Pattern Matching Concept EA | Page 3 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Trading Equity Millipede Concept by Concept 99 replies

pattern matching 97 replies

Two matching highs trend breakout? 3 replies

Matching metatrader times to alerts 0 replies

Oanda's NOT a matching engine...? 4 replies

Attachments: Pattern Matching Concept EA
Exit Attachments

Pattern Matching Concept EA

  • #41
  • Mar 30, 2010 10:05am Mar 30, 2010 10:05am
Quoting Ronald Raygun
[edit] You could change the distribution of the 255 'characteristic slots' so one feature is more weighted than another.
I was thinking about the same thing. You can count how many different values of High - Open, Close - Low and Close - Open occured since 2000.

You will possibly get something like this:

High - Open - 50 different values
Close - Low - 50 different values
Close - Open - 100 different values

In this example you can assign ca. half of the 'characteristic slots' to "Close - Open".
  • #42
  • Apr 1, 2010 10:52am Apr 1, 2010 10:52am
Will pay attention to this thread.
福 子 Victor V
  • #43
  • Apr 1, 2010 11:33am Apr 1, 2010 11:33am
Thanks for April Offer You Cannot Refuse.

I refuse (but it is really really hard)!

Subscribe.

RR
  • #44
  • Apr 1, 2010 12:12pm Apr 1, 2010 12:12pm
Ronald,

Let me know if I'm off base...

Patterns need to be generalizations, so that they could be skewed by a certain percentage but still fit in the pattern... so that if a bars' range was 10 pips or 100 pips, it could still fit within a pattern.

Would it be any use to catagorize a bar by its range? An example... a bar range is always 100%. So, the open would be at 50% of its range (center of the bar), and the close could be at 75% (3/4 of the bar) of its range.
Then you could classify the low of the current bar to the low of the previous bar as <100%, low of previous bar within the range of the current bar, etc.

I am unsure at this time how to handle if the previous bar is part of engulfing pattern...

Just throwing out an idea...
Carl
  • #45
  • Apr 1, 2010 12:32pm Apr 1, 2010 12:32pm
Quoting crodzilla
Ronald,

Let me know if I'm off base...

Patterns need to be generalizations, so that they could be skewed by a certain percentage but still fit in the pattern... so that if a bars' range was 10 pips or 100 pips, it could still fit within a pattern.

Would it be any use to catagorize a bar by its range? An example... a bar range is always 100%....
Hey Carl,

Your method identifies bars based on their relative shape. I'm not sure if you intended for this circumstance, but let's assume we have two bars which are perfect scaled replicas of each other. When looking at a chart, would you visually classify them as the same bar?

I'm classifying bars based on their absolute shape, it would be easier to differentiate between patterns where the price is ranging and patterns where the price is trending.

Ideally, I'd rather just write a neural network which could take into account all of this data. 2 layer is enough to come up with every possible continuous mathematical function, but I suspect adding more layers and nodes will allow the NN to create a conditional step function. It's similar to our own trading thinking..

-If the market price is moving sideways use this mindset
-If the market is moving up, use this mindset
-If there is undefinable chaos, use a different mindset.

That's the hope anyway
  • #46
  • Apr 4, 2010 2:19pm Apr 4, 2010 2:19pm
Quoting Ronald Raygun
Ideally, I'd rather just write a neural network which could take into account all of this data. 2 layer is enough to come up with every possible continuous mathematical function, but I suspect adding more layers and nodes will allow the NN to create a conditional step function. It's similar to our own trading thinking..
Hi Ronald,

Are you writing a neural network based on FANN?
Will you share a NN?

If NN is based on FANN I hope this info may be useful.
  • #47
  • Apr 23, 2010 11:39pm Apr 23, 2010 11:39pm
Quoting Ronald Raygun
...This is a newer concept EA which is programmed to find its own candle patterns. The pattern finding algorithm is simple: it checks the previous x closed bars and generates a string something like "LLDSSL" that means the bars from right (most recent) to left were long, short, short, doji, long, long.

The EA then scans the chart history looking for a pattern of bars with that exact same sequence. The EA checks x bars after the sequence to see if that bar was long or short. The results are then added to the statistics generated by the EA...
Hey there Ronald... This reminds me of something I did a number of years ago for equities. My set-up was essentially a generalization of the same idea you are working with.

The way I generalized the approach was that rather than trying to categorize candles into traditional patterns (long, doji, short, etc.) and then match sequences, I just generated sequences of percentage-based differentials for each of OHLC on each bar, as compared to the previous bar. By expressing all of these values from bar to bar in terms of similarity to the previous bar and then setting a specific number of bars to examine, I ended-up with a single mathematically derived value that expressed the overall "similarity" of all of OHLC for a set number of bars to the current condition.

Then I would check the subsequent bars from the close historical match to see if it was indicative of a good move and to get an immediate term signal to trade as per the previous matching pattern. Anyway, I just thought I would mention this as this kind of set-up it might be something to add to your pattern recognition to give a generalization effect.

(To elaborate, my system did something like this, for the past n bars being examined:

open_2% = (open_2-open_1)/open_1
open_3% = (open_3-open_2)/open_2
open_n% = (open_n-open_n-1)/open_n-1

This was done also for n bars for high, low, and close. The next step was to find and sum the differences between each bar position in the historical data to the same offset bar within the current bars (with the same bar to bar percentages computed for the current bars.) Finally, the sum of the sums of the differences for the past and current OHLC would be taken. Then it was a simple matter to evaluate that final result -- the closer the number was to zero, with each of OHLC results also being close to zero, indicated similarity between the current bars and the historical bars.

Once similarity was determined, large moves either up or down in the past data were of interest as the current conditions might be expected to behave in a similar manner.)

Anyway, I thought I would mention that, as you might be able to make your pattern matching more abstract and less dependent on human-recognizable constructions using a set-up like that.

Theo Buitendyk.
  • #48
  • Apr 26, 2010 9:18am Apr 26, 2010 9:18am
The,

can you tell us the result of your implementation ?
  • #49
  • Apr 26, 2010 1:13pm Apr 26, 2010 1:13pm
I wonder, what was the result of the implementation of this method ? The idea sounds very interesting..
  • #50
  • Apr 26, 2010 1:24pm Apr 26, 2010 1:24pm
Quoting tbuitendyk
Hey there Ronald... This reminds me of something I did a number of years ago for equities. My set-up was essentially a generalization of the same idea you are working with.

The way I generalized the approach was that rather than trying to categorize candles into traditional patterns (long, doji, short, etc.) and then match sequences, I just generated sequences of percentage-based differentials for each of OHLC on each bar, as compared to the previous bar. By expressing all of these values from bar to bar in terms of similarity to the previous...

That's an interesting approach Theo,

I did something similar with a coordinate-based system. Each individual bar is given 3 coordinates. Those coordinates can be plotted in a 3-D region. The next step is finding a way to group those bars so bars within a specific coordinate region are identified as the same bar.

It's this last step where I'm having trouble.

Theoretically, I could expand this coordinate system to as many dimensions as I could possibly want. But, before I can do that, I need to figure out how to combine various groups together.

If you know anyone who is good with clustering algorithms, I would love to speak with them.
  • #51
  • Apr 26, 2010 5:05pm Apr 26, 2010 5:05pm
Ray and others,

you definitely don't want to limit yourself to just classifying individual bars. As Theo and others have alluded to, you also need some classification of the current bar in relation to the previous bar and possibly to the bar before that one too. If you plan on feeding this derived data into a backprop NN, you also want to normalise your derived data into deltas across the bars, much like Theo suggested.

Good luck!

Ohka
  • #52
  • Apr 26, 2010 5:12pm Apr 26, 2010 5:12pm
The purpose of uniquely identifying and grouping bars is so I can generate a "pattern string".

I introduced an EA at the beginning of this thread which had three types of bars. (Long, Short, and Doji) those classifications form a string. I'm trying to expand the number of characters that I can use.
  • #53
  • Apr 26, 2010 8:23pm Apr 26, 2010 8:23pm
Am not sure I understand what you are trying to achieve, but I think the implementation should be very simple. Each pattern will have four different variables (Open,close,high,low).

Say you have three consequent candles, you would need 4 different strings:
Open "+12.1-01.5+30.9-20.4"
High "+02.1-01.5+30.9-20.4"
Low "+52.1-01.5+30.9-20.4"
Close "+22.1-01.5+30.9-20.4"

The numbers can be anything. The first number on the left (+12.1) is the percentage change of the first candle in comparison with the previous one. The second number is for the OPEN change of the second candle in the pattern in comparison with the previous candle. And so on..

You might want to add a tolerance value when searching for patterns. So, if a pattern is almost similar to the pattern you are looking for (the difference in the percentage is less than a previously defined threshold), you would still count it as a HIT.

This would require a large processing power to be implemented of course, but it is doable.
  • #54
  • Apr 26, 2010 8:38pm Apr 26, 2010 8:38pm
I have skimmed through the past four pages to try to understand what you are trying to achieve. If my understanding is correct, it seems like you are trying to find a representation for each candle on its own. Well, this is not that much usefull in my opinion. You should care more about the relationships between candles through the percentages method I described above.


If you want to implement a 1 candle pattern, then you would also use percentages by computing the difference between open,close,high and low and dividing them with the maximum between high and low or something like that. It's that simple.
  • #55
  • Apr 26, 2010 8:45pm Apr 26, 2010 8:45pm
Quoting semelak1
I have skimmed through the past four pages to try to understand what you are trying to achieve. If my understanding is correct, it seems like you are trying to find a representation for each candle on its own. Well, this is not that much usefull in my opinion. You should care more about the relationships between candles through the percentages method I described above.


If you want to implement a 1 candle pattern, then you would also use percentages by computing the difference between open,close,high and low and dividing them with the maximum between...
You're half there.

I'm trying to generate representations of individual candles. Each candle type is represented by a character. When multiple characters are put together forming a 'word', I have a pattern which I can look for back through the chart history.

The old model has 3 characters. (L, S, and D). This new model uses 255 characters. (A, B, C, D, E, F....). Hopefully this enlarged 'alphabet' will result in more words (candle variations) leading to a larger vocabulary (more unique patterns).
  • #56
  • Apr 27, 2010 3:19am Apr 27, 2010 3:19am
Quoting Ronald Raygun
You're half there.

I'm trying to generate representations of individual candles. Each candle type is represented by a character. When multiple characters are put together forming a 'word', I have a pattern which I can look for back through the chart history.

The old model has 3 characters. (L, S, and D). This new model uses 255 characters. (A, B, C, D, E, F....). Hopefully this enlarged 'alphabet' will result in more words (candle variations) leading to a larger vocabulary (more unique patterns).
And what can we conclude for our trading? For example, your algo found out that after the word / the five candles "R-O-N-A-L" there is a likely chance for a candle 'D' - how would that influence your trading decision?

Another problem might be that you use such a huge alphabet. Regarding my example above, when we assume all candles are nearly equally distributed - then the chance for the candle "D" might be around 0.4%... cause we stated, that our algo proved "D" more likely in that case the chance might increase to maybe 1%... Nevertheless not much...
  • #57
  • Apr 27, 2010 4:36am Apr 27, 2010 4:36am
Quoting Ronald Raygun
You're half there.

I'm trying to generate representations of individual candles. Each candle type is represented by a character. When multiple characters are put together forming a 'word', I have a pattern which I can look for back through the chart history.

The old model has 3 characters. (L, S, and D). This new model uses 255 characters. (A, B, C, D, E, F....). Hopefully this enlarged 'alphabet' will result in more words (candle variations) leading to a larger vocabulary (more unique patterns).
I still don't understand why you care about naming/representing each candle with a character!! It's much easier (and more generic) to simply use numbers instead. For example, each candle would be represented by the following characteristics:
First, compute the difference between High and Low (High - Low) in pips (call it Diff). Then the following features are extracted for each candle:
# (Open - High ) / Diff
# (Close - Low) / Diff
# (Open - Close)/ Diff

Only the numbers resulting from the above three characteristics are needed for each candle. Again, you would need to take into account the relevance of each candle with it's following or preceding candle. For example, A gap in a downtrend following by a Hammer is a bullish signal and a pattern.
  • #58
  • Apr 27, 2010 4:59am Apr 27, 2010 4:59am
Quoting Ronald Raygun
That's an interesting approach Theo,

I did something similar with a coordinate-based system. Each individual bar is given 3 coordinates. Those coordinates can be plotted in a 3-D region. The next step is finding a way to group those bars so bars within a specific coordinate region are identified as the same bar.

It's this last step where I'm having trouble.

Theoretically, I could expand this coordinate system to as many dimensions as I could possibly want. But, before I can do that, I need to figure out how to combine various groups together....
Hi Ronald,

Is a genious idea what you are trying to achieve. a dynamic way to generate clusters is the SOM (Self Organized Map), as the name states, is a way the where the sata get self organized and the result is clusters with a set of parametrs for each one. By having the SOM implemented, your system will generate on real time clusters for a specific set and you can detects if the current pattern if any belongs to a high probabliity profit cluster.

And to make it more realistic then you can apply the fuzzy logic when deciding about a high probablity cluster.

I am not sure if my message is clear but plz don't hesitate to ask if you have anu question. googling SOM or SOM + fuzzy logic will get you more information but the challenge is to code the theory into mq4. another way is to use excel handle the SOM clustering and use the result.

Cheers,

Way2Freedom
  • #59
  • Apr 27, 2010 6:32am Apr 27, 2010 6:32am
Quoting nightyhawk
And what can we conclude for our trading? For example, your algo found out that after the word / the five candles "R-O-N-A-L" there is a likely chance for a candle 'D' - how would that influence your trading decision?

Another problem might be that you use such a huge alphabet. Regarding my example above, when we assume all candles are nearly equally distributed - then the chance for the candle "D" might be around 0.4%... cause we stated, that our algo proved "D" more likely in that case the chance might increase to maybe 1%... Nevertheless...
Using the alphabet actually allows for more trade opportunities than simply using straight numbers. With the alphabet, there are only 255 candle types, and assuming two candle patterns, the likelihood of finding an identical pattern is 1/65025.

When using straight numbers, there are theoretically infinite variations. You could try to create a margin of error of sorts, and work things that way. At least with the alphabet, it takes care of the margin of error for me.


Quoting semelak1
I still don't understand why you care about...
The main problem is neither system fully accounts for gaps. Using the relational method you mentioned earlier, a gap is the same as a large candle. With the way I've structured the code, I should be able to add more dimensions/parameters to the three I already have unless I misunderstood something in multi-variable calculus. For now, the only relation between candles that I'm really concerned about is whether an 'R' candle comes after a 'P' candle more often than it does a 'Q' candle. In live trading, we'll very rarely catch the gap anyway, so I don't bother.
  • #60
  • Apr 27, 2010 6:32am Apr 27, 2010 6:32am
Quoting way2Freedom
Hi Ronald,

Is a genious idea what you are trying to achieve. a dynamic way to generate clusters is the SOM (Self Organized Map), as the name states, is a way the where the sata get self organized and the result is clusters with a set of parametrs for each one. By having the SOM implemented, your system will generate on real time clusters for a specific set and you can detects if the current pattern if any belongs to a high probabliity profit cluster.

And to make it more realistic then you can apply the fuzzy logic when deciding about a high...
A self organizing map would be far easier in theory to use. I just don't know how to make one, but you mentioned it's possible in Excel... could you elaborate?
0 traders viewing now