Thread: Patroonaanpassingsconcept EA

Originally Posted by ;

En wat kunnen we concluderen voor onze handel? Uw algoritme kwam er bijvoorbeeld achter dat er na het woordde vijf kaarsen R-O-N-A-L een waarschijnlijke kans is op een kaars 'D' - hoe zou dat uw handelsbeslissing be�nvloeden? Een ander probleem kan zijn dat je zo'n enorm alfabet gebruikt. Wat betreft mijn voorbeeld hierboven, als we aannemen dat alle kaarsen bijna gelijk verdeeld zijn - dan zou de kans voor de kaars D ongeveer 0,4% kunnen zijn... 1%... Toch...

Het gebruik van het alfabet biedt eigenlijk meer handelsmogelijkheden dan alleen het gebruik van gewone cijfers. Met het alfabet zijn er slechts 255 soorten kaarsen en uitgaande van twee kaarspatronen is de kans op het vinden van een identiek patroon 1/65025. Bij het gebruik van rechte getallen zijn er theoretisch oneindig veel variaties. Je zou kunnen proberen een soort foutmarge te cre�ren en de dingen op die manier te werken. Het alfabet zorgt in ieder geval voor de foutmarge voor mij.

Originally Posted by ;

Ik begrijp nog steeds niet waarom je om...

Het grootste probleem is dat geen van beide systemen volledig rekening houdt met hiaten. Met behulp van de relationele methode die u eerder noemde, is een opening hetzelfde als een grote kaars. Met de manier waarop ik de code heb gestructureerd, zou ik in staat moeten zijn om meer dimensies/parameters toe te voegen aan de drie die ik al heb, tenzij ik iets verkeerd heb begrepen in multi-variabele calculus. Voor nu is de enige relatie tussen kaarsen waar ik me echt zorgen over maak of een 'R'-kaars vaker na een' P'-kaars komt dan na een 'Q'-kaars. In live trading zullen we hoe dan ook zelden het gat dichten, dus ik doe geen moeite.

Originally Posted by ;

[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.

Will pay attention to this thread.

Thanks for April Offer You Cannot Refuse. I refuse (but it is really really hard)! Subscribe. RR

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 agorize 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 lt;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

Originally Posted by ;

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 agorize 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

Originally Posted by ;

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
http://articles.mql4.com/868 may be useful.

Originally Posted by ;

...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 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 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 indiive 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, indied 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.

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

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