> dfc7 >&bjbjUU .^77>"l
2
2EGGGGGG$ #~kk2222bE2E2*2\)^nG
TV)0,$$)2
Candlestick Model Forecast
This article describes the study on forecasting abilities of a model based on math methods of Fuzzy Logic and Neural Network. The name of this model in Timing Solution software is the candlestick model. This model is a closed one (which means that the model makes a price forecast based on the price itself, i.e. on internal factors; another type of a closed model in this program is Spectrum model). Closed models are opposed to open models that use some external factors and make a forecast comparing the price and other physical processes. A sample of open models in Timing Solution software is any astro model. External factors for the astro models are astro cycles.
To evaluate and compare the models, the correlation (and only the correlation) has been applied. It is quite possible that the correlation coefficient does not provide full and (what is more important) obvious enough information regarding the quality of the projection line. However, it is quite clear to me that the bigger correlation coefficient means the better model, and not the contrary. Also, it is important to remember that this model tries to forecast every price movement (daily for daily data and hourly for hourly data). So, if we concentrate on forecasting the direction of the change for the next price bar, the starting point is 50% probability. We take into account the general direction of the forecast and its swings as well.
When is this model useful? Many traders diversify their portfolios to diminish the risk; they do it by distributing their means among several financial instruments (or several types of financial instruments). We can apply the same approach to forecasting: we can use several forecasts with low correlation and make our trading decisions when the majority of these forecasts point at the same direction of the future price movements. Therefore, the discussed model may serve as addition tool in your trading decisionmaking.
Timing Solution software provides a ready solution for candlestick models; see the solution TS7. The correlation provided by this model is about 10% (on top of those 50%), so we can state that the model does a forecast of future price movements. However, we can do better than that, we can get better forecast. The following is a description of the tests that I have done for futures EuroFX (Euro/USD) traded at Globex, data from January 2005 to August 2006. The LBC (Learning Border Cursor) has been shifted 80 times, each time for 5 price bars. I compared the correlation coefficient for the first 5 price bars in every sample.
There is one more important thing to remember (it is based on my previous experience with other models suggested by Timing Solution): the length of the training interval. It is the important parameter. Moreover, I think that we need to apply small training intervals while making shortterm forecasts. Big training intervals are necessary for longterm forecasts. Thus, as a first step, I have tested the model candlestick 1 on different training intervals.
Here are the results of Neural Net forecast:
candlestick_1.hypcandlestick_1.hypcandlestick_1.hypcandlestick_1.hypcandlestick_1.hypcandlestick_1.hyp32 hidden32 hidden32 hidden32 hidden32 hidden32 hidden50 before LBCtrain 10000 steps100 before LBCtrain 10000 steps200 before LBCtrain 10000 steps500 before LBCtrain 10000 steps1000 before LBCtrain 10000 steps2000 before LBCtrain 10000 steps+54/27 (66.7%) ChSq=4.5+56/25 (69.1%) ChSq=5.9+54/27 (66.7%) ChSq=4.5+51/30 (63.0%) ChSq=2.7+48/33 (59.3%) ChSq=1.4+50/31 (61.7%) ChSq=2.2r=0.234 dev=0.1759r=0.195 dev=0.1753r=0.171 dev=0.1647r=0.108 dev=0.1641r=0.120 dev=0.1625r=0.102 dev=0.1573
We can create a diagram that illustrates this table:
The results of my test confirm that the small training interval used by the Neural Net provides better forecast.
Timing Solution provides several types of the candlestick model. Here is the comparison of all these models:
Modelcandlestick_1.hypcandlestick_1_a.hypcandlestick_1_b.hypcandlestick_1_c.hypNN Topology32 hidden32 hidden32 hidden32 hiddenTraining Mode50 before LBCtrain 5000 steps50 before LBCtrain 5000 steps50 before LBCtrain 5000 steps50 before LBCtrain 5000 steps+/ Statistics+54/27 (66.7%) ChSq=4.5+60/21 (74.1%) ChSq=9.4+54/27 (66.7%) ChSq=4.5+58/23 (71.6%) ChSq=7.6Average (r,dev)R=0.221 dev=0.1655r=0.290 dev=0.1516r=0.226 dev=0.1616r=0.219 dev=0.1631
Modelcandlestick_2.hypcandlestick_2_a.hypcandlestick_2_b.hypcandlestick_2_c.hypNN Topology32 hidden32 hidden32 hidden32 hiddenTraining Mode50 before LBCtrain 5000 steps50 before LBCtrain 5000 steps50 before LBCtrain 5000 steps50 before LBCtrain 5000 steps+/ Statistics+52/29 (64.2%) ChSq=3.3+61/20 (75.3%) ChSq=10.4+55/26 (67.9%) ChSq=5.2+36/15 (70.6%) ChSq=4.3Average (r,dev)r=0.193 dev=0.1612r=0.294 dev=0.1577r=0.207 dev=0.1717r=0.203 dev=0.2184
The models candlestick_1_a and candlestick_2_a provide the highest correlation of the forecasted projection line. Now, let us explore whether this result is occasional and what these two models have in common.
First of all, I have to mention that, in spite of the name, these models are not a traditional candlestick models. The program does not look for popular formations (such as Doji Star or Harami Cross) within the price data. Instead, it evaluates each candlestick inside the chosen interval using Fuzzy Logic methods. There are three parameters there to be evaluated, three components: the candlesticks height (High Low), its body (Open Close) and its lower shadow (Open Low). We can set ourselves the level of accuracy for the program to do this classification. Also, we can define as well the amount of candlesticks to be analyzed and how the candlesticks parameters are interfering among themselves. This info becomes a base for the prognosis.
For example, if we choose the model candlestick_1, the candles height is defined as one of the four possibilities: Lowest, Low, Medium, and High. Its body can be described as one of the three Low, Medium, and High. The lower shadow also takes into account 3 levels. I did analysis for 24 candlesticks and chose the minimum interference of these parameters for different candles. The more evaluating criteria are used, the more events for the Neural Net to consider.
If we compare the models candlestick_1_a and candlestick_2_a, we see that they have the same range for the candlesticks parameters and the same amount of candles being analyzed. The only difference is the way of how the parameters of analyzed candlesticks affect the parameters of the projected ones. The rules for the model candlestick_2_a are more variable than for the model candlestick_1_a. However, the number of these rules for the model candlestick_2_a is also bigger. Thus, we can state that these models are very similar.
Now, let us look again at the length of the training interval for these two best models. See yourself:
Both models show good results for all chosen training intervals. The best results are for the training interval of 40 60 price bars.
I checked different types of Relative Price Oscillators and Oscillators as well. My research leads to these general conclusions:
It does not matter for the tests what we use as the price fields. The results are pretty much the same for Close and Mean Price (H+L+C/3);
If we would like to get a higher correlation, it is better to use RPO with parameters from (1, 3, 3) to (1, 5, 5);
If we use the Oscillators with bigger periods, the training interval needs to be extended;
If we apply the averaged RPO (with the first parameter 2), the prediction ability goes down;
In regards to the NN outputs, the results are better for RPOs than for Oscillators themselves;
Backtest results change when we use symmetric parameters for the oscillator (due to future leaks);
I did not find any special settings for the Neural Net that have impact on the rules for the projection lines (I have checked the number of hidden neurons, etc.);
When we vary the models parameters (like Grade or number of rules), it means extra job for the Neural Net without any significant positive changes of the projection line
The results are very similar if we apply these models to the analysis of other financial instruments;
The maximum correlation received is approximately 30% +/ systematic error, for the training interval of 5 to 7 price bars after LBC.
See below the test results received by Ben Price. He did this test for Euro/USD. The Neural Net gives this correlation for 7 price bars, RPO (3, 10, 0, Sym1), first LBC set at 5.01.2005, 80 shifts*5 bars:
v{ ,&'/0BCUVhi{NOFGEFVhi{.B*mH phsH jUB*CJaJphCJaJCJaJCJaJmH sH CJaJmH sH 56\]mH sH 6]mH sH mH sH G ,&'9K]oFf$If$a$>&3Tv/ \Ff$IfFf/0CVi/)))$If$$IfTֈ6 'K.~4W9L_%0'634abi)$$IfTֈ6 'K.~4W9L_%0'634ab$IfNO$If^`$If(#$If$$IfTr^!)lw,
20)634ab(2<F$IfFGUt#,$If$$IfTr^!)lw,
20)634abt$If#$If$$IfTr^!)lw,
20)634ab,EFVi5t$$IfTr^!)lw,
0)634ab$Ifi53$$IfTr^!)lw,
0)634ab$If$If#$If$$IfTrg$&v Y
20)634ab#$If.<[#,$If$$IfTrg$&v Y
20)634ab[z$If..>PQcdvw_`RSAB4 5 $i%:&;&>& j@U6]mH sH jUmH sH B*phCJaJ #$If$$IfTrg$&v Y
20)634ab.>Q5t$$IfTrg$&v Y
0)634ab$IfQdw53$$IfTrg$&v Y
0)634ab$If_`RS,AB4 5 B!!"m""0##$$k%l%9&
&F9&:&<&=&>&,1h. A!"R#n$n%$$IfTֈ6 'K.~4W9L_% <0'634ab$$IfTֈ6 'K.~4W9L_% <0'634ab$$IfTֈ6 'K.~4W9L_% <0'634ab!Dd0
#A"+r dY2yY@=q+r dY2@
@TEz&?xY_hEflښxԤmjꅲF#!BHI+
R(d)TCBB@
B4 h^ԇ P<Ы
PoӤ.7o;ޮ"a&Zg?K[e'N箺O
}S8@@ݠveTVs7慛Z (ݙi+ۑ\^$ˏ@G6U,pz@i{㠝+F\fmq1iA/IO
*!j5Ex'>!Ef/pՖ7@a[n[k7Y0Ϫgo3DgMW;kf'_kU?bs'z}ܷlP9*\u_3!.^Ky
u¼+hg3
QG#}45]Әw Yau G"}4T5?ü1b S'"[ax(3v x\olV?4TSlUR
nڮyP][eXq8VeJXn7 !&e LD5Lwss>=*>{ο{sE
!#SZV&z)&QskUB.S'2B܈4\F/0 nD9su2'E5#yO)kB^ˈ9d\"ٜran$ɉE:.
*:=H!]_`;w9q?}m<^t3!uGo^ǳ`wqѷ6;}]Q_3{Zq UӕMgFJ< zŽW5b=2`VxY:~߳^9/*E"ά\7^+U~\U<+R$zO+ǹ 0)z52
2k3TuAVR+hɶ*+ދcdۗ]~gTXeʈ*=cR]1gsT72Uo:qĴk
ڗU?9j'oaݴ̈M]T=MEDT?OQg;Ie \Ko?
oA.9V]?[&5Qe[1ma=11C3.1juXx'io^?j4v^} RtsKz{=lffWl\qfccÇgxofx/DݯLD=~~~^#J7_Q2$O2,2ʛW_fx7GRΫz>iJ
:TQrc]oںַM/ҟ̣1H;
}iA!:zCП_MBb^z7
gxfu2dNֲ:ieu2V':Y<S';,})җ:8͟fkHGi_4'vg6EYVި1fH#=FȽga'91
߂6.xgF8F 8H#}t'q/{gyǱM.8GA.r8* ;@wqu[Es
r+>~}"G.8$82y7An)]:qp9"ø<{'Vvୁ"G~rqP@
@w^qJE'cA>9 ]q{᪱r7qÐ:9^x+[6];E:^1m/]5dX/jTT.o\Z?:EqQqF?Av;E]qR"Ez_r{q\Yp/E.qxE[!vޫj\=깷q,K/qY{/,iU1X=;RcfN@#)~\iľ7#eW
Js/UYy 5&O\պ^Kc߉`V;;5"{S {KǯmHoiJ~o_c"UEU 5$w1樤TRb''}#ֱJc^iCZIhb=i{$ݥ)Y 6X0Ur]mح+,vmm=m b;HyG,v]f;Fyb;rfWQ ]F"kS/o/ۯm4Zk,~Cdc"o''l}f6;1Ui=7Yoace~ȏ}_e/?6οIE}!\t@/BD{B4У18QwckA+1UlI+݈zNřmb#F a;W3.3dza&7KgǏ&x]g)gfT$%꿇Cߠ;f~Ɽ߃ZON^J3F2GT]~ԴQAc^q6Fՙ85Anȅ{!?%t}x
>R[ lŔ Dd +/00
#A", ,oUi3 Y@= ,oUi3 <`j7x[klE>WZBbB(Z@"JɅ҄(HyZc0E1$"UR%cBB&VB gvϴٻBacts<3;+`(,z3r,a*4N?y#0dH~`剸l*. g#Ћ
g_@:l][JNˉh"qht7LuՓ>^۽[=d8.D'/
4=\~m7#h9rP`.'axo3zI~71ucx1Ë
]}KPF(u_[3.e3<_?c9)ߣyG< BX`u+FIj0ckOO`ةWQJRM.Gy&ziG8t9՚fv/mgV0λvy=gK2y㼔q~P_%TWCtqN3/ZA}\ev֧.N;~86혺l㽡c1 6=0rTV98^;0jաV/4(E9q [#G5_xq͒LvY(b@ԖQ́^}4iiG<cm9ռqzzՋl9ű +cvB1.AؖQ_^%n%[\r9*ҋ
nƿc\Ko(ƸնFO8?j9~QNCn{?Qځ5_CaX ~/.Ѹm]>H?=*=!k^\QBVëvZ'{q&wZ{q~j=ㄳ#?EOɹCExqYKk#/ť@F.jOˋ#+=χڏ"]pE%zq~RKXR߃s^瓾~9m>xVNH~5]lι~kAEFle5чZz[Cg[
{֓N
N71Fw`чVz7%7> do^_ݏ10%8UIXY)'AAՅһҾoлѧ G)G}h{j{ѧ;Oz'^yl[BizKvҾiЫziވi^tfп÷WH%SOpteݷW=G~yE:~^\ms`::yg=q$LB~=ys*8$ϧ4JLz~>L۹"'gcޕ̴#e?s0V
뛢O˵5K^xVBS
]cq.Nm^C6~Yޮ]Zk]K_oR~OR=gh'懬ۺ7~g5OQowQLzhí~Ƞ>;S֭LekTW@
U
i8@8NormalCJ_HaJmHsHtHN@N Heading 3$<@&5CJOJQJ\^JaJ<A@<Default Paragraph Font>"^
,

&'9K]o
3
T
v
/0CViNO(2<FGUt,EFVi#.<[z.>Qdw_`RS,AB45Bm0 k!l!9":"<"="@"(0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000 0 0 0 0 0 0 0 0 0 0 00@0@0@00.>&)/i(Fti[Q9&>& !"#$%&'(*+,.>&
2
3
S
T
u
v
.0BCUVhi{"#,.;<Z[yz,.=>PQcdvw@"
2
3
S
T
u
v
.0BCUVhi{"#,.;<Z[yz,.=>PQcdvw@"
/0..@"Tanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanya.C:\TS Articles\English\MikhailCandle final.docTanyanC:\Documents and Settings\Tatiana\Application Data\Microsoft\Word\AutoRecovery save of MikhailCandle final.asdTanyanC:\Documents and Settings\Tatiana\Application Data\Microsoft\Word\AutoRecovery save of MikhailCandle final.asdG .8 ^`OJPJQJ^Jo(^`OJQJo(opp^p`OJQJo(@@^@`OJQJo(^`OJQJo(o^`OJQJo(^`OJQJo(^`OJQJo(oPP^P`OJQJo(G ^P '9K]o
3
T
v
/0CVi(2<FGUt,EFVi#.<[z.>Qdw@"@="="B="=">"@UnknownGz Times New Roman5Symbol3&z Arial?5 z Courier New;Wingdings"h0ff7<<Jw4nr0d">"2QHX 40==>9 AB0BL5 O 8AA;54CN ?@>3=>AB8G5A:85 2>7<>6=>AB8 <>45;8, >A=>20==>9 =0 <0B5<0B8G5A:>< <5B>45 =5G5B:>9 ;>38:8 8 =59@>A5B8 B0: =07K205<0O candlestick modelMikeTanyaOh+'0
<H
T`hpxВ данной статье я исследую прогностические возможности модели, основанной на математическом методе нечеткой логики и нейросети – так называемая candlestick model MikeаikeikeNormalTanya26yMicrosoft Word 9.0@
@@.`G<՜.+,0hp
home?<"
В данной статье я исследую прогностические возможности модели, основанной на математическом методе нечеткой логики и нейросети – так называемая candlestick modelTitle
!"#$%&'()*+,./123456789:;<=>?ABCDEFGHIJKLMNOPQRTUVWXYZ\]^_`abeRoot Entry FGgData
01Table@$WordDocument.^SummaryInformation(SDocumentSummaryInformation8[CompObjjObjectPoolGG
FMicrosoft Word Document
MSWordDocWord.Document.89q