Machine learning in trading: theory, models, practice and algo-trading - page 3457

New comment
 
Ivan Butko #:
We add a fourth price - times more.

On the 14th candle we get 100,000,000,000 possible patterns.

Add the comparison of high and low of these 14 candlesticks and we get another 100,000,000,000 patterns.

There are as many patterns as there are degrees of freedom on the chart

And you think that all these trillions of patterns will have multiple repeatability/integrity?

 
Grigori.S.B #:

And you think all those trillions of patterns will have multiple repeatability/integrity?

Specifically - the above/below comparison are too clumsy patterns. Any of them tends to be 50% working out in different directions.

The rarer the occurrence, the higher the probability of working out, but this is a small sample problem.

I did such an analysis, collected statistics. Nothing good. The problem here is in linking to the time period, perhaps. Traders do not trade at 7:15 exactly when the alarm clock rings for the third time.

But Zigzag is also a problem. But the trouble is interesting.

If you take a figure, a specific one. And open a position in one direction, placing TP and SL in certain points of the zigzag - the performance when testing on all pairs in the "All Symbols" mode in the optimiser is distributed almost evenly from the worst to the best.

Roughly speaking - one figure works on one pair, another figure works on another pair. Although Petya and Vasya trade on both currencies.

That is, as if the market works according to one formula, where all currencies are interconnected and balanced.

Because if one figure (simple) will work on all currency pairs - then triangles will break, which are known as lock in technical arbitrage.



In short, I took a figure, ran it, let's say on Eurodollar. I set TP on the nearest point of the zigzag, set SL on the opposite nearest point of the zigzag.

And so, I got a scribble on the balance chart. I take - change the location of the TP to... the third point of the zigzag. And, voila - the balance curve became better, but still shit.

I take and add a new point (thus reducing the repeatability in N,N times), set the condition "if this rear - fifth point (by count) is greater than that and that, and less than that and that..." and once - the balance has improved.

Another condition - still improved. In the end there are 100 inputs left for 10 years.


Then I take another zigzag pattern. I set TP SL, there and there, run - again 1000 inputs, chaos. I change TP SL to different points, add conditions of points - and now the new pattern is more stable.



I do this with 5-10 patterns, and here are 10 years of stable profit. We add 28 more pairs, pick up these dirty patterns on each pair - filter them, "wash", create a set of patterns for all pairs - and now the strategy works on all pairs.



This is how I imagine the creation of all-pairs trading. But I got so bored with zigzag that I temporarily abandoned it.

A little later I will check it all on forwards (I was afraid that it would not work, but the process of selection was entertaining).



The machine should do it all on the machine, to search everything, and on the smallest TF, where the spread does not kill profitability. Well, M5, for example.

In general, there is still a field for research.


The only way to prove that these actions are fitting is to check these "sets" on the forward.

And so, yes, all this looks like perverted optimisation.

 
You need good repeatability, preferably 100+ per year. Otherwise, it is not a pattern, but a statistical error
 
Ivan Butko #:

Specifically, the above/below comparisons are too clumsy patterns. Any of them tends to work out 50% in different directions.

The rarer the occurrence, the higher the probability of working out, but this is a problem of a small sample.

I did such an analysis, collected statistics. Nothing good. The problem here is in linking to the time period, perhaps. Traders do not trade at exactly 7:15 when the alarm clock rings for the third time.

Hallelujah fucking hallelujah!

 

The blue dot is the first candle on which the pattern gave a signal.

the blue dot is the first candle on which the pattern gave a signal, everything after the dot is like the future....


Here are the signals from the pattern on new (test) data, those examples where the pattern did not guess the pullback.


-------------------------------------------------------------------------------

And here are examples where the pattern guessed the rebound on test data.


 

Even the false signals show that there is a pattern.

And here's the pattern itself.

[1] "(EURGBP.M15.High)1" "(CADCHF.M1.Open)0"  "(EURGBP.M5.Low)0"   "(EURGBP.M1.Low)0"  
 [5] "(USDCHF.M1.High)-1" "(EURUSD.M1.Open)0"  "(EURCAD.M15.High)1" "(USDCHF.M1.Open)0" 
 [9] "(EURNZD.M15.High)0" "(NZDCAD.M1.Open)1"

This is a certain sequence of events that must occur in a huge vector of all kinds of events.

There are about 20 thousand such events, each event is already correctly compressed information about the market and in the correct time sequence.

 

You can beautifully search for patterns through any regressor that has the ability to prune trees.

You can even use linear regression.

 
No, you can't.
 
mytarmailS #:
No, you can't.
Yes, you can.
And here we go again with this unhealthy topic about events.

Pile the signs into the bousting, pick out what it predicts well, cluster it into groups. There are your patterns. How much more?

What are you using?
 
Maxim Dmitrievsky #:
We have to
And here we go again with this unhealthy topic about some events.

Pile the signs into the bousting, pick out what it predicts well, cluster it into groups. There's your patterns. How much more?

What are you using?
I don't know how to boost at least 100 million features, do you?

I have a simple function that searches for a sequence in a large vector.

You'd have to sit through an hour of live conversation and pictures to explain it.
1...345034513452345334543455345634573458345934603461346234633464...3591
New comment