Timing Solution 2008

Timing Solution Table of Contents

•

Timing Solution - General Idea Definitions

•

Interface

•

Easy Start

•

Spectrum Module - Forecast based on fixed cycles

•

Composite Module - Forecast based on astronomical cycles

•

Obect Oriented !eural !et"or# $OO!!% - uni&ersal Forecasting system

•

Obect Oriented !eural !et"or# - Inputs $"'at forecasting is based on(((%

•

Obect Oriented !eural !et"or# - Outputs $"'at "e are forecasting(((%

•

Timing Solution )or#s'eet

•

Timing Solution Styles

•

*ac# Testing

•

Turning Tu rning +oints ,nalyer

•

)or# *oo# • •

Composite Module Examples *ac# Testing Testing Examples

Miscellaneous • •

Description of t'e Models in t'e pac#age Detrended .igag Index - t'e best indicator to re&eal turning points

Trusted by over 1 million members

Try Scribd FREE for 30 days to access over 125 million titles without ads or interruptions! Start Free Trial Cancel Anytime.









































•

Some 'ints to create models for intraday trading

*oo#mar#s •

Training Regime

Timing Solution - General Idea o o

*asic Idea Timing Solution and Tec'nical Tec'nical ,nalysis

*asic Idea

Timing Solution is Solution is the program designed to get a reliable projection line for any consistent set of data. o

o

The program creates projection lines for different financial instruments. The project projection ion lines lines are based based on models models and built by the specialized Neural Net. The models are based on different ideas - presumably Math methods and some other techniques. We wor only with the models that ha!e pro!ed their usefulness. The program uses many parameters. The default settings show the best !alues for parameters recei!ed in the process of e"tensi!e #ac Testing. This process $i.e. #ac Testing Testing of the models% is our o ur wor-in-progress. We do constantly the search for best &default& !alues. Thus' it is quite possible that some discrepancy may occur between the default parameters' written documen documentat tation ion and the progra program m itself itself.. Watch for the upgrad upgrades es - each upgr upgrad adee of the the prog progra ram m is cumu cumula lati ti!e !e and and cont contai ains ns the the most most recen recentt









































o

o

the correlation coefficient between the price data and the projection line and !ariations of this method $lie ne"t day price changes%. The best practical results are achie!ed when se!eral projection lines are conside considered red $such $such as the project projection ion lines lines obtain obtained ed through through *pectru *pectrum' m' (stronomical model and the model based on price bar patterns+,apanese andlestics method%. We constantly constantly do the #ac Testing of forecasting models trying to find the best models applicable to different financial instruments. instruments.

Timing Solution and Tec'nical Tec'nical ,nalysis

Timing Solution is not an illustration or application of Technical (nalysis (nalysis ideas. n brief' the main goal of Technical Technical (nalysis $T(% $T(% is to build the effecti!e trading strategy that optimizes profit+ris. The final result of this system is the accurately calculated entr entry y+e"i +e"itt sign signal alss $Rob $Rober ertt /ard /ardo' o' ,ohn ,ohn Wiley0 ley0*o *ons ns'' nc. nc. Design, Testing Testing and Optimization of Trading Systems' Systems' N1 2334%. Mostly' these systems are based on different price indicators $including !olume and open interest%. The main idea of Timing Solution is to get the projection line based on different e!ents that actually tae place in our world. We can tae any e!ent any e!ent for the analysis' and the price itself might be considered as the e!ent as well $for e"ample' in the auto regression model%. The 5bject 5riented Neural Networ models the price mo!ement in respect to the e!ents that we choose. n this case' it is more important to get the fitness between the price and the projection line6 thus' we need a correlation coefficient to e!aluate the fitness and then #ac Testing to get the e!idence of the model7s effecti!eness. *ometi *ometimes mes users users as me why Timing Solution does not generate buy+sell signals and does not estimate the model7s performance this way. To answer this question' let us consider the simplest Technical (nalysis (nalysis trading system) the buy+sell signals generated by intersection of the two mo!ing a!erages - short and long ones. The profit is the natural measure to estimate the performance of this system. We cannot apply other criteria $lie correlation% correlation% for this system' system' because they are simply not applicable applicable here. This approach does not allow to &seeing& the future' all it does is just a statement that the current price configuration indicates the possible turning point. The The ques questi tion on is) is) what what is the the real real fore forecas casti ting ng abili ability ty of trad tradin ing g syst system emss base based d on indicators8  thin the best answer can be found in this boo) Robert W. olby &The 9ncyclop 9ncyclopedi ediaa of Techni echnical cal Maret Maret ndica ndicator tors&' s&' Mc :row-; :row-;il ill' l' where where the tradin trading g strategies strategies based on 24< indicators indicators are analyzed. analyzed. The detailed detailed research conducted conducted by Raden Research :roup $/.5. #o" 2=>3' Madison *quare *tation' New 1or' N1 2>2?3% shows that the forecast based on these indicators pro!ides usually >@-2>@ $if we use >@-2>>@ scale%' and it is mostly closer to zero. Moreo!er' if we analyze the short term forecast with >-4> days horizon' the forecast ability is no more than A@ $see Robert W.









































is the natural measure for this tas' it shows the closeness of these cur!es. To estimate the performance of these models' we use intensi!ely the #ac Testing Testing procedure. ;ere is a piece from a real #ac Testing Testing report) Model

The #est /arameters

*econd #est /arameters

!eural !et $/0 Neural Net $24 'idden% hidden% train last /111 bars train last 2>>> bars C(MD2?D/osD(spD:eo.hyp forecast on 21 bars forecast on 2>> bars ,&( (!. Correlation31(/224 orrelationE>.2AA? $FAB+-2<% $5607-/8%

C(MD2>D/osD:eo.;1/

Ginear Model train last 2>>> bars forecast on H>> bars (!. orrelationE>.2?4? $FBH+-?%

Ginear Model train last 2>>> bars forecast on A>> bars (!. orrelationE>.2BA< $FA?+-2H%

C(MD2?D/osD:eo.hyp

Neural Net $24 hidden% train last 2>>> bars forecast on H>> bars (!. orrelationE>.>=2> $FA>+-42%

Neural Net $24 hidden% train last 2>>> bars forecast on H>> bars (!. orrelationE>.><=< $FA2+-4>%

t shows that one of the models $mared by red color% pro!ides the a!erage correlation >.2??= for ?> price bars ahead. t means that this model e"plains 2?@ $>@-2>>@ scale% of price mo!ements on ?> price bars ahead. The following diagram is the typical output of Timing *olution program. Goo at it. ;ere you can see three possible scenarios of future price mo!ements for some security)









































1ou can locate zones with a strong probability of the trend change for this stoc. What is also good' you are able to now them a way ahead. f this information is confirmed by Technical (nalysis methods or by some other fundamental factors' this is a serious sign for the trader. Thus' Timing *olution gi!es you hints on future price mo!ements. f these hints are confirmed by other methods' you may apply them as a part p art of your trading strategy.

Timing Solution - Definitions o

o

Training and Testing inter&als( 9earning *order Cursor $9*C% and ,*C notation( *ac# Testing Concept $'o" to a&oid :information lea#s:%

o

Final Optimiation

o

)'at is t'e Correlation Coefficient;









































#lue inter!al inter!al is Trainin Training g nter!al' nter!al' Red Red inter!al inter!al is Testing nter!al. The border between them is Gearning #order ursor - G#. 1ou 1ou can set the G# where!er you want. To do this' just clic on this icon) clic the left mouse button at the place where you want to set G#. (lso you can set G# clicing on this button) to set G# on any price point manually)

and then

. n this case' you will get the window











































*ac# Testing Concept $'o" to a&oid :information lea#s:%

When you create any forecasting model' the program taes the points from T<,I!I!G inter!al only. t does not use the price points from TESTI!G inter!al' these points ser!e to estimate the modelIs performance only. Goo at this e"ample)

This is one of the models to forecast Jow ,ones ndustrial $J,% inde". This model is based on fi"ed cycles plus Neural Net. The *lac# line line is the oscillator for J,' the red line is the forecasting cur!e calculated by the Neural Net. t shows good fitness on the blue $training% inter!al. This is not surprising' because the program uses these points to train our Neural Net. The fitness fitness on red $testing% inter!al of the e"ample is not too good. ;ere the Neural Net wors in the forecasting regime. The price points on the testing inter!al are necessary to









































Now the Main window loos)









































the correlation between projection line and price calculated on T9*TN: $% inter!al is >.24?3.

Final Optimiation

Cinal optimization is when you use ,99 a!ailable a!ailable price points to mae the forecastin forecasting g model. When we do this' there is no price points left that can ser!e as a testing inter!al to do #ac Testing' because now the program uses all of the price points to create the final model. Thus' final optimization is recommended only when you trust your model. To mae the final optimization' mo!e the G# to the last a!ailable price bar $by clicing on this icon %. ;ere is how the forecasting model loos lie)









































What is the Correl Co rrelation ation Coefficient? The correlation correlation coefficient concept from from statistics statistics is a measure measure of how well trends trends in the predicted values follow trends in the actual values in the past. t is a measure of how well the predicted values from a forecast model !fit! with the real"life data. The correlation correlation coefficien coefficientt is a num#er #etween $ and %. f there is no relationsh relationship ip #etween the predicted values and the actual values the correlation coefficient is $ or very low &the predicted predicted values are no #etter than random num#ers'. num#ers'. (s the strength strength of the relat relation ionshi ship p #etwee #etween n the predic predicted ted values values and actual actual values values incre increase asess so does the correlati correlation on coefficient. coefficient. ( perfect perfect fit gives a coefficient coefficient of %.$. Thus the higher higher the correlation coefficient the #etter. #etter.

Cor practical usage' you should now that) 2 - Means ideal coincidence between some data. > - No correlation. Two sets of data are not related. -2 - This is anti-correlation' which means that the predicted !alues &mirror& the actual !alues $or one data set is the &mirror& for another one%. These are e"amples) /ositi!e correlation $E>.?%6 these two cur!ed lines show the same price mo!ement $most of the time%. n other words' price goes up or down for both lines)









































Negati!e correlation $E->.B%6 we obser!e the &mirror& effect $when one cur!ed line goes go es up' the other one goes down in most cases' and !ice !ersa%)

What correlation is good enough8 The more the better. sually' the models that we analyze pro!ide >.2->.4 correlation. *ometimes it is more than that' but these results are not stable. To be sure that this result is not accidental' it is necessary to ha!e a sufficient amount of price bars for calculating the correlation.











































Timing Solution - Interface

o

Timing Solution Interface at a First Glance Timing Solution Interface at t'e Second Glance - Do"nloading t'e price 'istory data

o

+rice C'art Manipulations

o

Creating Indicators

o

Timing Solution Interface at a First Glance

When you run the Timing Solution program' you will get this window)











































- These buttons allow downloading price data' adding one or more new price bar+bars' sa!e+open your wor' setting the analyzed time frame.

- This set of buttons helps to manipulate the price chart' i.e. magnify any an y area of the chart' shift it' and shift+reset the Gearning #oarder ursor $G#%. - These buttons ser!e to create the indicators $lie M(J' R*' and 5scillators% on the chart. This button)

acti!ates the &Main Window Window Liew&) Liew&)











































To downlo download ad the the price price histo history ry data' data' cli clic c on this this butt button on displayed)

. This This wind window ow will will be











































This information is necessary if you use the model based on astronomical cycles. n this case' it is necessary to synchronize the trade time with the real time.











































Now we are almost ready for the analysis. lic on &Calculate & button to produce the necessary calculations before the analysis)

Now all ready *olutions of the program are a!ailable for you. They are combined into some some grou groups ps P modu module less $see see Modu Module less desc descri ript ptio ions ns in the the prop proper er part partss of this this Jocumentation%)













































training and learning purposes while the data to the right will be used for testing the models. 5ther buttons buttons related related to G# are) . The first first button button ser!es to shift shift G# on one price bar ahead. The second button shifts G# on one price ahead and it shift shiftss the !iewed !iewed inter! inter!al al at the same same time. time. The third third button button allows allows to set G# position manually' while the last button sets the G# on the last price bar. This option is especially !aluable when you do the final optimiation $not the regular training and *ac# Testing of the models%6 in other words' when you use ,99 a!ailable price bars to produce the projection line. Magnifying =ie"ed Inter&al To magnify magnify any inter!al' inter!al' clic on this button button and drag the mouse cursor. cursor. The other way to acti!ate acti!ate this button button is by clicing clicing and then then pressing the rig't mouse button and dragging the cursor6 thus you select the time inter!al













































































































































































































































































4% Daily Intraday data - we ha!e intraday data stream during the day. Get say it is measured by ? min tics. We will get for just one day so many ? minutes price ba rs. n other words' this type of data allows obser!ing the intraday dynamics of the price change' and this dynamics is limited by trading hours $i.e.' from 3)A> am to B)>> pm for N1*9%.



























































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































:=ie": panel A This panel allows specifying indices displayed in the Main window)

(s an e"ample' let7s create the composite for lose inde". The green line in the Main window corresponds to the projection line for the chosen astro cycle)

hecing &Target for omposite& option' you can display in the Main window the inde" that is used to calculate the omposite. Cor e"ample' if we research the (nnual $*un% cycle for lose inde"' the program normalizes the lose using relati!e price oscillator with periodE
,cti&e .ones Get us e"plain what &(cti!e Oones& are considering one simple e"ample. Jownload the crude oil price from 23=A to 4>>B. *et these options in the &(cti!e Oones& panel)









































*et &critical change& parameter to 24.=@ $this is the minimum height for a zigzag wing. n our e"ample' it corresponds to H? turning points%. n the &Terms& &Terms& panel' set the Moon-*un pair because we will analyze the Moon phases)

;ere is the composite diagram for these settings)









































points $H?% found for the changing angle between the Moon and the *un and draws lines corresponding to these angles $red lines for top points' blue lines for bottom points%. Thus' the acti!e zones show where most of the turning points are located. f there are no turning points in some area $or just a few%' the area is called &uiet Oone&. :,cti&e .ones: readingA Cor this particular e"ample' we can mae some conclusions. The crude oil price reaches its top 4-B days after the New Moon. 5n the diagram abo!e' >V corresponds to the New Moon. The &Region 2& $there are many top turning points% is located within the inter!al 2>V-?>V. Then' there is one more acti!e zone) &Region 4&. t is not so strong as the pre!ious one and corresponds to the bottom turning points. The &Region A& corresponds to the relati!ely quiet zone' it corresponds to 4A>V and happens B? days after the Cull Moon lasting for 4-A days. n the quiet zone' the price does not reach the turning points' but just follows the e"isting trend. :,cti&e .ones: criterionA We can do this analysis for the turning points and for the points when big price mo!ements happen happ en as well. n this case' set the parameters of the &(cti!e Oones& as follows)

Cor these settings' the program will loo for all days when price changes at least for B@ $within a day range of the ;G5 price bar6 otherwise' it compares today7s price and the price one day before%. The minimum change is B@. The diagram follows $this is the annual cycle' or *un-*un' composite%)

The region between Gibra and (ries is mared by the yellow bar. This time period corresponds appro"imately to the inter!al from the end of *eptember to the end of March. We can see that strong mo!e points are located more often in this region. n other words' from 5ctober to March' the crude oil price maes more strong mo!ements $may be because it is a cold period in the Northern hemisphere...%. analyzee the long-ter long-term m data data series series'' it might might be :,cti&e :,c ti&e .ones: .ones: 'ist 'istogr ogramA amA f we analyz impossible to see all related lines. Goo at this diagram6 these are turning points for Jow,ones inde" from 23>> to 4>>B' minimal change B@)









































We cannot see it clearly because there are 3H< turning points. #ut we can display this information as a histogram and loo at the density of the turning points. *et options as follows)

Cor these settings' the (cti!e Oones ;istogram for the annual cycle $*un-*un% will loo lie this)

The brighter red zones correspond correspond to the areas where turning points happen more often. often. Cor e"ample' such a zone corresponds to the period starting from the end of *eptember till the beginning of No!ember. n other words' the !ery strong possibility e"ists that there might be trend changes at this period' especially for bottom turning points. *ome useful statistics options for this feature can be found in the &5ptions& tab) :o to the &*tatistics& page)









































e"act' it means) &With probability =3@' we can assume that turning points happen more often from the end of *eptember to the beginning of No!ember&. n the &*tatistics& window' we recommend you to eep the parameters for minimal sample size and the amount of the control groups as they are and change the &ritical hi *q& option only. o nly.

+redictable .ones (s mentioned before' the main idea behind the predictable zones is to calculate the composite for a few independent inter!als and see how the projection line is changing in respect to each inter!al. Cor this program' we ha!e applied and de!eloped special math methods that will calculate these zones !ery fast. t is this &now-how& of ours that maes this this progra program m so ad!ance ad!anced. d. #efore #efore any discus discussio sion' n' some some useful useful option optionss should should be e"plained)

Split on x inter&als option allows specifying the number of independent inter!als to use. The greater greater amount of independent independent inter!als' inter!als' the more narrow narrow predictable predictable inter!als inter!als are $howe!er' these inter!als are more reliable%. (djust this parameter according to the data length for your analysis. CalculateA if you do not need to use predicable zones' you can disable this option. The program will then calculate the composites significantly faster. faster. +roect +roectionA ionA if this this option option is acti!a acti!ated ted'' the program program uses uses predic predictabl tablee zones zones to mae mae projection cur!es. (s an e"ample' here is the projection cur!e based on Moon phases cycle for crude oil)

;owe!er' as we ha!e mentioned abo!e' the Moon phases gi!e the best forecast around the Cull Moon. hec the &/rojection& option' and you will get this picture)











































This diagram shows a part of the projection cur!e around the Cull Moon. 5ther points are simply sipped. especially for the omposite omposite #o" $when we combine many Be careful using this option' especially cycles%. t is !ery difficult to distinguish between the true price gap and the area within the end of the pre!ious predictable zone and the beginning of the ne"t one. The predictable zones are displayed in the bottom part of the diagram) #right red regions correspond to predictable zones' gray ones - to unpredictable. ;ere you can specify what ind of the diagram you would lie to see. see. &*plit &*plit&& allows allows displa displayi ying ng of the compos composite ite diagra diagrams ms calcul calculate ated d on differ different ent independent inter!als. &*ummary& corresponds to the composite diagram $colored one% calculated for all price points of the optimizing $red% inter!al. #y the way' when the predictable zone is significant' the program displays this)

Composite *oxA Forecast based on Composites (fter finishing the analysis of different planetary pairs' you can put them in a special place called the &omposite #o"&. #o" &. t is easy to wor with the omposite #o". #o ". Get7s say that' while analyzing some set of data' we ha!e found that the annual cycle is important6









































;ere the blac line is the summary cur!e - it shows the summary influence of two cycles) the annual cycle and the Moon phase cycle. We can magnify any part of this diagram dragging the mouse. t is also possible to sa!e this model into a file and download it in the future. The program allows allows to manipulate by composite pairs) delete any term' delete se!eral uncheced terms' clear model $i.e.' delete all terms%' rate these terms by weights' download+sa!e the model' and put this model into the clipboard $&-Xl& button%. When the model is sa!ed into the clipboard' it is sa!ed in G9 format $the format of ni!ersal Ganguage of 9!ents%. (fter that' the model is a!ailable for the process of creating a forecast based on Neural Net $NN% technology. To create the inputs for NN' clic on this button) $the program will tae the model out of the clipboard%. n this window' you can see how the projection p rojection cur!e $a blac line% fits the analyzed price inde")









































1ou can specify any inde" to be used to calculate the correlation $to do it' clic on the &9dit& button%. /lease pay attention to the time inter!al used to calculate correlation $see details in the section &Definitions & of the documentation%.

Composite








































;ere' in this window' define parameters of the report.








































can use this model as inputs for the Neural Net forecasting model. The technology of creating such a forecast is described in &Timing *olutions& module. $C(M model stands for Cloating (ngle Model6 it is the planetary cycle model con!erted into G9 format%. These are parameters for the +arameters of F,M model C(M model. Cirst of all' a few words about what actually the C(M model is. ts purpose is to e"pl e"plor oree the the acti acti!e !e point pointss of the the Oodi Oodiac ac.. The The &5rb &5rb&& para parame mete terr indi indica cate tess the the thoroughness of this e"ploration. n other words' if there is+are any acti!e point+s within the orb7s limits' the program will use this area as one input for the Neural Net. Thus' if the program chooses = composites as the important ones' the number of the NN inputs is more than ?>> $??>' to be e"act6 it depends on &5rb& and C !alues%. The bigger 5rb maes this model7s diagram smoother' but some details might be sipped $when it is too smooth%. The C parameter defines the ability of this model to see details) the smaller C' the more details this model can see. ;owe!er' the Neural Net will contain more inputs for smaller C' and it can cause the o!er-training effect. *o' it is necessary to find a balance nce between the resolu olution of this model and o!erer-training. One practical practical recommend recommendation ation for creating creating )(* model ) f' while creating the C(M model' you get too many e!ents for the Neural Net $lie 4.>>> inputs%' the Neural Net will be training too long. n this case' we recommend increasing the Orb parameter $for e"ample' mae it 2? degrees%. d egrees%. (lso (lso you can try different Filter criteria $see below%. Filters OptionsA Filters OptionsA This is a !ery important option that allows defining the criterion in regards to the composite cycle $we can specify this cycle as an important or nonimportant%. :Min !umber of Cycles:A ;ere you can specify what composites you tae into account. t shows the minimum number of cycles to be analyzed to create a composite' regarding the analyzed data. Cor e"ample' we analyze the annual cycle6 this period should be at least A years' otherwise we ha!e not no t enough information to re!eal this cycle. :Filter: There are se!eral criteria to estimate the degree of importance for each cycle. These criter criteria ia are define defined d by Cilter Cilter settin setting. g. The progra program m will will choose choose the compos composite itess that that correspond to the filter. The The illustration shows three of them. 2% The first one is used as the default option).











































oscill oscillato ator% r% and the projec projectio tion n line line pro!id pro!ided ed by this this compos composite ite is calcul calculate ated d on this this inter!al) 4% This filter is based on another criterion)

This filter allows us to choose the composites that are produced within (cti!e Oones $(O% only. n other words' we can choose $for e"ample% the composites that are related to the turning points' and acti!e zones will show some areas where the turning points happen more often. We recommend using these criteria if you create the model to predict turning points. *ee the e"planation for (cti!e Oones $(O%. A% 5ne more filter)

nder these criteria' the program will choose only predictable composites' i.e. those composites that pro!ide the same diagram on different time inter!als. *ee the e"planation for /redictable Oones $/O%. n other words' this type of composites means that the











































(s an e"ample' the program has found < important cycles. This is for crude oil data' 23=A-4>>B. This feature of the program sa!es you a lot of time.











































Obect Oriented !eural !et"or# $OO!!% as a uni&ersal Forecasting system

o

*asic Idea !eural !et at a First GlanceA t'e Japanese Candlestic# Model $+rice *ar +atterns !eural !et at t'e Second Glance Going into Dept'

o

=ie" Options

o

)'en to Stop t'e Training +rocess

o o

o

o o

Training
*asic Idea :enerally speaing' the Timing *olution program is based on two concepts. The first one is the idea of G9' or ni!ersal Ganguage of 9!ents. (nother one will be discussed now. t is the 5bject 5riented Neural Net $or 55NN%. #oth concepts were designed to wor











































Net% does its job and maes its conclusions. #ased on these conclusions' we can produce the forecasting model and project it into the future. This simple scheme allows us to sol!e a wide !ariety of different forecasting problems. !eural !et at a First GlanceA t'e Japanese Candlestic# Model Get us create the simplest forecasting model with the help of Neural Net. Get this model be based on the ,apanese andlestic technique $or price bar patterns%. t is quite simple to do that. Jownload price data of your choice $ did it for crude oil prices' from 233? to May 4>>B%)











































The testing inter!al is the first attempt to e!aluate the forecasting ability of our model. 5ur e"perience has shown that to find a model that fits !ery well within the learning inter!al is not really hard. (t this stage' it is mostly due to the sills and professionalism of software designers. $t e"plains why so many products on the maret are so good at e"plaining e"plaining the past...% past...% The testing testing inter!al inter!al demonstrates demonstrates how this model actually actually wors. wors. 5nly if it fits well on the testing inter!al' can we apply this model to forecasting. (ctually' it is possible to apply the more complicated scheme and di!ide the price history in three inter!als) 2% training inter!al - to train Neural Net6 4% !alidating inter!al - to stop training when we ha!e a good fitness between the price and the projection line on this inter!al6 A% testing inter!al - to test the model7s performance. #ut we use this scheme !ery seldom as we ha!e other methods to e!aluate the model7s effecti!eness. There is one more feature in the program that reduces the possibility of mistaes. This featur featuree is called called &#ac &#ac Testin esting g /roced /rocedure ure&' &' and it allows allows you to !erify !erify the model7 model7ss performance on a!ailable data history. history. 5nly after that testing'  would recommend to apply the model and its projection line to the real trading. Get us return to the problem that we ha!e decided to sol!e6 i.e.' let us create the forecasting model based on ,apanese andlestics technique. The main idea we will wor with is that the price bar structure now has some impact on a future price. /lease loo at this picture)













































is the reason why we can use them to mae a forecast. t is possible to do it because they can be recorded by means of ni!ersal Ganguage of 9!ents $G9%. !eural !eu ral !et Out Output putss

A The future price price The endi ending ng of this sentenc sentencee $... $... the future

moement % bel belongs to Outputs of the the Neur Neural al Net. Net. We mar mar it as $som $somee conclusion goes out of the system' based on the analysis and testing pro!ided by the Neural Net%. This is what we would lie to predict6 this this is our final target. target. Therefore' we ha!e here two types of things) 2% things that go into the system $they always belong to the outer world' are pro!ided by the outer world' and possess some information in respect to this outer world%6 and 4% things that are going out of the system $these are conclusions made by the Neural Net regarding to the processes in the outer world%. The Neural Networ is a uni!ersal tool $ would say - an incredible tool% that allows the user to find relationships between inputs and outputs6 it is the middle part of that same sentence) $&The price bar $which represents the relationships between the ;igh'













































lic on the button that corresponds to &5utputs&)

( 1ou will get this window)



















































































































































































results' i.e. forecast% lines correspond well - as our purpose was to find &future price mo!ements&6 the change of the direction coincides well. This particular model produces a forecast only for a wee ahead' and this is its ma"imum. t is because we choose to create such a model - we set this parameter as) *o th

e















































2% lic on this button) button) . The projection projection line produced produced by this Neural Net model will appear in the Main Window Window of the program. 4% lic lic  on this this butto button' n' Window.

' to disa disabl blee the the Resu Result ltss /ane /anell and and mae mae !isib !isible le the the Main Main















































1ou can create as many Neural Net models as you need. The new panels will be added to the Main *creen. (s an e"ample  ha!e created the *pectrum model that uses fi"ed cycles to mae forecast. Now there are two Neural Net panels in the Main *creen)















































!eural !et at t'e Second Glance When a Neural Net model is created and its performance is of an acceptable standard' you can sa!e this model. lic on this button)































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































t seems that this pair has the ?>>d-4 years stochastic cycle. This fact has been pro!ided by one of Timing Solution users who obtained this result using #ac Testing module of the program.









































Timing Solution - *ac# Testing This module is for adanced users, so we recommend stud$ing it onl$ if $ou are comfortable with the !eural !et module

o o

o

*ac# Testing at a First GlanceA Main +rinciple *ac# Testing at t'e Second GlanceA Example *ac# TestingA Different =a =ariances riances

*ac# Testing at a First GlanceA Main +rinciple

#ac Testing allows you to) 9stimate the performance of any Neural Net for any particular financial o instrument6 Cind the best Neural Net model for any particular financial instrument6 o Cind the best training inter!al for any model6 o Cind the best forecasting inter!al for any model. o This is a !ery hard tas to do manually. The #ac Testing module will mae your life easier. Get us show you how Timing Solution can help you with the first first tas $i.e.' $i.e.' estimatin estimating g the model performance%. The scheme loos lie this)









































;ere is the illustration for one typical e"ample) *tep 2) The program trains the Neural Net using the price points from the yellow region $see the upper picture%. (fter training' it checs the Neural Net performance using the points from the red region. n other words' the program creates the Neural Net projection line line usin using g the the point pointss from from the the yell yellow ow inte inter! r!al al and then then it obse obser! r!es es how how well well this this projection line fits to the real price data from the red inter!al. These two inter!als $yellow and red% are independent' so any &future leas& or Kinformation leas are e"cluded. *tep *tep 4) We shift shift the Gearning Gearning #order #order ursor ursor $G#% $G#% for se!eral se!eral price price bars bars ahead. ahead. Therefore' we shift the yellow and red regions as well and repeat the whole procedure again $optimization and e!aluation of performance%.









































;ere G# will be shifted ?> times for 2>> price bars ahead. n this case' the program will estimate the model7s performance for ?> different inter!als. The results of the #ac Testing Testing procedure loo lie this) ModeA !eural !et +rice E&entsA
Model

dynamicDmodel.hpp

NN Topology Topology

A4 hidden

NN Training Training

last 24?>6 2?>>> st.

*tatistics

FA? + -2? h*qEB.>

(!erage

>.2BH

42.><.23<= >3)A>

->.2=?

43.>2.23<3 >3)A>

->.>>?

>H.>=.23<3 >3)A>

>.2?A

24.>4.23=> >3)A>

>.A=?

4>.>=.23=> >3)A>

>.><2

....

...

This is the sample report of the #ac Testing procedure we ha!e actually pro!ided. t means that the program has trained the Neural Net ?> times. Cor each training inter!al and G# position' it compared the projection line and the price $to be precise' the price oscillator% and calculated the correlation coefficient between them using AH? price points after after G#. A? times the correlat correlation ion happened happened to be positi!e positi!e while while 2? times it was negati!e. The a!erage correlation is >.2BH. The e"planation regarding the correlation coefficient read in t'e :Definitions: c'apter . t means that this model can be accepted as a woring one. #esides the correlation' there are other criteria to estimate the model7s performance) the direction of price mo!ement for the ne"t day and the direction for price mo!ement for @" days ahead.









































*ac# Testing at t'e Second Glance - Example

Now we will demonstrate the #ac Testing Testing procedure itself using one e"ample. Cirst of all' download the price history file. The amount of price data should be enough to produce the #ac Testing. Testing. n this e"ample' we use Jow ,ones inde" from 23B> to No!ember 4>>B. Run the Neural Net module' clic button and choose . *ac# Testing Scenario Script( 1ou will get the window that allows you to create the ;ere it is)

Goo closely at this window. The following describes the components) *ac# Testing Scenario - Outputs for !eural !et $"'at "e "ant to forecast%A t corresponds to this part of the window)











































lic on module)

button6 button6 you will go to e"actly e"actly the same window as seen in the Neural Net

t wors e"actly the same way as in the Neural Net. ;ere you can define any price indicator you want' and the Neural Net will create the projection line for this indicator. 1ou 1ou can create se!eral outputs lie this)

n this case' the Neural Net will produce the projection lines for 5scillator and R*. *ac# Testing Scenario - Models $Inputs% for t'e !eural !et $"'at t'is forecast is based on%A *et the inputs in this part of the window)









































;ere define these parameters) n the upper part of this window' choose models $a standard model or the model o created and sa!ed earlier6 see S.hyp or S.hpp files located in c)\Timing*olution\Gib directory%. (lso there are e"tremely important models that are created by the program in









































recalculate these cycles after each changing of G#. n other words' in this case we analyze not a fi"ed model' but rather a ind of algorithm &to e"tract an astro cycle' use these parameters and then use these cycles to create Neural Net projection line based on those parameters&. n this case' the in!ol!ed cycles depend on G# position' and we try to catch the strongest recent cycles. 1ou can define the parameters for *pectrum and (stro modules here)

These parameters are described in :Timing Solution Styles:(

o

The training regime. 1ou can use all price points before G#' or you can use a part of them $closer to G#%. 5ur research shows that !ery often a lesser amount of price points pro!ides better results. Cor e"ample' while bac testing some model for Jow ,ones inde"' we ha!e found that training on B>>> price bars $2H years history% gi!es worse results for this model than its training on the last 2>>> price bars $B years history%. t is !ery important to find the best training inter!al for each model. *ee notes regarding this issue at the end. 1ou can also define se!eral training inter!als' lie this)

$2>>>'2?>>'4>>>'A>>>'B>>> price bars%. n this case you may find the best training inter!al for the particular particular model $do not forget to to clic on &5Y& button%.

o

*et *et

the









































are able to mae a forecast based on linearity only. #ut it loos lie the Neural Net gi!es the better results. (s models' the program uses S.;1/ and S.;// files located in \G# subdirectory. 1ou can create these files through the inputs editor $clic on &*a!e& when the files are created%)









































1ou 1ou can also define se!eral testing inter!als' lie this)

$?>'2>>'2?>'4>> and 4?> price bars after G#%. n this case you may find the best forecasting inter!al for the particular model $do not forget to clic on &5Y& button%. f you set this parameter this way' the program will calculate the correlation coefficient between the projection line and the price $or' more precisely' precisely' price indicators defined as outputs%. t uses @ " price bars after G# $2>>' in our e"ample%.

f you choose this item' the program will analyze the price mo!ement for the ne"t bar after G#. This wors well when we are more interested in catching the ne"t price mo!ement.

*etting this parameter' the program loos for the price mo!ement for @ " price bars $2>>











































Constant S'iftA in this e"ample' we constantly shift G# for 2>> price bars ahead. (nd we do it ?> times. !ormaliedA as G#' we use Kequilibrium points P points where the oscillator crosses zero)

1ou 1ou can also sa!e this scenario into the file $clic &*a!e&%)











































Now we will demonstrate other possible !ariants of #ac Testing. Testing.

Finding t'e *est model for any particular security

hoose se!eral models. Gie this)

We will will chec three models at the same time.

Finding t'e *est training inter&al for any particular model and any particular security

hoose the same model' but use different parameters of the NN)











































;ere we chec the dynamic models performance for ?> price bars after G#' also 2>> bars and 2?> bars.

Finding t'e *est training and forecasting inter&al $particular model 5 particular security%











































Optimie E&eryt'ing

(lso' you can play with different price e!ents)











































General Idea

The main idea of this module is finding the price le!els where the price mo!ement changes its trend. To do that' we create the zigzag and loo for these le!els by analyzing different proportions of zigzag swings. (ctually' (ctually' this idea is !ery close to 9lliot Wa!e Theory' but our goal is to pro!ide the uni!ersal tool that allows to re!eal hidden patterns in zigzag. Get7s consider the zigzag created for #M shares. This picture shows = months inter!al)













































#oth approaches are a!ailable in Timing Solution software' though we fa!or the second one. t gi!es you a clue to a real process. Moreo!er' you will be able to see when the classical ratios wor.

Easy Start

Get us loo how this module actually wors. (s an e"ample' we use *0/ ?>>.  ha!e downloaded *0/ ?>> inde" data from 23?> to the end of ,anuary ,anuary 4>>H. While downloading the historical data' set the &Corecast ;orizon& big enough to be able maing a forecast for a long period ahead)

























































































A% The central issue of this module is the histogram - &;igh probability zones&. n the program' the same histogram is shown as a !ertical one and a horizontal one. ,ust looing at the !ertical histogram' we can assume the price le!els where the T5/ turning points are more probable. Gooing at the details on the horizontal histogram' we can easily see four main clusters there corresponding to four different price le!els)







































































































































that trend changes occur in these cluster zones. (nd the program can calculate these zones. #y the way' checing the &Cib& option you can see how the most used in technical analysis ratios for retracement swings correspond to real price mo!ement)





























































































Going into Dept'

The Timing *olution allows analyzing zigzag from different points of !iew. These















































:Order:A in all e"amples abo!e we ha!e analyzed the ratios between neighboring swings. ;owe!er' using Timing *olution we can mae a forecast based on different ratios' lie ratio between swing D-E and *-C $picture abo!e%. #oth these swings correspond to downtrend mo!ements in our case.















































*uppose we ha!e point &2& as a last turning point $as in the picture abo!e% and we want to find the bottom turning point 4. *etting =th wa!e model' the program will analyze the ratio between swings /-07C-/ $using 9lliot notation%. #ut we need much more price history to pro!ide this analysis.

















































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































(s you see' the higher cur!ature is' the more we are focused on turning points pointed by this indicator. Now our zigzag loos this way)

/lay with this indicator as a target $output% for Neural Net module. (lso it is interesting to use this inde" for *pectrum module.  applied this e"perimental technique for C(M models for some stocs and for *pectrum models. Results are really interesting. 0% +r +rop opor ortio tion n InA This This param paramet eter er defi define ness the the ind ind of proportion we use for detrended zigzag. f we use the Julian Day option' the zigzag $with cur!at cur!atur ureE eE2% 2% loos loos as conn connect ected ed stra straig ight ht line lines' s' if we use use trade days days option' the proportion will be calculated in relati!e trade days $trade bars%.

Some 'ints to create models for f or intraday trading Timing *olution operates with three types of price data. t is !ery important to distinguish these data because the manner of calculating the projection line by Neural Net module strongly depends on it. 2% Daily data - when we ha!e one price bar per one day. The sample of such data loos lie this)









































Woring Woring with this type of data' we need n eed to e"clude weeends and holidays in &5ptions&)









































n this case' the program generates a projection line on a daily basis and sips weeends and holidays automatically $see red stripes on the time scale%.

4% Daily Intraday data - when we ha!e intraday data stream during the day. Cor e"ample' if it is measured by ? min tics' we will get for just one day many ? minutes price bars. n other words' this type of data allows obser!ing the intraday dynamics of the price change' and this dynamics is limited by trading hours $i.e.' from 3)A> am till B)>> pm%. This is the typical e"ample of such ind of data)









































Cor any day $e"cept $e"cept weeends and holidays%' holidays%' we ha!e H)A> hours bunches of price data. n this case' you should set up the time when trading starts and ends during the day)

Goo how the Neural Net will generate the projection line under these op tions)

The program will sip non-trading days and non-trading hours $these periods are mared by red stripes under unde r the time scale%. f you try to set daily options for this ind of data' you will not get the intraday projection cur!e. The program must now what ind of data it operates with. A% )ee#ly Intraday data - this ind of data corresponds to intraday data within trading hours. The trading begins $as an e"ample% at 3)A> am Monday and stops at B)>> pm Criday. Juring this fi!e day period' we ha!e continuing intraday data stream $lie for









































;ere we ha!e the data where the trading begins at A)>> am e!ery Monday and ends *aturday at <)>> am $local time% e!ery wee. Cor this particular e"ample' set these options)











































We ha!e series of continuous ? days data. f you try to use Jaily-ntraday for this ind of data' you will not get the projection line for non-trading hours $lie on Wednesday from B)>> pm to 3)>> pm%.

Description of t'e Models in t'e pac#age Spectrum Model

This is the model based on fi"ed cycles. To apply this we recommend using price history data of at least 4 years' though this model gi!es better results for 2> years data. "ea#nesss of this model is that it is based on fi"ed cycles $i.e.' these cycles are The "ea#nes supposed to be the same all the time%. ;owe!er' in reality' fi"ed cycles are not fi"ed fi"ed their their phase phase and their their period period are subjec subjects ts to change. change. $(ctual $(ctual fi"ed fi"ed cycles cycles might be









































The true picture might be shifted' stretched and squeezed. To deal with these phenomena $i.e.' cycles changing their period%' we pro!ide the waelet diagram . The lifetime of the fi"ed cycle is defined by the length of the &cycle mode& $,ohn C. 9hlers 9hlers'' *2S( and Trading Trading *aret Cycles%. Cycles%. (fter that the playing cycles change their period or disappear6 the phases of these cycles can mae crazy jumps.

Dynamic and ,stro Model

We recommend using at least B years of price history data for these models. These models are related to astronomical cycles. cycles. The difference between Jynamic and (stro models lies in the methods of describing these cycles. The worst character "ea#ness of these models is in&ersions . Goo at this e"ample)









































Japanese Candlestic#s and ,uto regression Models

These models deal with price related e!ents. We recommend using at least <>> price bars history $or A years for end of day data% for these models. #oth types of models pro!ide the forecast < price bars ahead after G#. ;ere you can see the results of the #ac Testing procedure for ,apanese andlestics model) htpp)++www. htpp)++www.timingsolution.com+T*+#T+andle?+. timingsolution.com+T*+#T+andle?+.

%mportant note Woring Woring with all these models' you should remember the e"istence of some -factor. 1ou 1ou can call T by any names you want' the fact is that any forecast can fail almost any moment - thans to this -factor. To To understand when, why and why and how this how this factor enters our li!es might be an unbelie!ably hard tas. The least we can do is to try to catch the moments when moments when it has happened in the past. This is the reason for our attention to #ac Testing Testing procedure. Today' it is the only nown way to enter this territory. territory. We will continue the #ac Testing procedure for all these models. Crom time to time' the new information will be released on our website http)++www.timingsolution.com+T*+#T+

NTiming Solution 011B-011K

9dition Cebruary 4A' 4>>H

Timing Solution 2008

Recommend Documents