that requires a process of differencing as +DM because the current closing price is higher than the
the highs and lows of two consecutive periods. It is an previous closing price.
inconsistency to say that a direction of a market movement can Now, if we observe the bar patterns in scenarios C, D, I,
be determined by its high and low, since they only represent and J thoroughly, we will find out that they are similar to the
the trading activity within one period of time. In reality, the bar pattern in scenarios A and B in the original DM’s scenario.
majority of market participants pay more attention to the This is the most deceptive part. In the original DM, scenarios
closing value to determine the main direction of a trading C and I will be assigned as + DM, while scenarios D and J are
period—it is a daily, weekly, monthly, yearly, or even the tiny assigned as –DM. The problem is when we compare the closing
incremental period of time, such as seconds, minutes, and prices. Scenarios C and I have lower current closing price values
hours. This presumption is aligned with the Dow Theory, as compared to the previous period’s values, showing a negative
one of its basic tenets states that only closing price is used.1 direction. Scenarios D and J have higher current closing price
Therefore, a modification needs to be made to the input by values compared to the previous period’s values, showing a
utilizing the closing prices rather than highs and lows, and we positive direction. Based on the rules of the Modified ADX,
should call this Modified indicator the Modified ADX. scenarios C and I should be assigned with –DM and + DM for
scenarios D and J.
The two last scenarios, G and H, are similar to scenarios C
Introducing the Modified ADX and D in the original DM. The good thing is that the Modified
The Modified ADX consists of several calculations similar to DM provides a simpler DM calculation than the original DM.
the original ADX. Basically, the only difference between them is Instead of comparing the difference as the original DM does,
the DM input, but this only makes a significant difference within this Modified DM calculates only the difference between closing
the output value. Instead of using the comparison between prices in order to get the DM’s value.

IFTA.ORG PAGE 75
 IFTA JOURNAL 2014 EDITION

Figure 3: Comparison between ADXs on XAUUSD daily data Through the DM scenarios in the
earlier section, the difference between
the original ADX and the Modified ADX
arises. The original ADX will assign +DM
value when there is higher difference
between highs than the difference
between the lows, even though the bar
is closed lower than the previous bar’s
close, and vice versa. It means that the
calculation of the original ADX includes
many inappropriate DM values; thus,
biases will occur in the output value,
the original ADX value. These biases
conceptually translate as the higher
original ADX value when being compared
Figure 4: Comparison between ADXs on LQ45 index daily data
to the Modified ADX in the exact same
period. In other words, we can say that
the Modified ADX will be less sensitive
to the price volatility compared to the
original ADX. In the following charts,
we will see the differences between the
values of the ADXs.
The difference between the ADXs is
expected to occur when the price tends
to move in a sideways manner (Figures
3 and 4). Although the prices showed
such erratic movement on the XAUUSD
and LQ45 indexes, the original ADX
value did not fully confirm the event
Figure 5: Comparison between ADXs on XAUUSD daily data because it tends to move both above and
below the buffer level (Wilder suggests
to stop using a trend-following trading
system when the ADX value is below 20).
Interestingly, the Modified ADX value
gave a valid confirmation of the event, as
its value moved below 20. It proves that
the erratic movements within the price
action have made the ADX value biased
by a wide trading range, as it uses highs
and lows as its input.
Meanwhile, the Modified ADX could
capture the start of trending phase
earlier than the original ADX on XAUUSD
(Figure 5). On 8/28/10, we can see that
Figure 6: Wilder’s Directional Movement system on XAUUSD daily data
the Modified ADX value has breached the
level of 20 on 9/21/10. More than three
weeks passed before the original ADX
breached that level, indicating that the
original DM input to calculate the ADX
value has been biased. As a consequence
of the biases within the original DM
input, the trading signal resulted by the
Directional Indicator tends to be late
in capturing better timing on entering
trades, as shown on the charts below.
The orange vertical lines show the
entry point of the Directional Movement
System that used closing prices as its DM

PAGE 76 IFTA.ORG
 IFTA JOURNAL 2014 EDITION

input. The blue vertical lines show the Figure 7: Wilder’s Directional Movement system on LQ45 index daily data
entry point of the original Directional
Movement System, and the purple
vertical lines show the exit point of
both systems. All the examples (Charts
6, 7 and 8) show that the Directional
Movement System that used closing
prices, as its DM input generated trading
signals with earlier entry and entered the
trades with a lower price than the original
Directional Movement System one, while
the exit timing is about the same.
In the next sections, I will conduct
both the statistical significance test
and trading performance comparison Figure 8: Wilder’s Directional Movement system on LQ45 index daily data
on Gold Spot Price (XAUUSD) and LQ45
index, an index that consisted of the top
45 companies and is considered as the
most followed index in the Indonesia
Stock Exchange. The data will span from
8/1/05–8/31/13, or eight years of daily
data from eSignal. I believe the time
span is sufficient enough to comply with
the fitness of the test due to the huge
amount of analyzed data, which consists
of 2,108 daily trading data of XAUUSD
and 1,958 daily trading data of LQ45
index.
Figure 9: Comparison among price, ADXs, and price volatility on XAUUSD
Slope Comparison Test daily data
Both of the ADXs should be valued
relative to other variables that can
objectively measure the existence of a
trend. This paper utilizes 14-day price
volatility, specifically the Standard
Deviation, as the variable to measure
whether the price has sufficient trends
to be traded profitably or not. The basic
assumption is that when the volatility
is low, then the price tends to move in
a narrow range; therefore, there will
be insufficient trends to be traded
profitably. Conversely, there will be
sufficient trends to be traded when the
volatility is high. Figure 10: Comparison among price, ADXs, and price volatility on LQ45 index
We can see that both the volatility daily data
and ADX tend to rise significantly when
the price moves in a strong trend, either
bullish (Figure 9) or bearish (Figure 10).
It suggests that strong trends tend to be
followed with a rise in volatility along
with the ADX value. A rising ADX value
indicates the development of a strong
trend within price action; thus, the
instrument can be traded with better
statistical edge. Conversely, a falling
ADX value indicates the tendency of
a weak trend; thus, many whipsaws

IFTA.ORG PAGE 77
 IFTA JOURNAL 2015 EDITION

Table 3: Coefficients of regression equation between

are expected to occur, and the statistical edge to trade will be volatility (standard deviation) and original ADX value on
diminished. XAUUSD daily data
Since we need statistical evidence, both ADXs should act as
Estimated Std. Error t value Pr(>|t|)
the explanatory variable to the simple linear regression equation
of the price volatility, the dependent variable. There are two Intercept 0.1206854 0.004959 24.34 <2E-16 ***
simple linear regression equations with different explanatory ADX 0.0024551 0.000184 13.38 <2E-16 ***
variables. The first equation uses the original ADX as its Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
explanatory variable, while the second one uses the Modified
ADX as its explanatory variable. The slope of both regression
Table 4: Coefficients of regression equation between
lines will be compared using the Analysis of Covariance volatility (standard deviation) and Modified ADX value
(ANCOVA) method to test the statistical significance of slope’s on XAUUSD daily data
difference between them. The ANCOVA method is an extension
Estimated Std. Error t value Pr(>|t|)
of the Analysis of Variance (ANOVA) model for providing a way
to statistically control the effects of the covariates within a Intercept 0.1530967 0.005007 30.576 <2E-16 ***
model that includes both quantitative (the value of the ADX) and ADX 0.0014548 0.000229 6.6367 2.36E-10 ***
qualitative (the type of the ADX) regressors. Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
The objective is to prove that the biases within the original
ADX formula unnecessarily enhance the sensitivity of the
ADX value to the price volatility. This supports the belief Figure 11: A graph of regression line of ADX and Modified
about adjusting the ADX value with closing prices, as its input ADX (ADX_mod) of XAUUSD daily data
will lower the ADX sensitivity to the price volatility, thereby
reducing more trading whipsaws.
The first ANCOVA model of both XAUUSD (Table 1) and LQ45
index (Table 5) shows that volatility (represented as std) is
modeled as a dependent variable with type (type 1 represents
the original ADX and type 2 represents Modified ADX) as its
factor and ADX value as its covariate. Based on the first ANCOVA
models, there is sufficient evidence to claim that there is a
significant impact of type and ADX value. Furthermore, we
can tell that the interaction between the regression lines is
significant as it is shown in the graph above (Figures 11 and 12).
Table 5: First ANCOVA model of LQ45 index daily data
We can see that the sensitivity of original the ADX value to the
volatility has led the type 1’s regression line to cross upward to Df Sum Sq Mean Sq F value PR(>F)
the type 2’s regression line. ADX 1 5.52 5.517 328.184 <2E-16 ***
The second model (Tables 2 and 6) shows that the ADX type 1 0.14 0.144 8.573 0.00343 **
value has significant effect on the dependent variable. In other
words, we can tell that there is significant difference within the ADX:type 1 0.16 0.16 9.529 0.00204 **
intercepts of both regression lines. Therefore, we may come up Residuals 3912 65.76 0.017
with a conclusion that the original ADX value is significantly more Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
sensitive to the volatility compared to the Modified ADX value.
Table 6: Second ANCOVA model of LQ45 index daily data
Table 1: First ANCOVA model of XAUUSD daily data
Df Sum Sq Mean Sq F value PR(>F)
Df Sum Sq Mean Sq F value PR(>F)
ADX 1 5.52 5.517 327.47 <2E-16 ***
ADX 1 1.216 1.2164 188.55 <2E-16 ***
type 1 0.14 0.144 8.555 0.00347 **
Type 1 0.096 0.0965 14.96 0.000112 ***
Residuals 3913 65.92 0.017
ADX:type 1 0.076 0.0756 11.72 0.000624 ***
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residuals 4212 27.172 0.0065
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Table 7: Coefficients of regression equation between
volatility (standard deviation) and original ADX value on
Table 2: Second ANCOVA model of XAUUSD daily data LQ45 index daily data
Df Sum Sq Mean Sq F value PR(>F) Estimated Std. Error t value Pr(>|t|)
ADX 1 1.216 1.2164 188.07 <2E-16 *** Intercept 0.1236632 0.007479 16.54 <2E-16 ***
type 1 0.096 0.0965 14.92 0.000114 *** ADX 0.0046388 0.000284 16.36 <2E-16 ***
Residuals 4213 27.248 0.0065 Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

PAGE 78 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Table 8: Coefficients of regression equation between

volatility (standard deviation) and original ADX value on ƒƒ There is no stop-loss used. The position will be closed when a
LQ45 index daily data trading signal occurs.
Estimated Std. Error t value Pr(>|t|) ƒƒ A trade is executed with market order one day after the
trading signal occurs.
Intercept 0.1672815 0.008052 20.774 <2E-16 ***
ADX 0.0032527 0.00035 9.286 <2E-16 *** The trading summary consists of some key factors, such as
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Total Net Profit, Profit Factor, Return on Starting Equity, Total
Number of Trades, Percent Profitable, Average Winning Trade,
Average Losing Trade, Average Trade, and Average Drawdown.
Figure 12: A graph of regression line of ADX and Modified Each factor will be used to evaluate each system and compare
ADX (ADX_mod) of LQ45 index daily data between the one that uses original ADX inputs and the one
that uses closing prices. The summaries are shown in tables 9
throught 12 below.
Table 9: Trading summary of Directional Movement
System as applied to the XAUUSD daily data
DMI DMImod
Total Net Profit $24,692.69 $79,825.58
Profit Factor 1.119 1.376
Return on Starting Equity 24.69% 79.83%
Total Number of Trades 175 231
Trading Implication Analysis Percent Profitable 35.43% 27.71%
I believe the statistical evidence about the ADXs should
be translated into the validity of their trading performances Average Winning Trade $3,755.09 $4,565.73
through backtesting the two trading systems below: Average Winning Trade (%) 3.76% 3.85%
Average Losing Trade ($1,841.80) ($1,271.74)
a. Directional Movement System. This trading system is
originally created by Wilder in his book New Concepts in Average Losing Trade (%) -1.75% -1.05%
Technical Trading Systems. When +DI14 crosses above –DI14, Average Trade (Expectation) $141.10 $345.57
a long position is taken. The position is reversed when –DI14 Average Trade (%) 0.20% 0.31%
crosses above +DI14.2
b. Donchian’s Dual Moving Average Crossover Filtered With Average Drawdown (%) 10.48% 5.91%
Directional Movement System (Donchian’s MA). This trading
system utilizes one of the alternative rules provided in the Table 10: Trading summary of Directional Movement
same book written by Wilder. When +DI14 crosses above –DI14, System as applied to the LQ45 index daily data
take only the long trades; when +DI14 crosses above –DI14, DMI DMImod
take only the short trades.3 I combine the system with one
of the most commonly known trend-following trading Total Net Profit $100,804.62 $217,002.18
systems—dual moving average—which consists of 5- and Profit Factor 1.22 1.344
20-day simple moving averages (SMA). So, the system will Return on Starting Equity 100.80% 217.00%
take long positions when the 5-day SMA crosses above 20-day
Total Number of Trades 165 192
SMA and the +DI14 is above the –DI14. The short position will
be taken as the 20-day SMA crosses above the 5-day SMA and Percent Profitable 30.91% 29.69%
the –DI14 is above the +DI14. When the system is in conflict Average Winning Trade $10,973.71 $14,877.95
(bullish SMA crossover happens when the –DI14 is above the
Average Winning Trade (%) 6.09% 6.21%
+DI14 or the opposite), then no trade will be taken.
Average Losing Trade ($4,025.04) ($4,674.38)
The constraints of the backtest are: Average Losing Trade (%) -1.86% -1.53%
Average Trade (Expectation) $610.94 $1,130.22
ƒƒ Each system is capitalized with US$100,000 as a starting equity.
ƒƒ I assume that there is no trading cost (and slippage) and that Average Trade (%) 0.60% 0.77%
the instrument can be bought or shorted with its price as a Average Drawdown (%) 11.72% 8.09%
nominal value. This means that if the LQ45 index value is 500,
then the value of one contract is US$500.
ƒƒ The system will always use the whole cash (approximately
100% of the equity value) to buy or short each instrument.

IFTA.ORG PAGE 79
 IFTA JOURNAL 2015 EDITION

Table 11: Trading summary of Donchian’s MA system as

applied to the XAUUSD daily data lower expected losses, since they have larger Average Winning
Trades and Average Losing Trades than the other systems with
MA_DMI MA_DMImod
original DM inputs. Thus, the result of these two key factors
Total Net Profit $9,741.64 $24,692.69 may fully comply with my earlier hypothesis of better trading
Profit Factor 1.078 1.119 performance from the trading systems using the closing price as
Return on Starting Equity 9.74% 24.69%
its DM input.
To validate the trading performances further, I decided to
Total Number of Trades 95 175 apply Student’s t-test to the return of each system so that we
Percent Profitable 35.79% 35.43% can know the probability of a system producing positive return
Average Winning Trade $3,945.44 $3,755.09
or greater than zero, and then compare it to the similar system
with Modified DM input.
Average Winning Trade (%) 4.99% 3.76%
Average Losing Trade ($2,039.40) ($1,841.80) Table 13: Statistical significance test summary of
Average Losing Trade (%) -2.48% -1.75% Directional Movement System as applied to the XAUUSD
daily data
Average Trade (Expectation) $141.10 $141.10
DMI DMImod
Average Trade (%) $102.54 0.20%
Number of Trades 175 231
Average Drawdown (%) 19.52% 10.48%
Number of Degrees of Freedom 165 221
Average trade at 95.00% $141.10 +/- $345.57 +/-
Table 12: Trading summary of Donchian’s MA System as confidence 498.84 439.01
applied to the LQ45 index daily data
Worst-case average trade at
($357.74) ($93.44)
MA_DMI MA_DMImod 95.00% confidence
Total Net Profit $93,377.88 $111,683.01 Probability that average trade is
67.56% 90.18%
greater than zero
Profit Factor 1.317 1.403
Return on Starting Equity 93.38% 111.70%
Total Number of Trades 82 85 Table 14: Statistical significance test summary of
Directional Movement System as applied to the LQ45
Percent Profitable 41.46% 41.18% index daily data
Average Winning Trade $11,409.44 $11,117.39
DMI DMImod
Average Winning Trade (%) 7.80% 7.80%
Number of Trades 165 192
Average Losing Trade ($6,136.32) ($5,548.51)
Number of Degrees of Freedom 155 182
Average Losing Trade (%) -3.57% -3.40%
Average trade at 95.00% $610.94 +/- $1130.22 +/-
Average Trade (Expectation) $1,138.75 $1,313.92 confidence 1623.84 1809.79
Average Trade (%) 1.15% 1.21% Worst-case average trade at
95.00% confidence ($1,012.91) ($679.57)
Average Drawdown (%) 13.19% 14.79%
Probability that average trade is
greater than zero 73.04% 83.72%
The result shows a typical pattern in most of the key factors.
The systems that used closing prices as their DM input (DMImod
and MA_DMImod) tend to perform significantly better than Table 15: Statistical significance test summary
the ones that used original inputs (DMI and MA_DMI), as seen of Donchian’s MA system as applied to the XAUUSD
on the Total Net Profit, Return on Starting Equity, and Average daily data
Trade factors. The other supporting key factors, such as Average MA_DMI MA_DMImod
Drawdown and Average Losing Trade, show a similar result,
except the Donchian’s MA system that was applied on LQ45 Number of Trades 95 104
index (Table 12), showing slightly higher Average Drawdown on Number of Degrees of Freedom 85 94
MA_DMImod system. Average trade at 95.00% $102.54 +/- $344.56 +/-
The most interesting part is that there are two factors that confidence 638.91 744.74
generate opposite results—the Total Number of Trades and
Worst-case average trade at
Percent Profitable factors. Systems that used closing prices 95.00% confidence ($536.37) ($400.18)
as their DM input tend to produce more trading signals, yet
they came up with a result of slightly lower winning rates. This Probability that average trade is
greater than zero 60.47% 77.24%
means that these systems are able to capture more winning
trades with larger expected profits and losing trades with

PAGE 80 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Table 16: Statistical significance test summary of

Donchian’s MA system as applied to the LQ45 index DM in order to adjust the ADX output value with the proper
daily data interpretations of price actions.
MA_DMI MA_DMImod The slope of the ADX with closing price as its Modified
DM input has shown statistically much less sensitivity to the
Number of Trades 82 85
volatility as compared to the slope of original ADX over eight
Number of Degrees of Freedom 72 75 years of trading data for the XAUUSD and LQ45 indexes. It
Average trade at 95.00% $1138.75 +/- $1313.92 +/- proved that the original ADX value has included much biased
confidence 2442.02 2155.02 information regarding the directional movement resulting from
Worst-case average trade at its DM inputs. Furthermore, it has successfully translated its
95.00% confidence ($1,303.27) ($841.10) bias reduction within the DM into the significantly better overall
trading performances. The statistical analysis on the average
Probability that average trade is
greater than zero 77.42% 83.24% trade of each system has provided us with strong indications
that the likelihood of generating a positive average trade for
trading systems that used closing prices as their input is better
All of the results above show that the systems that used than for the standard trading systems.
closing price as their DM input tend to have a larger probability Nevertheless, this paper is constrained by some limitations,
of producing returns greater than zero. Similarly, the average as I did not provide comparisons of the trading performances
trade at 95% confidence level provides us with evidence that to any benchmark, and they were only tested with standard
systems with Modified DM input have a better range, higher position sizing that always invests all of its equity. The
profits on the winning trades and lower losses on the losing sole purpose of this paper is not to criticize one of Wilder’s
trades, except for the Directional Movement System, as applied masterpieces, but rather to provide the reader with an
on the LQ45 index and the Donchian’s MA on the XAUUSD, which alternative way of using ADX. As diamond stones need to be
have slightly higher losses on the losing trades. refined through time, there are still spaces for those who are
The greatest spread of the probability occurs between the willing to further explore the reliability of these systems by
Directional Movement System as applied on XAUUSD (Table optimizing the variables within the ADX, such as its period
13), as the DMI_mod system has 90.18% probability to generate and ADX buffer level, or to develop proper money management
an average trade greater than zero compared with 67.56% techniques that fit the trading systems, in the hope of
probability in the original DMI system. The smallest probability generating maximized risk-adjusted return.
spread occurs on the Donchian’s MA System as applied on LQ45
index (Table 16), since the MA_DMImod system only has 83.24% Software and Data
probability compared with the MA_DMI system, which has Multicharts (www.multicharts.com)
77.42% probability. Market System Analyzer (www.adaptrade.com/MSA)
Another interesting factor that I observed is the worst-case R (www.r-project.org)
average trade occurring on the 95% confidence level. As we can Microsoft Excel
see, the Modified DM input has provided us with a better result, eSignal (www.esignal.com)
even in the expected worst condition. Although the Donchian’s
MA System on LQ45 index has the lowest probability spread, Notes
it has better worst-case average trade of the MA_DMImod 1. Edwards, Robert D. and John Magee, Technical Analysis of Stock Trends 9th ed.,
FL: Taylor & Francis Group, 2006, p. 22.
system, only losing $841.10 compared to $1,303.27 on the
2. Wilder, Welles J., New Concepts in Technical Trading Systems, NY: Trend
MA_DMI system. Therefore, there are strong reasons to believe Research, 1978, p. 47.
that systems with closing price as their Modified DM input will 3. ibid., p. 48.
provide significantly better trading performances than systems
that used original DM input, due to the higher probability of References
generating average trades greater than zero as well as the Aronson, David R., Evidence-Based Technical Analysis: Applying the Scientific
Method and Statistical Inference to Trading Signals, NY: John Wiley & Sons, 2007.
capability of producing higher profits on the winning trades,
Edwards, Robert D. and John Magee, Technical Analysis of Stock Trends 9th ed., FL:
lower losses on the losing trades, and lower losses on the worst- Taylor & Francis Group, 2006.
case average trade factor. Gujarati, Damodar. Basic Econometrics 4th ed., NY: The McGraw-Hill, 2004.
Kaufman, Perry J., Trading Systems and Methods 3rd ed., NY: John Wiley &
Conclusion Sons, 1998.
Marci, Glenn. “Is Average True Range a Superior Volatility Measure?” IFTA
This paper shows that the original DM input has caused Journal (2012): pp. 45–56.
biases within the ADX output value because it utilized the Schaap, Charles B., ADXcellence: Power Trend Strategies, NV:
components of a trading range, which are the highs and lows StockMarketStore.com, 2006.

between two consecutive bars. The biases occurred as the Wilder, Welles J., New Concepts in Technical Trading Systems, NY: Trend Research, 1978.

scenarios to determine the original DM value are fundamentally

flawed and caused many misinterpretations of the proper
direction of the price movements. Through the statistical
analysis conducted in this paper, I now believe we have strong
reasons to utilize closing price as the main input of Modified

IFTA.ORG PAGE 81
 IFTA JOURNAL 2015 EDITION

Fergal A. Walsh
[email protected]
The Alternative Head and Shoulders: A New 39 Clarinda Park East

Perspective on A Pre-Eminent Pattern Dunlaoighre

Co. Dublin 01
Ireland
By Fergal Walsh, MFTA
+3530863235805

Abstract pertain to a range of polities and derive from three distinct

The Head and Shoulders formation is arguably the most asset classes: currencies, commodities, and equities. I was
well-known market pattern in the field of Technical Analysis. especially eager to include currencies due to their especially
It features prominently in Technical Analysis textbooks and high liquidity, 24-hour markets, and lack of a centralized
has been the topic of a research paper produced by the Federal exchange, factors highlighted by the authors of the Federal
Reserve Bank of New York. More importantly though, it is a Reserve research paper.1
common feature on charts right across the asset spectrum,
from precious metals to stock indices to currency markets, Defining Patterns
readily observable from even a precursory glance. Although it Technical analysis textbooks provide basic templates with
has been demonstrated to be a fairly accurate predictor of price which to identify head and shoulders patterns but rarely
reversals, I intend to demonstrate that an alternative to the provide quantitative parameters. Indeed none of the sources
traditional analytical approach exists, and that this alternative I referenced, namely Technical Analysis,2 Technical Analysis of
yields considerably superior results with regard to market Financial Markets,3 Technical Analysis of Stock Trends,4 Technical
prognostication. Analysis Explained and Encyclopaedia of Chart Patterns,5
In this paper, I catalogue all head and shoulders and provided specific quantitative guidelines regarding pattern
alternative head and shoulders patterns across four asset formation. I have adhered to the basic tenets of a pattern
classes (GBP/USD, AUD/USD, SPX and XAU/USD) from the formation in that each recorded pattern appears at the end of
period November 2001–April 2013 and conduct a comparative a trend and possesses two distinct shoulders connected to an
analysis between the two patterns across the aforementioned intervening head by a neckline. The measured target, derived
securities over the same time period. Subsequently, I from measuring the distance from the apex of the head to the
investigate whether extreme Relative Strength Index (RSI) neckline projected from the neckline, is likewise retained. The
readings have any predictive value to successful pattern invalidation level, for which there appears to be no general
completion. I also examine pattern size to determine if consensus amongst market technicians is, for the purpose
there is any correlation between size and successful pattern of this study, the apex of the head. Unlike the authors of The
completion. My results show that, over the test period in Head and Shoulders: not just a flaky pattern,7 I have not included
question, the alternative head and shoulders is a better symmetry between shoulders as a necessary criterion. While it
market prognosticator and yields superior nominal profits is evident that many patterns do exhibit this tendency, I concur
compared to the traditional pattern. Given the nature of with Edwards and Magee in concluding that “symmetry is not
my undertaking, and lacking access to more sophisticated essential to a significant Head and Shoulders development,”
software for back testing, manual measurement was utilized and since this is not an intrinsically vital aspect of the pattern
for this project. As such, the figures perused in constructing (as opposed to, for example, the tenet that the head must be, in
the data tables throughout this paper are approximate; in all cases, greater in height than both shoulders) I don’t believe
addition, the drawing of lines and the evaluation of whether or its exclusion is detrimental to this study.
not a particular pattern constitutes a Head and Shoulders are Chart pattern analysis is perhaps the most unsuitable field
subjective matters and opinions may vary. With all this in mind, of technical analysis to quantify. As Chang & Osler observe,
and given the minor errors contained within this work, for they are “…highly non-linear and complex, trading rules based
anyone seeking to utilise the concepts/data located within this on these patterns normally cannot be expressed algebraically.”
paper, I ask that you conduct your own back tests beforehand I have attempted to address the issue of quantitative absence
and proceed only if you are pleased with your results. cited above by utilizing a series of guidelines (located below
this paragraph), in addition to standard rules, for determining
The Selection Process a head and shoulders pattern. However, in the realm of pattern
When deliberating over which assets to include in my analysis and, in particular, the more convoluted patterns
study, I focused primarily on liquidity, asset class and such as the head and shoulders, a degree of qualitative
geopolitical spread. I feel the selections conform to these interpretation is necessary.
principles in that all securities chosen are highly liquid,

PAGE 82 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

1. During pattern formation, price must not breach the I have applied in an attempt to provide a more quantitative
neckline by more 10% of the head height. basis for observation. Deviations from idealised patterns
2. Both shoulders must be at least 33.3 % of the head height, as across varying technical analysis fields are not uncommon.
measured from their respective positions on the neckline. The important factor is not the aesthetic quality but rather
3. The height measured from the start of a pattern formation to whether the pattern itself conforms to the underlying tenets
the apex of the head must not exceed 60% of the total price of its construction. Guidelines 2 and 3 are illustrated below for
movement from the preceding trough/peak unless a new further clarification in Figures 1 and 2 respectively.
high in the current overall trend is established
The Alternative Head and Shoulders
Discussion of these guidelines is necessary: An alternative head and shoulders (my own terminology)
pattern is identical in formation to the traditional version and
No. 1: I feel that it is overly pedantic to nullify a pattern occurs when either A) price breaks through the neckline of a
because a perfect neckline could not be drawn connecting the traditional H&S formation but does not meet the measured
two shoulders and head. However, it is also true that too much target (the length of the base to the apex of the head) and
leniency in this regard renders the pattern suspect. 10% is, I instead continues onward in the direction of the prior trend,
believe, an appropriate compromise, in that it provides a degree rising above the highest point of the head, or B) when a
of interpretative flexibility without damaging the integrity of traditional head and shoulders pattern develops but fails
the pattern. to activate with a close beyond the neckline and advances
No. 2: The situation is more nuanced where there exists a beyond the apex of the head. Once price surpasses the head, the
sloping neckline. Essentially, the height of the head at the base pattern is initiated.
of the shoulder and not the total head height as measured from The target is derived from the same method used to
the neckline, is the point of reference. From this point, the calculate that of the traditional pattern, but instead of
shoulder (measured from the neckline to the shoulder apex) in projecting the measurement forward from the neckline, it is
question must be at least 33.3% of the distance to the apex of projected forward from the apex of the head. If price breaks
the head. (See Figure 1) back below the lowest point (in the case of regular patterns)
No. 3: This might be considered contentious. The crux of or above the highest point (in the case of inverted patterns) of
the issue here is whether or not there is enough of a trend to the pattern, it is considered invalidated. As necklines at times
reverse for a head and shoulders pattern to be considered a have a tendency to slope, I judged it prudent to substitute
genuine reversal formation, or whether it should be classified this invalidation criteria in place of a movement beyond the
as a continuation pattern. Continuation patterns themselves neckline. Take, for example, the case of a downward sloping
are a rather tentative element of head and shoulders literature. neckline for a regular alternative head and shoulders pattern.
They occur at the apex of an upward price movement following With each passing day, the neckline descends, necessitating
a larger decline or at the trough of a downward price movement an ever larger stop loss. For particularly protracted price
following a larger advance. Sometimes, they simply appear at action following the activation of an alternative pattern that
the apex of uptrends in the form of inverse patterns, and at the subsequently fails, the resulting loss could be egregious. As
trough of downtrends in the form af regular patterns. They are such, establishing a concrete and readily observable point of
mentioned fleetingly in Tech Analysis Explained10 and Technical pattern failure was, in my opinion, a prudent measure. So as
Analysis of Financial Markets 11 but receive no attention in to avoid any confusion, I have provided an illustration of the
Technical Analysis,12 Technical Analysis of Stock Trends13 and alternative pattern in Figure 3.
Encyclopaedia of Chart Patterns.14
Kirkpatrick merely states that some patterns appear in Comparing Results
times of consolidation. The rules with respect to formation, Once I had recorded and analysed the results of my tests,
targets and activation are identical to those of reversal it was clear that the alternative head and shoulders pattern
patterns. As such, and provided that there is no universally was a superior prognosticator of market direction relative to
accepted conclusion of their nature, I have treated patterns the traditional pattern. Across all four assets, the alternative
that have formed after only a brief price movement as reversal version had a higher success percentage. The success rate
rather than continuation patterns, provided they meet the ranged between 58.33% (SPX) and 78.78 % (XAU/USD) for the
criteria of not exceeding 60% of the total price movement from former and between 45.83% (AUD/USD) and 53.19% (XAU/
the preceding trough in the case of uptrends, and peak in the USD) for the latter. Furthermore, trading the alternative
case of downtrends. (See Figure 2) pattern in accordance with the measurement target described
These parameters were used to establish clear boundaries earlier in this paper yielded considerably higher nominal gains
for defining head and shoulder patterns, and give a level compared to the traditional head and shoulders (Table 3).
of precision not found in mainstream technical analysis In all four tests, the alternative pattern recorded profits; in
handbooks. I have applied these rules across the test samples contrast, only two of the traditional patterns were profitable.
in an unbiased manner, cataloguing patterns which, at times, The alternative pattern outperformed the traditional in AUD/
do not conform to the idealized, immediately observable USD, GBP/USD, XAU/USD and the SPX. A full appraisal of my
head and shoulders pattern but that nonetheless fit the basic findings can be found below. All nominal figures are exclusive
criteria of technical analysis manuals, and also that which of transaction costs.

IFTA.ORG PAGE 83
 IFTA JOURNAL 2015 EDITION

Table 1. Traditional Head and Shoulders Pattern Nominal

And Percentage Success Figures capability when an extreme reading occurs simultaneously with
Nominal Nominal % Success pattern formation.
Success Failure
Table 4. Traditional Patterns with Extreme Rsi Readings
GBP/USD 24 24 50%
During Formation (T = Total S= Successful F= Failed)
AUD/USD 22 26 45.83%
GBP/USD AUD/USD XAU/USD SPX
XAU/USD 25 22 53.19%
T= 23. S=12 F=11 T=18. S=9 F=9 T=18. S=8 F=10 T=11. S=7 F=4
SPX 18 19 48.64%

Table 5. Alternative Patterns With Extreme Rsi Readings

Table 2. Alternative Head and Shoulders Pattern During Formation (T = Total S= Successful F= Failed)
Nominal and Percentage Success Figures
Nominal Nominal % Success GBP/USD AUD/USD XAU/USD SPX
Success Failure T=13. S=9 F=4 T=16. S=12 F=4 T=12. S=10 F =2 T=10. S=5 F=5
GBP/USD 20 11 64.51%
AUD/USD 32 12 72.72% Pattern Size
XAU/USD 26 7 78.78% The results for each asset class were divided into three
SPX 21 15 58.33% categories based on height. Rather than assign an arbitrary
range to each category, I took the smallest and largest size
Table 3. Profits In Price Units For Each Security patterns in each instrument and divided the intervening sum
GBP/USD TRADITIONAL -619 pips into three equal ranges, which were subsequently termed small,
GBP/USD ALTERNATIVE +1724 pips intermediate, and large, respectively. The vast majority of
patterns in each test period fell under categories 1 and 2, with
AUD/USD TRADITIONAL -1348pips
3 accounting for only a minority of patterns, both traditional
AUD/USD ALTERNATIVE +5624pips
and alternative. For alternative head and shoulders, the success
XAU/USD TRADITIONAL +213.70 $
rate of small and intermediate patterns was high. Given that a
XAU/USD ALTERNATIVE +469.23 $ considerable majority of patterns formed in these two categories,
SPX TRADIONAL +68.8pts I feel this is particularly noteworthy and encouraging.
SPX ALTERNATIVE +229.4 pts
Table 6. Traditional Head and Shoulders Pattern Size
Distribution
RSI Application
I applied the RSI to each asset class to ascertain whether Small % Small Inter- Inter- Large % Large
of Total Success mediate mediate of Total Success
there was any predictive significance in overbought (>70) or % % of Total Success % %
oversold (<30) levels in the RSI occurring during the formation GBP/USD 77.08% 48.64% 18.72 55.55% 4.16% 50%
of both the traditional and alternative head and shoulders AUD/USD 60.86% 53.57% 26.07% 58.33% 13.03% 16.66%
patterns. In the case of normal patterns, each reading above 70
XAU/USD 65.21% 50% 28.24% 46.15% 6.51% 100%
was noted, while for inverted patterns, each reading below 30
SPX 69.44% 44% 22.16% 50% 8.31% 66.66%
was noted. For this I used the 14-day setting. This was the period
utilized by the indicator’s creator, Welles Wilder, and continues
to be the most frequently used setting.15 The results were mixed. Table 7. Alternative Head and Shoulders Pattern Size
Across the four test periods, less than half of regular patterns Distribution
were accompanied by an extreme RSI reading, an outcome Small % Small Inter- Inter- Large Large
of Total Success mediate mediate % of Success
mirrored by the alternative patterns. This indicated that there % % of Total Success % Total %
was no positive correlation between extreme RSI readings and
GBP/USD 62.50% 70% 31.20% 70% 6.30% 0%
head and shoulders patterns, both traditional and alternative.
AUD/USD 80.95% 76.47% 16.66% 85.71% 2.38% 100%
Examining those patterns that were associated with extreme
readings presented some interesting findings, however. While XAU/USD 54.54% 83.33% 27.27% 88.88% 18.18% 66.66%
analysis of traditional patterns accompanied by extreme SPX 86.48% 56.25% 10.80% 75% 2.70% 0%
readings (Table 3) indicated only a minor connection between
successful pattern completion and an oversold or overbought
RSI, extreme RSI readings were significant in forecasting the Pattern Variety
success of subsequent alternative H&S patterns (Table 4). An The distribution of inverted and regular patterns across
RSI above 70 during any stage of a regular pattern formation all assets for both the traditional and alternative head and
constituted one instance; similarly, an RSI below 30 during any shoulders was quite consistent, with regulars accounting for the
stage of an inverted pattern formation also constituted one vast majority of occurrences, while inverted patterns accounted
instance. Regardless of how many times the RSI registered an for a minority. This was an unsurprising finding; Bulkowski,
extreme reading during each pattern development, only one in his vast survey of head and shoulders patterns, found that
instance was recorded for each individual pattern. This suggests regular patterns were far more frequent than inverted.16
that for alternative patterns, the RSI has some predictive

PAGE 84 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Figure 1

Figure 2

Figure 3

IFTA.ORG PAGE 85
 IFTA JOURNAL 2015 EDITION

Table 8. Pattern Variety for Traditional Head And

Shoulders Patterns Conclusion
Regular Inverted My objective for this paper was to introduce the alternative
GBP/USD 32 16 head and shoulders pattern; catalogue its occurrences across
AUD/USD 32 19 GBP/USD, AUD/USD, XAU/USD and SPX; and detail specific
XAU/USD 33 15 aspects, such as size and variety, so that it could be compared and
SPX 27 10 contrasted with similar data pertaining to the traditional head
and shoulders pattern, in order to prove its superiority as both a
profitable trading pattern and market prognosticator. I believe
Table 9. Pattern Variety for Alternative Head And
I have demonstrated the validity of its superiority over the test
Shoulders Patterns
period in question; my data clearly show that the alternative
Regular Inverted pattern surpasses the traditional pattern in both success rate
GBP/USD 21 10 and nominal gains accrued. I believe further investigation across
AUD/USD 32 12 a wider selection of asset classes and over a longer timeframe is
XAU/USD 28 4 warranted and will provide encouraging results.
SPX 30 6
Software and Data
Practical Applications FXCM
Having demonstrated the superiority of the alternative Marketscope
pattern, turning briefly to practical applications of the findings Incrediblechar
would seem logical. To this end, there are, in my opinion, three
methods of utilizing the data presented above. The first would Notes
be to simply trade the alternative pattern as it appears across 1. Chang, K.P.H & Osler, C.L, ‘Head and Shoulders: Not just a flaky pattern’,
Federal Reserve Bank of New York, 1995
the AUD/USD, GBP/USD, XAU/USD and SPX. As demonstrated, 2. Kirkpatrick II, C.D & Dahlquist, J.R, ‘Technical analysis’, FT Press, New Jersey, 2010
this approach would have been profitable had a prospective 3. Murphy, J.J, ‘Technical analysis of the financial markets’, Penguin Group, New
trader been following this pattern from the beginning of the York, New York, 1999
4. Edwards, R.D & Magee, J, ‘Technical Analysis of stock trends’, BN Publishing, 2008
data series (May 2001). Alternatively, the pattern could be useful 5. Pring, M.J, ‘Technical Analysis Explained’, McGraw-Hill, 2002
for those already committed to the market. If, for example, 6. Bulkowski, T.N, ‘Encyclopaedia of chart patterns’, Wiley, Hoboken, New Jersey,
a trader is currently holding a long position in the AUD/USD, 2005
7. Chang, K.P.H & Osler, C.L, ‘Head and Shoulders: Not just a flaky pattern’, Federal
and an alternative head and shoulders pattern is activated,
Reserve Bank of New York, 1995
they would be aware that the probability is that the current 8. Edwards, R.D & Magee, J, ‘Technical Analysis of stock trends’, BN Publishing, 2008
trend will continue until at least a point equal to the height of 9. Chang, K.P.H & Osler, C.L, ‘Head and Shoulders: Not just a flaky pattern’, Federal
Reserve Bank of New York, 1995
the pattern, as measured from the neckline to the apex of the
10. M.J, ‘Technical Analysis Explained’, McGraw-Hill, 2002
head, projected upward from the apex. Finally, it could serve 11. Murphy, J.J, ‘Technical analysis of the financial markets’, Penguin Group, New
as a market prognosticator for those awaiting a turn in price, York, New York, 1999
12. Kirkpatrick II, C.D & Dahlquist, J.R, ‘Technical analysis’, FT Press, New
informing them that a reversal in trend is unlikely until, at least,
Jersey, 2012
the measured target of the alternative pattern is met. 13. Edwards, R.D & Magee, J, ‘Technical Analysis of stock trends’, BN Publishing, 2008
14. Bulkowski, T.N, ‘Encyclopaedia of chart patterns’, Wiley, Hoboken, New
Analysing Failure Jersey, 2005
15. Kirkpatrick II, C.D & Dahlquist, J.R, ‘Technical analysis’, FT Press, New
Why did the traditional head and shoulders pattern— Jersey, 2010
arguably the most prominent and supposedly the most reliable 16. Bulkowski, T.N, ‘Encyclopaedia of chart patterns’, Wiley, Hoboken, New
Jersey, 2005
of all technical chart patterns—deliver such mediocre results? 17. Levy, R.L, ‘The predictive significance of five-point chart patterns’, Journal of
The test period could simply have been a particularly bad business, Volume 44, Issue 3, 1971
timeframe for the pattern; however, that it performed thusly 18. Bulkowski, T.N, ‘Encyclopaedia of chart patterns’, Wiley, Hoboken, New
Jersey, 2005
across diverse asset classes and over a period encompassing 19. Chang, K.P.H & Osler, C.L, ‘Head and Shoulders: Not just a flaky pattern’, Federal
both bull and bear markets is not encouraging. Furthermore, Reserve Bank of New York, 1995
previous detailed studies have given a distinctly mixed picture 20. Pring, M.J, ‘Technical Analysis Explained’, McGraw-Hill, 2002
as to the reliability of the head and shoulders. For example,
Levy assigned no predictive value to the pattern,17 while References
Bulkowski rated it highly in his comprehensive study of equity Bulkowski, T.N, ‘Encyclopaedia of chart patterns’, Wiley, Hoboken, New Jersey, 2005
chart patterns.18 The Federal Reserve Bank of New York paper, Edwards, R.D & Magee, J, ‘Technical Analysis of stock trends’, BN Publishing, 2008
conducted using data from several FX pairs, concluded that Kirkpatrick II, C.D & Dahlquist, J.R, ‘Technical analysis’, FT Press, New Jersey, 2011

the pattern had predictive value for some currencies but not Murphy, J.J, ‘Technical analysis of the financial markets’, Penguin Group, New
York, New York, 1999
for others.19 Meanwhile Pring observes that in recent years the Pring, M.J, ‘Technical Analysis Explained’, McGraw-Hill, 2002
pattern has become less reliable.20 It would appear that further Chang, K.P.H & Osler, C.L, ‘Head and Shoulders: Not just a flaky pattern’, Federal
comprehensive studies are required to establish whether the Reserve Bank of New York, 1995
head and shoulders pattern deserves the pre-eminence it Levy, R.L, ‘The predictive significance of five-point chart patterns’, Journal of
business, Volume 44, Issue 3, 1971
currently holds amongst the technical analysis community.

PAGE 86 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Know Your System! – Turning Data Mining David Walton

[email protected]

From Bias to Benefit Through System 1741 Dunkeld Ln.

Folsom, CA 95630
Parameter Permutation USA
001-916-458-4330
By Dave Walton, MBA

Abstract interaction of system rules, portfolio effects and market data

The goal of the paper is to assist the trader in answering two that other methods do not. Thus, SPP enables a much deeper
questions: 1) “What is a reasonable performance estimate of the understanding of potential risks and rewards prior to allocating
long-run edge of the trading system?” and 2) “What worst-case capital to a system.
contingencies must be tolerated in short-run performance
in order to achieve the long-run expectation?” With this Basic Requirements, Definitions and
information, the trader can make probabilistic, data-driven Methods
decisions on whether to allocate capital to the system, and once
actively trading, whether the system is “broken” and should The only requirement for applying SPP as defined in this
cease trading. paper is that the trading system must be completely rules based
To achieve this goal, a method called System Parameter and use parameters that are optimized during the development
Permutation (SPP) is introduced that enables realistic process. This requirement is necessary because SPP makes
contingency planning based on probabilities. Many traders and use of the parameter optimization process and corresponding
system developers go to great lengths to avoid the effects of data. Thus SPP is most applicable to trading systems based on
randomness in simulated trading results, knowing the large technical analysis.
impact it may cause. In contrast, SPP embraces randomness as The following definitions and methods are used heavily
a tool to help uncover what may probabilistically be expected throughout this paper:
from a trading system in the future. The method is simple to Quantitative Trading System: A trading system defined by
apply yet very effective. clear, unambiguous, and comprehensive entry and exit rules
Application to an example trading system shows how which can be machine coded. The results of a quantitative
SPP fully leverages available historical data to enable deep trading system can be independently reproduced and verified.
understanding of potential risks and rewards prior to allocating Any mention of the term trading system in this paper implies it
capital to a trading system. is quantitative.
The trading system used in this study is long only, meaning
Introduction no shorting or inverse ETFs were used. In the context of a
Prior to putting capital at risk, every trader desires an historical simulation, long-only systems greatly reduce the
accurate estimate of the potential risks and rewards expected number of assumptions made regarding ability to borrow
from a trading system and often employs historical simulation shares, share call-backs, dividends paid, and interest charged.
to gain such an understanding. Unfortunately, many traders The trading system was developed based on published research
are subsequently frustrated by poor realized trading system and simulated using realistic portfolio simulation that accounts
performance that does not live up to overly optimistic for margin available, prioritization of trading signals, position
expectations. One large and prevalent source of overly sizing, and portfolio heat limits.
optimistic expectations that remains largely misunderstood Trading Universe: A total of 10 ETFs were used: SPDR S&P
and underestimated is the data mining bias (DMB). 500 ETF (SPY), iShares Russell 2000 ETF (IWM), iShares MSCI
Even though DMB tends to have a large impact on historical Emerging Markets ETF (EEM), iShares Core S&P Mid-Cap ETF
simulation results, mitigation tools available to the average (IJH), PowerShares QQQ ETF (QQQ), SPDR Gold Shares ETF
trader are relatively crude. More advanced tools are available (GLD), iShares MSCI EAFE ETF (EFA), iShares 20+ Year Treasury
to academics and quantitative professionals but are largely too Bond ETF (TLT), iShares US Real Estate ETF (IYR), and iShares 1-3
complex for the average trading system developer. This paper Year Treasury Bond ETF(SHY). These ETFs were chosen because
attempts to change that by introducing System Parameter they are the most liquid ETFs covering major asset classes.
Permutation (SPP). With SPP, the average trader is armed with a Historical Simulation Timeframe: The trading system
simple yet powerful tool to effectively mitigate DMB and more historical simulation period for this analysis begins 11/19/2005
accurately estimate future trading system performance. and ends 5/31/2013. The start date was selected to allow roughly
The power of SPP extends beyond mitigating DMB, one year of market data history for all traded ETFs (GLD started
however. SPP explores facets of the trading system due to the trading 11/19/2004) prior to any entry signal.

IFTA.ORG PAGE 87
 IFTA JOURNAL 2015 EDITION

Input Data: Daily OHLC market data were used from Norgate walk component is equally likely to be favorable (good luck)
Premium Data and adjusted for splits. Market data were not or unfavorable (bad luck). Thus, realized trading system
adjusted to include dividends in order to avoid non-linear, performance consists of two components of unknown relative
a posteriori distortion of technical indicator-based trading magnitude: the inherent edge and luck. Periods of good and
signals that use percentages (Kaufman 2013). To be as realistic bad luck cause variability around the long-run expected
as possible, historical dividend data were used from Yahoo! performance due to the system edge.
Finance,1 and dividend payments were injected into the portfolio The second precondition of DMB is the multiple comparison
as cash per the applicable ex-dividend date. and selection process inherent to the typical system
Transaction Costs and Fees: A $0.01 per share per side development process. At each stage of development, system
allowance was made for commissions as well as a 0.05% rules and parameters exhibiting the best performance are
estimate per side for slippage. Where applicable, margin interest selected from historical simulation results. This selection
was charged daily at a rate of 1.5% + the Fed Funds daily rate2 process is known as data mining. Because of the random
(varied between 0.04% and 5.41% in the simulation period). Data component in measured performance, the selected rules
for the Fed Funds daily rate was taken from the public website are guaranteed to have taken advantage of good luck. The
of the Federal Reserve Bank of New York.3 All order types used probability that a favorable result is due to chance alone
were Market-On-Open, Market-On-Close, or Market-On-Stop. increases with the number of combinations tested.
Thus, fees and slippage were modeled toward what a retail Almost all trading system development platforms support
trader might expect to see. multiple types of search optimization algorithms and thus lead
Output Data: Four different system metrics were evaluated: the developer, perhaps unknowingly, into a data mining venture.
1) compounded annual return including dividends, 2) max The process of data mining to find the best performing system
drawdown, 3) annualized information ratio4 (vs. dividend rules is not the problem however. Data mining in attempt to
re-invested SPY ETF), and 4) annualized standard deviation of find the best (in meeting the objectives of the trader) entry/exit
daily returns. For cross-validation of historical simulations, rules and best combination of parameters is a natural, intuitive
traditional Out-Of-Sample (OOS) testing was used for process. In fact, Aronson (2007) mentions that data mining
comparison to the SPP performance estimation method is the “preferred method of knowledge acquisition” when
discussed in this paper. In OOS testing, market data was split employing technical analysis.
into 80% training and 20% validation sets, with the validation The real problem is not considering that the performance
set comprising the most recent data. of the chosen system rules is inflated by good luck and that the
same amount of good luck is not likely to repeat in the future.
Data Mining Bias In fact, the statistical law of regression toward the mean5
Many traders are familiar with the idea that future trading indicates that extreme performance in historical simulation
system performance is likely to be worse than was seen in will be probabilistically followed by performance closer to the
historical simulation. However, the origins of this performance unknown, long-run level of performance of the inherent edge.
degradation are often not well understood. One significantly This is illustrated in Figure 1.
large cause is the DMB, also commonly known by other names
such as curve-fitting, over-fitting, data snooping, or over- Figure 1: Impact of Luck on Trading Results
optimization. DMB is built-into the typical system development
process and yet largely remains unknown, misunderstood, and/
or ignored.
This may be understandable for retail traders with limited
knowledge of statistics. However, Bailey et al. (2013) note that
professional publications also tend to disregard or gloss over
the effects of DMB. Unfortunately, ignoring the problem doesn’t
eliminate the consequence, which is that the trading system
fails to live up to performance expectations in cross-validation
or worse, in live trading.

Understanding Data Mining Bias

To understand DMB, one must first recognize its two
preconditions, which are inherent to the system development
process: 1) randomness and 2) a multiple comparison procedure
in the search for the best system rules. The interaction of The Consequences of Data Mining Bias
randomness and the search process is unique to the system DMB has two consequences: inflated performance metrics
rules evaluated and the historical market data and results in and inability to perform statistical inference using standard
inflated performance metrics. methods. Both consequences can lead to improper decision-
The first precondition of DMB—randomness—means the making.
random walk component of market data. In any sequence A logical question is how large DMB might be. Although
of trades, the result of system rules acting on the random the magnitude of the DMB is specific to the analyzed trading

PAGE 88 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

rules and market data, it can be quite large. For 6,402 simple significance testing. SPP is not subject to data mining bias6 and
trading rules data mined on the S&P500 index over 25 years of uses standard trading system optimization approaches that
historical data, Aronson (2007) found that the level of annual are already built-into generally available system development
return needed to overcome DMB was approximately 15% at the software packages.
significance level of α = 0.05 and none of the examined rules had SPP provides much more than a method to mitigate DMB,
any statistically significant edge. however. SPP enables the trader to objectively determine: 1)
Further, attempting to test the statistical significance of the performance of the inherent edge expected in the long-run,
performance metrics using standard statistical inference and 2) the worst-case performance expected in the short-
procedures is not valid when the data contain systematic run. With this information, the trader can make data-driven
error (DMB is systematic error). Sound statistical inference decisions on whether to allocate capital to the system and once
in the context of data mining requires the use of a sampling actively trading, whether the system is “broken” and should
distribution that includes the effect of good luck. cease trading.

Mitigating Data Mining Bias System Parameter Permutation Defined

DMB is systematic; it is inherent to the typical system In the simplest of terms, SPP generates sampling
development process. DMB cannot be lessened or eliminated distributions of system performance metrics by leveraging the
by evaluating via the “best” system performance metrics or system optimization process. Each point in a given distribution
by performing a “perfect” historical simulation (e.g., properly is the result of a historical simulation run that accurately
modeled transaction costs, clean and accurately adjusted modeled portfolio effects. Via sampling distributions, the trader
market data, no look-ahead bias, no hindsight bias, no may evaluate a system based on any desired performance
survivorship bias, properly modeled portfolio effects, omissions metrics. SPP then uses the descriptive statistics of the sampling
and contingencies considered). The only viable methods to distributions to arrive at performance estimates and measures
estimate performance or test significance in the presence of of statistical significance.
DMB are those that consider systematic error. Unlike standard optimization, SPP does not simply
One such method is to estimate performance and perform choose the best set of parameters but rather uses all of the
significance testing on an independent data sample, which performance data available for all sets of parameters evaluated
effectively is looking at system performance after regression during optimization. Whereas traditional optimization picks the
toward the mean has occurred; this is known as cross- best set of parameters and discards the rest, SPP makes use of
validation. Another method is to perform significance testing all available information. Figure 2 illustrates the difference.
by creating a sampling distribution of maximum means that
reflects the role that good luck plays in data mining; this is Figure 2: SPP Compared to Traditional Optimization
known as bias compensation. Yet another method is to calculate
a deflation factor for data mining bias, which is applied to
measured performance metrics.
Aronson (2007) explains each of these methods in detail. The
key strengths and weaknesses of each are summarized in Table 1.
Table 1: Comparison of Methods to Mitigate DMB
Strengths Weaknesses
Cross- Ease of use, allows Inefficient use of market
Validation statistical inference, data, smaller sample size
provides performance reduces accuracy
estimate
Bias Allows statistical Complex, special
Compensation inference, efficient use software + large
of market data database required, no
performance estimate
For each system metric of interest, the output of SPP is
Bias Deflation Provides performance Possibly inaccurate,
estimate, efficient use of large database required, a sampling distribution that includes trade results from all
market data statistical inference not system variants (combinations of parameter values) where the
possible median serves as the best estimate of true system performance.
This is very different from cross-validation or data mining
bias compensation, which use the result of a single sequence of
System Parameter Permutation trades to estimate system performance.
Each method of DMB mitigation described in the previous The median performance is used as the best estimate of
section has certain limitations or complexities. This section future performance for several reasons: 1) the median is not
offers an alternative method called System Parameter subject to data mining bias because no selection is involved; 2)
Permutation (SPP). SPP provides a practical means of estimating no assumptions of the shape of the distribution are required;
the performance of a trading system as well as statistical and 3) the median is robust in the presence of outlier values.

IFTA.ORG PAGE 89
 IFTA JOURNAL 2015 EDITION

Steps of System Parameter Permutation SPP Estimate of the Long–Run Performance of the Trading
To generate sampling distributions of system variant System
performance metrics, the set of parameter ranges under which The trader would like to answer the question: “What is a
the trading system is expected to function is determined ex reasonable performance estimate of the long-run edge of the
ante in preparation for optimization. Methods for choosing the system?” SPP can effectively answer this question.
parameter ranges and observation points are beyond the scope To generate long-run performance estimates, sampling
of this paper; however, Kaufman (2013) and Pardo (2008) are distributions are produced as described above using all available
suggested for further research into these topics. SPP follows market data. The use of all available market data enables the
these general steps: best approximation of the long-run, so the more market data
available, the more accurate the estimate. For each performance
1. Parameter scan ranges for the system concept are metric of interest, the median value is used as the best, unbiased
determined by the system developer. performance estimate.
2. Each parameter scan range is divided into an appropriate The trader may also be interested in testing the statistical
number of observation points (specific parameter values). significance of the SPP long-run performance estimates, either
3. Exhaustive optimization (all possible parameter value in terms of absolute returns or relative to a benchmark. Because
combinations) is performed using a realistic portfolio-based SPP generates complete sampling distributions, estimated
historical simulation over the selected time period. p-values and confidence levels may be observed directly from
4. The simulated results for each system variant are combined the CDF, as illustrated in Figure 4.
to create a sampling distribution for each performance
metric of interest (e.g., CAR, max drawdown, Sharpe ratio). Figure 4: Using the Cumulative Distribution Function for
Statistical Inference
Each point on a distribution is the result of a historical
simulation run from a single system variant.
Figure 3: General Steps of System Parameter Permutation

The example in Figure 4 shows that in 95% of cases, the

true value lies in the confidence interval above the level of
5% compounded annual return (CAR); this is equivalent to a
p-value of 0.05. Depending on the objectives of the trader, this
may or may not be satisfactory. If the trader is interested in
Figure 3 illustrates the process. In this case, sampling outperforming a benchmark with a CAR of 10%, the picture is a
distributions for four performance metrics are shown for bit different. In only 59% of cases does the true value lie in the
illustration. Any number of specific performance metrics confidence interval above the benchmark return; this is not
may be selected by the trader for his specific objectives. The statistically significant.
cumulative distribution function (CDF) for each metric may be
examined directly and may be used for performance estimation Short-Run Performance Estimate and Worst-Case
and statistical inference. Contingency Analysis
To ensure the SPP result is not biased, care must be taken Whereas the long-run performance estimate indicates what
to thoughtfully select parameter scan ranges ex ante. If SPP may be expected from the system edge long term, short-run
is repeated multiple times by changing the parameter scan variability may be significant. Thus, the trader would also like to
ranges in an attempt to get a better result, data mining is at answer the question: “What worst-case contingencies must be
work and the SPP estimate may become positively biased. Since tolerated in short-run performance in order to achieve the long-
the intent of SPP is the avoidance of bias, such a practice would run expectation?” Once the short-run time period is specified,
be counterproductive. Thus, it is important that the system SPP can effectively answer this question.
developer start the system development process with this The duration of the short-run time period is dependent on
consideration in mind. the preferences and psychology of the trader and/or clients.
Chekhlov et al. (2003) mention that the typical drawdown
duration tolerated by clients of managed account practitioners
ranges from 1-2 years at the most. In any case, the trader needs

PAGE 90 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Figure 6: Increased Variation in Short-Run Performance

to determine the duration of the short-run time period that best
50%
fits the trading objectives. In general, shorter duration periods
45%
have wider ranges of expected performance. 40%
The following steps explain how to perform SPP for the short- 35%
run time period: 30%

Frequency
25% Long-run
20%
1. All available market data is split into blocks equal in length Short-run
15%
to the short-run time period (t). Each time block may overlap 10%
with the previous block depending on the timeframe of 5%
trading signals (such as any month within a year or any hour 0%

-5
-30
-25
-20
-15
-10

0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
within a day). This results in some number of time blocks (m).
CAR
2. Steps 1 through 4 of the general SPP process are performed
on all m time blocks separately. Thus, if a system has
n combinations of parameter values, a total of m • n With sampling distributions, the trader may make a
optimization permutations are performed on a historical probabilistic, data-driven decision of whether to risk capital on
time period of length t in order to generate the sampling the system. To do so, the trader determines a probability level he
distribution for each performance metric of interest over the determines to be highly improbable but tolerable as his worst-
selected short-run timeframe. case (common levels are 5% or 1%). Alternatively, the trader
may specify the worst-case in terms of the least favorable but
Figure 5 illustrates the process. Again, sampling tolerable level of performance. Whatever worst-case probability
distributions for four performance metrics are shown for or level of performance is chosen, the CDF of the short-run
illustration. Any number of specific performance metrics may system metrics of choice are examined as in Figure 7. If the
be selected by the trader for his specific objectives. worst-case contingency at the respective probability cannot be
tolerated by the trader or clients, capital should not be allocated
Figure 5: SPP Application for Short-Run Performance to the system.
Estimation
Figure 7: Evaluating the “Worst-Case Contingency”
Compounded Annual Return
100

80

60

Worst-Case Contingency Probability = 5% (p-val = 0.05)

40

20

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
-5.51% CAR Worst Case Contingency (@ SPP 5%)
-20

-40

The example in Figure 7 indicates that if the trader cannot

tolerate a 5% probability of a realized compounded annual
return of -5.51% over the short-run time period chosen, the
The sampling distributions resulting from this process each system should not be traded. If the worst-case contingency
contain many more individual samples with a higher degree of is tolerable and capital is allocated to the trading system, the
variation than were generated via the SPP long-run performance same (or different) worst-case probability(s) or level(s) of
estimate process. However each sample has a shorter simulation performance may be used to determine whether the system is
timeframe and thus a fewer number of closed trades contained “broken” and if trading should cease.
in each sample. With fewer closed trades per sample, the The stop trading decision should be made when the system has
standard error associated with each sample increases. As the been traded for the duration of the short-term period selected.
standard error per sample increases, so does the variation of the Thus, if one year was selected as the short-term time period, the
sampling distribution. The increased variation can be seen in the stop trading decision should only be made at the one year mark.
respective probability density functions as shown in Figure 6. Again, using the example in Figure 7, if realized performance is
worse than -5.51% over a year of trading, the trader may decide
to stop trading the system because the ex ante worst case
contingency was violated. Any timeframe or probability may

IFTA.ORG PAGE 91
 IFTA JOURNAL 2015 EDITION

be used in this decision. Thus SPP enables the trader to add an of entry and exit signals where other similar combinations may
objective method of risk control to his trading plan. generate much less favorable signals on the same market data.
SPP generates a distribution of performance results from
Why System Parameter Permutation Is Effective a large number of individual historical simulation runs that
Using traditional optimization, all performance metrics use the same market data applied to different combinations of
for the system are derived from the single (best) sequence of parameter values. The distribution includes the results from
trades selected during the optimization process. To generate many slightly different entry/exit signal combinations across
a distribution of contingencies, randomization techniques simulation runs. With a large number of samples, the impact of
employing resampling, such as bootstrap or Monte Carlo regression toward the mean is seen to varying degrees over the
Permutation (MCP), are commonly used. distribution of system variant performance results, as shown in
There are several problematic assumptions made by Figure 8 below.
resampling methods, but two are of particular interest here: 1)
the result of a single historical simulation is representative of Figure 8: Sampling Distribution Generated by SPP
the future distribution of trade results; 2) real world portfolio
effects combined with position sizing are accurately modeled. Median performance
The discussion of data mining bias already showed that
assumption number one is problematic. Assumption number
two is also problematic; portfolio effects such as buying power,
dynamic inter-symbol correlation, and autocorrelation would
likely not allow some of the resampled results to occur in real
trading. Likewise, this type of randomization does not explore
trades unseen in the original, single sample sequence of trades
Bad luck Good luck
that may have occurred under slightly different conditions. This Deflates Inflates
is a natural consequence of random resampling. Performance Performance
Unlike random resampling, the random variation in SPP
originates from the application of a set of slightly varied entry/
exit rules on actual market data, where trading signals are
evaluated using a realistic simulated portfolio. In effect, SPP
explores facets of the trading system that would otherwise
remain hidden yet are possible in real trading. Another way luck affects system performance is through
SPP produces reliable estimates of trading system the interaction of the timing of market entry/exit signals and
performance by: 1) leveraging the statistical law of regression portfolio effects such as buying power, dynamic inter-symbol
toward the mean, and 2) extracting maximum information correlation, and autocorrelation. As demonstrated by Krawinkel
from available market data. For #1, the use of a large number of (2011) randomly skipped trades can have a large impact on
combinations of parameter values thoroughly examines various realized system performance. Yet, this phenomenon remains
ways randomness may affect the system and thus estimates largely unrecognized and underexplored. SPP thoroughly and
the effects of regression toward the mean. For #2, the use of all realistically explores this effect through the distribution of
available market data ensures that performance results contain performance results.
the smallest standard error possible and that the system has In SPP, one combination of parameter values may capture a
been exposed to the most varied market conditions possible. certain set of trades, whereas a slight variation in parameter
Both are explored in more detail. values may capture trades not previously seen and/or skip
others that were previously captured. Through this interaction,
How SPP Leverages Regression Toward the Mean SPP includes the effects of randomly skipped and included
In system optimization, regression toward the mean trades. Again, the impact of regression toward the mean is seen
indicates that the specific combination of optimized parameter to varying degrees over the distribution of performance results.
values that led to extreme performance in historical simulation
will probabilistically not retain a level of extreme performance How SPP Extracts Maximum Information From Available
in the future. The section on DMB showed that extreme Market Data
performance tends to regress toward the mean level over SPP minimizes standard error of the mean (SEM) by using
time as the impact of luck tends to change. It is instructive to all available market data in the historical simulation. As sample
examine the mechanics of how luck affects system variant size increases, SEM decreases proportionally to the square root
performance. of the sample size due to the mathematical identity: SEM = s/ .
In general, good luck involves some combination of Although the use of all available market data is not a unique
catching favorable market moves and avoiding adverse market feature of SPP, it is one of its strengths. In contrast, traditional
moves. One way luck affects system performance is through cross-validation methods split market data in some way. The
the interaction of parameters on market data. Parameter effect on SEM of such a split can be large. Table 2 shows the
values control the exact timing of entry and exit signals; one approximate percentage increase of SEM for various data-
combination of parameters may generate a very favorable set splitting schemes over SPP.

PAGE 92 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Table 2: Increase of SEM for Market Data Splits

The ROC indicator was chosen to measure momentum as
50/50 Split 80/20 Split 90/10 Split
the percentage change over the look-back period. The timing
In-sample 41% 12% 5%
of asset rotation was chosen to be once per month on a specific
Out-of-sample 41% 124% 216% day in relation to the last trading day of month. Finally a
catastrophic stop loss as a percentage of the entry price was
Further, Inoue and Kilian (2002) found that OOS and IS introduced for risk management.
tests are equally reliable in the presence of data mining once The parameter scan ranges were defined in light of the
proper critical values are used and that IS (using all market generalized system concept. The portfolio composition was
data) tests have power advantages when there is “unmodelled limited to the top two to five assets out of 10 in a balance
structural change in the parameter of interest” (a change between momentum and diversification. The ROC look-back
in market conditions). The use of all available market data length was varied in increments of 10% starting from 60 trading
ensures that the system has been exposed to the most varied days (~3 months) up to 251 trading days (~1 year). The date of
market conditions possible in historical simulation. Doing so entry/exit rotation chosen was the last trading day of month +/-
cannot guarantee that future market conditions will be similar 5 trading days. Finally the stop loss was varied from 10% to 20%
to those seen historically, but any sort of data split ensures in increments of 2%. The system details are shown in Table 3.
loss of information and thus less representative performance
results. The most information-rich historical simulation uses all Table 3: Relative Momentum Trading System Details
available market data. System Indicator Minimum Maximum Step # Values
Component
Practical Example of SPP Applied to a # Assets
N/A 2 5 1 4
Held
Model Trading System
Momentum
ROC(a) 60 251 10% 16
The relative momentum concept in the style of Blitz and Van Rank
Vliet (2008) was chosen to create an example system because Rotation Last DOM
-5 5 1 11
significant research has validated these types of strategies Time Period +b
within and across many different asset classes (Asness et al. Stop Loss % of entry
10% 20% 2% 6
2009). Further, a large amount of post-publication, out-of- Point price
sample validation exists for relative momentum (Asness et al.
2009), thus confirming its viability. Exhaustive optimization of the above scan ranges resulted in
4,224 combinations of parameter values. The method of position
Generalized System Model sizing used was equal margin per position. A standard 100%
The relative momentum trading system concept is based on maintenance margin requirement was used along with a 5%
the observation that the best performing assets or asset classes cash safety buffer. This allowed up to 95% of trading capital to
in the current period tend to continue their outperformance in be used to take entry signals.
the next period. Research indicates that momentum measured
over 3–12 months tends to show the largest edge. Optimization Results and Out-of-Sample Calibration
The generalized system model defines how momentum This section discusses traditional OOS testing applied to
is measured, the number of assets to comprise the portfolio, the trading system. The OOS analysis is used for comparison
and the timing of asset selection. In the interest of risk purposes to SPP. The trading system was optimized using the
management, a catastrophic stop-loss is added to the general annualized information ratio (vs. the dividend reinvested SPY
model as well. Thus, the generalized system model shown in ETF) as the fitness function in order to maximize benchmark
Figure 9 contains four parameters. outperformance. The OOS test used 80% of available market
data in-sample, and the remaining 20% was reserved for out-of-
Figure 9: Relative Momentum Generalized System Model sample calibration. Results are shown in Table 4.
Table 4: Relative Momentum System OOS Results

IS OOS OOS %
of IS
Compounded Annual Return 22.41% 14.47% 65%
Maximum Drawdown -18.06% -9.6% 53%
Annualized Standard Deviation 19.08% 11.96% 63%
Annualized Information Ratio 0.70 -1.20 -171%

Using this method, the OOS performance metrics serve

as the only unbiased estimates in setting expectations for
future performance and thus also serve as the determinant

IFTA.ORG PAGE 93
 IFTA JOURNAL 2015 EDITION

Table 5: Long-Run SPP Estimate vs. OOS and Benchmark

of whether to risk capital on the system. Standard practice in
OOS testing dictates that a system passes cross-validation if Compounded Maximum Annualized Annualized
Annual Drawdown Information Standard
OOS performance is >= 50% of IS performance. In this case, the Return Ratio Deviation
majority of the system metrics are above the desired threshold, Cross-
yet the information ratio for the OOS segment is much below. Validation 14.47% -9.60% -1.20 11.96%
Therefore, this system fails traditional cross-validation due to OOS Estimate
the unacceptable OOS information ratio. SPP Estimate
of Long-Run 8.94% -24.22% 0.06 15.61%
SPP Long-Run Estimate of System Performance Perf.
Next, SPP was performed on the system as specified in the SPY
Benchmark 6.54% -55.05% N/A 24.97%
section titled “SPP Estimate of the Long–Run Performance of
the Trading System.” In contrast to the previously described
Equiv. P-Val
method, SPP uses all available market data, and when applied for Outper- 0.10 0.00 0.25 0.00
to the same system, provides much more information. Table 5 formance
shows the traditional OOS results/estimates compared to the
respective SPP estimates and to the buy-and-hold benchmark The results in Table 5 are taken from specific points
(SPY ETF with dividends reinvested). The goal in employing along the SPP sampling distributions. For the four system
this system is to outperform the benchmark, and thus, the metrics examined in this example, the CDFs (blue) are
statistical significance of outperformance for each system shown in Figure 10, from which the trader can make further
metric (via the equivalent p-value) is also shown. probabilistic estimates. The system metrics are shown on the
The data in Table 5 may be used by the trader to decide y-axis of the charts and the cumulative probabilities on the
whether to allocate capital to the system. For example, the x-axis.
trader may ask “Is an unbiased estimate of realizing a 8.94% Additionally, the SPP estimate (red), OOS estimate (green)
CAR sufficient reward to compensate for the risk of a -24.22% and benchmark (purple) are overlaid onto each CDF chart.
drawdown and an annualized 15.61% standard deviation? Is The vertical black line highlights the median of the sampling
a p-value of 0.10 for CAR significant enough to be confident distribution. The intersection point of the CDF and the
in outperforming the benchmark?” These questions may be benchmark as measured along the x-axis is the value of the
answered via the SPP generated sampling distributions. equivalent p-value from Table 5.

Figure 10: Long-Run SPP Generated CDFs for Selected System Metrics

Compounded Annual Return Maximum Drawdown

25 0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
-9.60% OOS Estimate
-10
20

-20
15
14.47% OOS Estimate -24.22% SPP Estimate

-30
10 8.94% SPP Estimate
-40

5 6.54% SPY Benchmark

-50
-55.05% SPY Benchmark
0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% -60

Information Ratio (Annualized) Standard Deviation (Annualized)

0.6
24 24.97% SPY Benchmark
0.5
22
0.4

20
0.3

0.2 18

0.1 0.06 SPP Estimate 16

0
15.61% SPP Estimate
14
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
-0.1
12
11.96% OOS Estimate
-0.2 -1.20 OOS Estimate
10
-0.3 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

PAGE 94 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

SPP Worst-Case Contingency Analysis for Calendar Year compared to the SPP worst-case contingency and to the range
Performance of the buy-and-hold benchmark over each of the seven full
The next analysis uses the calendar year as the short-term calendar years in the historical simulation period.
time period of interest. The historical market data were divided The trader must decide whether the worst-case contingency
into seven blocks, for each of the full calendar years present in for calendar year performance shown in Table 6 is tolerable in
the data. The process from Section 4.2.2 was completed on this order to achieve the long-run SPP performance expectations
data to evaluate the expected worst-case contingency for any of the trading system shown in Table 5 (previous section). For
calendar year period. example, the trader must be prepared to accept a 5% probability
of realizing a -1.45 annualized information ratio (significantly
Table 6: Calendar Year SPP Worst-Case Contingency vs. underperforming the benchmark) in any given calendar year,
OOS and Benchmark
while at the same time achieving negative absolute returns
Compounded Maximum Annualized Annualized (-13% CAR).
Annual Drawdown Information Standard
Return Ratio Deviation Figure 11 shows the CDFs (blue) for the four chosen system
Cross- metrics as well as the SPP estimate (red) and OOS estimate
Validation 14.47% -9.60% -1.20 11.96% (green) overlaid. The vertical black line highlights the 5th
OOS Estimate percentile of the sampling distribution (worst-case contingency
Worst-Case probability chosen), and the calendar year range of the
Contingency -12.98% -23.95% -1.45 21.67% benchmark is shown by a purple bar on the y-axis.
(@SPP 5%)
SPY Discussion of Results
Benchmark -36.27% -47.04% N/A 10.03% The above example showed that, compared to standard
Minimum
OOS cross-validation, SPP provides the trader with much more
SPY
Benchmark 22.8% -7.63% N/A 41.92% information. SPP creates long-run and short-run sampling
Maximum distributions of system metrics using all available historical
market data, whereas traditional OOS cross-validation provides
In this case, the SPP 5th percentile (equivalent to p-value = only a point estimate on a subset of historical market data. SPP
0.05) was chosen as the worst-case contingency probability. enables probabilistic decision-making, whereas traditional
Table 6 shows the same OOS results/estimate from above OOS necessitates a binary pass/fail decision. Thus, SPP enables

Figure 11: Calendar Year SPP Generated CDFs for Selected System Metrics

Compounded Annual Return Maximum Drawdown

70 0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
60 -5
-9.60% OOS Estimate
50 -10
40
-15
30
-20
20 14.47% OOS Estimate
-25
10
-23.95% Worst-Case Contingency (@ SPP 5%)
-30
0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% -35
-10

-20 -12.98% Worst-Case Contingency (@ SPP 5%) -40

-30 -45
-47.04% to -7.63% SPY Benchmark Range
-36.27% to 22.8% SPY Benchmark Range -50
-40

Information Ratio (Annualized) Standard Deviation (Annualized)

3 45

40
10.02% to 41.92% SPY Benchmark Range
2
35

30
1
25
21.67% Worst Case Contingency (@ SPP 5%)
0 20
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
15
-1 -1.20 OOS Estimate 11.96% OOS Estimate
10

-2 -1.45 Worst-Case Contingency(@ SPP 5%) 5

0
-3 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

IFTA.ORG PAGE 95
 IFTA JOURNAL 2015 EDITION

a much deeper understanding of how the trading system may Notes

perform going forward. 1. http://finance.yahoo.com/
SPP applied to the relative momentum trading system 2. From the current (January 2014) schedule of fees for margin interest from
Interactive Brokers <$500,000 borrowed.
demonstrates outperformance over the buy-and-hold 3. http://www.newyorkfed.org/
benchmark in the long-run with varying degrees of statistical 4. Defined as expected active return (system return – benchmark return)
significance for different system metrics. Specifically, an divided by tracking error: .
5. “Regression toward the mean” is a statistical law, not to be confused
equivalent p-value of 0.10 for CAR outperformance is marginally with the financial term “mean reversion,” which assumes that
significant. In contrast, an equivalent p-value of 0.00 for max observed high and low prices are temporary and that price will tend to
move to the average over time.
drawdown outperformance is highly significant, and these 6. To ensure the absence of DMB, SPP must be conducted as a standalone
results together indicate that a strength of relative momentum process (not to compare systems). Any ex post selection based on
performance has the potential to introduce DMB as discussed in
is avoidance of large drawdowns. Section 3.
However, the SPP worst-case contingency analysis for
calendar year performance demonstrated that, in order to References
achieve long-term outperformance, the trader must be willing Aronson, David R., 2007, Evidence-based Technical Analysis: Applying the
Scientific Method and Statistical Inference to Trading Signals, John Wiley & Sons,
to accept the possibility of significant underperformance in Inc., Hoboken, NJ, 544 p.
any calendar year. With this information, the capital allocation Asness, Clifford S., Tobias J. Moskowitz, and Lasse J. Pedersen, 2009, “Value
decision may be made probabilistically. and Momentum Everywhere,” AFA 2010 Atlanta Meetings Paper, http://ssrn.com/
abstract=1363476 (December 1, 2013).
Bailey, David H., Jonathan M. Borwein, Marcos López de Prado, and Qiji Jim Zhu,
Conclusion 2013, “Pseudo-mathematics and Financial Charlatanism: The Effects of Backtest
It is essential for any trader to thoroughly understand what Overfitting on Out-of-sample Performance,” http://ssrn.com/abstract=2308659
(December 20, 2013).
to expect from a trading system before allocating capital. Blitz, David C. and Pim Van Vliet, 2008, “Global Tactical Cross-Asset Allocation:
Without knowledge of the probable ranges of performance Applying Value and Momentum Across Asset Classes,” Journal of Portfolio
Management 35 (1), p. 23–38.
expected in the future, the trader or client is prone to abandon a
Chekhlov, Alexei, Stanislov, Uryasev, and Michael Zabarankin, 2003, “Portfolio
good system in the stress of an unexpected drawdown or period Optimization with Drawdown Constraints,” http://www.ise.ufl.edu/uryasev/
of underperformance. Even worse, capital may be allocated files/2011/11/drawdown.pdf (December 20, 2013).
on the basis of inflated expectations gained from traditional Inoue, Atsushi and Lutz Kilian, 2002, “In-Sample or Out-of-Sample tests of
Predictability: Which One Should We Use?,” Working Paper No. 195, European
evaluation methods when the system should be discarded in the Central Bank Working paper Series (November 2002).
light of the probabilistic information that SPP is able to provide. Kaufman, Perry J., 2013, Trading Systems and Methods, + Website, 5th Edition, John
The majority of traditional system development approaches Wiley & Sons, Inc., Hoboken, NJ, 1232 p.
Krawinkel, Thomas, 2011, “Buying Power – The Overlooked Success Factor,”
provide a single point estimate of performance and/or measure NAAIM Wagner Award 2011.
of statistical significance based on a single sequence of trades. Pardo, Robert, 2008, The Evaluation and Optimization of Trading Strategies 2nd
With the limited information from such a point estimate, the Edition, John Wiley & Sons, Inc., Hoboken, NJ, 334 p.
capital allocation decision is difficult at best. In contrast, SPP
produces sampling distributions of system metrics that allow Note from the Editor
more realistic contingency planning based on probabilities. The author, Dave Walton, is the winner of the NAAIM Wagner
Ultimately, SPP offers a simple, easy to use, yet realistic Award 2014. This paper was originally submitted for this
method for estimating future system performance. It is the contest. IFTA is thankful to Greg Morris and the National
balance of these three factors that is the true strength of the Association of Active Investment Managers for the permission
method. Thus, SPP is broadly applicable by traders and system to print this document.
developers of varying backgrounds and adds value in real-life
practice.
The trading system example showed that SPP provides
a clear, balanced picture of expected system performance,
whereas standard cross-validation did not. The example also
demonstrated that the relative momentum trading system is
likely to outperform the buy-and-hold benchmark over the long
run but that in order to achieve long-term outperformance, the
trader must be willing to accept the possibility of significant
underperformance in any given year. With this information, the
capital allocation decision may be made probabilistically.

PAGE 96 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Crowd Money: A Practical Guide to Macro Behavioural

Technical Analysis—By Eoin Treacy
Reviewed By David Hunt

Let me start by saying that I have been a big fan of David 1. Price action/crowd psychology
Fuller since I attended two of his Chart Seminars on point and 2. Liquidity
figure charting in the mid to late 1980s in Sydney. His calm 3. Governance
personality and array of knowledge and capacity to follow many 4. Theme/fundamental value
markets set him apart from others I was studying at that time.
And so, when I saw this book in an airport bookstore, I Eoin is upfront: “Today, of all the technical indicators out
figured I’ll take a look at it. there, the only one I use with any kind of regularity is the
Yet, I had never really understood 200-day moving average (MA) because it
David’s macro-behavioural represents the trend mean.”
approach—probably due to the fact As a result, this book is probably not
that I was a recovering economist going to greatly increase an expert’s
who had just discovered charts to knowledge of technical analysis. But
actually enjoy movements in the what it does do is give you a greater
financial markets both on a short- understanding of the drivers behind
term and long-term basis. the concepts that fund managers and
So reading this book by one of his larger investors incorporate into their
associates was both a blast from the decision-making (e.g., trends, fads), plus
past and highly educational for me. a good dollop of well-grounded trade
What I enjoyed about this book was management.
seeing the fusion of charting and WD Gann talked about campaigns in
macro-behavioural economics. the markets, and this book allows you
It is simply a well-crafted, holistic to understand themes in the markets
approach to markets and future and campaigns on which you can trade
trends that is well worth the read. technically based on the themes. Plus, it
As a bonus, Eoin discusses future gives you a bonus of trade and stop loss
themes for 2015–2025. Whether they management and investor and trader
come true is not important; it’s the thought and action patterns (crowd
thought process that goes into the psychology) from the perspective of how
development of the themes that is you hang in for the big moves.
very useful. One of the chapters in the book is
Now, what shocked and stunned “Governance is Everything.” I found this
me about this book was the lack chapter particularly challenging in that
of point and figure charts! When I with good governance, while we might
expressed amazement about this fact to one of my advisory get bull markets, there is less chance of a complete routing.
clients who used to work with David Fuller at Fullermoney, she We all believe the American example is probably one of good
explained that while she was working there, they dropped the governance, and of course the global financial crisis was a
point and figure charts, as they took a lot more effort. She added failure of regulation and governance. But the reality is, it wasn’t
that the point and figure software options they looked at were stealing—it was just a typical catastrophe, which we get every
not quite right, so they went for the more standard charts. This couple of decades!
broke my heart a little to see not one point and figure chart in an All in all, it was a lovely way to see some charts along with
entire David Fuller derived book, but it would not be a surprise behavioural and macro-economic reasons why these particular
to subscribers. stocks were chosen as part of their themes.
The first thing you will notice is that David Fuller and Eoin Like it or lump it (as a technical analyst), the people who
Treacy use chart reading (his form of technical analysis) in only drive the markets are the biggest crowd we have—they are the
one part of their four investing pillars: fund managers. They also tend to be value investors.

IFTA.ORG PAGE 97
 IFTA JOURNAL 2015 EDITION

If you are a technical analyst whose head is always in the

price action chart-only detail, then PLEASE READ THIS BOOK.
It will help you get inside the minds and hearts of people who
IFTA Board of Directors
regard technical analysts as aliens, enemies or an occasionally
useful addenda. President
The section on management and stops is something anyone Rolf Wetzer, Ph.D. (SAMT)

managing money or advising others should spend time studying. Vice-President—the Americas
William Chin, MBA (CSTA)
The final part of the book has a section on what David Fuller
calls Autonomies (from page 138): Vice-President—Asia
Akira Homma, CFA, CIIA, CFTe, FRM, CMA, CMT (NTAA)
“David Fuller christened such large multinationals “mobile
principalities” or “Autonomies” because they are independent, Vice-President—Europe and Conference Director
Deborah Owen, FSTA, CFTe (STA)
powerful, mobile “mini countries” that focus where the best
potential for profit exists. What country they consolidate Vice-President—Middle East, Africa
Mohamed Ashraf Mohfauz, MFTA, CFTe, CETA (ESTA)
earnings in is less important than the source of their income.”
This is the only chapter in which I found myself wanting more Treasurer
Ralph Böckel, CFA (VTAD)
detail. What the reader gets are charts with some 150–300
word descriptions of what they are and do. Now, while this is a Education Director
Gregor Bauer, Ph.D., CFTe (VTAD)
great concept, it left me hungry to know more about why David
Fuller and Eoin Treacy like these stocks and how they would Accreditation Director
Roberto Vargas, CFTe, MSTA (TSAASF)
handle them in the future.
This is a deep and well-thought-out book on Fuller’s methods, Marketing and Membership Director
Dan Valcu, CFTe (AATROM)
which have been honed over 40 years. I recommend it to anyone
Exam Management Director
who dips their toes into the markets—no matter how deep. Jeannette Schwarz-Young CFP®, CMT, M.S. (AAPTA)
Technical analysts who want to know how to have greater
Strategic Development Director
appeal to investors should read Crowd Money five or six times— Robert Grigg, CPA (ATAA)
or at least as many times as you read WG Gann’s Tunnel Thru
Online Development Director
the Air. Buy Crowd Money—it certainly made me think in bigger Roman Bogomazov (TSAASF)
terms, and it will help you too.
Journal and Newsletter Director  
Aurélia Gerber, MBA, CFA (SAMT)

Website Content Director  

Akihiro Niimi, MFTA (NTAA)

Sponsorship Director  
Mohamed El Saiid, CFTe, MFTA (ESTA)

IFTA Staff
Executive Director
Beth W. Palys, FASAE, CAE

Vice President, Meetings

Grace L. Jan, CAE, CMP

Senior Member Services Manager

Linda Bernetich

Marketing Director
Julie Hill

Director of Editorial Services

Lynne Agoston

Senior Graphic Designer

Jon Benjamin

Accounting
Dawn Rosenfeld

PAGE 98 IFTA.ORG
 IFTA JOURNAL 2015 EDITION

Author Profiles

Adam Cox, MFTA Shawn Lim, CFTe, MSTA

Adam Cox has more than 20 years of investment Shawn Lim is currently an analyst at Morgan
management, trading, and banking experience. Stanley in the London office. His research
He is currently director and head consultant for interests lie in asset pricing and portfolio
Prime Consulting, a specialist consulting theory, and he has written and published a
company operating from New Zealand. Mr. Cox’s number of articles in journals such as the
passion lies in trading system development, International Journal of Economics and Finance
systematic FX trading, and quantitative finance, including and the Journal of Finance and Investment Analysis. He is a
econometric-based system development and behavioral finance Certified Financial Technician (CFTe) and a member of the
applications to trading. His current emphasis is based on US ETF, Society of Technical Analysts UK (MSTA). He also graduated with
Futures (commodities, FX and indices) and macro strategies, as well a first class honours degree in economics from University College
as intraday trading. Mr. Cox is currently working on a Ph.D. in London and holds an advanced diploma in data and systems
finance (proposal stage), aiming to explain trader behavior and analysis from the University of Oxford.
provide a model for real-time applications. He has developed a
range of customized indicators, which borrow from econometrics, Andrew J.D. Long, MFTA
time series analysis, and inter-market and investment analysis to Andrew J.D. Long, MFTA, is a professional
discretionary trading, as well as trading risk management systems technical analyst and trader. The founder and
(scoring systems). Among these systems and analytical approaches, publisher of TRIGGER$.ca, an economic and
Mr. Cox has extensive and hands-on experience applying a wide technical analysis publication for active traders
range of methodologies, including wavelet de-noising and compres- and investors, he and his venture partner,
sion techniques to trading, as well as mandani fuzzy logic, Gordon T. Long.com, regularly publish market
structural equation modeling, and SVM. Mr. Cox’s current emphasis and economic forecasts with uncanny accuracy. Mr. Long’s
in wavelet applications includes forecasting methodologies and in experiences vary and include working for a private fund
particular, multivariate approaches to short trading system design researching and developing proprietary technical analysis
aimed at capturing inter-market relationships and trading methods. Researching and trying to understand the markets
behavioral patterns across markets. This work also extends to has been a lifelong pursuit and journey—starting in early high
(mother) wavelet design for application to de-nosing applications to school with P/E ratios and balance sheets, to the present day,
specific markets Mr. Cox is also very interested in emerging market where he continues to research and develop advanced technical
Asia equities markets, and in particular, the Vietnamese equities analysis techniques. In his publication, global economic and
market. As a side note, Mr. Cox has a Vietnamese and Cantonese fundamental analysis is integrated with advanced technical
language background and has extensively traveled, having lived in analysis, offering unique and often correct market perspectives.
the United States, Canada, Australia, New Zealand and New The purely technical method discussed in his MFTA paper is
Caledonia. He is currently located in Singapore. Adam can be practiced daily with published forecasts, tracking the progress
reached at acox @ primeconsulting.co.nz and is open to any queries, of the system where it continues to achieve 80–90% accuracy.
feedback and business/work opportunities.
Stanislaus Maier-Paape
David Hunt Professor Stanislaus Maier-Paape has taught
David Hunt has 32 years of experience covering mathematics at RWTH Aachen University since
hedging and trading for Macquarie Funds Group. 2001. His specialties include differential
He is CIO of private equity fund PHG Investments equations, nonlinear analysis, stochastics, and
and is Australian charting advisor to financial applications of mathematics in other fields.
advisors, professional investors and traders Since 2010, he has been CEO of SMP Financial
through his Profit Hunter Group. He is an author Engineering GmbH (www.smp-fe.de). His research fields in
of the Wiley Trading Guide II. Mr. Hunt is director and past quantitative finance include mechanical trading systems and
president of the Australian Professional Technical Analysts (APTA) portfolio and money management.
as well as co-founder of APTA and the Australian Technical Analysts
Association. He is a regular on Fox Business TV in Australia and the
Australian Financial Review. He can be contacted through http://
adest.com.au.

IFTA.ORG PAGE 99
 IFTA JOURNAL 2015 EDITION

Alex Neale, MSTA, CMT, MFTA Samuel Utomo, CFTe, MFTA

Alex Neale is a professional technical analyst. He works for Samuel Utomo is currently a final-year
Cantor Fitzgerald Europe providing technical analysis and undergraduate student at Prasetiya Mulya
market commentary to high net worth CFD traders. Mr. Neale Business School, Indonesia, majoring in
has 15 years of experience in writing analysis reports. finance. He previously worked at PT
He started his career in 1998 at GNI, the company that Astronacci International as a technical
pioneered CFDs for individuals, before moving to Cantor analyst, and then as head of research and
Fitzgerald Europe in 2012. He is based in London and can be lecturer, dealing with individual and
contacted through his LinkedIn profile. institutional clients in Indonesia. Mr. Utomo frequently appears
on local television and radio stations as well as on several online
David Price BSc, MSc, CFTe, MSTA, MFTA media channels, providing market outlook. He also passed the
David Price has over a decade of experience final level of the Chartered Market Technician (CMT)
working for financial firms, including Goldman examination given by the Market Technician Association (MTA)
Sachs, Brevan Howard and JPMorgan, across as well as the local investment manager representative
equity, global macro, FX, quantitative and examination. He is now shifting his independent academic
systematic investment strategies. Mr. Price has research focus to the development of a quantitative equity
produced technical analysis research of equity, portfolio management platform for the South East Asian equity
fixed income, FX and commodity markets for discretionary market.
portfolio managers. He is a Certified Financial Technician (CFTe), a
Master of Financial Technical Analysis® (MFTA®), a Member of the Fergal Walsh, CFTe, MFTA
Society of Technical Analysts (MSTA) and holds an MSc in financial Fergal Walsh has been a part-time trader in
management. As someone who is passionate about investment and financial markets since 2010, focusing
technical analysis, he is keen to research and identify new technical primarily on CFDs and currencies. He is
alpha-generating and risk-reducing investment strategies. currently completing a bachelor of arts degree
in economics and history at University College
Andreas Thalassinos, BSc, MSc, MSTA, Dublin. Subsequently, he hopes to focus full
CFTe, MFTA time on trading, further his experience in
Andreas Thalassinos, BSc, MSc, MSTA, CFTe, financial markets, and continue to develop trading strategies
MFTA, is a highly respected lecturer in the based primarily on technical analysis.
education of traders, investors and FOREX
professionals. His passion for trading led him Dave Walton, MBA (Wagner Award)
to study the markets from a mathematical and Dave Walton has been involved in active
mechanical point of view. He is a dynamic investing since 1999; he has focused on system
advocate of algorithmic trading and has developed hundreds of trading since completing the Van Tharp
automated systems, indicators, and trading tools. His trading Institute’s Super Trader program.
products are being used today all over the world by traders, A computer engineer by training, Mr. Walton
investors, FOREX brokers and investment firms. Mr. Thalassinos applies system validation principles and
founded FxWizards (www.fxwizards.com), an educational statistical methods to trading. He spent 18
company specialized in FOREX trading with the goal of years ensuring the quality and reliability of cutting-edge
providing the real truth about FOREX trading through high-level technology products for one of the world’s largest computer chip
education to beginners and professionals alike. He emphasizes manufacturers. Mr. Walton received his B.S. in computer
that capital preservation is imperative for traders to survive in engineering from Virginia Tech and his MBA from UC Davis.
the financial markets, and he guides traders on employing swing He is co-founder and principal of StatisTrade, LLC, a trading
trading and locking profits based on his research findings and system evaluation firm that works with fund managers to
results. Mr. Thalassinos is currently travelling around the globe evaluate and improve trading system performance using
giving seminars as a guest speaker at FOREX conferences. advanced statistical techniques. He is a proponent of applying a
scientific approach to trading system development and
Douglas Stridsberg evaluation, which includes rigorous validation principles
Douglas Stridsberg is currently a student gleaned from his engineering background. Mr. Walton won the
studying mechanical engineering at University NAAIM (National Association of Active Investment Managers)
College London (UCL). His interests span 2014 Wagner Award for one of his creative system validation
multiple disciplines, but his passion lies in Big methods. He primarily focuses on long-term trading systems
Data and how it can improve financial that have significant academic research behind them. His
forecasts. He is a keen trader and programmer primary trading vehicles are ETFs and U.S. stocks.
and is currently working on building an automated intraday FX
trading algorithm. In the past, he founded UCL’s first discussion
group on developments in the financial markets and has been
involved in the running of several large-scale conferences about
the financial industry.

PAGE 100 IFTA.ORG

Master of Financial Technical
Analysis (MFTA) Program
Examinations
IFTA’s Master of Financial Technical Analysis In order to complete the MFTA and receive your
(MFTA) represents the highest professional Diploma, you must write a research paper of no
achievement in the technical analysis community, less than three thousand, and no more than five
worldwide. Achieving this level of certification thousand, words. Charts, Figures and Tables may
requires you to submit an original body of be presented in addition.
research in the discipline of international
technical analysis, which should be of practical Your paper must meet the following criteria:
application. • It must be original
• It must develop a reasoned and logical
The MFTA is open to individuals who have argument and lead to a sound conclusion,
attained the Certified Financial Technician (CFTe) supported by the tests, studies and analysis
designation or its equivalent, e.g. the Certified contained in the paper
ESTA Technical Analyst Program (CETA) from the • The subject matter should be of practical
Egyptian Society of Technical Analysts (ESTA) application
For those IFTA colleagues who do not possess • It should add to the body of knowledge in the
the formal qualifications outlined above, but discipline of international technical analysis
who have other certifications and/or many
years experience working as a technical analyst,
the Accreditation Committee has developed
Timelines & Schedules
There are two MFTA sessions per year, with the
an “alternate path” by which candidates,
following deadlines:
with substantial academic or practical work in
technical analysis, can bypass the requirements Session 1
for the CFTe and prequalify for the MFTA. “Alternative Path” application deadline
The alternate path is open to individuals who February 28
have a certification, such as: Application, outline and fees deadline
May 2
• Certified Market Technician (CMT) or a Society Paper submission deadline
of Technical Analysts (STA) Diploma, plus three October 15
years experience as a technical analyst; or
• a financial certification such as Certified Session 2
Financial Analyst (CFA), Certified Public “Alternative Path” application deadline
Accountant (CPA), or Masters of Business July 31
Administration (MBA), plus five years Application, outline and fees deadline
experience as a technical analyst; or October 2
• a minimum of eight years experience as a Paper submission deadline
technical analyst. March 15 (of the following year)
A Candidate who meets the foregoing criteria
may apply for the “alternate path”. If approved,
they can register for the MFTA and submit their To Register
research abstract. On approval, the candidate Please visit our website at http://www.ifta.org/
will be invited to submit a paper. certifications/master-of-financial-technical-
analysis-mfta-program/ for further details and to
register.

Cost
$900 US (IFTA Member Colleagues);
$1,100 US (Non-Members)
Certified Financial Technician
(CFTe) Program
IFTA Certified Financial Technician (CFTe) consists of the CFTe I and
CFTe II examinations.
Curriculum
The CFTe II program is designed for self-study, however,
Successful completion of both examinations culminates in the award IFTA will also be happy to assist in finding qualified
of the CFTe, an internationally recognised professional qualification in trainers. Local societies may offer preparatory courses
technical analysis. to assist potential candidates. Syllabuses, Study Guides
and registration are all available on the IFTA website at
http://www.ifta.org/certifications/registration/.
Examinations
The CFTe I exam is multiple-choice, covering a wide range of technical To Register
knowledge and understanding of the principals of technical analysis; it Please visit our website at http://www.ifta.org/
is offered in English, French, German, Italian, Spanish and Arabic; it’s certifications/registration/ for registration details.
available, year-round, at testing centers throughout the world, from
IFTA’s computer-based testing provider, Pearson VUE.
The CFTe II exam incorporates a number of questions that require
Cost
essay-based, analysis responses. The candidate needs to demonstrate IFTA Member Colleagues Non-Members
a depth of knowledge and experience in applying various methods CFTe I $500 US CFTe I $700 US
of technical analysis. The candidate is provided with current charts CFTe II $800* US CFTe II $1,000* US
covering one specific market (often an equity) to be analysed, as though *Additional Fees (CFTe II only):
for a Fund Manager. $250 US translation fee applies to non-English exams
The CFTe II is also offered in English, French, German, Italian, Spanish $100 US applies for non-IFTA proctored exam locations
and Arabic, typically in April and October of each year.