Fractals, ZZ & the Pissing Dog | Page 29 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

ECN && STP && Scalping && Hedging - Broker Review - Oct 2009 24 replies

Alert on Fractals (MT) 2 replies

Zigzag & Fractals 10 replies

EA wanted for sound alert depending on fractals condition 2 replies

Fractals Edge Fx 3 replies

Attachments: Fractals, ZZ & the Pissing Dog
Exit Attachments

Fractals, ZZ & the Pissing Dog

  • #561
  • Mar 29, 2014 4:36am Mar 29, 2014 4:36am
[Quote = candero; 7371159] {quote} Gracias por el enlace, G! Cerebritos de matemáticas, me pregunto qué fractales y ZZS siguen la curva de campana? Mandelbrot ciertamente dejó en claro que el precio no lo hace. [/ Quote]
You can see a good example with the use of Gauss in the calculation of probability of length .
The binomial probability can not be used because it is for smaller samples of 30 elements.
The data do not fit a normal distribution .
Bayes usually used in samples with independence between the variables .
There is a Bayes method for dependency between variables rather interesting .
T -student is a derivation of normal distribution . Data need to fit a normal distribution .
I like the logarithmic , but the log should be applied elsewhere in prices.
In conclusion, we have a continuous random process with dependence between variables. Is continuous because it always depends on ( t). There dependence , because there is dependence between sequences bars.
The probabilistic model is not perfect . There will always be market conditions that alter the odds. We saw performances in Graph gg53 . During a certain period , was very successful , but at a certain point I had an error that resulted in losses. Do not obsessions with perfect entries During all time . We are doing calculations with past data. Extrapolation in time are not accurate.
thank you very much
  • #562
  • Apr 3, 2014 3:50am Apr 3, 2014 3:50am
I really would like to keep this thread alive. I have put so much time into this system and I know many others who have done the same and it would be sad if we leave here without achieving our goal.

I know that there has to be an error up to my third step as I do not get enough signals. (start,fractal prediction, length)
http://www.forexfactory.com/showthread.php?p=7363655#post7363655

Anyone got any ideas?
  • #563
  • Apr 3, 2014 4:17am Apr 3, 2014 4:17am
I dont get enough signals too.
At first it was because of a buggy routine which resets the prediction when a new zzleg appears.

But now its because of an unfinished database.
Not enough occurences and not enough balance.

I have build in some error exception routine to find out why an arrow has not be painted. (waiting for another fractal or next bar not in opposite direction)
Maybe this can help.

here is my screen:
Attached Image (click to enlarge)
Click to Enlarge Name: eurusdm5.png Size: 49 KB
  • #564
  • Apr 3, 2014 5:42am Apr 3, 2014 5:42am
Been busy with other projects and the nightmare of the new build so have not started yet.
One thing I am considering is keeping track of the Fib value since last zigzag or fractal as a possible edge in predicting next bar.
How often is there a fractal at or near a fib level on a zigzag leg? That is just another field in the data structure (database).

As I mentioned I will be building a data structure ( considering binary tree or trinary tree to also hold dojis) that holds information at each node similar to a database but the structure holds the 0-1 data automatically with a cound of how often visited. I will also look at all my notes and attempt to code from rules of this thread.

Has anyone looked at the 3 level zigzag semaphore? It might help to use different settings for ZigZag as is done in that indicator.
Attached File(s)
File Type: mq4 3_Level_ZZ_Semafor.mq4   8 KB | 575 downloads
  • #565
  • Apr 3, 2014 10:16am Apr 3, 2014 10:16am
Quoting Kilian19
I really would like to keep this thread alive. I have put so much time into this system and I know many others who have done the same and it would be sad if we leave here without achieving our goal. I know that there has to be an error up to my third step as I do not get enough signals. (start,fractal prediction, length) http://www.forexfactory.com/showthread.php?p=7363655#post7363655 Anyone got any ideas?
It was you who brought gg's amazing indi to my attention in Steve Hopwood's forum. Since then, I'm hooked. And you know what, during your brief hiatus from this thread, I really thought you've nailed it and off busy earning your pot of gold with it. I damn well wish you (and everyone actively involved) nailed this system to prove that despite scattered pieces of puzzles, there are enough pieces here to complete the puzzle.

I don't have much time to code this. I'm barely 30% done. But I'm afraid I might end up with 999 pieces in a 1000-piece jigsaw puzzle.

Having said that, this is still my one and only system in FF I'll work on indefinitely. It's worth the time and effort. Worth failing for.
  • #566
  • Apr 3, 2014 12:29pm Apr 3, 2014 12:29pm
Quoting MrPip
Been busy with other projects and the nightmare of the new build so have not started yet. One thing I am considering is keeping track of the Fib value since last zigzag or fractal as a possible edge in predicting next bar. How often is there a fractal at or near a fib level on a zigzag leg? That is just another field in the data structure (database). As I mentioned I will be building a data structure ( considering binary tree or trinary tree to also hold dojis) that holds information at each node similar to a database but the structure holds the...
If you're dealing with cold hard metadata and statistics, I wouldn't add the voodoo of fib levels. Rather, I'd prefer to bin up retracements (bins of 5 or 10) and let the data speak for itself as it has so far.
"Mrrraggglhlhghghlgh"
  • #567
  • Apr 3, 2014 1:30pm Apr 3, 2014 1:30pm
I plan on doing this in steps as described in the thread, the Fib piece will be applied later to see if it helps.
From my math background I have found Fibs at work in many areas, even traffic flow studies. Fibs also work well combined with fractal geometry. I have also seen price turn at fib levels, just a possible way of improving the odds.
For example if you are at fib 62.8 and a possible fractal is forming based on 1 prior bar being a local low/high it might give an addition level of confidence(statistical not mental).
Robert
  • #568
  • Apr 3, 2014 2:41pm Apr 3, 2014 2:41pm
Most steps are published.
1 - probability of fractal.
must compare occurrences in the past in order to find high probability of fractal otherwise.
2 - probability of length
I think that is the least important step. You have to find the number of (-1) bar.
3 - probability of next bar and current bar
is the step where I have doubts.
We know that no probabilistic calculations are simply occurrences are compared.
The operation of this system is equal to that of a genetic algorithm. A genetic algorithm works as the human mind, "learn from past mistakes."
This system is complemented with strength, for sure pips.
  • #569
  • Apr 3, 2014 3:32pm Apr 3, 2014 3:32pm
If I recall correctly, gg stated in the previous thread that his program at the time took 2.5 to 3.5 minutes to do a prediction. How long does your program take to do a prediction compared to his? Why the difference?
  • #570
  • Apr 3, 2014 5:17pm Apr 3, 2014 5:17pm
Quoting Kilian19
I really would like to keep this thread alive. I have put so much time into this system and I know many others who have done the same and it would be sad if we leave here without achieving our goal. I know that there has to be an error up to my third step as I do not get enough signals. (start,fractal prediction, length) http://www.forexfactory.com/showthread.php?p=7363655#post7363655 Anyone got any ideas?
Yes, I know there are a few of us who are going after this in a serious way. I don't post that much but am committed to replicating G's final results. If one can do it, others can too. I'm fighting with MT4 at the moment.

With flx23's document it is now possible to get the basic version of this coded up in less than a week for someone who has the data and has done work with time series before. Having a custom testing platform helps a lot too. Code review, bug fixes and interfacing take longer than it takes to code the underlying logic.

I have a version that had really good in-sample performance with poor out-of-sample performance. Has anyone had a similar experience? It is almost like it memorizes the in-sample data but then fails miserably on new data. I'm working on getting the data into MT4 so I can see what's happening. I have some ideas about which parts of the basic algorithm are important and which are not but will hold those thoughts until I can get a better picture of exactly what is happening.
  • #571
  • Apr 3, 2014 5:48pm Apr 3, 2014 5:48pm
Quoting Xaphod
If I recall correctly, gg stated in the previous thread that his program at the time took 2.5 to 3.5 minutes to do a prediction. How long does your program take to do a prediction compared to his? Why the difference?
It depends on how one has built it and the amount of data, for me it's less than 5 secs on a 10+ year old machine and there's more time to find, but it doesn't really matter if you haven't got the results you're after... And further code optimization makes debugging a bugger hell for me.

Quoting FXEZ
I have a version that had really good in-sample performance with poor out-of-sample performance. Has anyone had a similar experience? It is almost like it memorizes the in-sample data but then fails miserably on new data.
Yep, it happened to me. In-sample was ideal, OOS destroyed the dream...

Quoting burnssss
2 - probability of length I think that is the least important step.
I agree. It's too primitive me thinks. I'm contemplating of adding another dimension to ZZ pattern - its length or height in pips. Has anyone tried this? I think G talked about it a while back - universal connection between the pattern and its height. My rough idea is to retain the length in bars and add height.
  • #572
  • Apr 4, 2014 12:32am Apr 4, 2014 12:32am
Yes, ZigZagHeight is on my list and will be stored as a percentage of adr as well as pips.
I am also looking at different settings for zigzag in an earlier phase to see if that helps as well.
Larger ZigZags are made up of smaller ZigZags, even on the same time frame.
Once the data structure is created it will simply be a matter of following the most recent path(s) that lead to the current candle pattern. Then check the pre-calculated percentage of up/down from that point, etc.
It is much faster than using a database,less overhead, and is a form of chess opening book.
I will try to draw a picture when I have the time.
  • #573
  • Apr 4, 2014 1:50am Apr 4, 2014 1:50am
Is Renko/Range bars not the way to go, or has this already been discussed?
  • #574
  • Apr 4, 2014 2:30am Apr 4, 2014 2:30am
To answer a few recent thoughts:
Running time - 2.5-3.5 minutes run was on early stages of development. Lots of not-optimized loops...
Fibs - I acknowledge it's accuracy, but it's voodoo for me - I don't use it.
Renko - Looks very promising, but requires major changes in this algo. Not finished yet.
Length (bar count) - yes, it's the least important, BUT...
You have to take into account large bars (pips wise) and not just the number of them, they are equivalent to two or three "regular" bars in other ZZ's.

G.
  • #575
  • Apr 4, 2014 2:32am Apr 4, 2014 2:32am
Quoting gg53
Length (bar count) - yes, it's the least important, BUT...
You have to take into account large bars (pips wise) and not just the number of them, they are equivalent to two or three "regular" bars in other ZZ's.

Didn't I suggested that already a few times in the past? Yes that sounds about right. So if a bar is bigger than 2x the atr of the past we count it twice?

Do we consider that already during the creation of our database? Are Dojis ignored length wise? Hmm there is a lot to try out
  • #576
  • Apr 4, 2014 2:42am Apr 4, 2014 2:42am
Quoting Kilian19
{quote} Didn't I suggested that already a few times in the past? Yes that sounds about right. So if a bar is bigger than 2x the atr of the past we count it twice? Do we consider that already during the creation of our database? Are Dojis ignored length wise? Hmm there is a lot to try out
Kilian19,
I mentioned it several times in that thread or previous one - but maybe failed to be mentioned in the docs.
To be practical - If you got large bars (pips wise) - they are equivalent to larger bars count in the database.

G.
  • #577
  • Apr 4, 2014 3:10am Apr 4, 2014 3:10am
Quoting gg53
{quote} Kilian19, I mentioned it several times in that thread or previous one - but maybe failed to be mentioned in the docs. To be practical - If you got large bars (pips wise) - they are equivalent to larger bars count in the database. G.
Hmm, I dont recall reading anything about handling large bars..I'll check again.
  • #578
  • Apr 4, 2014 3:11am Apr 4, 2014 3:11am
Hi all,

First time poster on FF. I've only recently started looking at FX (or any trading - noob alert), having come from a career of machine learning, signal processing and time series analysis. What caught my eye here was how similar this concept is to one I've been considering, and it's brilliant to see it develop with the expert guidance from gg53 et al.

The database work you're doing, to me, looks effectively like a neural network. I have a long way to catch up (thanks to flx23 and Crodzilla for the brilliant documentation), but I hope to be able to contribute my experience with these and other machine learning techniques to this thread. I'll refrain from clogging the thread any further until I'm at a point I can be helpful.

Cheers

p.s. Congrats gg53 on your successful trial!
  • #579
  • Apr 4, 2014 8:47am Apr 4, 2014 8:47am
Quoting Thoughts
{quote} Hmm, I dont recall reading anything about handling large bars..I'll check again.
I see what you were trying to say GG53 >> http://www.forexfactory.com/showthread.php?p=7222781#post7222781

I had assumed from your reply to Kilian19, you were speaking of counting large bars more than once.
  • #580
  • Apr 5, 2014 4:05pm Apr 5, 2014 4:05pm
Quoting IamBadger
Hi all, First time poster on FF. I've only recently started looking at FX (or any trading - noob alert), having come from a career of machine learning, signal processing and time series analysis. What caught my eye here was how similar this concept is to one I've been considering, and it's brilliant to see it develop with the expert guidance from gg53 et al. The database work you're doing, to me, looks effectively like a neural network. I have a long way to catch up (thanks to flx23 and Crodzilla for the brilliant documentation), but I hope to be...
Welcome, we have a similar background in a way as your career was a piece of mine.
I did a lot with real time processing, multi tasking OS on 8 bit back in the 80's.
I am currently starting development of a type of neural net similar to one I did in the past for fingerprint images.
Less complex because only need a binary tree. The one for the prints required more of a network with varying nodes at eack level.

Someone asked for a picture so am attaching.

Attached Image (click to enlarge)
Click to Enlarge Name: treeforzz.png Size: 17 KB



The structure holds the 0,1 information that is currently stored in a sql database. THANKS FOR THAT. I already use it to store information for other systems instead of my old way of using log files.
Each node will hold data like number of times visited, size, is it a fractal, is it a zigzag, what is the fib value as well as statistical information. Then when following the current leg of a zigzag there will we information about the most likely path down the tree. As I said, similar to a chess open book.
Robert
0 traders viewing now