Equity markets are tumbling as this issue goes to press and most of the

articles were submitted well before the latest collapse in share prices. No-

one is prepared to call how far the markets will fall in the current

environment but the head and shoulders pattern that has been traced out

on a number of charts looks very ominous. With so much gloom and doom

around, it is perhaps salutary to remind ourselves that bear markets end

when everyone is at their most bearish. Cycle theory also teaches us that

trends do not go on forever and the current bear market has been in place

for two and a quarter years, making it one of the most prolonged

downturns since the great depression. It should also be remembered that

there has been a positive signal on the Coppock indicator and two articles

in this issue focus on this remarkably consistent long term buying signal.

Full credit for spotting the extent of the downturn must go to Peter Beutell

of MTS Research Ltd. who warned us what could be in store in his article

‘Great Bear Markets and Some Time Targets for a Major Low’ which appeared

in last October’s issue of the Market Technician. Peter will be giving us an

update of his view on the market in the next issue.

The October IFTA conference, which this year is being hosted by the STA in

London, looks like being a very successful occasion. You will by now all have

received a brochure about the conference and will have seen that

we have managed to put together a most impressive range of first-rate

speakers. We are aware that it is difficult for some members to put aside

three days to attend a conference and so we have made available a limited

number of day tickets. Further details can be obtained from the website or

by phoning Kate Connors of Concorde Services on 0208 743 3106 who are

helping with the administrative side of the conference. Anyone thinkingof attending should note that the early booking discount ends on 9th August.

* * *

The financial sector has always been well represented north of border. In

recent years the established fund management houses have been joined

by a number of hedge funds. The Scottish Chapter of the Society has

flourished on the back of the increasing use of technical analysis.

Membership has risen steadily, drawing new members from both the

professional and private sides of the market, as well as the academic

community.

The Chapter has been galvanised by Gerry Celaya’s return to his Scottish

roots. We are endeavouring to hold regular monthly meetings as well as

the now established one-day Scottish Seminar in November. We are very

grateful to Standard Life Investments for the use of their facilities each

month in Edinburgh and we are also hoping to hold at least one meeting in

Glasgow during the summer.

We have been lucky to have had some top class speakers recently who have

covered a broad range of disciplines from orthodox pattern recognition

through to Elliott Wave and the cutting edge of Neural Networks. We have

also hosted a joint meeting with the Society of Business Economists where

the debate was particularly vigorous and enjoyable. Each speaker was

presented with some quality Scottish malt whisky to help them in their

future analysis!

We are keenly aware that members and prospective members would like to

establish local educational courses in Scotland in order to facilitate

preparation for the STA Diploma exam. We are working with John Cameron

and colleagues at a local university and look forward to being able to hold

our own diploma courses in the future.

The Scottish Chapter is always on the look out for anyone interested in

attending or speaking at meetings so if you find yourself in the area around

the third Thursday of each month and want to come along then please get

in touch with Murray Gunn or Gerry Celaya. Details of upcoming meetings

are always posted on the website.

IN THIS ISSUE

STA Exam Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2

D. Murrin The physics, behaviour of markets and the investor . . . . . . . . . . 3

C. Mack Coppock understood . . . . . . . . . . . . . 6

S. Griffiths The simple ABC correction . . . . . . . . 8

Z. Harland Using support vector machinesto trade aluminium on the LME . . . . 9

S. Warner Beat the index using the index . . . 13

R. Marshall The Coppock turns up . . . . . . . . . . . 16

July 2002 The Journal of the STAIssue No. 44 www.sta-uk.org

FOR YOUR DIARY

Wednesday, 11th September Monthly Meeting

Wednesday, 2nd October Monthly Meeting

10-12th October IFTA Conference, London

Wednesday, 13th November Monthly Meeting

Wednesday, 4th December Monthly Meeting

N.B. The monthly meetings will take place at theInstitute of Marine Engineers

80 Coleman Street, London EC2 at 6.00 p.m.

MARKET TECHNICIAN

COPY DEADLINE FOR THE NEXT ISSUE

30th AUGUST 2002

PUBLICATION OF THE NEXT ISSUE

OCTOBER 2002

MARKET TECHNICIAN Issue 44 – July 20022

STA DIPLOMA EXAM RESULTS

NOVEMBER 2001

PASS LIST

E. Blake

A. Dowdell

J. Hanson

S. Ng

S. Rossos

S. Rowe

C. Smith

M. Torres

OVERSEAS CANDIDATES

P. Gaba

F. Tam

CHAIRMAN

Adam Sorab,Deutsche Bank Asset Management,1 Appold Street, London EC2A 2UU

TREASURER

Vic Woodhouse. Tel: 020-8810 4500

PROGRAMME ORGANISATION

Mark Tennyson d'Eyncourt.Tel: 020-8995 5998 (eves)

LIBRARY AND LIAISON

Michael Feeny. Tel: 020-7786 1322

The Barbican library contains our collection. Michael buys new books for itwhere appropriate, any suggestions for new books should be made to him.

EDUCATION

John Cameron. Tel: 01981-510210 Clive Hale. Tel: 01628-471911 George Maclean. Tel: 020-7312 7000

EXTERNAL RELATIONS

Axel Rudolph. Tel: 020-7842 9494

IFTA

Anne Whitby. Tel: 020-7636 6533

MARKETING

Simon Warren. Tel: 020-7656 2212 Kevan Conlon. Tel: 020-7329 6333

MEMBERSHIP

Simon Warren. Tel: 020-7656 2212 Gerry Celaya. Tel: 020-7730 5316 Barry Tarr. Tel: 020-7522 3626

REGIONAL CHAPTERS

Robert Newgrosh. Tel: 0161-428 1069 Murray Gunn. Tel: 0131-245 7885

SECRETARY

Mark Tennyson d’Eyncourt.Tel: 020-8995 5998 (eves)

STA JOURNAL

Editor, Deborah Owen,108 Barnsbury Road, London N1 OES

Please keep the articles coming in – the success of the Journal dependson its authors, and we would like to thank all those who have supportedus with their high standard of work. The aim is to make the Journal avaluable showcase for members’ research – as well as to inform andentertain readers.

The Society is not responsible for any material published in The MarketTechnician and publication of any material or expression of opinionsdoes not necessarily imply that the Society agrees with them. The Societyis not authorised to conduct investment business and does not provideinvestment advice or recommendations.

Articles are published without responsibility on the part of the Society,the editor or authors for loss occasioned by any person acting orrefraining from action as a result of any view expressed therein.

NetworkingWHO TO CONTACT ON YOUR COMMITTEE

Exam results

ANY QUERIES

For any queries about joining the Society, attending one of the

STA courses on technical analysis or taking the diploma examination,

please contact:

STA Administration Services (Katie Abberton)

Dean House, Vernham Dean, Hampshire SP11 0LA

Tel: 07000710207 Fax: 07000710208

www.sta-uk.org

For information about advertising in the journal, please contact

Deborah Owen

108 Barnsbury Road, London N1 OES

Tel: 020-7278 4605

Dynamic Technical Trader/Analyst seeks Position

Skilled Technical Trader /Analyst seeks new challenge preferably

within a new or established Fund Mgt Set-up. Skilled at

applying real-time analysis and trading decisions, Futures and

Equity markets.

Experienced in both Equity Hedge and Managed Futures Funds

from both a Short & Long Term perspective.

Would suit both Technical & Fundamental Trading Strategies.

University Educated, SFA & Eurex Qualified.

For more detailed information, please contact

Sta_trader@hotmail.com\fs96

Issue 44 – July 2002 MARKET TECHNICIAN 3

The purpose of this paper is an attempt to instigate a change in your

perceptions of markets and how you interact with them to reach your

goals. To achieve this I shall pose some deep questions and provide some

answers that I have gathered through 15 years of direct experience. I hope

that this information will be applicable in whatever specific approach you

have chosen to involve yourselves in the markets.

My rationale starts by breaking the problem down into basic principals to

develop a model for markets. If we take the essential components of

markets we have the following:

i. A market transaction takes place within the universe and is subject to

all its laws and processes. Therefore, an examination of the laws of

physics that might pertain to the mechanics of a market is essential to

our model.

ii. Next there is the market, which in essence is a process for exchanging

an asset with a group of other human beings that comprise a closed

system.

iii. Finally, there is the investor, a human being and a product of thousands

of years of evolution, with all its prejudices, strengths and weaknesses.

Following this train of thought, I will start with the universe and the five

theories that might apply to markets, as this should help us transact our

business more successfully

1.1 Chaos theory

The recent research into chaos theory has provided an insight into the

essence of order and disorder in nature. The red spot of Jupiter best

exemplifies this, being a gaseous planet in a constant state of change. The

border of the red spot exists within the chaotic gaseous atmosphere that

shows great disorder, and then, within the domain of the spot, the gases

also exist in a state of flux. In short, disorder, order and disorder lie side by

side.

This explains why people and systems, which have attuned to a certain type

of order in the market place, suddenly encounter periods of losing trades.

The key lesson to be derived from this theory is the importance of knowing

when your system is working in the ordered phases and to run the winners,

but when you encounter chaos, to have a methodology to stop trading and

conserve the war chest.

1.2 Fractal theory

The universe shows fractal properties, as described in fractal theory. In

simple terms, this is the repetition of geometry within nature across

different degrees of size. This means that patterns that are apparent on a

small scale are also generated on larger scales within the same system.

Examples of such fractals are:

● The distribution of star clusters● The definition of coastlines● The structure of a simple cauliflower● The structures in the human body and brain

Any model that is used in the markets should recognise fractal structures

across different time frames, which allows us to develop short, medium and

long-term time frames.

1.3 Fibonacci theory

This well-known theory, derived from the Fibonacci number sequence,

recognises that there are natural harmonic ratios that exist in nature, the most

applicable to the market being 23%, 38%, 50%, 62% and 76% as a function of

the retracements of moves. However, the essence of applying such a theory

requires us to be able to recognise the inter-relationships of a sequence of

highs and lows before the retracement aptitudes can have any meaning.

As I will explain later, it is my finding that the application of Fibonacci theory

is related to the Elliott wave structure, so wave 2 retraces wave 1 according

to Fibonacci ratios.

1.4 Quantum theory

The current thinking on the universe is governed by the theories of

quantum mechanics, which imply a non-deterministic nature to the

unfolding of events within the time continuum. This 20th century theory

changed the old Newtonian world of predictability to a state of pure

potential and infinite possibility.

Furthermore, as found during the early experiments in particle physics, the

very act of observing an experiment changed the outcome of the

experiment. Thus, awareness of an event changes the outcome of the

universe. In effect, the infinite state of all possibilities of a quantum particle

collapsed into a set entity as soon as it was observed; in other words, the

process of observation was akin to solidifying the jelly of probability.

This is a shattering observation, that the consciousness of the observer

brought the observed into being. Nothing in the universe exists as an

actual thing independent of the perception of it. This applies to all aspects

of our existence, but a particularly pure example is our participation in the

markets, where our internalised model of emotions and expectations are

played out through interacting with the market.

Thus, any model developed that was designed to predict market behaviour

should be best described in the context of a probability field, with each

event being given a probability to its outcome.

1.5 The zero point field

One of the big questions posed in physics over the years has been the

nature of space. Over the centuries, the concept of the ether has waxed and

waned, but the newest concept is the zero point field.

This theory describes the basic sub-structure of the universe as a sea of

quantum fields like the one described in Quantum theory. It is this

revolutionary concept that starts to change our view of the universe, which is

in fact a seething maelstrom of sub-atomic particles popping in and out of

existence,within which sub-atomic waves and information can be propagated.

Why is this significant? Well, it would allow for a faster than light

communication between human beings and, therefore, a mechanism to

share a group consciousness. It further implies that the past, present and

future are all closely entwined in the zero point field and that the universe

is holographic in nature and much, much more!

2.0 The behavioural patterns of markets

Having examined the universe and some of current theories used to

describe our current understanding of its mechanisms; I shall now examine

the human component, primarily you and the market.

The physics, behaviour of markets and theinvestor

by David Murrin

This is a summary of a talk given to the Society in February 2002

MARKET TECHNICIAN Issue 44 – July 20024

2.1 The market is a closed systemThis means that for every winner there is a loser as there are only a finite

number of people in the market place. This is key as we can assume that on

the basis of natural selection, the most successful predators, or players, live

longer and become bigger. The implications of this are that the success of

a large player would require the failure of many small and medium size

players. As an inexperienced player to the market, the key is to enact a

strategy of learning that allows one to survive long enough to grow bigger.

A closed system like a market needs to be balanced and maintain its

equilibrium of both long and short positions, which is achieved by moving

the price to a balanced position. This starts to explain the process where

patterns in close proximity alternate in form to confuse expectations. For

example, as more than a sustainable percentage of the system believe that

a certain outcome is inevitable, the pattern evolves into a more complex

form to re-assert equilibrium. This phenomenon is seen in Elliott wave

alternation between waves 2 and 4.

This is why simple behavioural or 1st degree patterns will be apparent in

markets that are not sophisticated. However, in more developed markets,

where complex trading systems become the norm, and move to the 2nd

degree, the price patterns should be harder to capitalise on.

2.2 A fishy questionOne question that I have always wanted to ask a fish who is a member of a

shoal is:

What is it like to be a part of the shoal, to be sublimated to the group

consciousness and not even be aware of so many choices that have been

made by the group consciousness?’

And it is my guess that the fish, if he could,would answer that he had not even

been aware of the influence of the shoal’s consciousness upon its choices.

2.3 Tribal behaviourIn our current social structure in the western world, it is easy to forget our

tribal origins. In the mid-eighties, I was lucky enough to spend a couple of

years in the swamps of the upper Sepik basin in Papua New Guinea. This

was at a time when the local population had been scarcely touched by

civilised society, living in small tribes of 60 to 200 in a difficult environment.

During his period, I made the following observations:

The emotions of one person could be transferred to the rest of the tribe, so

that one person’s anger towards an object outside the tribe could be

resonated in the rest of the tribe.

The tribe’s emotions appeared to be stored in a similar way to the charge in

a capacitor, with an initial high level of feeling that decayed over a few hours

to a point where the majority of the tribe forgot their original grievance.

This may be the way all human emotion behaves with time, but what was

notable about this observation was that one person could influence so

many with his emotion. The extremity of this behaviour was initially

attributed to a lower threshold of individuality that existed within the tribe

as the result of a small inter-dependent community. However, subsequent

observations on institutional trading floors revealed that similar behaviour

patterns exist in our society and govern the majority of our actions, just as

they did in the jungles of New Guinea.

It is my belief that humankind evolved within a tribal structure, and

developed strong group behavioural patterns. Despite the fact that our

tribes have expanded from small groups into thousands within companies

and millions in countries, these basic behavioural patterns still govern a

large portion of our thoughts and actions.

2.4 Individuality Despite living in a society that values individuality, we are all, to some

degree, deeply linked to the rest of society and its collective consciousness

by our lower brain functions. Only a few people are able to maintain their

individuality in the face of strong group perceptions and it is these people

who are most suited to directional risk-taking in markets.

For purposes of market analysis, we may conclude that the market has a

collective consciousness that processes information and then responds

predominantly on an emotional basis, reflecting fear and greed. For the

individual, this equates to the fear of losing money, employment and self-

respect and the greed associated with a higher standard of living, freedom

of action and increased self-esteem. Furthermore, as I will later explain, the

dominant thought process within markets at present that comprises this

collective consciousness is the right-eyed, left-brain participant. This has

some important implications as developed below.

3.0 You the investor

Having determined that the market comprises a collection of investors, our

model now requires us to look more closely at the individuals within it and

their decisions.

3.1 Individual thought process and brain development

Understanding the decision-making process starts with knowing the

development and structure of the human brain. In simple terms, the

earliest section of the brain to develop is the stem, which governs the

majority of the body’s basic functions, but was also responsible for

assessing the linkage and acceptability of an individual within a tribe.

This means that each word and action was assessed for acceptability within

the tribe to ensure continued inclusion in the group. One must remember

that, in the early stages of human existence, to be ejected from the tribe was

to reduce one’s chances of survival considerably, thus all of our forefathers

would have developed this trait. As a result, the tribe developed a collective

consciousness to which all its members were connected by varying degrees.

Only in more recent history did man’s brain develop frontal lobes, allowing

a more sophisticated thought process. However, as a more recent addition

to the brain, they have not had time to develop complete control of our

thoughts and actions and, therefore, the older sections of our brain

dominate and produce highly emotional behavioural patterns that can, at

times, override the thoughts from the higher centres of the brain. Thus

human history has not been governed by the vulcan process of logic, but

prominently influenced by emotional perceptions.

3.2 Style of trading

The majority of market participants are involved in a mechanistic pricing

business and complex structuring of products that rely on spreads for

revenue generation. Only a few are involved in the rather predatory activity

of speculation or directional trading. This area is very demanding and,

within the context of group patterns, there are some important insights

with respect to individual characteristics that facilitate success.

i. Contra entry point trading: where the initiation of the trade is contra to

the prevailing trend, which, if correctly located, is also at the end point

of the trend. From this advantageous entry point the duration of the

trade can be varied across all time frames.

ii. Trend Trading:where the initiation of the trade only occurs a considerable

time after the new trend has commenced. To take advantage of such

trades the duration of the trade has to match the duration of the trend.

This style tends to be that of longer time frame traders.

3.3 Master eye dominanceTo those not familiar with a master eye, all of us use one eye more than theother, and this is the dominant eye. This eye is then cross-linked to theopposite side of the brain, which governs the principal mode of thinking.Thus, a left-handed person in 95% of occasions will have a left dominanteye, and right brain hemisphere dominance in thought process. In oursociety, the word ‘sinister’ meant a left-handed person and, because ofprejudice, many were forced to become right-handed. So it is possible thatpeople with left dominant eyes are right-handed. These dominance

Issue 44 – July 2002 MARKET TECHNICIAN 5

patterns represent a hard wiring difference in the neurological connectionsin and between the two hemispheres of the brain.

It is our theory that these differences in the male brain for a left master eyeconfer an advantage in Contra entry point trading, as follows:

i. In practical terms, the use of the right side of the brain confers non-linearity and creative processes, while the left side of the brain is linearand logical. As the structure of market predictions is not one of linearextrapolation but more one of a non-liner creative process, this givesthe right side of the brain an operating advantage.

ii. On the basis that more people are right-handed and use the left side ofthe brain, any collective market consciousness described above is morelikely to be based on the left side of the brain, allowing those with leftmaster eyes to be less affected and more objective in their analysis.

iii. Furthermore, left eye dominance tends to produce a more visualperception of information, so systems like wave counting are moreeasily assimilated.

Conversely, traders with right eye dominance should show a more linearthought process that is more suited to systematic trend trading.

Ideally the best combination would be enough sensitivity to enter at thelows, but then to ignore all the shorter term signals and run the trade for aslong as possible to reach the highs. Such a style would require a combinationof both approaches.

3.4 A study of the brain and brain damageUntil I read a recent study of brain damage and recovery rates, I could notunderstand why these traits were more apparent in men. The study madethe following observations:

i. Men recovered better than women,

ii. Left-handed men and women recovered faster than right-handed menand women.

Thus, at the extremes, a left-handed man would have the best recovery rateand a right-handed woman the worst chances of recovery. The reason forthe differing recovery rates is to do with the number of cross-linkagesbetween the hemispheres; the greater number of cross linkages in awoman facilitates greater re-routing of neural connections once damagehas occurred. Conversely men have more polarized hard wiring in theirbrains than women, using one or the other hemisphere more actively.

I believe that major decisions should be taken using a combination of skillsand approaches, and Wavecount’s services are designed to compliment theinstitutionalised thought process, which, by its very nature, will tend to beright eyed and left brained.

3.6 Bushido zen and enhancing separatenessThere is one more important aspect of discretionary analysis – the mindsetwith which the analyst enters the market place, with and learning to liveunder the stresses generated by taking risk on a daily basis.

One interesting example that stands out is the description of a veteran U.S.8th army bomber crew member in World War 2 who flew missions overGermany as described in the book ‘Combat Crew’ by John Comer. Hedescribed his own transition from a combat air gunner to a veteran, whenafter 50 to 60 missions, he moved from being a green crew member, with allthe terrors and fears of such an unforgiving and brutal environment andadjusted to the point where another mission became just part of daily life.Two lessons are apparent from this example: the first is that humans, if givenenough time, can adapt to the most challenging of environments; andsecond, that the adaptation does not take place as fast as we might wish!

In a methodology such as Elliott Wave Analysis, the thought process of theanalyst is vital to the success of the analysis. While a black box systemreduces the emotional connection with the decision-making process, morediscretionary approaches require a method of clearing the mind.

The most important point of this is to remove the emotional aspects ofmarket involvement, both on the profitable side and on the losing side,make it as clinical as possible and aim to be in a state of detachment. Asmany experiments have now shown, we are more able to influence oursurroundings when we are in such a detached state, as per the theories inquantum mechanics. Conversely, stress will only strengthen the individual’ssusceptibility to the markets behavioural patterns.

Above my desk, I have a picture of a Japanese Bushido warrior, and the mostpowerful lesson is, just like the warrior, it does not matter how many battlesyou have fought, the next one may be your last if you forget your disciplineand mental process. He must clear his mind before each battle, and hissurvival will depend on seeing the world for what it is rather than what hethinks it is.

Trading is no different.

4.0 Conclusion – Why Elliott Wave? In this paper I have explored the universe and some of the laws that I thinkare important in the development of a profitable market model. The formof analysis that I have chosen that conforms to all of these theories is the Elliott Wave model. A detailed explanation of how to apply this method is outside the scope of this article, but a complete guide is given onwww.wavecounts.com. However, to summarise, the key points are:

● A closed system – alternation, equilibrium and evolution as abiological system

● Chaos theory – knowing when you know, chose only the clearpatterns

● Fractal theory – seen throughout a wave count in different timeframes

● Fibonacci – requires a wave count to evaluate the interrelationshipbetween waves

● Quantum theory – a system that copes with a world of probabilitiesand demands a model that ascribes probabilities to the outcome ofdifferent future paths of the market place.

● Zero point field – the means by which a group consciousness, in thiscase a market, is able to communicate instantly across space.

I have also explored the human brain and the effect of markets onindividuals, which, once fully appreciated, can be integrated into one’smodel of markets to enhance profitability.

● A market is a closed system – it is a finite system, with boundaryconditions

● Price is the net product of the market’s perception of value at anygiven time.

● A market is a biological entity and acts as a dynamic organism thathas a short and long-term memory and has a degree of self-awareness of its past and present behaviour. This accounts foralternation in adjacent corrective structures.

As an investor you need to:

● Understand the way your brain is hard wired● Find ways of optimising the way you trade and how your brain

operates● Optimise your separation from the market consciousness.

David Murrin, www.wavecounts.com

The Coppock Indicator’s buy signals around the start of 2002 on the main

European and U.S. share indices led me to think about the mechanics of the

indicator. It is widely used on Stock Indices, and there is plenty of published

work on its reliable record, when it signalled, and how much markets rose

afterwards. But why might it (or might it not) work as a long-term trend

indicator?

Coppock’s original concept was to devise a momentum indicator, with the

period chosen to reflect our period of adjusting to life’s traumas like death

and divorce; this was determined at 11-14 months. Presumably he reasoned

that financial loss following a serious stock market fall (he studied the Dow

index) would be such a trauma, and need a similar period for sentiment to

recover. So he devised the indicator based on monthly prices compared

with their 11-14 months-ago figures. He averaged those periods’

momentum figures, then produced a weighted moving average of the

answer. This gave a smoothed indicator with apparently faster response to

more recent months’ data.

The result is a curve, which cycles up and down over several months, being

visually straightforward with smooth major trends but occasional small

blips. The major turns are the signal points – the absolute number of the

indicator is not important. The indicator number will usually (but not

always) swing into negative territory before turning up, and it usually rises

to a positive figure greater than its deepest negative before rolling over

again. Coppock said it must only be used for market recoveries (when the

indicator bottoms), and presumably this is in line with his trauma thinking

– a rising market towards a top has a different psychological effect on us, so

he did not deem it suitable for tops. Finally, he said a low-risk ‘buy’ signal

was given by the indicator turning up from a below-zero reading.

However, some people do use the topping of the indicator as a ‘sell’ signal,

and also recognise a ‘buy’ signal when it turns up from just above zero. How

valid might all this be?

Calculating the indicator figure – conventionally

Price – Coppock used (maybe for ease of calculation in the pre-computer

era) only a month-end figure for the Dow index. Nowadays we can as easily

take averages of weekly or daily closes within each month, to represent a

month’s index performance better. Whichever representative ‘price’ is

chosen, the calculation then continues.

Momentum – for each month, the price is divided by that of 11 months

previously, and also by that of 14 months previously, and each is expressed

as a percentage increase. The ‘momentum’ for the month is the average of

these two percentage increases. Some software shortcuts this and just uses

a single 12 month figure. Whichever ‘momentum’ figure is used, the

smoothing is then the same.

Smoothing – The momentum figure is both smoothed using a moving

average, and weighted for more responsiveness to recent months. The

Coppock Indicator is calculated as a linear-weighted 10-month moving

average of the momentum. If, for example, the momentum in month one is

called M1, then the indicator number is calculated as the left-hand column

of the table.

Coppock calculated in this way for the FTSE100 in the 1990s is shown

plotted against the FTSE itself on Chart A. One can observe :

● the three ‘buy’ signals from indicator-below-zero were all successful; they

followed extended sideways ranges or triangle patterns in the share

index.

● a short sharp correction at the start of 1994 was not enough to trigger a

Coppock turn up; a longer period of consolidation was required first.

● using Coppock as a ‘sell’ indicator, you would have had mixed success

and two failures.

● the 1997 ‘buy’ from indicator-above-zero was successful, if you allowed it

as a signal, though late in an already well-established trend.

● for Coppock to reach below zero requires the share index to be below its

previous peaks for an extended time.

Thus far there are no new insights, but the workings of Coppock may be

seen in a clearer light through the following, alternative method of

calculation.

Calculating the indicator – alternative method

Whilst most people calculate the indicator by using the full formula each

month, the alternative way is to calculate in full for only the first month to

get a starting number, then add the calculated change each month to get

the next month’s figure. The Table’s column ‘Change in Coppock’ shows an

interesting result. The Change in the Indicator figure from this month to

next is simply:

– next month’s ‘momentum’ percentage, minus this month’s simple

moving average of ‘momentum’ for the 10 months to date.

This gives us the clue to understanding the Coppock Indicator’s response:

– Coppock turns up whenever the ‘Change’ becomes positive after

falling. That is, when next month’s momentum exceeds its 10-month

simple moving average this month. And that is, when momentum cuts

its own moving average from below – the classic moving average

crossover signal.

So to find when Coppock will signal ‘buy’, you do not need to calculate or

plot the indicator at all – merely chart say the12 month momentum of the

share index (the purist will use the average of the 11 and 14 month-end

MARKET TECHNICIAN Issue 44 – July 20026

Coppock understoodBy Christopher Mack

This month’s Coppock Next month’s Coppock Change in Coppock– 1.0 x M11 + M11

1.0 x M10 0.9 x M10 -0.1 x M100.9 x M9 0.8 x M9 -0.1 x M90.8 x M8 0.7 x M8 -0.1 x M80.7 x M7 0.6 x M7 -0.1 x M70.6 x M6 0.5 x M6 -0.1 x M60.5 x M5 0.4 x M5 -0.1 x M50.4 x M4 0.3 x M4 -0.1 x M40.3 x M3 0.2 x M3 -0.1 x M30.2 x M2 0.1 x M2 -0.1 x M20.1 x M1 – -0.1 x M1

total = indicator for total = indicator for total = M11 minus (this) month 10 (next) month 11 0.1(M1.....+M10)

Issue 44 – July 2002 MARKET TECHNICIAN 7

figures, as Coppock did), superimpose the simple 10-month moving

average of that momentum, and observe the crossovers from below. Chart

B shows Coppock, compared with momentum and its simple10-month

moving average, for the same years as Chart A, to demonstrate this.

Coppock’s response now becomes clear:

– underlying the Coppock curve is the chart of momentum with its own

simple 10-month moving average. Coppock’s signals (turns up or

down) correspond exactly with momentum’s moving average

crossovers.

– the usual features of a momentum chart apply. Momentum may show

definite trends which can be analysed with trendlines; breaches of

trendline slightly anticipate the Coppock signal. Likewise divergences

in momentum vs price give anticipatory signals. Overbought and

oversold levels can also be set for momentum, which may also

anticipate or confirm the moving average crossovers.

– a Coppock ‘below-zero-buy’ results when momentum has been in a

lengthy downtrend and is itself coming up from below zero. The buy

signal itself (at the moving average crossover) occurs near the

conventional buy signal given when momentum crosses up through

its zero line.

– a ‘buy-just-above-zero’signal comes when momentum has been falling

less sharply and is still above zero. In that sense, the signal is weaker

(less oversold) than a turn up from below zero. Likewise, a turn up from

well above zero is weaker still as a signal.

– a deep ‘spike’ or V-bottom reversal in the share index may be expected

to reverse momentum sharply enough to foreshadow a Coppock

signal – as it did in September 2001.

– the small blips on the Coppock curve – turning down in a rise, then

up again within a month or two – are now seen as whipsaws in the

moving average crossover of momentum. It is not the 11-14 month

momentum period which controls this, but rather the 10 month

smoothing and moving average period. With a longer moving average

period there would be less false signals like this but at the expense of

a later crossover signal.

● Coppock’s Indicator has a smooth curve because he only took a single

month-end figure for price to calculate momentum. All the potential

whipsaws within a month were thereby eliminated – but the month-

end figure itself might still produce a whipsaw for a month or two.

These whipsaws represent the 2-3 month intermediate corrections

often present in share and index prices.

● To see whether a Coppock turn is potentially a real signal or a

whipsaw, you might therefore wait a month for confirmation. You

might also look for other evidence in the price chart – for example, is

there evidence of completion of an Elliott five waves, anticipating a

major correction, or is this only the end of a first or third wave, in

which case a Coppock blip is more likely?

● You can anticipate Coppock signals using a more sensitive weekly

chart for momentum, for example a 52-weekly momentum chart

with a 44-week simple moving average. The chart shows momentum

trends, moving average crossovers and whipsaws very well. A monthly

chart of 12-month momentum with 10-month moving average will

be smoother and closely approximate pure Coppock signals. You can

get a feeling for whether momentum has finished trending and is

therefore likely to produce a valid moving average crossover signal.

Chart C is a recent unsmoothed example of FTSE momentum data

which illustrates this. The February 2002 breach of the downtrend

line from the 1999 momentum high helps confirm the Coppock signal

in January; but a breach of the longer-term downtrend line from the

1998 high would help confirm it more strongly. Whether the upward

move will last and for how long is, as always, difficult to predict.

Not valid for downturns?

There remains the question of why the Coppock Indicator might work

better for upward than for downward turns in a share index. In principle,

momentum is symmetrical – ‘sell’ signals should be the converse of ‘buys’ –

momentum crosses over its 10-month moving average from above, and

Coppock turns down.

The answer must lie in the shape of the trends of the price/index chart and

the response of the momentum chart to them. For example, a V-top spike

reversal should be as good a reversal signal from a momentum point of

view as it would be at a bottom.They just don’t occur very often at tops. But

a sharp, deep turn-down in price would be one component of a valid

Coppock sell signal. Conversely, a short shallow price correction will stall

momentum’s rise and may turn it down – but is only likely to cause a

Coppock blip. This will be even more likely if the correction is only a small

percentage of the previous wave up, not reaching usual Fibonnacci levels.

The timespan of the price correction is clearly a key factor. If there is a

shallow but enduring price correction in a range, momentum will stall and

its moving average will rise towards it. This will increase the chance of an

moving average crossing from above and hence a Coppock ‘sell’. But such

trading ranges must also be examined on a long-term price chart for signs

of whether they portend long-term rises or falls. Again, Elliott may help to

determine the type and degree of correction in a long-term pattern.

Another factor will be the duration of the price correction compared with

the chosen momentum period – in Coppock’s case, 11-14 months. Brief

inspection of the price chart shows that a price correction lasting at least 11

months and at least 10 per cent in depth seems necessary for a good

Coppock ‘buy’ signal. FTSE and other main indices display that. If so, it

ought to be true in reverse – a good sell signal needs a preceding price rise

of at least that extent and size in the 11-14 month preceding period (not

counting any extended rise before that).

However, upward trends in share indices last longer than downward trends.

If there has been a price rise lasting much longer than the chosen

momentum period, at the end of which the price has flattened out, then the

above condition may not be met. So despite a good overall foregoing rise,

a price high and a Coppock ‘sell’ signal, momentum may be telling you this

is just a pause in a longer rise. You may get a much better view by looking

at momentum charts for longer periods – for example, a 150-week

Continues on page 8

MARKET TECHNICIAN Issue 44 – July 20028

I have been involved in the speculative markets for over 15 years, and one

thing I have learnt, particularly in these days of ever more complicated and

sophisticated analysis techniques, is that sometimes the simple approach to

analysis is all that is needed to be able to uncover some great trading

opportunities. One of the simplest trade formations that I use is the simple

ABC correction. Although this is not a new pattern, I believe that its profit

potential has been largely overlooked recently, mainly due to this simplicity.

A simple, ABC correction is a three swing correction against the main

trend, where the third swing exceeds the price extreme of the first swing

as shown in chart 1 below.

Chart 1

The chart shows how the euro/dollar exchange rate was in a strong

impulsive rally.This was followed by a simple three swing correction, where

the third swing exceeded the low (in this case) of the first corrective swing.

This is a simple ABC correction against the main up trend.

Why is the identification of the simple ABC corrections so important? Well,

once the simple ABC correction is complete, the main trend normally

resumes. If we look at the above example a few days later, we can see how,

once the simple ABC correction was completed, the euro resumed its up

trend moving up to new highs. (see chart 2 below.)

Chart 2

As such, the end of the simple ABC correction represents a great place to look

to enter the market to take advantage of a resumption in the main trend.

However, in the past the main difficulty has been identifying where the

simple ABC correction was most likely to end. This is resolved by looking at

the price relationships of the individual swings within the simple ABC

correction itself. The most common relationship is where the Wave C (third

swing) is equal in price to the Wave A (first swing). Armed with this simple

piece of knowledge, we can anticipate where the price area (Wave C), and,

as such, the entire ABC correction is most likely to end. I term this the typical

Wave Price Target (WPT).

Chart 3

Chart 3 shows how this simple ABC correction ended right in the typical

Wave C WPT support area. Knowing the most likely support area where the

entire ABC correction is most likely to end allows us to enter a trade with a

very small risk, allowing us to be in a very good position to take advantage

of the main trend once it resumes.

We can also look at price relationships between Wave C and Wave B to

obtain narrow price clusters which will help indicate where Wave C is

anticipated to end. However, the most important relationship for projecting

the end of the Wave C is related to the Wave A swing. As such, the price

cluster given by the Wave C = Wave A price swing in is the most important.

In a subsequent article I will show how the simple ABC correction can be

used to identify the start of a strong impulsive swing. In Elliott waves terms,

this means identifying the start of a Wave 3 type swing, which can be one

of the most profitable trades in a complete 5 wave sequence.

Steve Griffiths is the developer of the MT Predictor software program, which is

designed to scan, and automatically identify ABC corrections as well as specific

Type 1 and 2 trades and all of the Elliott wave sequences. For more information,

please visit Steve’s web site at www.MTPredictor.com

continued from page 7

momentum chart will give you a better feeling of whether momentum is

well overbought, which may also be a pre-condition for a good Coppock

‘sell’ signal. It will also give you divergences from price, which are not

apparent from the Coppock Indicator alone.

Conclusion

Though the Coppock Indicator is quite reliable and widely followed, it will

never be a perfect indicator. This analysis shows that it is just a ‘subset’ of

ordinary momentum, with its success due to judicious choice of the

momentum period (11-14 months) and the smoothing period (10 months).

Coppock signals actually translate into signals from an ordinary

momentum chart with a simple moving average crossover system, using

the same periods. Though the indicator signals ‘buys’ quite well after

deepish downtrends lasting around a year, chosing different periods may

tune it to give better ‘sell’ signals for a given market. In any event, a weekly

or monthly momentum chart with a period of at least 11 months can give

much more guidance than Coppock alone, whilst also giving the exact

Coppock signal time at its 10-month moving average crossover point.

The simple ABC correctionBy S. E. Griffiths

Issue 44 – July 2002 MARKET TECHNICIAN 9

Abstract

This paper describes and evaluates the use of support vector regression to

trade the three month Aluminium futures contract on the London Metal

Exchange, over the period June 1987 to November 1999.The Support Vector

Machine is a machine learning method for classification and regression and

is fast replacing neural networks as the tool of choice for prediction and

pattern recognition tasks, primarily due to their ability to generalise well on

unseen data. The algorithm is founded on ideas derived from statistical

learning theory and can be understood intuitively within a geometric

framework. In this paper we use support vector regression to develop a

number of trading sub-models that, when combined, result in a final model

that exhibits above-average returns on out of sample data, thus providing

some evidence that the aluminium futures price is less than efficient.

Whether these inefficiencies will continue into the future is unknown.

Motivation

Is it possible to design quantitative trading models that result in above-

average, risk-adjusted returns? The Efficient Markets Hypothesis (EMH) rules

this out as a possibility.This is not surprising; the arguments supporting the

EMH are extremely persuasive (Fama, 1965), none more so than the

contention that any predictable component will be traded out of the

markets by ‘rational’ arbitragers, rendering them efficient once again.

Few would disagree with the idea that a visible discrepancy in price, for the

same commodity in two markets, would quickly disappear due to the

effects of arbitrage. The reality, however, is that profit opportunities, when

they exist, are not as obvious as the arbitrage argument might suggest.

They tend to be statistical in nature and, though they may represent a

favourable bet, they are not riskless, possibly requiring infinite capital to

remove completely (Zhang, 1999). Moreover, the effectiveness of ‘zero risk’

(as opposed to statistical) arbitrage relies on the availability of a perfect (or

close) substitute. This is not always the case.

If we accept that obvious, predictable components will be traded out of the

markets with relative ease then we must conclude that any remaining

inefficiencies, if they exist, are complex in nature to a degree that they are

not easily exploited with methods used by the majority of market

participants1. With this mind, we employ a relatively new machine learning

method, support vector regression, in an attempt to extract possible

regularities in the market price of aluminium – a market that is arguably

less scrutinized than others such as the stock markets.

Aluminium on the LME

The London Metal Exchange (LME) was established in 1877 and is the

world’s largest non-ferrous metals derivatives market with a turnover value

of approximately US$2000 billion per annum. It is a 24-hour market trading

through a combination of continuous inter-office dealing and open-outcry

sessions at certain fixed time slots during the day. Of this, hedging

represents 75-85% of turnover (Martinot et al., 2000). Aluminium began

trading on the LME in 1978 though it was only in the mid-eighties that the

contract became widely used when the LME price was adopted as the

industry marker price (Figuerola-Ferretti & Gilbert, 2000). At the time of

writing the LME price forms the effective price basis for the international

base metals market.

The LME three month Aluminium futures contract (liquidity is mainly

concentrated in the three month and cash contracts) is a forward contract

between buyer and seller for delivery of 25 tonnes of the metal on a

specified three month “prompt” date in the future at a specified price. The

majority of positions are closed before prompt by trading offsetting

contracts, replacing delivery obligations with monetary differences,

officially quoted in USD – though sterling, euro, mark and yen can be used

for clearing purposes.

The method of trading on the LME differs to that on most standard futures

exchanges partly due to the LME’s close links with the physical metals

industry and its status as a wholesale market (Gilbert, 1996). While initial

and variation margins are called during the term of the contract, profits and

losses are not realised until the contract prompt date or until it is closed out

– deemed an advantage to the physical users of the exchange. Moreover,

when a trade is entered it is at the current three month price, however,

when exiting the trade, an adjustment has to be made to take into account

the possible contango or backwardation of the contract which depends on

demand, supply and interest rate factors for contracts in different delivery

periods. This means that the quoted historical three month price being

modelled is not what would be experienced in real-time trading.Whilst it is

important to bear this in mind, it is not a serious problem as the differences

will tend to even out in the long term, with long (short) positions affected

adversely (favourably) in conditions of backwardation and vice versa in

conditions of contango.

Support vector regression

The Support Vector Machine (SVM) is a powerful machine learning method

for classification and regression which is fast replacing neural networks as

the tool of choice for prediction and pattern recognition tasks, primarily

due to its ability to generalise well on unseen data. Although the SVM, as a

learning method, has only recently gained in popularity, the underlying

principles of the algorithm date back to work done by Vapnik in the early

60’s (Vapnik & Chervonenkis, 1964; Vapnik & Lerner, 1963) and are based on

ideas derived from statistical learning theory. This recent increase in

popularity is due to advances in methods and theory which include the

extension to regression from the original classification formulation. For a

thorough treatment see (Vapnik, 1998; Vapnik, 1995), the tutorials (Burgess,

1998; Smola & Scholkopf, 1998) and the introduction (Cristianini & Shawe-

Taylor, 2000).

SVM Regression involves a non-linear mapping of an n-dimensional input

space into a high dimensional feature space. A linear regression is then

performed in this feature space. SVMs use the structural risk minimisation

(SRM) induction principle which differentiates the method from many

other conventional learning algorithms based on empirical risk

minimisation (ERM) alone, for example standard neural networks. This is

equivalent to minimizing an upper bound in probability on the test set

error as opposed to minimising the training set error, which should result in

better generalisation. Importantly for practitioners, recently published

research has shown successful application of the SVM methodology in a

wide variety of fields (Barabino et al., 1999; Joachims, 1997; Mukherjee et

al.,1999; Trafalis & Ince, 2000).

The method has a number of advantages over other techniques; the

parameters that need to be fitted are relatively low in number and, unlike

Using support vector machines to tradealuminium on the LME

by Zac Harland

1Fundamental and/or traditional technical analysis

MARKET TECHNICIAN Issue 44 – July 200210

other methods such as neural networks, they do not suffer from local

minima. The two main features of SVMs are their theoretical motivation

from statistical learning theory and the use of kernel substitution to

transform a linear method into a general non-linear method, with little

added complexity.

Model design and methodology

There are different approaches when it comes to deciding how much data

to use when designing trading models. One view is that the market is

always changing and therefore one does not want to use data too far back

in history, as there is a danger that much of it will be redundant. The other

approach is to use as much data as is available, reasoning that the only way

to have confidence in the model’s final results is if it has acceptable

performance over as long a data history as possible. We subscribe to the

latter approach and use the entire available data set.

To tackle the problem we invoke the principle of divide and conquer in that

we start by attempting to build a number of trading sub-models, each using

a different set of input features. These sub-models are then combined with

the intention of creating a final model that is more effective than any one

model used in isolation – constituting a type of committee machine.

Transaction costs are not used as a constraint on model selection when

building the sub-models; the rationale being that in some cases strong short

term regularities contained within price, but not tradeable in isolation due

to transaction costs, might be exploitable if derived models are combined

with other longer term sub-models and the use of majority voting methods.

At the final stage, a trading wrapper – which takes into account commission,

slippage and order type – is added to simulate real trading.

The data used in this study consist of the “LME provisional closing price” –

the daily 5pm close of the second kerb session of the three month LME

Aluminium futures contract, covering the period from 11 June 1987 to 4

Nov 19992/3. Data from 5 Nov 1999 to the present date is excluded from this

study as it is to be used in a final trading meta-model (that may or may not

include the results from this paper) which will require further out of sample

testing. This will help to alleviate the problem of ‘cherry picking’ the best

models.

The data are divided into training, validation and out of sample sets. The

training period covers 2136 days from 11 June 1987 to 17 Aug 1995, the

validation set 400 days from 18 Aug 1995 to 27 Feb 1997 and the out of

sample from 28 Feb 1997 to 4 Nov 1999, 700 days (see Table 1).

Table 1

Inputs

Finding a good representation of the data to use as inputs and outputs is

very important, especially when building trading models.The objective is to

find a representation that will render the signal (if one exists) more explicit

and/or attenuate the noise component. The number of possible

transformations of price to arrive at potential input candidates is infinite, so

for the sake of simplicity we considered inputs of the form:

Xt = log (pricet-0 /pricet-N) , N = 1,...,10 (1)

in addition to associated lags up to a maximum of 10 and, finally, a

proprietary input based on price seasonality derived from initial

exploratory data analysis. Training examples with values beyond the

±3j range were clipped and all values were then scaled to zero mean, unit

variance. The number of inputs per model was kept to a maximum of six in

order to limit complexity. Standard log returns were chosen as the target or

dependent variable:

Xt = log (pricet+1 /pricet-0) (2)

The objective was to choose input candidates that exhibited correlation to

future changes in price i.e., the target. To do so, a number of techniques to

measure correlation were used including non-parametric correlation,

mutual information and a technique using ANOVA also used in (Harland,

2000a; Harland, 2000b ) which is similar to that originally proposed by

(Burgess & Refenes, 1995). The main criteria was that any correlation

exhibited by a potential input candidate had to be as constant as possible

over the full length of the training set.

The above procedure resulted in multiple candidate SVR input sets which

we restricted in number to ten. SVR was then used to build a model for each

input set. The choice of kernel determines the type of the resulting learning

machine. Common kernel functions include polynomial, radial basis

functions (RBF) and sigmoid kernels. In this experiment we used RBF

kernels which have the form:

K(x, y) = exp 2j 2 (3)

where j 2 is the width of the kernel. The SVM regression method has a

number of ‘tunable’ parameters that need to be determined by the user: C

a regularisation parameter, e , and in this case the width of the RBF kernel

j 2. Table 2 shows the values that were tested.

Table 2

To simulate trading the following rule was used on the continuous output

of the SVM:

IF SVM_OUTPUT>0 THEN LONG ELSE SHORT.

A measure of each model’s performance was calculated over the validation

set by dividing total daily log returns by the daily standard deviation

(similar in nature to the Sharpe ratio).Those parameters that resulted in the

highest value for this statistic were chosen for the final sub-model, with the

proviso that the gradient of the in-sample training performance was

reasonably similar to that exhibited over validation data. The above model

design procedure resulted in ten sub-models. These were combined

together using majority voting to arrive at the actual daily trading signal –

see further details below. For the sake of brevity we provide the results for

two of these sub-models in addition to the final model.

Sub-Model 1

Figure 1 depicts the performance over the whole dataset and is based on

trading one contract. The three month aluminium price is re-based to zero

at the start of the period and multiplied by 25 (contract is for 25 tonnes)

and represents the equity stream experienced by buying and holding one

contract. The Cumulative Equity Curve (CEC) represents the profit

gained/lost by following the output of the SVM following the rule; If

output>0 then long else short. Figure 2 shows the performance over the

out of sample data.

2In order to satisfy concerns regarding robustness we also tested the methodology using official am/pm “fixes” and gained similar results.3Prior to 1988 the data is the “P.M. Unofficial” price. The start date is the earliest available for CSI data providers.

Set Dates Length

Training Set 11 June 1987 to 17 August 1995 2136 Days

Validation Set 18 August 1995 to 27 February 1997 400 Days

Final out of 28 February 1997 to 4 November 1999 700 Dayssample set

Parameters Values

C 10, 100, 1000, 10000,100000

e 0.1, 0.01, 0.001, 0.0001

j 2 0.001, 0.01, 0.25, 0.5, 0.75

( )_ II x – y II 2

Issue 44 – July 2002 MARKET TECHNICIAN 11

Figure 1

Figure 2

Table 3

A number of trade statistics for this sub-model can be seen in table 3. The

results are also included for the second half of the in-sample training set to

give a clearer picture of overall performance.

As mentioned above, no transaction costs have been included so the

numbers at this stage should be seen as general statistics rather than real

trading results. As expected, the training and validation results are good.

Most important are out of sample results, which, in this case, exhibit a

generally upward sloping CEC along with a similar winning percentage as

the rest of the data. The CEC is relatively smooth over the whole data set

from in-sample to out of sample, although the average trade figure

decreases somewhat in the out of sample data to $101, from $118 in the

validation set. If we assume transaction costs of $100 per round turn the

model is not tradeable in its current state, however, the results suggest that

it is managing to detect some form of predictable structure in the price of

aluminium and is therefore used in the final model.

Sub-Model 2.

This model uses inputs that were designed to result in longer average trade

duration and higher average trade returns than the previous model. Results

can be seen in Figures 3, 4 and table 4.

Figure 3

Figure 4

Table 4

Model All 2nd Half Validation Out of Training Training data sample data

Start Date 870611 910716 950818 970228

End Date 950817 950817 970227 991104

No. of Traders 939 469 168 300

No. of Winners 526 252 101 164

Pot. Winners 56 53 60 54

Gross Profit $403,912 $130,287 $46,437 $73,500

Net Profit $191,000 $56,975 $19,925 $30,525

Average Trade $203 $121 $118 $101

Average Winner $767 $517 $459 $448

Average Loser $515 $337 $395 $315

Max. Draw $12,950 $5,675 $4,400 $4,250

Avg. bars in Wins 3 3 3 3

Avg. bars in Loses 3 3 3 3

Avg. bars in Trades 3 3 3 3

Model All 2nd Half Validation Out of Training Training data sample data

Start Date 870611 910716 950818 970228

End Date 950817 950817 970227 991104

No. of Traders 386 164 61 108

No. of Winners 187 73 27 50

Pot. Winners 48 44 44 46

Gross Profit $284,450 $79,400 $23,287 $35,350

Net Profit $164,450 $46,625 $10,700 $17,600

Average Trade $426 $284 $175 $162

Average Winner $1,521 $1,087 $862 $707

Average Loser $603 $360 $370 $306

Max. Draw $11,750 $10,575 $4,600 $4,425

Avg. bars in Wins 8 11 9 10

Avg. bars in Loses 4 4 6 4

Avg. bars in Trades 6 7 7 7

MARKET TECHNICIAN Issue 44 – July 200212

Final Model

The final model is the result of combining the ten sub-models, using a

special rule to trade the result. The rule is explained below, the parameters

of which were arrived at via optimising the Sharpe ratio over the in-sample

training and validation sets subject to the constraint that the average trade

was greater than $250 excluding transaction costs.

Majority Trading Rule:

1) Each SVM sub-model output is assigned 1 if >= 0.0 and -1 if < 0.0.

2) The result for each model on each day is then summed, producing a

number representing the majority decision which can range from -10

to 10.

3) A long (short) trade is only taken if this majority is above (below) a

certain (-)threshold. In this case the threshold was ± 4.

4) Each trade is then held until the majority decision signals a trade in the

opposite direction. This results in what is commonly called a stop &

reverse system and is always in the market.

Figure 5

Figure 6

The results can be seen in Figures 5, 6, and in table 5. Slippage & transaction

costs of $100 per trade, that is per round turn, are included.

The results are encouraging; the Sharpe Ratio for the training and validation

sets is just over 1.6, rising in the out of sample period to 2.0 (the Figure of

1.41 for the second half on the training data was due to a higher than

average standard deviation of returns). It has an average trade of $258 and a

maximum drawdown (MD) of $4725 on the out of sample data. A more

sobering $14,900 MD occurs at the beginning of the training period from 19

Oct 1987 to 17 Dec 1987 – during the stock market crash of the same year.

The average trade in the validation set is $175, lower than both the training

and out of sample sets. At the start of the training data in Figure 5 the CEC

exhibits a sharp rise and then has a relatively constant gradient from the

early-nineties to the end of the data. One possible explanation is that the

market was less efficient at the start of the period, though this is by no

means certain. There is a flat period from 8 Mar 1995 to 18 Jun 1996, which

would be difficult to trade through, however, the fact that it occurred in the

training period suggests the model is not overfitting to any great extent.

Table 5

No trade exit strategy has been incorporated such as stop loss exits, profit

limits etc., as their addition did not result in any discernable improvement.

We find this to be generally the case with trading models of this type –

probably due to the model signal itself indicating the ideal time to exit a

trade. Having said that the final trading system would have emergency

stops placed at a distance where they would only be hit in extreme price

moves – whether they would be filled at that level is another matter.

Conclusion

An effective methodology has been developed for the application of

support vector regression to trade three-month Aluminium futures on the

LME. Ten sub-models were designed and then combined using a majority

voting trading rule to obtain a final model that, as a first attempt, exhibits

profitable performance over out of sample data.This suggests that the three

month aluminium price contains inefficiencies that can be exploited using

machine learning.Combining the Final Model with other models built by the

author using different methods should result in a usable trading system for

aluminium futures. Of course, there is no guarantee that the relationships

detected over the time period analysed will continue into the future – it may

be that these inefficiencies have been traded out of the market.

Acknowledgements

It is a pleasure to thank Terence Roopnaraine, Martin Sewell and Paul Teetor

for useful discussions and corrections.

Zac Harland, Krueger Research

References

Barabino, N., Pallavicini, M., Petrolini, A.k, Pontil, M., & Verri, A. (1999). ‘Support

Vector Machines vs Multi-layer Perceptrons in Particle Identification.’

European Symposium on Artificial Neural Networks, 1999, Bruges, Belgium

Burges, C.J.C. (1998). A Tutorial on Support Vector Machines for Pattern

Recognition .’ Data Mining and Knowledge Discovery, 2(2):1-47.

Burgess, A.N. & Refenes, A.N. (1995). ‘Modelling Non-linear Cointegration in

International Equity Index Futures.’ Neural Networks in Financial

Engineering (Proceedings of the Third International conference on neural

networks in the Capital Markets 1995). World Scientific.

Continues on page 15

Model All 2nd Half Validation Out of Training Training data sample data

Start Date 870611 910716 950818 970228

End Date 950817 950817 970227 991104

No. of Traders 369 162 75 98

No. of Winners 203 76 41 58

Pot. Winners 55 46 54 59

Gross Profit $280,450 $78,537 $29,762 $46,550

Net Profit $172,250 $38,675 $13,125 $25,325

Sharpe Ratio 1.64 1.41 1.61 2.01

Average Trade $466 $238 $175 $258

Average Winner $1,381 $1,033 $725 $802

Average Loser $651 $463 $489 $530

Max. Draw $14,900 $7,175 $6,600 $4,725

Avg. bars in Wins 7 8 4 9

Avg. bars in Loses 6 6 8 5

Avg. bars in Trades 6 7 6 8

Issue 44 – July 2002 MARKET TECHNICIAN 13

After years of searching for an investment/trading methodology which is

both fully automated and able to beat the stock market, we developed and

settled upon the Balance of Sentiment System (BoSS).

While much of Warner Prince’s equity, bond, currency and commodity

market research has historically utilized price momentum as the primary

driver of automated buy/sell signals, we felt we could improve our results.

Late in the 1970s we hit upon a combination of daily New York Stock

Exchange breadth statistics which worked quite well, but there were

several drawbacks. First, we did not have a computer; so we had to hand

calculate our formula from all the numbers which was quite tedious.

Second, we learned that there was tremendous volatility in the end result

of the formula, and, having not started our charts on semi-log paper, found

it took an enormous height of graph paper on the wall to keep the charts

up-to-date. Frequently, we found ourselves standing on chairs to post new

plots on the charts. Looking back on it now from our state-of-the-art P.Cs, it

seems completely ridiculous. Such is the life of an independent,

undercapitalized research firm.

The lack of a computer made it virtually impossible to sample large

amounts of historical data; so we were stymied in trying to back-test our

formulas to determine their long-term veracity.

With the advent of the desktop computer in the early 1980s we obtained

graphic software first for our Apple IIe and then for our IBM P.C. and, in the

early 1990s, we came upon SuperCharts´®, which has served us well ever

since.

Several years ago I had some spare time to do more research. I felt I really

wasn’t getting anywhere until one night I suddenly awoke from a dream with

what I thought was the answer and immediately wrote down the idea that

would keep the volatility within a useable range. The next day, after some

considerable tussle with my maths and the programming language I was able

to produce what is now known as the Balance of Sentiment Signal or BoSS.

BoSS simply measures the difference between buying pressure and selling

pressure as indicated by breadth statistics. When BoSS crosses the

threshold on the upside it gives a buy signal and a sell/sell short signal on

a cross to the downside.

Chart 1

It was immensely satisfying when, without any tweaking or testing

whatsoever, I applied the formula to fifteen years of daily New York Stock

Exchange data and discovered that to December 29 2000, 51% of the trades

had been profitable, the average gain per trade had been 2.04x the average

loss, and the model had reaped 917.18 NYSE Composite Index points profit,

versus the bought and held index’s gain of 553.94 points… without any

short sales!1

Chart 1 shows the market index with signals (1=buy;0=sell) and the daily

Balance of Sentiment plot over the last three years of this period. It also

graphs the cumulative results as “System Equity”.

Table 1 shows the quantitative summary of the activity and results over

those 15 years.

Chart 2 brings results up to the date of writing this paper. Note that since

December 29, 2000 the index has been in a bear market, dropping from

656.85 to 581.14 or by 75.71 points (-11.5%), but BoSS’ profit has been

24.07 points (+3.7%).

Despite this success, one problem to using this research is that over the

past 17 years period the model has made 906 trades, or slightly more than

one a week.This is no problem for traders, but for me and my accounts who

are not fast traders, it is too much action.2 Since the BoSS model uses

statistics for the exchange as a whole it is best applied to trading certain

broad market indices. Our NYSE BoSS model is most applicable to trading

the very broad NYSE Composite Index. It works, but not quite as well on the

S&P 500 and DJIA. Nonetheless, there are several derivatives available to

trade the NYSE Composite Index.

We have also applied BoSS to NASDAQ statistics with outstanding results.

Again, the problem from our perspective being too many trades, we slowed

things down using moving averages. We back-tested various lengths of

moving average together with various hurdle and trip levels. In the final

analysis, profits were best when the moving average BoS has to climb over

a .7 reading before it will give a buy signal and drop below -3.2 before it will

give a sell signal.

Beat the index using the indexBy Seth Warner

1Past performance is not necessarily indicative of future performance.2The BoSS model’s results assume that all trades are made at the closing price on the day of the signal and do not deduct commissions, which under the right circumstances in theU.S.A. have become almost negligible.

Table 1BoSS System on the NYSE Composite Index 3/19/1985 – 12/29/2000

Gross Profits 1737.08

Gross Losses 819.90

Net Profit 917.18

Total # Trades 835

Number of Profitable Trades 425

Number of Losing Trades 410

Percent Profitable 51%

Average Profitable Trade 4.09

Average Losing Trade 2.00

Ratio avg. profit/losing trade 2.04 to 1

Average # days profitable trades held 4

Average # days losing trades held 1

MARKET TECHNICIAN Issue 44 – July 200214

Chart 2

Chart 3 presents the results of this application over the past ten years and

shows the moving average which triggers the buy/sell signals.

Chart 3

During the period October 6, 1992 to April 24, 2002, the number of trades

on the slowed NASDAQ model was reduced to about 7.6 a year, and 55%

have been profitable.The all trades win/loss ratio has improved to 2.59 to 1.

The model has taken out 3,753.15 index points of profit vs. 1,142.79 for the

index itself. System Equity shows that it has not lost money in this bear

market.

Throughout the part of our research career that has been centered on

timing market indices rather than individual stocks, we have been the

subject of some criticism. People would say “you can’t trade the market”.

We’d say “we know that but, just as you’ll cover more ground flying with

the wind at your back rather than in your face, you will want to be long

most stocks in a bull market and out of most stocks in a bear market.”

Now that there are highly liquid exchange traded funds (ETFs) which mimic

most important indices, our research can be directly applied to markets as

a whole, and we have begun managing funds employing the BoSS models.

Some of the ETFs are highly liquid. For example, QQQ (the NASDAQ 100

Trust, a quasi proxy for U.S. hi-tech companies) has been trading in excess

of $2 billion a day for some time and is now one of the most actively traded

stocks in the world. Chart 4 shows our smoothed BoSS on QQQ since June

4, 19993.

Chart 4

Table 2 presents the date and price of each trade and the cumulative profit.

Table 2

BoSS System on QQQ – NASDAQ 100 Trust 6/4/1999 – 4/24/2002

Date Action Price Entry P/L Cumulative

Profit

06/04/99 Buy 52.56

06/14/99 Sell 50.80 -1.76 -1.76

06/16/99 Buy 54.21

07/26/99 Sell 55.88 1.67 -.09

08/20/99 Buy 57.97

09/22/99 Sell 62.59 4.63 4.53

10/07/99 Buy 63.09

10/15/99 Sell 59.88 -3.21 1.32

10/29/99 Buy 65.75

03/17/00 Sell 110.81 45.06 46.38

05/03/00 Buy 89.13

05/05/00 Sell 91.50 2.37 48.75

06/07/00 Buy 94.00

06/23/00 Sell 92.00 -2.00 46.75

06/30/00 Buy 93.44

07/05/00 Sell 91.38 -2.06 44.69

07/12/00 Buy 97.25

07/24/00 Sell 95.00 -2.25 42.44

08/15/00 Buy 93.00

09/12/00 Sell 91.25 -1.75 40.69

11/02/00 Buy 82.25

11/07/00 Sell 82.50 .25 40.94

01/09/01 Buy 57.13

02/05/01 Sell 61.50 4.37 45.31

04/17/01 Buy 41.25

3This is the same as the model run on the NASDAQ Composite for the past ten years.

Issue 44 – July 2002 MARKET TECHNICIAN 15

Table 3 summarizes signals on the NASDAQ 100 Index between 4/6/1999

and 24/4/2000.

One can easily see that over the above period BoSS has been wrong more

than it has been right but, as successful traders say, it has tended to cut

losses quickly and let profits run. From its sell signal near QQQ’s peak in

March 2000 at 110.81, QQQ itself has lost 74.17 points or 66.9%. The BoSS

model has actually made 4.03 points during this period without ever

selling short.

Given the still significant number of trades, the BoSS models are most

appropriate for U.S. tax exempt or off-shore accounts. Our more

aggressive models, which also go short and about double the results of

the long only models, are best suited to aggressive off-shore accounts

and hedge funds.

Seth C. Warner, MSTA, is President of Warner Prince Incorporated, Litchfield, CT,

U.S.A., which has provided technically based research to institutions and

managed private client accounts since 1975.

®SuperCharts is a registered trademark of Trade Station Group, Inc.

sm Balance of Sentiment, Balance of Sentiment System, and BoSS are

servicemarks of Warner Prince, Incorporated

©Copyright, Warner Prince Incorporated, April 25, 2002

continued from page 12

Cristianini, N. & Shawe-Taylor, J. (2000). An Introduction to Support Vector

Machines: And Other Kernel-Based Learning Methods, Cambridge

University Press.Fama, E. F. (1965).“The Behavior of Stock Market Prices.”

Journal of Business 38:34-105.

Figuerola-Ferretti I. & Gilbert C.L. (2000). ‘Has Futures Trading Affected

the Volatility of Aluminium Transactions Prices?’ AEA International

Conference on Industrial Econometrics, Luxembourg City, July 5,6 and 7,

2000.

Gilbert C.L. (1996). ‘Manipulation of Metals Futures: Lessons from

Sumitomo.’ CEPR Working Paper.

Harland Z. (2000a). ‘Using Nonlinear Neurogenetic Models with Profit

Related Objective Functions to Trade the US T-bond Future.’ In

Computational Finance 1999 (Proceedings of the Sixth International

Conference on Computational Finance). Leonard N. Stern School of

Business, New York University, January 6-8, 1999. Edited by Yaser S.

Abu-Mostafa, Blake LeBaron, Andrew W. Lo, and Andreas S. Weigend,

327 – 342, MIT Press, Cambridge, MA.

Harland Z. (2000b). ‘Trading a 2 Year Old – the Real-Time Performance of a

Neurogenetic T-bond Futures Trading System.’ The Market Technician,

Journal of the Society of Technical Analysts. 38, 14-15

Joachims, T. (1997). ‘Text Categorization with Support Vector Machines.’

Technical Report, LS VIII Number 23, University of Dortmund.

Martinot, N., Lesourd J., Morard B. (2000). ‘On the Information Content of

Futures Prices – Application to LME Nonferrous Metals Futures.’ AEA

International Conference on Industrial Econometrics, Luxembourg City,

July 5,6 and 7, 2000.

Mukherjee, S., Tamayo, P., Slonim, D., Verri, A., Golub, T., Mesirov, J.P. & Poggio,

T. (1999). ‘Support Vector Machine Classification of Microarray Data.’ AI

Memo 1677, Massachusetts Institute of Technology.

Smola, A. & Scholkopf, B. (1998). ‘A Tutorial on Support Vector Regression.’

NeuroCOLT2 Technical Report Series, NC2-TR-1998-030.

Trafalis, T.B. & Ince, H. (2000). ‘Support Vector Machine for Regression and

Applications to Financial Forecasting.” Proceedings of the IEEE-INNS-

ENNS International Joint Conference on Neural Networks (IJCNN’00).

Vapnik, V. (1998). Statistical Learning Theory. John Wiley & Sons.

Vapnik, V. (1995). The Nature of Statistical Learning theory. Spring-Verlag,

New York.

Vapnik, V.N. & A.Y. Chervonenkis (1971). ‘On the Uniform Convergence of

Relative Frequencies of Events to their Probabilities.’ Theory of

Probability and its Applications 16(2), pp. 264ó281.

Vapnik, V. & Chervonenkis, A. (1964). ‘A Note on One Class of Perceptrons.’

Automation and Remote Control, 25.

Vapnik,V. & Lerner, A. (1963).‘Pattern Recognition Using Generalized Portrait

Method.’ Automation and Remote Control, 24.

Zhang,Y. (1999).‘Toward a Theory of Marginally Efficient Markets.’ Physica A,

269, 30-44.

Table 2 Continued

Date Action Price Entry P/L Cumulative

Profit

05/18/01 Sell 48.04 6.79 52.10

05/21/01 Buy 51.05

05/30/01 Sell 44.43 -6.62 45.48

05/31/01 Buy 44.73

06/08/01 Sell 47.35 2.62 48.10

08/01/01 Buy 43.10

08/06/01 Sell 42.50 -.60 47.50

10/08/01 Buy 31.86

10/30/01 Sell 33.38 1.52 49.02

10/31/01 Buy 33.90

12/13/01 Sell 39.76 5.86 54.88

12/14/01 Buy 40.11

12/20/01 Sell 38.79 -1.32 53.56

12/27/01 Buy 40.01

01/16.02 Sell 38.78 -1.23 52.33

03/06/02 Buy 37.60

03/20/02 Sell 36.06 -1.54 50.79

03/21/02 Buy 37.02

03/22/02 Sell 36.64 -.38 50.41

Table 3

BoSS System on QQQ – NASDAQ 100 Trust 06/04/1999 – 04/24/2002

Gross Profits 75.13

Gross Losses -24.72

Net Profit 50.41

Total # Trades 22

Number of Profitable Trades 10

Number of Losing Trades 12

Percent Profitable 45%

Average Profitable Trade 7.51

Average Losing Trades -2.06

Ratio avg. profit/losing trade 3.65 to 1

Average # days profitable trades held 24

Average # days losing trades held 8

MARKET TECHNICIAN Issue 44 – July 200216

After falling steadily for 25 months the IRC Coppock Indicator has turned up

and given a major buy signal for UK equities. The last buy signal was in 1995

following which the FTSE All-Share Index rose over 75% to its peak in 1998.

How much weight should we give to this signal? Given its track record, a

great deal.

So what is the Coppock and how does it work? The Coppock indicator was

conceived by a Texan, Edwin Coppock, who, the story goes, was approached

by the Episcopalian church and asked to devise a system which would

disregard short term fluctuations in the market and provide long term buy

signals. The church funds could then be invested just when the

stockmarket was due for a significant rise.

Coppock came up with an indicator which uses a weighted momentum

calculation to establish major turning points in markets. It was apparently

based upon the amount of time it takes to get over a bereavement,which was

judged to be between 11 to 14 months. After this, optimism could return.

Edwin Coppock founded the Trendex Research Organisation in San Antonio

where they published his timing indicators. Sadly we understand that

Trendex stopped publishing in the 1980s and Coppock himself died in 1989

aged 83. However, his indicator lives on still giving signals, and has

remained one of the most reliable long term buy indicators for both Wall

Street and the FTSE All-Share.

The IRC Coppock is based on the weighted percentage changes of the FTSE

All-Share Index monthly average over a twelve month period. The

weighting ensures that progressively more emphasis is placed on the most

recent months. Buy signals are given when the indicator turns up after

falling below the mean line (in this case 1000). Sell signals for the Coppock,

i.e. when the indicator turns down, are not so reliable and should therefore

be ignored, however, most major bear markets, including the recent one,

have been signalled by this indicator.

Tracing the Coppock back on the Dow Jones Industrial Average to 1901 on

our copies of the Trendex charts, reveals 30 buy signals during the past 100

years (assuming 3 additional buy signals since 1988 when our records from

Trendex ceased).

Of the 30 buy signals, only three failed to work and produce a subsequent

rise on the Dow, being 1901, 1914 and 1941. Notably the last two were

when there was a world war. A buy signal in 1947 produced only a minor

advance and a sideways trading range followed until another buy signal in

1949 resulted in the usual bull market run. Also in 1911 and 1938 as war

approached there was only a small rally. All the other 24 signals were

subsequently followed by strong advances in the next year or two.

This outstanding track record has also been seen on the FTSE All-Share

Index when the IRC Coppock gives a buy signal. Using a strict rule of

calculation, six of the previous eight buy signals on the IRC Coppock since

1965 produced a rise of not less than 16% on the monthly average of the

All-Share Index before the Coppock turned down again. 1965 and 1991

were exceptions, but interestingly the latter of these was also at the time of

war – the Gulf War.

On many occasions,of course, the index eventually rose significantly higher,as

between 1995 and 1998 and we have not included 1980 and 1992 when the

indicator turned up as it touched the mean line (rather than turning up from

below it). Including these in the calculations gives an average rise of 32%

on the All-Share monthly average

following an IRC Coppock buy signal.

One of the interesting features of

the recent buy signal, given in

March 2002, is that the indicator

had fallen to its lowest level since

1975 and therefore a strong rally

could be anticipated from this

oversold position.

Over the years the IRC Coppock has

been a reliable indicator for calling

the UK equity market higher.

Since 1965 it has signalled every

significant bull market move. As in

the past, there is no guarantee that

the current signal will subsequently

result in another major bull market

but, given its excellent track record,

when a buy signal is given on the

Coppock, it is worth sitting up and

taking note.

Richard Marshall,Senior Technical Analyst

Investment Research of Cambridge

The Coppock turns upBy Richard Marshall

Buy Signal Date

IRC COPPOCK INDICATORBuy Signal Analysis

October 1965July 1967June 1970

February 1975April 1977

November 1988March 1991May 1995

March 2002Equivalent on

FTSE 100 Index

104.4106.6125.7125.3179.8933.5

1193.01632.5

2557.4

5271.8

109.5166.8224.2167.8223.1

1204.01266.21889.7

(3350.2)

(6906.0)

+5%+56%+79%+34%+24%+29%+6%

+16%(+31 average

rise)

10 months15 months23 months13 months10 months15 months11 months13 months

FTA All-ShareIndex at beginning

of the followingmonth

Subsequent highof FTA All-Share

monthly averagebefore next sell

signal

% riseTime until next

“Sell” signal