Thread: แนวคิดการจับคู่รูปแบบ EA

I wonder, what was the result of the implementation of this method ? The idea sounds very interesting..

Originally Posted by ;

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 egorize 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.

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 classifiion 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!

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 classifiions form a string. I'm trying to expand the number of characters that I can use.

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.

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.

Originally Posted by ;

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).

Originally Posted by ;

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 wordthe 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...

Originally Posted by ;

คุณอยู่ครึ่งหนึ่งแล้ว ฉันกำลังพยายามสร้างตัวแทนของเทียนแต่ละแท่ง เทียนแต่ละประเภทจะแสดงด้วยอักขระ เมื่อนำอักขระหลายตัวมารวมกันเป็น 'คำ' ฉันมีรูปแบบที่สามารถค้นหาย้อนหลังได้จากประวัติแผนภูมิ รุ่นเก่ามี3ตัว (L, S และ D) โมเดลใหม่นี้ใช้อักขระ 255 ตัว (A, B, C, D, E, F....). หวังว่า 'ตัวอักษร' ที่ขยายใหญ่ขึ้นนี้จะส่งผลให้มีคำมากขึ้น (รูปแบบแท่งเทียน) ซึ่งนำไปสู่คำศัพท์ที่ใหญ่ขึ้น (รูปแบบที่ไม่ซ้ำกันมากขึ้น)

ฉันยังไม่เข้าใจเลยว่าทำไมคุณถึงสนใจเรื่องการตั้งชื่อ/แทนตัวอักษรแต่ละแท่ง!! ง่ายกว่ามาก (และทั่วไปกว่า) เพียงใช้ตัวเลขแทน ตัวอย่างเช่น แท่งเทียนแต่ละแท่งจะมีลักษณะดังต่อไปนี้: ขั้นแรก คำนวณความแตกต่างระหว่างสูงและต่ำ (สูง - ต่ำ) ในหน่วย pip (เรียกว่าส่วนต่าง) จากนั้นคุณลักษณะต่อไปนี้จะถูกดึงออกมาสำหรับแต่ละแท่งเทียน: # (เปิด - สูง )ส่วนต่าง # (ปิด - ต่ำ)ส่วนต่าง # (เปิด - ปิด)/ส่วนต่าง เฉพาะตัวเลขที่เกิดจากลักษณะสามประการข้างต้นเท่านั้นที่จำเป็นสำหรับแต่ละแท่งเทียน อีกครั้ง คุณจะต้องคำนึงถึงความเกี่ยวข้องของแท่งเทียนแต่ละแท่งด้วยแท่งเทียนที่ตามมาหรือก่อนหน้า ตัวอย่างเช่น ช่องว่างในแนวโน้มขาลงตามด้วยค้อนเป็นสัญญาณรั้นและรูปแบบ

Originally Posted by ;

The, can you tell us the result of your implementation ?

In back testing it seemed somewhat successful, but I got busy with other projects and never got it fully off the ground and trading live. Personally I think that if implemented properly it would be very successful.